cc
Browse files- .gitignore +1 -0
- configs/autoencoder/autoencoder_kl_32x32x4.yaml +24 -0
- lightning_logs/version_0/hparams.yaml +1 -0
- lightning_logs/version_1/hparams.yaml +1 -0
- lightning_logs/version_2/hparams.yaml +1 -0
- lightning_logs/version_3/hparams.yaml +1 -0
- lightning_logs/version_4/hparams.yaml +1 -0
- lightning_logs/version_5/hparams.yaml +1 -0
- lightning_logs/version_6/hparams.yaml +1 -0
- lightning_logs/version_6/metrics.csv +8 -0
- main.py +19 -0
- swim/attention_blocks.py +0 -315
- swim/autoencoder.py +0 -247
- swim/blocks.py +0 -227
- swim/codeblock.py +0 -74
- swim/discriminator.py +0 -45
- swim/lr_scheduler.py +98 -0
- swim/models/autoencoder.py +219 -0
- swim/modules/attention.py +277 -0
- swim/modules/dataset.py +145 -0
- swim/{__init__.py → modules/diffusionmodules/__init__.py} +0 -0
- swim/modules/diffusionmodules/model.py +1010 -0
- swim/modules/diffusionmodules/openaimodel.py +1012 -0
- swim/modules/diffusionmodules/util.py +296 -0
- swim/modules/discriminators/n_layer_discriminator.py +88 -0
- swim/modules/distributions/__init__.py +0 -0
- swim/modules/distributions/distributions.py +92 -0
- swim/modules/ema.py +76 -0
- swim/modules/losses/__init__.py +1 -0
- swim/modules/losses/contperceptual.py +181 -0
- swim/modules/x_transformer.py +641 -0
- swim/unet.py +0 -169
- swim/utils.py +297 -0
- train.py +0 -72
.gitignore
CHANGED
@@ -1,3 +1,4 @@
|
|
1 |
__pycache__
|
2 |
datasets
|
|
|
3 |
*.zip
|
|
|
1 |
__pycache__
|
2 |
datasets
|
3 |
+
wandb
|
4 |
*.zip
|
configs/autoencoder/autoencoder_kl_32x32x4.yaml
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 4.5e-6
|
3 |
+
target: swim.models.autoencoder.AutoencoderKL
|
4 |
+
params:
|
5 |
+
monitor: "val/rec_loss"
|
6 |
+
embed_dim: 4
|
7 |
+
lossconfig:
|
8 |
+
target: swim.modules.losses.LPIPSWithDiscriminator
|
9 |
+
params:
|
10 |
+
disc_start: 50001
|
11 |
+
kl_weight: 0.000001
|
12 |
+
disc_weight: 0.5
|
13 |
+
|
14 |
+
ddconfig:
|
15 |
+
double_z: True
|
16 |
+
z_channels: 4
|
17 |
+
resolution: 512
|
18 |
+
in_channels: 3
|
19 |
+
out_ch: 3
|
20 |
+
ch: 128
|
21 |
+
ch_mult: [1, 2, 4, 4] # num_down = len(ch_mult)-1
|
22 |
+
num_res_blocks: 2
|
23 |
+
attn_resolutions: []
|
24 |
+
dropout: 0.0
|
lightning_logs/version_0/hparams.yaml
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{}
|
lightning_logs/version_1/hparams.yaml
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{}
|
lightning_logs/version_2/hparams.yaml
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{}
|
lightning_logs/version_3/hparams.yaml
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{}
|
lightning_logs/version_4/hparams.yaml
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{}
|
lightning_logs/version_5/hparams.yaml
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{}
|
lightning_logs/version_6/hparams.yaml
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{}
|
lightning_logs/version_6/metrics.csv
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
aeloss_step,discloss_step,epoch,step,train/d_weight,train/disc_factor,train/disc_loss,train/g_loss,train/kl_loss,train/logits_fake,train/logits_real,train/logvar,train/nll_loss,train/rec_loss,train/total_loss
|
2 |
+
2953.825927734375,0.0,0,49,247.98703002929688,0.0,0.0,0.4574800729751587,3.8113327026367188,-0.4574800729751587,0.25408393144607544,0.0,2953.825927734375,0.9615318775177002,2953.825927734375
|
3 |
+
2997.572021484375,0.0,0,99,582.9641723632812,0.0,0.0,0.30489930510520935,9.548039436340332,-0.30489930510520935,-0.17654460668563843,0.0,2997.572021484375,0.9757721424102783,2997.572021484375
|
4 |
+
3229.461181640625,0.0,0,149,153.92608642578125,0.0,0.0,0.2801606059074402,38.64592742919922,-0.2801606059074402,0.49510449171066284,0.0,3229.461181640625,1.0512568950653076,3229.461181640625
|
5 |
+
2391.34716796875,0.0,0,199,171.67630004882812,0.0,0.0,0.3865272104740143,29.355663299560547,-0.3865272104740143,-0.2872551679611206,0.0,2391.34716796875,0.7784333229064941,2391.34716796875
|
6 |
+
2008.2764892578125,0.0,0,249,223.92453002929688,0.0,0.0,0.1616521179676056,75.05839538574219,-0.1616521179676056,-0.43370532989501953,0.0,2008.2763671875,0.6537358164787292,2008.2764892578125
|
7 |
+
2754.87353515625,0.0,0,299,370.9005432128906,0.0,0.0,0.08434182405471802,16.720157623291016,-0.08434182405471802,-0.3120231032371521,0.0,2754.87353515625,0.8967687487602234,2754.87353515625
|
8 |
+
1949.1402587890625,0.0,0,349,579.2057495117188,0.0,0.0,0.4615066647529602,88.49298095703125,-0.4615066647529602,-0.39516788721084595,0.0,1949.14013671875,0.6344857215881348,1949.1402587890625
|
main.py
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from omegaconf import OmegaConf
|
3 |
+
from swim.utils import instantiate_from_config
|
4 |
+
from torchinfo import summary
|
5 |
+
from swim.modules.dataset import SwimDataModule
|
6 |
+
from lightning import Trainer
|
7 |
+
from lightning.pytorch.loggers import WandbLogger
|
8 |
+
|
9 |
+
config = OmegaConf.load("configs/autoencoder/autoencoder_kl_32x32x4.yaml")
|
10 |
+
|
11 |
+
model = instantiate_from_config(config.model)
|
12 |
+
model.learning_rate = config.model.base_learning_rate
|
13 |
+
|
14 |
+
datamodule = SwimDataModule(img_size=32)
|
15 |
+
|
16 |
+
logger = WandbLogger(project="swim", name="autoencoder_kl")
|
17 |
+
|
18 |
+
trainer = Trainer(max_epochs=10, devices=[0], logger=logger, log_every_n_steps=10)
|
19 |
+
trainer.fit(model, datamodule)
|
swim/attention_blocks.py
DELETED
@@ -1,315 +0,0 @@
|
|
1 |
-
from typing import Optional
|
2 |
-
|
3 |
-
import torch
|
4 |
-
import torch.nn.functional as F
|
5 |
-
from torch import nn
|
6 |
-
|
7 |
-
|
8 |
-
class SpatialTransformer(nn.Module):
|
9 |
-
"""
|
10 |
-
## Spatial Transformer
|
11 |
-
"""
|
12 |
-
|
13 |
-
def __init__(self, channels: int, n_heads: int, n_layers: int, d_cond: int):
|
14 |
-
"""
|
15 |
-
:param channels: is the number of channels in the feature map
|
16 |
-
:param n_heads: is the number of attention heads
|
17 |
-
:param n_layers: is the number of transformer layers
|
18 |
-
:param d_cond: is the size of the conditional embedding
|
19 |
-
"""
|
20 |
-
super().__init__()
|
21 |
-
# Initial group normalization
|
22 |
-
self.norm = torch.nn.GroupNorm(
|
23 |
-
num_groups=32, num_channels=channels, eps=1e-6, affine=True
|
24 |
-
)
|
25 |
-
# Initial $1 \times 1$ convolution
|
26 |
-
self.proj_in = nn.Conv2d(channels, channels, kernel_size=1, stride=1, padding=0)
|
27 |
-
|
28 |
-
# Transformer layers
|
29 |
-
self.transformer_blocks = nn.ModuleList(
|
30 |
-
[
|
31 |
-
BasicTransformerBlock(
|
32 |
-
channels, n_heads, channels // n_heads, d_cond=d_cond
|
33 |
-
)
|
34 |
-
for _ in range(n_layers)
|
35 |
-
]
|
36 |
-
)
|
37 |
-
|
38 |
-
# Final $1 \times 1$ convolution
|
39 |
-
self.proj_out = nn.Conv2d(
|
40 |
-
channels, channels, kernel_size=1, stride=1, padding=0
|
41 |
-
)
|
42 |
-
|
43 |
-
def forward(self, x: torch.Tensor, cond: torch.Tensor):
|
44 |
-
"""
|
45 |
-
:param x: is the feature map of shape `[batch_size, channels, height, width]`
|
46 |
-
:param cond: is the conditional embeddings of shape `[batch_size, n_cond, d_cond]`
|
47 |
-
"""
|
48 |
-
# Get shape `[batch_size, channels, height, width]`
|
49 |
-
b, c, h, w = x.shape
|
50 |
-
# For residual connection
|
51 |
-
x_in = x
|
52 |
-
# Normalize
|
53 |
-
x = self.norm(x)
|
54 |
-
# Initial $1 \times 1$ convolution
|
55 |
-
x = self.proj_in(x)
|
56 |
-
# Transpose and reshape from `[batch_size, channels, height, width]`
|
57 |
-
# to `[batch_size, height * width, channels]`
|
58 |
-
x = x.permute(0, 2, 3, 1).view(b, h * w, c)
|
59 |
-
# Apply the transformer layers
|
60 |
-
for block in self.transformer_blocks:
|
61 |
-
x = block(x, cond)
|
62 |
-
# Reshape and transpose from `[batch_size, height * width, channels]`
|
63 |
-
# to `[batch_size, channels, height, width]`
|
64 |
-
x = x.view(b, h, w, c).permute(0, 3, 1, 2)
|
65 |
-
# Final $1 \times 1$ convolution
|
66 |
-
x = self.proj_out(x)
|
67 |
-
# Add residual
|
68 |
-
return x + x_in
|
69 |
-
|
70 |
-
|
71 |
-
class BasicTransformerBlock(nn.Module):
|
72 |
-
"""
|
73 |
-
### Transformer Layer
|
74 |
-
"""
|
75 |
-
|
76 |
-
def __init__(self, d_model: int, n_heads: int, d_head: int, d_cond: int):
|
77 |
-
"""
|
78 |
-
:param d_model: is the input embedding size
|
79 |
-
:param n_heads: is the number of attention heads
|
80 |
-
:param d_head: is the size of a attention head
|
81 |
-
:param d_cond: is the size of the conditional embeddings
|
82 |
-
"""
|
83 |
-
super().__init__()
|
84 |
-
# Self-attention layer and pre-norm layer
|
85 |
-
self.attn1 = CrossAttention(d_model, d_model, n_heads, d_head)
|
86 |
-
self.norm1 = nn.LayerNorm(d_model)
|
87 |
-
# Cross attention layer and pre-norm layer
|
88 |
-
self.attn2 = CrossAttention(d_model, d_cond, n_heads, d_head)
|
89 |
-
self.norm2 = nn.LayerNorm(d_model)
|
90 |
-
# Feed-forward network and pre-norm layer
|
91 |
-
self.ff = FeedForward(d_model)
|
92 |
-
self.norm3 = nn.LayerNorm(d_model)
|
93 |
-
|
94 |
-
def forward(self, x: torch.Tensor, cond: torch.Tensor):
|
95 |
-
"""
|
96 |
-
:param x: are the input embeddings of shape `[batch_size, height * width, d_model]`
|
97 |
-
:param cond: is the conditional embeddings of shape `[batch_size, n_cond, d_cond]`
|
98 |
-
"""
|
99 |
-
# Self attention
|
100 |
-
x = self.attn1(self.norm1(x)) + x
|
101 |
-
# Cross-attention with conditioning
|
102 |
-
x = self.attn2(self.norm2(x), cond=cond) + x
|
103 |
-
# Feed-forward network
|
104 |
-
x = self.ff(self.norm3(x)) + x
|
105 |
-
#
|
106 |
-
return x
|
107 |
-
|
108 |
-
|
109 |
-
class CrossAttention(nn.Module):
|
110 |
-
"""
|
111 |
-
### Cross Attention Layer
|
112 |
-
|
113 |
-
This falls-back to self-attention when conditional embeddings are not specified.
|
114 |
-
"""
|
115 |
-
|
116 |
-
use_flash_attention: bool = False
|
117 |
-
|
118 |
-
def __init__(
|
119 |
-
self,
|
120 |
-
d_model: int,
|
121 |
-
d_cond: int,
|
122 |
-
n_heads: int,
|
123 |
-
d_head: int,
|
124 |
-
is_inplace: bool = True,
|
125 |
-
):
|
126 |
-
"""
|
127 |
-
:param d_model: is the input embedding size
|
128 |
-
:param n_heads: is the number of attention heads
|
129 |
-
:param d_head: is the size of a attention head
|
130 |
-
:param d_cond: is the size of the conditional embeddings
|
131 |
-
:param is_inplace: specifies whether to perform the attention softmax computation inplace to
|
132 |
-
save memory
|
133 |
-
"""
|
134 |
-
super().__init__()
|
135 |
-
|
136 |
-
self.is_inplace = is_inplace
|
137 |
-
self.n_heads = n_heads
|
138 |
-
self.d_head = d_head
|
139 |
-
|
140 |
-
# Attention scaling factor
|
141 |
-
self.scale = d_head**-0.5
|
142 |
-
|
143 |
-
# Query, key and value mappings
|
144 |
-
d_attn = d_head * n_heads
|
145 |
-
self.to_q = nn.Linear(d_model, d_attn, bias=False)
|
146 |
-
self.to_k = nn.Linear(d_cond, d_attn, bias=False)
|
147 |
-
self.to_v = nn.Linear(d_cond, d_attn, bias=False)
|
148 |
-
|
149 |
-
# Final linear layer
|
150 |
-
self.to_out = nn.Sequential(nn.Linear(d_attn, d_model))
|
151 |
-
|
152 |
-
# Setup [flash attention](https://github.com/HazyResearch/flash-attention).
|
153 |
-
# Flash attention is only used if it's installed
|
154 |
-
# and `CrossAttention.use_flash_attention` is set to `True`.
|
155 |
-
# try:
|
156 |
-
# # You can install flash attention by cloning their Github repo,
|
157 |
-
# # [https://github.com/HazyResearch/flash-attention](https://github.com/HazyResearch/flash-attention)
|
158 |
-
# # and then running `python setup.py install`
|
159 |
-
# from flash_attn.flash_attention import FlashAttention
|
160 |
-
|
161 |
-
# self.flash = FlashAttention()
|
162 |
-
# # Set the scale for scaled dot-product attention.
|
163 |
-
# self.flash.softmax_scale = self.scale
|
164 |
-
# # Set to `None` if it's not installed
|
165 |
-
# except ImportError:
|
166 |
-
# self.flash = None
|
167 |
-
|
168 |
-
def forward(self, x: torch.Tensor, cond: Optional[torch.Tensor] = None):
|
169 |
-
"""
|
170 |
-
:param x: are the input embeddings of shape `[batch_size, height * width, d_model]`
|
171 |
-
:param cond: is the conditional embeddings of shape `[batch_size, n_cond, d_cond]`
|
172 |
-
"""
|
173 |
-
|
174 |
-
# If `cond` is `None` we perform self attention
|
175 |
-
has_cond = cond is not None
|
176 |
-
if not has_cond:
|
177 |
-
cond = x
|
178 |
-
|
179 |
-
# Get query, key and value vectors
|
180 |
-
q = self.to_q(x)
|
181 |
-
k = self.to_k(cond)
|
182 |
-
v = self.to_v(cond)
|
183 |
-
|
184 |
-
# Use flash attention if it's available and the head size is less than or equal to `128`
|
185 |
-
if (
|
186 |
-
CrossAttention.use_flash_attention
|
187 |
-
and self.flash is not None
|
188 |
-
and not has_cond
|
189 |
-
and self.d_head <= 128
|
190 |
-
):
|
191 |
-
return self.flash_attention(q, k, v)
|
192 |
-
# Otherwise, fallback to normal attention
|
193 |
-
else:
|
194 |
-
return self.normal_attention(q, k, v)
|
195 |
-
|
196 |
-
def flash_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
|
197 |
-
"""
|
198 |
-
#### Flash Attention
|
199 |
-
|
200 |
-
:param q: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
|
201 |
-
:param k: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
|
202 |
-
:param v: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
|
203 |
-
"""
|
204 |
-
|
205 |
-
# Get batch size and number of elements along sequence axis (`width * height`)
|
206 |
-
batch_size, seq_len, _ = q.shape
|
207 |
-
|
208 |
-
# Stack `q`, `k`, `v` vectors for flash attention, to get a single tensor of
|
209 |
-
# shape `[batch_size, seq_len, 3, n_heads * d_head]`
|
210 |
-
qkv = torch.stack((q, k, v), dim=2)
|
211 |
-
# Split the heads
|
212 |
-
qkv = qkv.view(batch_size, seq_len, 3, self.n_heads, self.d_head)
|
213 |
-
|
214 |
-
# Flash attention works for head sizes `32`, `64` and `128`, so we have to pad the heads to
|
215 |
-
# fit this size.
|
216 |
-
if self.d_head <= 32:
|
217 |
-
pad = 32 - self.d_head
|
218 |
-
elif self.d_head <= 64:
|
219 |
-
pad = 64 - self.d_head
|
220 |
-
elif self.d_head <= 128:
|
221 |
-
pad = 128 - self.d_head
|
222 |
-
else:
|
223 |
-
raise ValueError(f"Head size ${self.d_head} too large for Flash Attention")
|
224 |
-
|
225 |
-
# Pad the heads
|
226 |
-
if pad:
|
227 |
-
qkv = torch.cat(
|
228 |
-
(qkv, qkv.new_zeros(batch_size, seq_len, 3, self.n_heads, pad)), dim=-1
|
229 |
-
)
|
230 |
-
|
231 |
-
# Compute attention
|
232 |
-
# $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$$
|
233 |
-
# This gives a tensor of shape `[batch_size, seq_len, n_heads, d_padded]`
|
234 |
-
out, _ = self.flash(qkv)
|
235 |
-
# Truncate the extra head size
|
236 |
-
out = out[:, :, :, : self.d_head]
|
237 |
-
# Reshape to `[batch_size, seq_len, n_heads * d_head]`
|
238 |
-
out = out.reshape(batch_size, seq_len, self.n_heads * self.d_head)
|
239 |
-
|
240 |
-
# Map to `[batch_size, height * width, d_model]` with a linear layer
|
241 |
-
return self.to_out(out)
|
242 |
-
|
243 |
-
def normal_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
|
244 |
-
"""
|
245 |
-
#### Normal Attention
|
246 |
-
|
247 |
-
:param q: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
|
248 |
-
:param k: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
|
249 |
-
:param v: are the query vectors before splitting heads, of shape `[batch_size, seq, d_attn]`
|
250 |
-
"""
|
251 |
-
|
252 |
-
# Split them to heads of shape `[batch_size, seq_len, n_heads, d_head]`
|
253 |
-
q = q.view(*q.shape[:2], self.n_heads, -1)
|
254 |
-
k = k.view(*k.shape[:2], self.n_heads, -1)
|
255 |
-
v = v.view(*v.shape[:2], self.n_heads, -1)
|
256 |
-
|
257 |
-
# Calculate attention $\frac{Q K^\top}{\sqrt{d_{key}}}$
|
258 |
-
attn = torch.einsum("bihd,bjhd->bhij", q, k) * self.scale
|
259 |
-
|
260 |
-
# Compute softmax
|
261 |
-
# $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)$$
|
262 |
-
if self.is_inplace:
|
263 |
-
half = attn.shape[0] // 2
|
264 |
-
attn[half:] = attn[half:].softmax(dim=-1)
|
265 |
-
attn[:half] = attn[:half].softmax(dim=-1)
|
266 |
-
else:
|
267 |
-
attn = attn.softmax(dim=-1)
|
268 |
-
|
269 |
-
# Compute attention output
|
270 |
-
# $$\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$$
|
271 |
-
out = torch.einsum("bhij,bjhd->bihd", attn, v)
|
272 |
-
# Reshape to `[batch_size, height * width, n_heads * d_head]`
|
273 |
-
out = out.reshape(*out.shape[:2], -1)
|
274 |
-
# Map to `[batch_size, height * width, d_model]` with a linear layer
|
275 |
-
return self.to_out(out)
|
276 |
-
|
277 |
-
|
278 |
-
class FeedForward(nn.Module):
|
279 |
-
"""
|
280 |
-
### Feed-Forward Network
|
281 |
-
"""
|
282 |
-
|
283 |
-
def __init__(self, d_model: int, d_mult: int = 4):
|
284 |
-
"""
|
285 |
-
:param d_model: is the input embedding size
|
286 |
-
:param d_mult: is multiplicative factor for the hidden layer size
|
287 |
-
"""
|
288 |
-
super().__init__()
|
289 |
-
self.net = nn.Sequential(
|
290 |
-
GeGLU(d_model, d_model * d_mult),
|
291 |
-
nn.Dropout(0.0),
|
292 |
-
nn.Linear(d_model * d_mult, d_model),
|
293 |
-
)
|
294 |
-
|
295 |
-
def forward(self, x: torch.Tensor):
|
296 |
-
return self.net(x)
|
297 |
-
|
298 |
-
|
299 |
-
class GeGLU(nn.Module):
|
300 |
-
"""
|
301 |
-
### GeGLU Activation
|
302 |
-
|
303 |
-
$$\text{GeGLU}(x) = (xW + b) * \text{GELU}(xV + c)$$
|
304 |
-
"""
|
305 |
-
|
306 |
-
def __init__(self, d_in: int, d_out: int):
|
307 |
-
super().__init__()
|
308 |
-
# Combined linear projections $xW + b$ and $xV + c$
|
309 |
-
self.proj = nn.Linear(d_in, d_out * 2)
|
310 |
-
|
311 |
-
def forward(self, x: torch.Tensor):
|
312 |
-
# Get $xW + b$ and $xV + c$
|
313 |
-
x, gate = self.proj(x).chunk(2, dim=-1)
|
314 |
-
# $\text{GeGLU}(x) = (xW + b) * \text{GELU}(xV + c)$
|
315 |
-
return x * F.gelu(gate)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
swim/autoencoder.py
DELETED
@@ -1,247 +0,0 @@
|
|
1 |
-
from typing import List
|
2 |
-
|
3 |
-
import torch
|
4 |
-
import torch.nn.functional as F
|
5 |
-
from torch import nn
|
6 |
-
|
7 |
-
from .blocks import (
|
8 |
-
ResnetBlock,
|
9 |
-
AttentionBlock,
|
10 |
-
GroupNorm,
|
11 |
-
UpSampleBlock,
|
12 |
-
DownSampleBlock,
|
13 |
-
)
|
14 |
-
|
15 |
-
|
16 |
-
class Autoencoder(nn.Module):
|
17 |
-
|
18 |
-
def __init__(
|
19 |
-
self,
|
20 |
-
channels: int,
|
21 |
-
channel_multipliers: List[int],
|
22 |
-
n_resnet_blocks: int,
|
23 |
-
in_channels: int,
|
24 |
-
z_channels: int,
|
25 |
-
emb_channels: int,
|
26 |
-
):
|
27 |
-
super().__init__()
|
28 |
-
self.encoder = Encoder(
|
29 |
-
channels=channels,
|
30 |
-
channel_multipliers=channel_multipliers,
|
31 |
-
n_resnet_blocks=n_resnet_blocks,
|
32 |
-
in_channels=in_channels,
|
33 |
-
z_channels=z_channels,
|
34 |
-
)
|
35 |
-
self.decoder = Decoder(
|
36 |
-
channels=channels,
|
37 |
-
channel_multipliers=channel_multipliers,
|
38 |
-
n_resnet_blocks=n_resnet_blocks,
|
39 |
-
out_channels=in_channels,
|
40 |
-
z_channels=z_channels,
|
41 |
-
)
|
42 |
-
# Convolution to map from embedding space to
|
43 |
-
# quantized embedding space moments (mean and log variance)
|
44 |
-
self.quant_conv = nn.Conv2d(2 * z_channels, 2 * emb_channels, 1)
|
45 |
-
# Convolution to map from quantized embedding space back to
|
46 |
-
# embedding space
|
47 |
-
self.post_quant_conv = nn.Conv2d(emb_channels, z_channels, 1)
|
48 |
-
|
49 |
-
def encode(self, img: torch.Tensor) -> "GaussianDistribution":
|
50 |
-
# Get embeddings with shape `[batch_size, z_channels * 2, z_height, z_height]`
|
51 |
-
z = self.encoder(img)
|
52 |
-
# Get the moments in the quantized embedding space
|
53 |
-
moments = self.quant_conv(z)
|
54 |
-
# Return the distribution
|
55 |
-
return GaussianDistribution(moments)
|
56 |
-
|
57 |
-
def decode(self, z: torch.Tensor):
|
58 |
-
# Map to embedding space from the quantized representation
|
59 |
-
z = self.post_quant_conv(z)
|
60 |
-
# Decode the image of shape `[batch_size, channels, height, width]`
|
61 |
-
return self.decoder(z)
|
62 |
-
|
63 |
-
def forward(self, x: torch.Tensor, sample_posterior: bool = False):
|
64 |
-
posterior = self.encode(x)
|
65 |
-
if sample_posterior:
|
66 |
-
z = posterior.sample()
|
67 |
-
else:
|
68 |
-
z = posterior.mode()
|
69 |
-
decoded_x = self.decode(z)
|
70 |
-
return decoded_x, posterior
|
71 |
-
|
72 |
-
|
73 |
-
class Encoder(nn.Module):
|
74 |
-
def __init__(
|
75 |
-
self,
|
76 |
-
*,
|
77 |
-
channels: int,
|
78 |
-
channel_multipliers: List[int],
|
79 |
-
n_resnet_blocks: int,
|
80 |
-
in_channels: int,
|
81 |
-
z_channels: int
|
82 |
-
):
|
83 |
-
super().__init__()
|
84 |
-
|
85 |
-
# Number of blocks of different resolutions.
|
86 |
-
# The resolution is halved at the end each top level block
|
87 |
-
n_resolutions = len(channel_multipliers)
|
88 |
-
|
89 |
-
# Initial $3 \times 3$ convolution layer that maps the image to `channels`
|
90 |
-
self.conv_in = nn.Conv2d(in_channels, channels, 3, stride=1, padding=1)
|
91 |
-
|
92 |
-
# Number of channels in each top level block
|
93 |
-
channels_list = [m * channels for m in [1] + channel_multipliers]
|
94 |
-
|
95 |
-
# List of top-level blocks
|
96 |
-
self.down = nn.ModuleList()
|
97 |
-
# Create top-level blocks
|
98 |
-
for i in range(n_resolutions):
|
99 |
-
# Each top level block consists of multiple ResNet Blocks and down-sampling
|
100 |
-
resnet_blocks = nn.ModuleList()
|
101 |
-
# Add ResNet Blocks
|
102 |
-
for _ in range(n_resnet_blocks):
|
103 |
-
resnet_blocks.append(ResnetBlock(channels, channels_list[i + 1]))
|
104 |
-
channels = channels_list[i + 1]
|
105 |
-
# Top-level block
|
106 |
-
down = nn.Module()
|
107 |
-
down.block = resnet_blocks
|
108 |
-
# Down-sampling at the end of each top level block except the last
|
109 |
-
if i != n_resolutions - 1:
|
110 |
-
down.downsample = DownSampleBlock(channels)
|
111 |
-
else:
|
112 |
-
down.downsample = nn.Identity()
|
113 |
-
#
|
114 |
-
self.down.append(down)
|
115 |
-
|
116 |
-
# Final ResNet blocks with attention
|
117 |
-
self.mid = nn.Module()
|
118 |
-
self.mid.block_1 = ResnetBlock(channels, channels)
|
119 |
-
self.mid.attn_1 = AttentionBlock(channels)
|
120 |
-
self.mid.block_2 = ResnetBlock(channels, channels)
|
121 |
-
|
122 |
-
# Map to embedding space with a $3 \times 3$ convolution
|
123 |
-
self.norm_out = GroupNorm(channels)
|
124 |
-
self.conv_out = nn.Conv2d(channels, 2 * z_channels, 3, stride=1, padding=1)
|
125 |
-
|
126 |
-
def forward(self, img: torch.Tensor):
|
127 |
-
# Map to `channels` with the initial convolution
|
128 |
-
x = self.conv_in(img)
|
129 |
-
|
130 |
-
# Top-level blocks
|
131 |
-
for down in self.down:
|
132 |
-
# ResNet Blocks
|
133 |
-
for block in down.block:
|
134 |
-
x = block(x)
|
135 |
-
# Down-sampling
|
136 |
-
x = down.downsample(x)
|
137 |
-
|
138 |
-
# Final ResNet blocks with attention
|
139 |
-
x = self.mid.block_1(x)
|
140 |
-
x = self.mid.attn_1(x)
|
141 |
-
x = self.mid.block_2(x)
|
142 |
-
|
143 |
-
# Normalize and map to embedding space
|
144 |
-
x = self.norm_out(x)
|
145 |
-
x = F.silu(x)
|
146 |
-
x = self.conv_out(x)
|
147 |
-
|
148 |
-
return x
|
149 |
-
|
150 |
-
|
151 |
-
class Decoder(nn.Module):
|
152 |
-
|
153 |
-
def __init__(
|
154 |
-
self,
|
155 |
-
*,
|
156 |
-
channels: int,
|
157 |
-
channel_multipliers: List[int],
|
158 |
-
n_resnet_blocks: int,
|
159 |
-
out_channels: int,
|
160 |
-
z_channels: int
|
161 |
-
):
|
162 |
-
super().__init__()
|
163 |
-
|
164 |
-
# Number of blocks of different resolutions.
|
165 |
-
# The resolution is halved at the end each top level block
|
166 |
-
num_resolutions = len(channel_multipliers)
|
167 |
-
|
168 |
-
# Number of channels in each top level block, in the reverse order
|
169 |
-
channels_list = [m * channels for m in channel_multipliers]
|
170 |
-
|
171 |
-
# Number of channels in the top-level block
|
172 |
-
channels = channels_list[-1]
|
173 |
-
|
174 |
-
# Initial $3 \times 3$ convolution layer that maps the embedding space to `channels`
|
175 |
-
self.conv_in = nn.Conv2d(z_channels, channels, 3, stride=1, padding=1)
|
176 |
-
|
177 |
-
# ResNet blocks with attention
|
178 |
-
self.mid = nn.Module()
|
179 |
-
self.mid.block_1 = ResnetBlock(channels, channels)
|
180 |
-
self.mid.attn_1 = AttentionBlock(channels)
|
181 |
-
self.mid.block_2 = ResnetBlock(channels, channels)
|
182 |
-
|
183 |
-
# List of top-level blocks
|
184 |
-
self.up = nn.ModuleList()
|
185 |
-
# Create top-level blocks
|
186 |
-
for i in reversed(range(num_resolutions)):
|
187 |
-
# Each top level block consists of multiple ResNet Blocks and up-sampling
|
188 |
-
resnet_blocks = nn.ModuleList()
|
189 |
-
# Add ResNet Blocks
|
190 |
-
for _ in range(n_resnet_blocks + 1):
|
191 |
-
resnet_blocks.append(ResnetBlock(channels, channels_list[i]))
|
192 |
-
channels = channels_list[i]
|
193 |
-
# Top-level block
|
194 |
-
up = nn.Module()
|
195 |
-
up.block = resnet_blocks
|
196 |
-
# Up-sampling at the end of each top level block except the first
|
197 |
-
if i != 0:
|
198 |
-
up.upsample = UpSampleBlock(channels)
|
199 |
-
else:
|
200 |
-
up.upsample = nn.Identity()
|
201 |
-
# Prepend to be consistent with the checkpoint
|
202 |
-
self.up.insert(0, up)
|
203 |
-
|
204 |
-
# Map to image space with a $3 \times 3$ convolution
|
205 |
-
self.norm_out = GroupNorm(channels)
|
206 |
-
self.conv_out = nn.Conv2d(channels, out_channels, 3, stride=1, padding=1)
|
207 |
-
|
208 |
-
def forward(self, z: torch.Tensor):
|
209 |
-
# Map to `channels` with the initial convolution
|
210 |
-
h = self.conv_in(z)
|
211 |
-
|
212 |
-
# ResNet blocks with attention
|
213 |
-
h = self.mid.block_1(h)
|
214 |
-
h = self.mid.attn_1(h)
|
215 |
-
h = self.mid.block_2(h)
|
216 |
-
|
217 |
-
# Top-level blocks
|
218 |
-
for up in reversed(self.up):
|
219 |
-
# ResNet Blocks
|
220 |
-
for block in up.block:
|
221 |
-
h = block(h)
|
222 |
-
# Up-sampling
|
223 |
-
h = up.upsample(h)
|
224 |
-
|
225 |
-
# Normalize and map to image space
|
226 |
-
h = self.norm_out(h)
|
227 |
-
h = F.silu(h)
|
228 |
-
img = self.conv_out(h)
|
229 |
-
|
230 |
-
return img
|
231 |
-
|
232 |
-
|
233 |
-
class GaussianDistribution:
|
234 |
-
def __init__(self, parameters: torch.Tensor):
|
235 |
-
# Split mean and log of variance
|
236 |
-
self.mean, log_var = torch.chunk(parameters, 2, dim=1)
|
237 |
-
# Clamp the log of variances
|
238 |
-
self.log_var = torch.clamp(log_var, -30.0, 20.0)
|
239 |
-
# Calculate standard deviation
|
240 |
-
self.std = torch.exp(0.5 * self.log_var)
|
241 |
-
|
242 |
-
def sample(self):
|
243 |
-
# Sample from the distribution
|
244 |
-
return self.mean + self.std * torch.randn_like(self.std)
|
245 |
-
|
246 |
-
def mode(self):
|
247 |
-
return self.mean
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
swim/blocks.py
DELETED
@@ -1,227 +0,0 @@
|
|
1 |
-
from abc import abstractmethod
|
2 |
-
|
3 |
-
import math
|
4 |
-
import torch
|
5 |
-
from torch import nn
|
6 |
-
from torch.nn import functional as F
|
7 |
-
|
8 |
-
|
9 |
-
def get_timestep_embedding(
|
10 |
-
timesteps: torch.Tensor, emb_dim: int, max_period: int = 10000
|
11 |
-
) -> torch.Tensor:
|
12 |
-
half_dim = emb_dim // 2
|
13 |
-
|
14 |
-
emb = math.log(max_period) / (half_dim - 1)
|
15 |
-
emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
|
16 |
-
emb = emb.to(device=timesteps.device)
|
17 |
-
|
18 |
-
emb = timesteps.float()[:, None] * emb[None, :]
|
19 |
-
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
|
20 |
-
|
21 |
-
if emb_dim % 2 == 1:
|
22 |
-
emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
|
23 |
-
|
24 |
-
return emb
|
25 |
-
|
26 |
-
|
27 |
-
class GroupNorm(nn.Module):
|
28 |
-
def __init__(self, in_channels: int) -> None:
|
29 |
-
super().__init__()
|
30 |
-
|
31 |
-
self.group_norm = nn.GroupNorm(
|
32 |
-
num_groups=32, num_channels=in_channels, eps=1e-06, affine=True
|
33 |
-
)
|
34 |
-
|
35 |
-
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
36 |
-
return self.group_norm(x)
|
37 |
-
|
38 |
-
|
39 |
-
class UpSampleBlock(nn.Module):
|
40 |
-
def __init__(self, channels: int):
|
41 |
-
super().__init__()
|
42 |
-
self.conv = nn.Conv2d(channels, channels, 3, padding=1)
|
43 |
-
|
44 |
-
def forward(self, x: torch.Tensor):
|
45 |
-
x = F.interpolate(x, scale_factor=2.0, mode="nearest")
|
46 |
-
return self.conv(x)
|
47 |
-
|
48 |
-
|
49 |
-
class DownSampleBlock(nn.Module):
|
50 |
-
def __init__(self, channels: int):
|
51 |
-
super().__init__()
|
52 |
-
self.conv = nn.Conv2d(channels, channels, 3, stride=2, padding=0)
|
53 |
-
|
54 |
-
def forward(self, x: torch.Tensor):
|
55 |
-
x = F.pad(x, (0, 1, 0, 1), mode="constant", value=0)
|
56 |
-
return self.conv(x)
|
57 |
-
|
58 |
-
|
59 |
-
class TimestepBlock(nn.Module):
|
60 |
-
@abstractmethod
|
61 |
-
def forward(self, x: torch.Tensor, t_emb: torch.Tensor) -> torch.Tensor:
|
62 |
-
pass
|
63 |
-
|
64 |
-
|
65 |
-
class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
|
66 |
-
def forward(self, x: torch.Tensor, t_emb: torch.Tensor) -> torch.Tensor:
|
67 |
-
for layer in self:
|
68 |
-
if isinstance(layer, TimestepBlock):
|
69 |
-
x = layer(x, t_emb)
|
70 |
-
else:
|
71 |
-
x = layer(x)
|
72 |
-
return x
|
73 |
-
|
74 |
-
|
75 |
-
class ResnetBlock(nn.Module):
|
76 |
-
|
77 |
-
def __init__(
|
78 |
-
self,
|
79 |
-
in_channels: int,
|
80 |
-
out_channels: int = None,
|
81 |
-
t_emb_dim: int = None,
|
82 |
-
dropout: float = 0.0,
|
83 |
-
):
|
84 |
-
super().__init__()
|
85 |
-
|
86 |
-
if out_channels is None:
|
87 |
-
out_channels = in_channels
|
88 |
-
|
89 |
-
self.input_layers = nn.Sequential(
|
90 |
-
GroupNorm(in_channels),
|
91 |
-
nn.SiLU(),
|
92 |
-
nn.Conv2d(in_channels, out_channels, 3, padding=1),
|
93 |
-
)
|
94 |
-
|
95 |
-
if t_emb_dim is not None:
|
96 |
-
self.t_emb_layers = nn.Sequential(
|
97 |
-
nn.SiLU(),
|
98 |
-
nn.Linear(t_emb_dim, out_channels),
|
99 |
-
)
|
100 |
-
else:
|
101 |
-
self.t_emb_layers = None
|
102 |
-
|
103 |
-
self.output_layers = nn.Sequential(
|
104 |
-
GroupNorm(out_channels),
|
105 |
-
nn.SiLU(),
|
106 |
-
nn.Dropout(dropout),
|
107 |
-
nn.Conv2d(out_channels, out_channels, 3, padding=1),
|
108 |
-
)
|
109 |
-
|
110 |
-
if in_channels != out_channels:
|
111 |
-
self.skip = nn.Conv2d(in_channels, out_channels, 1)
|
112 |
-
else:
|
113 |
-
self.skip = nn.Identity()
|
114 |
-
|
115 |
-
def forward(self, x: torch.Tensor, t: torch.Tensor = None) -> torch.Tensor:
|
116 |
-
assert t is not None or self.t_emb_layers is None
|
117 |
-
|
118 |
-
h = self.input_layers(x)
|
119 |
-
|
120 |
-
if self.t_emb_layers is not None:
|
121 |
-
t_emb = self.t_emb_layers(t)
|
122 |
-
h = h + t_emb[:, :, None, None]
|
123 |
-
|
124 |
-
h = self.output_layers(h)
|
125 |
-
|
126 |
-
h = h + self.skip(x)
|
127 |
-
|
128 |
-
return h
|
129 |
-
|
130 |
-
|
131 |
-
class AttentionBlock(nn.Module):
|
132 |
-
def __init__(self, in_channels: int) -> None:
|
133 |
-
super().__init__()
|
134 |
-
|
135 |
-
self.in_channels = in_channels
|
136 |
-
|
137 |
-
# normalization layer
|
138 |
-
self.norm = GroupNorm(in_channels)
|
139 |
-
|
140 |
-
# query, key and value layers
|
141 |
-
self.q = nn.Conv2d(in_channels, in_channels, 1, 1, 0)
|
142 |
-
self.k = nn.Conv2d(in_channels, in_channels, 1, 1, 0)
|
143 |
-
self.v = nn.Conv2d(in_channels, in_channels, 1, 1, 0)
|
144 |
-
|
145 |
-
self.project_out = nn.Conv2d(in_channels, in_channels, 1, 1, 0)
|
146 |
-
|
147 |
-
self.softmax = nn.Softmax(dim=2)
|
148 |
-
|
149 |
-
def forward(self, x):
|
150 |
-
|
151 |
-
batch, _, height, width = x.size()
|
152 |
-
|
153 |
-
x = self.norm(x)
|
154 |
-
|
155 |
-
# query, key and value layers
|
156 |
-
q = self.q(x)
|
157 |
-
k = self.k(x)
|
158 |
-
v = self.v(x)
|
159 |
-
|
160 |
-
# resizing the output from 4D to 3D to generate attention map
|
161 |
-
q = q.reshape(batch, self.in_channels, height * width)
|
162 |
-
k = k.reshape(batch, self.in_channels, height * width)
|
163 |
-
v = v.reshape(batch, self.in_channels, height * width)
|
164 |
-
|
165 |
-
# transpose the query tensor for dot product
|
166 |
-
q = q.permute(0, 2, 1)
|
167 |
-
|
168 |
-
# main attention formula
|
169 |
-
scores = torch.bmm(q, k) * (self.in_channels**-0.5)
|
170 |
-
weights = self.softmax(scores)
|
171 |
-
weights = weights.permute(0, 2, 1)
|
172 |
-
|
173 |
-
attention = torch.bmm(v, weights)
|
174 |
-
|
175 |
-
# resizing the output from 3D to 4D to match the input
|
176 |
-
attention = attention.reshape(batch, self.in_channels, height, width)
|
177 |
-
attention = self.project_out(attention)
|
178 |
-
|
179 |
-
# adding the identity to the output
|
180 |
-
return x + attention
|
181 |
-
|
182 |
-
|
183 |
-
class AttentionBlock(nn.Module):
|
184 |
-
def __init__(self, channels: int):
|
185 |
-
super().__init__()
|
186 |
-
# Group normalization
|
187 |
-
self.norm = GroupNorm(channels)
|
188 |
-
# Query, key and value mappings
|
189 |
-
self.q = nn.Conv2d(channels, channels, 1)
|
190 |
-
self.k = nn.Conv2d(channels, channels, 1)
|
191 |
-
self.v = nn.Conv2d(channels, channels, 1)
|
192 |
-
|
193 |
-
self.proj_out = nn.Conv2d(channels, channels, 1)
|
194 |
-
|
195 |
-
# Attention scaling factor
|
196 |
-
self.scale = channels**-0.5
|
197 |
-
|
198 |
-
def forward(self, x: torch.Tensor):
|
199 |
-
# Normalize `x`
|
200 |
-
x_norm = self.norm(x)
|
201 |
-
# Get query, key and vector embeddings
|
202 |
-
q = self.q(x_norm)
|
203 |
-
k = self.k(x_norm)
|
204 |
-
v = self.v(x_norm)
|
205 |
-
|
206 |
-
# Reshape to query, key and vector embeedings from
|
207 |
-
# `[batch_size, channels, height, width]` to
|
208 |
-
# `[batch_size, channels, height * width]`
|
209 |
-
b, c, h, w = q.shape
|
210 |
-
q = q.view(b, c, h * w)
|
211 |
-
k = k.view(b, c, h * w)
|
212 |
-
v = v.view(b, c, h * w)
|
213 |
-
|
214 |
-
# Compute $\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)$
|
215 |
-
attn = torch.einsum("bci,bcj->bij", q, k) * self.scale
|
216 |
-
attn = F.softmax(attn, dim=2)
|
217 |
-
|
218 |
-
# Compute $\underset{seq}{softmax}\Bigg(\frac{Q K^\top}{\sqrt{d_{key}}}\Bigg)V$
|
219 |
-
out = torch.einsum("bij,bcj->bci", attn, v)
|
220 |
-
|
221 |
-
# Reshape back to `[batch_size, channels, height, width]`
|
222 |
-
out = out.view(b, c, h, w)
|
223 |
-
# Final $1 \times 1$ convolution layer
|
224 |
-
out = self.proj_out(out)
|
225 |
-
|
226 |
-
# Add residual connection
|
227 |
-
return x + out
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
swim/codeblock.py
DELETED
@@ -1,74 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
import torch.nn as nn
|
3 |
-
|
4 |
-
|
5 |
-
class SwimCodeBook(nn.Module):
|
6 |
-
def __init__(
|
7 |
-
self, num_codebook_vectors: int = 1024, latent_dim: int = 256, beta: int = 0.25
|
8 |
-
):
|
9 |
-
super().__init__()
|
10 |
-
|
11 |
-
self.num_codebook_vectors = num_codebook_vectors
|
12 |
-
self.latent_dim = latent_dim
|
13 |
-
self.beta = beta
|
14 |
-
|
15 |
-
# creating the codebook, nn.Embedding here is simply a 2D array mainly for storing our embeddings, it's also learnable
|
16 |
-
self.codebook = nn.Embedding(num_codebook_vectors, latent_dim)
|
17 |
-
|
18 |
-
# Initializing the weights in codebook in uniform distribution
|
19 |
-
self.codebook.weight.data.uniform_(
|
20 |
-
-1 / num_codebook_vectors, 1 / num_codebook_vectors
|
21 |
-
)
|
22 |
-
|
23 |
-
def forward(self, z: torch.Tensor) -> torch.Tensor:
|
24 |
-
# Channel to last dimension and copying the tensor to store it in a contiguous ( in a sequence ) way
|
25 |
-
z = z.permute(0, 2, 3, 1).contiguous()
|
26 |
-
|
27 |
-
z_flattened = z.view(
|
28 |
-
-1, self.latent_dim
|
29 |
-
) # b*h*w * latent_dim, will look similar to codebook in fig 2 of the paper
|
30 |
-
|
31 |
-
# calculating the distance between the z to the vectors in flattened codebook, from eq. 2
|
32 |
-
# (a - b)^2 = a^2 + b^2 - 2ab
|
33 |
-
distance = (
|
34 |
-
torch.sum(
|
35 |
-
z_flattened**2, dim=1, keepdim=True
|
36 |
-
) # keepdim = True to keep the same original shape after the sum
|
37 |
-
+ torch.sum(self.codebook.weight**2, dim=1)
|
38 |
-
- 2
|
39 |
-
* torch.matmul(
|
40 |
-
z_flattened, self.codebook.weight.t()
|
41 |
-
) # 2*dot(z, codebook.T)
|
42 |
-
)
|
43 |
-
|
44 |
-
# getting indices of vectors with minimum distance from the codebook
|
45 |
-
min_distance_indices = torch.argmin(distance, dim=1)
|
46 |
-
|
47 |
-
# getting the corresponding vector from the codebook
|
48 |
-
z_q = self.codebook(min_distance_indices).view(z.shape)
|
49 |
-
|
50 |
-
"""
|
51 |
-
this represent the equation 4 from the paper ( except the reconstruction loss ) . Thia loss will then be added
|
52 |
-
to GAN loss to create the final loss function for VQGAN, eq. 6 in the paper.
|
53 |
-
|
54 |
-
|
55 |
-
Note : In the first para of A. Changlog section of the paper,
|
56 |
-
they found a bug which resulted in beta equal to 1. here https://github.com/CompVis/taming-transformers/issues/57
|
57 |
-
just a note :)
|
58 |
-
"""
|
59 |
-
loss = torch.mean(
|
60 |
-
(z_q.detach() - z) ** 2
|
61 |
-
# detach() to avoid calculating gradient while backpropagating
|
62 |
-
+ self.beta
|
63 |
-
* torch.mean(
|
64 |
-
(z_q - z.detach()) ** 2
|
65 |
-
) # commitment loss, detach() to avoid calculating gradient while backpropagating
|
66 |
-
)
|
67 |
-
|
68 |
-
# Not sure why we need this, but it's in the original implementation and mentions for "preserving gradients"
|
69 |
-
z_q = z + (z_q - z).detach()
|
70 |
-
|
71 |
-
# reshapring to the original shape
|
72 |
-
z_q = z_q.permute(0, 3, 1, 2)
|
73 |
-
|
74 |
-
return z_q, min_distance_indices, loss
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
swim/discriminator.py
DELETED
@@ -1,45 +0,0 @@
|
|
1 |
-
import torch.nn as nn
|
2 |
-
|
3 |
-
|
4 |
-
class Discriminator(nn.Module):
|
5 |
-
"""PatchGAN Discriminator
|
6 |
-
|
7 |
-
|
8 |
-
Args:
|
9 |
-
image_channels (int): Number of channels in the input image.
|
10 |
-
num_filters_last (int): Number of filters in the last layer of the discriminator.
|
11 |
-
n_layers (int): Number of layers in the discriminator.
|
12 |
-
|
13 |
-
|
14 |
-
"""
|
15 |
-
|
16 |
-
def __init__(self, image_channels: int = 3, num_filters_last=64, n_layers=3):
|
17 |
-
super(Discriminator, self).__init__()
|
18 |
-
|
19 |
-
layers = [
|
20 |
-
nn.Conv2d(image_channels, num_filters_last, 4, 2, 1),
|
21 |
-
nn.LeakyReLU(0.2),
|
22 |
-
]
|
23 |
-
num_filters_mult = 1
|
24 |
-
|
25 |
-
for i in range(1, n_layers + 1):
|
26 |
-
num_filters_mult_last = num_filters_mult
|
27 |
-
num_filters_mult = min(2**i, 8)
|
28 |
-
layers += [
|
29 |
-
nn.Conv2d(
|
30 |
-
num_filters_last * num_filters_mult_last,
|
31 |
-
num_filters_last * num_filters_mult,
|
32 |
-
4,
|
33 |
-
2 if i < n_layers else 1,
|
34 |
-
1,
|
35 |
-
bias=False,
|
36 |
-
),
|
37 |
-
nn.BatchNorm2d(num_filters_last * num_filters_mult),
|
38 |
-
nn.LeakyReLU(0.2, True),
|
39 |
-
]
|
40 |
-
|
41 |
-
layers.append(nn.Conv2d(num_filters_last * num_filters_mult, 1, 4, 1, 1))
|
42 |
-
self.model = nn.Sequential(*layers)
|
43 |
-
|
44 |
-
def forward(self, x):
|
45 |
-
return self.model(x)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
swim/lr_scheduler.py
ADDED
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
|
3 |
+
|
4 |
+
class LambdaWarmUpCosineScheduler:
|
5 |
+
"""
|
6 |
+
note: use with a base_lr of 1.0
|
7 |
+
"""
|
8 |
+
def __init__(self, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0):
|
9 |
+
self.lr_warm_up_steps = warm_up_steps
|
10 |
+
self.lr_start = lr_start
|
11 |
+
self.lr_min = lr_min
|
12 |
+
self.lr_max = lr_max
|
13 |
+
self.lr_max_decay_steps = max_decay_steps
|
14 |
+
self.last_lr = 0.
|
15 |
+
self.verbosity_interval = verbosity_interval
|
16 |
+
|
17 |
+
def schedule(self, n, **kwargs):
|
18 |
+
if self.verbosity_interval > 0:
|
19 |
+
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}")
|
20 |
+
if n < self.lr_warm_up_steps:
|
21 |
+
lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start
|
22 |
+
self.last_lr = lr
|
23 |
+
return lr
|
24 |
+
else:
|
25 |
+
t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps)
|
26 |
+
t = min(t, 1.0)
|
27 |
+
lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * (
|
28 |
+
1 + np.cos(t * np.pi))
|
29 |
+
self.last_lr = lr
|
30 |
+
return lr
|
31 |
+
|
32 |
+
def __call__(self, n, **kwargs):
|
33 |
+
return self.schedule(n,**kwargs)
|
34 |
+
|
35 |
+
|
36 |
+
class LambdaWarmUpCosineScheduler2:
|
37 |
+
"""
|
38 |
+
supports repeated iterations, configurable via lists
|
39 |
+
note: use with a base_lr of 1.0.
|
40 |
+
"""
|
41 |
+
def __init__(self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0):
|
42 |
+
assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths)
|
43 |
+
self.lr_warm_up_steps = warm_up_steps
|
44 |
+
self.f_start = f_start
|
45 |
+
self.f_min = f_min
|
46 |
+
self.f_max = f_max
|
47 |
+
self.cycle_lengths = cycle_lengths
|
48 |
+
self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths))
|
49 |
+
self.last_f = 0.
|
50 |
+
self.verbosity_interval = verbosity_interval
|
51 |
+
|
52 |
+
def find_in_interval(self, n):
|
53 |
+
interval = 0
|
54 |
+
for cl in self.cum_cycles[1:]:
|
55 |
+
if n <= cl:
|
56 |
+
return interval
|
57 |
+
interval += 1
|
58 |
+
|
59 |
+
def schedule(self, n, **kwargs):
|
60 |
+
cycle = self.find_in_interval(n)
|
61 |
+
n = n - self.cum_cycles[cycle]
|
62 |
+
if self.verbosity_interval > 0:
|
63 |
+
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
|
64 |
+
f"current cycle {cycle}")
|
65 |
+
if n < self.lr_warm_up_steps[cycle]:
|
66 |
+
f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
|
67 |
+
self.last_f = f
|
68 |
+
return f
|
69 |
+
else:
|
70 |
+
t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle])
|
71 |
+
t = min(t, 1.0)
|
72 |
+
f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * (
|
73 |
+
1 + np.cos(t * np.pi))
|
74 |
+
self.last_f = f
|
75 |
+
return f
|
76 |
+
|
77 |
+
def __call__(self, n, **kwargs):
|
78 |
+
return self.schedule(n, **kwargs)
|
79 |
+
|
80 |
+
|
81 |
+
class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2):
|
82 |
+
|
83 |
+
def schedule(self, n, **kwargs):
|
84 |
+
cycle = self.find_in_interval(n)
|
85 |
+
n = n - self.cum_cycles[cycle]
|
86 |
+
if self.verbosity_interval > 0:
|
87 |
+
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
|
88 |
+
f"current cycle {cycle}")
|
89 |
+
|
90 |
+
if n < self.lr_warm_up_steps[cycle]:
|
91 |
+
f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
|
92 |
+
self.last_f = f
|
93 |
+
return f
|
94 |
+
else:
|
95 |
+
f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle])
|
96 |
+
self.last_f = f
|
97 |
+
return f
|
98 |
+
|
swim/models/autoencoder.py
ADDED
@@ -0,0 +1,219 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from lightning import LightningModule
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from contextlib import contextmanager
|
5 |
+
|
6 |
+
from swim.modules.diffusionmodules.model import Encoder, Decoder
|
7 |
+
from swim.modules.distributions.distributions import DiagonalGaussianDistribution
|
8 |
+
|
9 |
+
from swim.utils import instantiate_from_config
|
10 |
+
|
11 |
+
|
12 |
+
class AutoencoderKL(LightningModule):
|
13 |
+
def __init__(
|
14 |
+
self,
|
15 |
+
ddconfig,
|
16 |
+
lossconfig,
|
17 |
+
embed_dim,
|
18 |
+
ckpt_path=None,
|
19 |
+
ignore_keys=[],
|
20 |
+
monitor=None,
|
21 |
+
):
|
22 |
+
super().__init__()
|
23 |
+
self.encoder = Encoder(**ddconfig)
|
24 |
+
self.decoder = Decoder(**ddconfig)
|
25 |
+
self.loss = instantiate_from_config(lossconfig)
|
26 |
+
assert ddconfig["double_z"]
|
27 |
+
self.quant_conv = torch.nn.Conv2d(2 * ddconfig["z_channels"], 2 * embed_dim, 1)
|
28 |
+
self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
|
29 |
+
self.embed_dim = embed_dim
|
30 |
+
|
31 |
+
self.automatic_optimization = False
|
32 |
+
|
33 |
+
if monitor is not None:
|
34 |
+
self.monitor = monitor
|
35 |
+
|
36 |
+
if ckpt_path is not None:
|
37 |
+
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
|
38 |
+
|
39 |
+
def init_from_ckpt(self, path, ignore_keys=list()):
|
40 |
+
sd = torch.load(path, map_location="cpu")["state_dict"]
|
41 |
+
keys = list(sd.keys())
|
42 |
+
for k in keys:
|
43 |
+
for ik in ignore_keys:
|
44 |
+
if k.startswith(ik):
|
45 |
+
print("Deleting key {} from state_dict.".format(k))
|
46 |
+
del sd[k]
|
47 |
+
self.load_state_dict(sd, strict=False)
|
48 |
+
print(f"Restored from {path}")
|
49 |
+
|
50 |
+
def encode(self, x):
|
51 |
+
h = self.encoder(x)
|
52 |
+
moments = self.quant_conv(h)
|
53 |
+
posterior = DiagonalGaussianDistribution(moments)
|
54 |
+
return posterior
|
55 |
+
|
56 |
+
def decode(self, z):
|
57 |
+
z = self.post_quant_conv(z)
|
58 |
+
dec = self.decoder(z)
|
59 |
+
return dec
|
60 |
+
|
61 |
+
def forward(self, input, sample_posterior=True):
|
62 |
+
posterior = self.encode(input)
|
63 |
+
if sample_posterior:
|
64 |
+
z = posterior.sample()
|
65 |
+
else:
|
66 |
+
z = posterior.mode()
|
67 |
+
dec = self.decode(z)
|
68 |
+
return dec, posterior
|
69 |
+
|
70 |
+
def get_input(self, batch, k):
|
71 |
+
x = batch[k]
|
72 |
+
if len(x.shape) == 3:
|
73 |
+
x = x[..., None]
|
74 |
+
x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
|
75 |
+
return x
|
76 |
+
|
77 |
+
def training_step(self, batch, batch_idx):
|
78 |
+
opt_ae, opt_disc = self.optimizers()
|
79 |
+
|
80 |
+
# optimize the autoencoder
|
81 |
+
reconstructions, posterior = self(batch["images"])
|
82 |
+
|
83 |
+
ae_loss, log_dict_ae = self.loss(
|
84 |
+
batch["images"],
|
85 |
+
reconstructions,
|
86 |
+
posterior,
|
87 |
+
0,
|
88 |
+
self.global_step,
|
89 |
+
last_layer=self.get_last_layer(),
|
90 |
+
split="train",
|
91 |
+
)
|
92 |
+
|
93 |
+
opt_ae.zero_grad()
|
94 |
+
self.manual_backward(ae_loss)
|
95 |
+
opt_ae.step()
|
96 |
+
|
97 |
+
self.log(
|
98 |
+
"aeloss",
|
99 |
+
ae_loss,
|
100 |
+
prog_bar=True,
|
101 |
+
logger=True,
|
102 |
+
on_step=True,
|
103 |
+
on_epoch=True,
|
104 |
+
)
|
105 |
+
|
106 |
+
self.log_dict(
|
107 |
+
log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False
|
108 |
+
)
|
109 |
+
|
110 |
+
# optimize the discriminator
|
111 |
+
reconstructions, posterior = self(batch["images"])
|
112 |
+
|
113 |
+
disc_loss, log_dict_disc = self.loss(
|
114 |
+
batch["images"],
|
115 |
+
reconstructions,
|
116 |
+
posterior,
|
117 |
+
1,
|
118 |
+
self.global_step,
|
119 |
+
last_layer=self.get_last_layer(),
|
120 |
+
split="train",
|
121 |
+
)
|
122 |
+
|
123 |
+
opt_disc.zero_grad()
|
124 |
+
self.manual_backward(disc_loss)
|
125 |
+
opt_disc.step()
|
126 |
+
|
127 |
+
self.log(
|
128 |
+
"discloss",
|
129 |
+
disc_loss,
|
130 |
+
prog_bar=True,
|
131 |
+
logger=True,
|
132 |
+
on_step=True,
|
133 |
+
on_epoch=True,
|
134 |
+
)
|
135 |
+
|
136 |
+
self.log_dict(
|
137 |
+
log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False
|
138 |
+
)
|
139 |
+
|
140 |
+
def validation_step(self, batch, batch_idx):
|
141 |
+
reconstructions, posterior = self(batch["images"])
|
142 |
+
aeloss, log_dict_ae = self.loss(
|
143 |
+
batch["images"],
|
144 |
+
reconstructions,
|
145 |
+
posterior,
|
146 |
+
0,
|
147 |
+
self.global_step,
|
148 |
+
last_layer=self.get_last_layer(),
|
149 |
+
split="val",
|
150 |
+
)
|
151 |
+
|
152 |
+
discloss, log_dict_disc = self.loss(
|
153 |
+
batch["images"],
|
154 |
+
reconstructions,
|
155 |
+
posterior,
|
156 |
+
1,
|
157 |
+
self.global_step,
|
158 |
+
last_layer=self.get_last_layer(),
|
159 |
+
split="val",
|
160 |
+
)
|
161 |
+
|
162 |
+
self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
|
163 |
+
self.log_dict(log_dict_ae)
|
164 |
+
self.log_dict(log_dict_disc)
|
165 |
+
|
166 |
+
def configure_optimizers(self):
|
167 |
+
lr = self.learning_rate
|
168 |
+
opt_ae = torch.optim.Adam(
|
169 |
+
list(self.encoder.parameters())
|
170 |
+
+ list(self.decoder.parameters())
|
171 |
+
+ list(self.quant_conv.parameters())
|
172 |
+
+ list(self.post_quant_conv.parameters()),
|
173 |
+
lr=lr,
|
174 |
+
betas=(0.5, 0.9),
|
175 |
+
)
|
176 |
+
opt_disc = torch.optim.Adam(
|
177 |
+
self.loss.discriminator.parameters(), lr=lr, betas=(0.5, 0.9)
|
178 |
+
)
|
179 |
+
return [opt_ae, opt_disc], []
|
180 |
+
|
181 |
+
def get_last_layer(self):
|
182 |
+
return self.decoder.conv_out.weight
|
183 |
+
|
184 |
+
@torch.no_grad()
|
185 |
+
def log_images(self, batch, only_inputs=False, **kwargs):
|
186 |
+
log = dict()
|
187 |
+
x = batch["images"]
|
188 |
+
x = x.to(self.device)
|
189 |
+
if not only_inputs:
|
190 |
+
xrec, posterior = self(x)
|
191 |
+
if x.shape[1] > 3:
|
192 |
+
# colorize with random projection
|
193 |
+
assert xrec.shape[1] > 3
|
194 |
+
x = self.to_rgb(x)
|
195 |
+
xrec = self.to_rgb(xrec)
|
196 |
+
log["samples"] = self.decode(torch.randn_like(posterior.sample()))
|
197 |
+
log["reconstructions"] = xrec
|
198 |
+
log["inputs"] = x
|
199 |
+
return log
|
200 |
+
|
201 |
+
|
202 |
+
class IdentityFirstStage(torch.nn.Module):
|
203 |
+
def __init__(self, *args, vq_interface=False, **kwargs):
|
204 |
+
self.vq_interface = vq_interface # TODO: Should be true by default but check to not break older stuff
|
205 |
+
super().__init__()
|
206 |
+
|
207 |
+
def encode(self, x, *args, **kwargs):
|
208 |
+
return x
|
209 |
+
|
210 |
+
def decode(self, x, *args, **kwargs):
|
211 |
+
return x
|
212 |
+
|
213 |
+
def quantize(self, x, *args, **kwargs):
|
214 |
+
if self.vq_interface:
|
215 |
+
return x, None, [None, None, None]
|
216 |
+
return x
|
217 |
+
|
218 |
+
def forward(self, x, *args, **kwargs):
|
219 |
+
return x
|
swim/modules/attention.py
ADDED
@@ -0,0 +1,277 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from inspect import isfunction
|
2 |
+
import math
|
3 |
+
import torch
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from torch import nn, einsum
|
6 |
+
from einops import rearrange, repeat
|
7 |
+
|
8 |
+
from swim.modules.diffusionmodules.util import checkpoint
|
9 |
+
|
10 |
+
|
11 |
+
def exists(val):
|
12 |
+
return val is not None
|
13 |
+
|
14 |
+
|
15 |
+
def uniq(arr):
|
16 |
+
return {el: True for el in arr}.keys()
|
17 |
+
|
18 |
+
|
19 |
+
def default(val, d):
|
20 |
+
if exists(val):
|
21 |
+
return val
|
22 |
+
return d() if isfunction(d) else d
|
23 |
+
|
24 |
+
|
25 |
+
def max_neg_value(t):
|
26 |
+
return -torch.finfo(t.dtype).max
|
27 |
+
|
28 |
+
|
29 |
+
def init_(tensor):
|
30 |
+
dim = tensor.shape[-1]
|
31 |
+
std = 1 / math.sqrt(dim)
|
32 |
+
tensor.uniform_(-std, std)
|
33 |
+
return tensor
|
34 |
+
|
35 |
+
|
36 |
+
# feedforward
|
37 |
+
class GEGLU(nn.Module):
|
38 |
+
def __init__(self, dim_in, dim_out):
|
39 |
+
super().__init__()
|
40 |
+
self.proj = nn.Linear(dim_in, dim_out * 2)
|
41 |
+
|
42 |
+
def forward(self, x):
|
43 |
+
x, gate = self.proj(x).chunk(2, dim=-1)
|
44 |
+
return x * F.gelu(gate)
|
45 |
+
|
46 |
+
|
47 |
+
class FeedForward(nn.Module):
|
48 |
+
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
|
49 |
+
super().__init__()
|
50 |
+
inner_dim = int(dim * mult)
|
51 |
+
dim_out = default(dim_out, dim)
|
52 |
+
project_in = (
|
53 |
+
nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
|
54 |
+
if not glu
|
55 |
+
else GEGLU(dim, inner_dim)
|
56 |
+
)
|
57 |
+
|
58 |
+
self.net = nn.Sequential(
|
59 |
+
project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
|
60 |
+
)
|
61 |
+
|
62 |
+
def forward(self, x):
|
63 |
+
return self.net(x)
|
64 |
+
|
65 |
+
|
66 |
+
def zero_module(module):
|
67 |
+
"""
|
68 |
+
Zero out the parameters of a module and return it.
|
69 |
+
"""
|
70 |
+
for p in module.parameters():
|
71 |
+
p.detach().zero_()
|
72 |
+
return module
|
73 |
+
|
74 |
+
|
75 |
+
def Normalize(in_channels):
|
76 |
+
return torch.nn.GroupNorm(
|
77 |
+
num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
|
78 |
+
)
|
79 |
+
|
80 |
+
|
81 |
+
class LinearAttention(nn.Module):
|
82 |
+
def __init__(self, dim, heads=4, dim_head=32):
|
83 |
+
super().__init__()
|
84 |
+
self.heads = heads
|
85 |
+
hidden_dim = dim_head * heads
|
86 |
+
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
|
87 |
+
self.to_out = nn.Conv2d(hidden_dim, dim, 1)
|
88 |
+
|
89 |
+
def forward(self, x):
|
90 |
+
b, c, h, w = x.shape
|
91 |
+
qkv = self.to_qkv(x)
|
92 |
+
q, k, v = rearrange(
|
93 |
+
qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3
|
94 |
+
)
|
95 |
+
k = k.softmax(dim=-1)
|
96 |
+
context = torch.einsum("bhdn,bhen->bhde", k, v)
|
97 |
+
out = torch.einsum("bhde,bhdn->bhen", context, q)
|
98 |
+
out = rearrange(
|
99 |
+
out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w
|
100 |
+
)
|
101 |
+
return self.to_out(out)
|
102 |
+
|
103 |
+
|
104 |
+
class SpatialSelfAttention(nn.Module):
|
105 |
+
def __init__(self, in_channels):
|
106 |
+
super().__init__()
|
107 |
+
self.in_channels = in_channels
|
108 |
+
|
109 |
+
self.norm = Normalize(in_channels)
|
110 |
+
self.q = torch.nn.Conv2d(
|
111 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
112 |
+
)
|
113 |
+
self.k = torch.nn.Conv2d(
|
114 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
115 |
+
)
|
116 |
+
self.v = torch.nn.Conv2d(
|
117 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
118 |
+
)
|
119 |
+
self.proj_out = torch.nn.Conv2d(
|
120 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
121 |
+
)
|
122 |
+
|
123 |
+
def forward(self, x):
|
124 |
+
h_ = x
|
125 |
+
h_ = self.norm(h_)
|
126 |
+
q = self.q(h_)
|
127 |
+
k = self.k(h_)
|
128 |
+
v = self.v(h_)
|
129 |
+
|
130 |
+
# compute attention
|
131 |
+
b, c, h, w = q.shape
|
132 |
+
q = rearrange(q, "b c h w -> b (h w) c")
|
133 |
+
k = rearrange(k, "b c h w -> b c (h w)")
|
134 |
+
w_ = torch.einsum("bij,bjk->bik", q, k)
|
135 |
+
|
136 |
+
w_ = w_ * (int(c) ** (-0.5))
|
137 |
+
w_ = torch.nn.functional.softmax(w_, dim=2)
|
138 |
+
|
139 |
+
# attend to values
|
140 |
+
v = rearrange(v, "b c h w -> b c (h w)")
|
141 |
+
w_ = rearrange(w_, "b i j -> b j i")
|
142 |
+
h_ = torch.einsum("bij,bjk->bik", v, w_)
|
143 |
+
h_ = rearrange(h_, "b c (h w) -> b c h w", h=h)
|
144 |
+
h_ = self.proj_out(h_)
|
145 |
+
|
146 |
+
return x + h_
|
147 |
+
|
148 |
+
|
149 |
+
class CrossAttention(nn.Module):
|
150 |
+
def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
|
151 |
+
super().__init__()
|
152 |
+
inner_dim = dim_head * heads
|
153 |
+
context_dim = default(context_dim, query_dim)
|
154 |
+
|
155 |
+
self.scale = dim_head**-0.5
|
156 |
+
self.heads = heads
|
157 |
+
|
158 |
+
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
|
159 |
+
self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
|
160 |
+
self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
|
161 |
+
|
162 |
+
self.to_out = nn.Sequential(
|
163 |
+
nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
|
164 |
+
)
|
165 |
+
|
166 |
+
def forward(self, x, context=None, mask=None):
|
167 |
+
h = self.heads
|
168 |
+
|
169 |
+
q = self.to_q(x)
|
170 |
+
context = default(context, x)
|
171 |
+
k = self.to_k(context)
|
172 |
+
v = self.to_v(context)
|
173 |
+
|
174 |
+
q, k, v = map(lambda t: rearrange(t, "b n (h d) -> (b h) n d", h=h), (q, k, v))
|
175 |
+
|
176 |
+
sim = einsum("b i d, b j d -> b i j", q, k) * self.scale
|
177 |
+
|
178 |
+
if exists(mask):
|
179 |
+
mask = rearrange(mask, "b ... -> b (...)")
|
180 |
+
max_neg_value = -torch.finfo(sim.dtype).max
|
181 |
+
mask = repeat(mask, "b j -> (b h) () j", h=h)
|
182 |
+
sim.masked_fill_(~mask, max_neg_value)
|
183 |
+
|
184 |
+
# attention, what we cannot get enough of
|
185 |
+
attn = sim.softmax(dim=-1)
|
186 |
+
|
187 |
+
out = einsum("b i j, b j d -> b i d", attn, v)
|
188 |
+
out = rearrange(out, "(b h) n d -> b n (h d)", h=h)
|
189 |
+
return self.to_out(out)
|
190 |
+
|
191 |
+
|
192 |
+
class BasicTransformerBlock(nn.Module):
|
193 |
+
def __init__(
|
194 |
+
self,
|
195 |
+
dim,
|
196 |
+
n_heads,
|
197 |
+
d_head,
|
198 |
+
dropout=0.0,
|
199 |
+
context_dim=None,
|
200 |
+
gated_ff=True,
|
201 |
+
checkpoint=True,
|
202 |
+
):
|
203 |
+
super().__init__()
|
204 |
+
self.attn1 = CrossAttention(
|
205 |
+
query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
|
206 |
+
) # is a self-attention
|
207 |
+
self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
|
208 |
+
self.attn2 = CrossAttention(
|
209 |
+
query_dim=dim,
|
210 |
+
context_dim=context_dim,
|
211 |
+
heads=n_heads,
|
212 |
+
dim_head=d_head,
|
213 |
+
dropout=dropout,
|
214 |
+
) # is self-attn if context is none
|
215 |
+
self.norm1 = nn.LayerNorm(dim)
|
216 |
+
self.norm2 = nn.LayerNorm(dim)
|
217 |
+
self.norm3 = nn.LayerNorm(dim)
|
218 |
+
self.checkpoint = checkpoint
|
219 |
+
|
220 |
+
def forward(self, x, context=None):
|
221 |
+
return checkpoint(
|
222 |
+
self._forward, (x, context), self.parameters(), self.checkpoint
|
223 |
+
)
|
224 |
+
|
225 |
+
def _forward(self, x, context=None):
|
226 |
+
x = self.attn1(self.norm1(x)) + x
|
227 |
+
x = self.attn2(self.norm2(x), context=context) + x
|
228 |
+
x = self.ff(self.norm3(x)) + x
|
229 |
+
return x
|
230 |
+
|
231 |
+
|
232 |
+
class SpatialTransformer(nn.Module):
|
233 |
+
"""
|
234 |
+
Transformer block for image-like data.
|
235 |
+
First, project the input (aka embedding)
|
236 |
+
and reshape to b, t, d.
|
237 |
+
Then apply standard transformer action.
|
238 |
+
Finally, reshape to image
|
239 |
+
"""
|
240 |
+
|
241 |
+
def __init__(
|
242 |
+
self, in_channels, n_heads, d_head, depth=1, dropout=0.0, context_dim=None
|
243 |
+
):
|
244 |
+
super().__init__()
|
245 |
+
self.in_channels = in_channels
|
246 |
+
inner_dim = n_heads * d_head
|
247 |
+
self.norm = Normalize(in_channels)
|
248 |
+
|
249 |
+
self.proj_in = nn.Conv2d(
|
250 |
+
in_channels, inner_dim, kernel_size=1, stride=1, padding=0
|
251 |
+
)
|
252 |
+
|
253 |
+
self.transformer_blocks = nn.ModuleList(
|
254 |
+
[
|
255 |
+
BasicTransformerBlock(
|
256 |
+
inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim
|
257 |
+
)
|
258 |
+
for d in range(depth)
|
259 |
+
]
|
260 |
+
)
|
261 |
+
|
262 |
+
self.proj_out = zero_module(
|
263 |
+
nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
|
264 |
+
)
|
265 |
+
|
266 |
+
def forward(self, x, context=None):
|
267 |
+
# note: if no context is given, cross-attention defaults to self-attention
|
268 |
+
b, c, h, w = x.shape
|
269 |
+
x_in = x
|
270 |
+
x = self.norm(x)
|
271 |
+
x = self.proj_in(x)
|
272 |
+
x = rearrange(x, "b c h w -> b (h w) c")
|
273 |
+
for block in self.transformer_blocks:
|
274 |
+
x = block(x, context=context)
|
275 |
+
x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
|
276 |
+
x = self.proj_out(x)
|
277 |
+
return x + x_in
|
swim/modules/dataset.py
ADDED
@@ -0,0 +1,145 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Literal, List
|
2 |
+
|
3 |
+
import os
|
4 |
+
import json
|
5 |
+
import torch
|
6 |
+
import torchvision.transforms as T
|
7 |
+
from torch.utils.data import Dataset, DataLoader
|
8 |
+
from PIL import Image
|
9 |
+
from lightning import LightningDataModule
|
10 |
+
|
11 |
+
|
12 |
+
class SwimDataset(Dataset):
|
13 |
+
def __init__(
|
14 |
+
self,
|
15 |
+
root_dir: str = "./datasets/swim_data",
|
16 |
+
split: Literal["train", "val"] = "train",
|
17 |
+
img_size: int = 512,
|
18 |
+
):
|
19 |
+
super().__init__()
|
20 |
+
self.root_dir = root_dir
|
21 |
+
self.split_dir = os.path.join(root_dir, split)
|
22 |
+
self.img_size = img_size
|
23 |
+
|
24 |
+
if split == "train":
|
25 |
+
self.transform = T.Compose(
|
26 |
+
[
|
27 |
+
T.Resize(img_size), # smaller edge of image resized to img_size
|
28 |
+
T.RandomCrop(img_size), # get a random crop of img_size x img_size
|
29 |
+
T.RandomHorizontalFlip(),
|
30 |
+
T.ToTensor(),
|
31 |
+
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
|
32 |
+
]
|
33 |
+
)
|
34 |
+
elif split == "val":
|
35 |
+
self.transform = T.Compose(
|
36 |
+
[
|
37 |
+
T.Resize(img_size),
|
38 |
+
T.CenterCrop(img_size),
|
39 |
+
T.ToTensor(),
|
40 |
+
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
|
41 |
+
]
|
42 |
+
)
|
43 |
+
|
44 |
+
with open(os.path.join(self.split_dir, "labels.json"), "r") as f:
|
45 |
+
self.data = json.load(f)
|
46 |
+
|
47 |
+
# filter out images that are both at night and have adverse weather conditions
|
48 |
+
self.data = [
|
49 |
+
img
|
50 |
+
for img in self.data
|
51 |
+
if not (img["timeofday"] == "night" and img["weather"] != "clear")
|
52 |
+
]
|
53 |
+
|
54 |
+
def __len__(self):
|
55 |
+
return len(self.data)
|
56 |
+
|
57 |
+
def __getitem__(self, idx):
|
58 |
+
data = self.data[idx]
|
59 |
+
|
60 |
+
# load image
|
61 |
+
img_path = os.path.join(self.split_dir, "images", data["name"])
|
62 |
+
img = Image.open(img_path).convert("RGB")
|
63 |
+
img = self.transform(img)
|
64 |
+
|
65 |
+
# load style
|
66 |
+
if data["weather"] != "clear":
|
67 |
+
style_name = data["weather"]
|
68 |
+
elif data["timeofday"] == "night":
|
69 |
+
style_name = "night"
|
70 |
+
else:
|
71 |
+
style_name = "clear"
|
72 |
+
|
73 |
+
# true if image has any styles
|
74 |
+
style_flag = style_name != "clear"
|
75 |
+
|
76 |
+
# one-hot encode style
|
77 |
+
style = torch.zeros(4)
|
78 |
+
|
79 |
+
if style_flag:
|
80 |
+
style[self.get_stylenames().index(style_name)] = 1
|
81 |
+
|
82 |
+
return {
|
83 |
+
"image": img,
|
84 |
+
"style": style,
|
85 |
+
"style_flag": style_flag,
|
86 |
+
}
|
87 |
+
|
88 |
+
def get_stylenames(self) -> List[str]:
|
89 |
+
return ["rain", "snow", "fog", "night"]
|
90 |
+
|
91 |
+
|
92 |
+
class SwimDataModule(LightningDataModule):
|
93 |
+
def __init__(
|
94 |
+
self,
|
95 |
+
root_dir: str = "./datasets/swim_data",
|
96 |
+
batch_size: int = 1,
|
97 |
+
img_size: int = 512,
|
98 |
+
):
|
99 |
+
super().__init__()
|
100 |
+
self.root_dir = root_dir
|
101 |
+
self.img_size = img_size
|
102 |
+
self.batch_size = batch_size
|
103 |
+
|
104 |
+
def setup(self, stage=None):
|
105 |
+
if stage == "fit" or stage is None:
|
106 |
+
self.train_dataset = SwimDataset(
|
107 |
+
root_dir=self.root_dir, split="train", img_size=self.img_size
|
108 |
+
)
|
109 |
+
self.val_dataset = SwimDataset(
|
110 |
+
root_dir=self.root_dir, split="val", img_size=self.img_size
|
111 |
+
)
|
112 |
+
|
113 |
+
def train_dataloader(self):
|
114 |
+
return DataLoader(
|
115 |
+
self.train_dataset,
|
116 |
+
batch_size=self.batch_size,
|
117 |
+
shuffle=True,
|
118 |
+
num_workers=4,
|
119 |
+
collate_fn=self.custom_collate_fn,
|
120 |
+
)
|
121 |
+
|
122 |
+
def val_dataloader(self):
|
123 |
+
return DataLoader(
|
124 |
+
self.val_dataset,
|
125 |
+
batch_size=self.batch_size,
|
126 |
+
shuffle=False,
|
127 |
+
num_workers=4,
|
128 |
+
collate_fn=self.custom_collate_fn,
|
129 |
+
)
|
130 |
+
|
131 |
+
def test_dataloader(self):
|
132 |
+
return DataLoader(
|
133 |
+
self.val_dataset,
|
134 |
+
batch_size=1,
|
135 |
+
shuffle=False,
|
136 |
+
num_workers=4,
|
137 |
+
collate_fn=self.custom_collate_fn,
|
138 |
+
)
|
139 |
+
|
140 |
+
@staticmethod
|
141 |
+
def custom_collate_fn(batch):
|
142 |
+
images = torch.stack([item["image"] for item in batch])
|
143 |
+
styles = torch.stack([item["style"] for item in batch])
|
144 |
+
style_flags = [item["style_flag"] for item in batch]
|
145 |
+
return {"images": images, "styles": styles, "style_flags": style_flags}
|
swim/{__init__.py → modules/diffusionmodules/__init__.py}
RENAMED
File without changes
|
swim/modules/diffusionmodules/model.py
ADDED
@@ -0,0 +1,1010 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# pytorch_diffusion + derived encoder decoder
|
2 |
+
import math
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import numpy as np
|
6 |
+
from einops import rearrange
|
7 |
+
|
8 |
+
from swim.utils import instantiate_from_config
|
9 |
+
from swim.modules.distributions.distributions import DiagonalGaussianDistribution
|
10 |
+
from swim.modules.attention import LinearAttention
|
11 |
+
|
12 |
+
|
13 |
+
def get_timestep_embedding(timesteps, embedding_dim):
|
14 |
+
"""
|
15 |
+
This matches the implementation in Denoising Diffusion Probabilistic Models:
|
16 |
+
From Fairseq.
|
17 |
+
Build sinusoidal embeddings.
|
18 |
+
This matches the implementation in tensor2tensor, but differs slightly
|
19 |
+
from the description in Section 3.5 of "Attention Is All You Need".
|
20 |
+
"""
|
21 |
+
assert len(timesteps.shape) == 1
|
22 |
+
|
23 |
+
half_dim = embedding_dim // 2
|
24 |
+
emb = math.log(10000) / (half_dim - 1)
|
25 |
+
emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
|
26 |
+
emb = emb.to(device=timesteps.device)
|
27 |
+
emb = timesteps.float()[:, None] * emb[None, :]
|
28 |
+
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
|
29 |
+
if embedding_dim % 2 == 1: # zero pad
|
30 |
+
emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
|
31 |
+
return emb
|
32 |
+
|
33 |
+
|
34 |
+
def nonlinearity(x):
|
35 |
+
# swish
|
36 |
+
return x * torch.sigmoid(x)
|
37 |
+
|
38 |
+
|
39 |
+
def Normalize(in_channels, num_groups=32):
|
40 |
+
return torch.nn.GroupNorm(
|
41 |
+
num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
|
42 |
+
)
|
43 |
+
|
44 |
+
|
45 |
+
class Upsample(nn.Module):
|
46 |
+
def __init__(self, in_channels, with_conv):
|
47 |
+
super().__init__()
|
48 |
+
self.with_conv = with_conv
|
49 |
+
if self.with_conv:
|
50 |
+
self.conv = torch.nn.Conv2d(
|
51 |
+
in_channels, in_channels, kernel_size=3, stride=1, padding=1
|
52 |
+
)
|
53 |
+
|
54 |
+
def forward(self, x):
|
55 |
+
x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
|
56 |
+
if self.with_conv:
|
57 |
+
x = self.conv(x)
|
58 |
+
return x
|
59 |
+
|
60 |
+
|
61 |
+
class Downsample(nn.Module):
|
62 |
+
def __init__(self, in_channels, with_conv):
|
63 |
+
super().__init__()
|
64 |
+
self.with_conv = with_conv
|
65 |
+
if self.with_conv:
|
66 |
+
# no asymmetric padding in torch conv, must do it ourselves
|
67 |
+
self.conv = torch.nn.Conv2d(
|
68 |
+
in_channels, in_channels, kernel_size=3, stride=2, padding=0
|
69 |
+
)
|
70 |
+
|
71 |
+
def forward(self, x):
|
72 |
+
if self.with_conv:
|
73 |
+
pad = (0, 1, 0, 1)
|
74 |
+
x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
|
75 |
+
x = self.conv(x)
|
76 |
+
else:
|
77 |
+
x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
|
78 |
+
return x
|
79 |
+
|
80 |
+
|
81 |
+
class ResnetBlock(nn.Module):
|
82 |
+
def __init__(
|
83 |
+
self,
|
84 |
+
*,
|
85 |
+
in_channels,
|
86 |
+
out_channels=None,
|
87 |
+
conv_shortcut=False,
|
88 |
+
dropout,
|
89 |
+
temb_channels=512,
|
90 |
+
):
|
91 |
+
super().__init__()
|
92 |
+
self.in_channels = in_channels
|
93 |
+
out_channels = in_channels if out_channels is None else out_channels
|
94 |
+
self.out_channels = out_channels
|
95 |
+
self.use_conv_shortcut = conv_shortcut
|
96 |
+
|
97 |
+
self.norm1 = Normalize(in_channels)
|
98 |
+
self.conv1 = torch.nn.Conv2d(
|
99 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
100 |
+
)
|
101 |
+
if temb_channels > 0:
|
102 |
+
self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
|
103 |
+
self.norm2 = Normalize(out_channels)
|
104 |
+
self.dropout = torch.nn.Dropout(dropout)
|
105 |
+
self.conv2 = torch.nn.Conv2d(
|
106 |
+
out_channels, out_channels, kernel_size=3, stride=1, padding=1
|
107 |
+
)
|
108 |
+
if self.in_channels != self.out_channels:
|
109 |
+
if self.use_conv_shortcut:
|
110 |
+
self.conv_shortcut = torch.nn.Conv2d(
|
111 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
112 |
+
)
|
113 |
+
else:
|
114 |
+
self.nin_shortcut = torch.nn.Conv2d(
|
115 |
+
in_channels, out_channels, kernel_size=1, stride=1, padding=0
|
116 |
+
)
|
117 |
+
|
118 |
+
def forward(self, x, temb):
|
119 |
+
h = x
|
120 |
+
h = self.norm1(h)
|
121 |
+
h = nonlinearity(h)
|
122 |
+
h = self.conv1(h)
|
123 |
+
|
124 |
+
if temb is not None:
|
125 |
+
h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
|
126 |
+
|
127 |
+
h = self.norm2(h)
|
128 |
+
h = nonlinearity(h)
|
129 |
+
h = self.dropout(h)
|
130 |
+
h = self.conv2(h)
|
131 |
+
|
132 |
+
if self.in_channels != self.out_channels:
|
133 |
+
if self.use_conv_shortcut:
|
134 |
+
x = self.conv_shortcut(x)
|
135 |
+
else:
|
136 |
+
x = self.nin_shortcut(x)
|
137 |
+
|
138 |
+
return x + h
|
139 |
+
|
140 |
+
|
141 |
+
class LinAttnBlock(LinearAttention):
|
142 |
+
"""to match AttnBlock usage"""
|
143 |
+
|
144 |
+
def __init__(self, in_channels):
|
145 |
+
super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
|
146 |
+
|
147 |
+
|
148 |
+
class AttnBlock(nn.Module):
|
149 |
+
def __init__(self, in_channels):
|
150 |
+
super().__init__()
|
151 |
+
self.in_channels = in_channels
|
152 |
+
|
153 |
+
self.norm = Normalize(in_channels)
|
154 |
+
self.q = torch.nn.Conv2d(
|
155 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
156 |
+
)
|
157 |
+
self.k = torch.nn.Conv2d(
|
158 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
159 |
+
)
|
160 |
+
self.v = torch.nn.Conv2d(
|
161 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
162 |
+
)
|
163 |
+
self.proj_out = torch.nn.Conv2d(
|
164 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
165 |
+
)
|
166 |
+
|
167 |
+
def forward(self, x):
|
168 |
+
h_ = x
|
169 |
+
h_ = self.norm(h_)
|
170 |
+
q = self.q(h_)
|
171 |
+
k = self.k(h_)
|
172 |
+
v = self.v(h_)
|
173 |
+
|
174 |
+
# compute attention
|
175 |
+
b, c, h, w = q.shape
|
176 |
+
q = q.reshape(b, c, h * w)
|
177 |
+
q = q.permute(0, 2, 1) # b,hw,c
|
178 |
+
k = k.reshape(b, c, h * w) # b,c,hw
|
179 |
+
w_ = torch.bmm(q, k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
|
180 |
+
w_ = w_ * (int(c) ** (-0.5))
|
181 |
+
w_ = torch.nn.functional.softmax(w_, dim=2)
|
182 |
+
|
183 |
+
# attend to values
|
184 |
+
v = v.reshape(b, c, h * w)
|
185 |
+
w_ = w_.permute(0, 2, 1) # b,hw,hw (first hw of k, second of q)
|
186 |
+
h_ = torch.bmm(v, w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
|
187 |
+
h_ = h_.reshape(b, c, h, w)
|
188 |
+
|
189 |
+
h_ = self.proj_out(h_)
|
190 |
+
|
191 |
+
return x + h_
|
192 |
+
|
193 |
+
|
194 |
+
def make_attn(in_channels, attn_type="vanilla"):
|
195 |
+
assert attn_type in ["vanilla", "linear", "none"], f"attn_type {attn_type} unknown"
|
196 |
+
print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
|
197 |
+
if attn_type == "vanilla":
|
198 |
+
return AttnBlock(in_channels)
|
199 |
+
elif attn_type == "none":
|
200 |
+
return nn.Identity(in_channels)
|
201 |
+
else:
|
202 |
+
return LinAttnBlock(in_channels)
|
203 |
+
|
204 |
+
|
205 |
+
class Model(nn.Module):
|
206 |
+
def __init__(
|
207 |
+
self,
|
208 |
+
*,
|
209 |
+
ch,
|
210 |
+
out_ch,
|
211 |
+
ch_mult=(1, 2, 4, 8),
|
212 |
+
num_res_blocks,
|
213 |
+
attn_resolutions,
|
214 |
+
dropout=0.0,
|
215 |
+
resamp_with_conv=True,
|
216 |
+
in_channels,
|
217 |
+
resolution,
|
218 |
+
use_timestep=True,
|
219 |
+
use_linear_attn=False,
|
220 |
+
attn_type="vanilla",
|
221 |
+
):
|
222 |
+
super().__init__()
|
223 |
+
if use_linear_attn:
|
224 |
+
attn_type = "linear"
|
225 |
+
self.ch = ch
|
226 |
+
self.temb_ch = self.ch * 4
|
227 |
+
self.num_resolutions = len(ch_mult)
|
228 |
+
self.num_res_blocks = num_res_blocks
|
229 |
+
self.resolution = resolution
|
230 |
+
self.in_channels = in_channels
|
231 |
+
|
232 |
+
self.use_timestep = use_timestep
|
233 |
+
if self.use_timestep:
|
234 |
+
# timestep embedding
|
235 |
+
self.temb = nn.Module()
|
236 |
+
self.temb.dense = nn.ModuleList(
|
237 |
+
[
|
238 |
+
torch.nn.Linear(self.ch, self.temb_ch),
|
239 |
+
torch.nn.Linear(self.temb_ch, self.temb_ch),
|
240 |
+
]
|
241 |
+
)
|
242 |
+
|
243 |
+
# downsampling
|
244 |
+
self.conv_in = torch.nn.Conv2d(
|
245 |
+
in_channels, self.ch, kernel_size=3, stride=1, padding=1
|
246 |
+
)
|
247 |
+
|
248 |
+
curr_res = resolution
|
249 |
+
in_ch_mult = (1,) + tuple(ch_mult)
|
250 |
+
self.down = nn.ModuleList()
|
251 |
+
for i_level in range(self.num_resolutions):
|
252 |
+
block = nn.ModuleList()
|
253 |
+
attn = nn.ModuleList()
|
254 |
+
block_in = ch * in_ch_mult[i_level]
|
255 |
+
block_out = ch * ch_mult[i_level]
|
256 |
+
for i_block in range(self.num_res_blocks):
|
257 |
+
block.append(
|
258 |
+
ResnetBlock(
|
259 |
+
in_channels=block_in,
|
260 |
+
out_channels=block_out,
|
261 |
+
temb_channels=self.temb_ch,
|
262 |
+
dropout=dropout,
|
263 |
+
)
|
264 |
+
)
|
265 |
+
block_in = block_out
|
266 |
+
if curr_res in attn_resolutions:
|
267 |
+
attn.append(make_attn(block_in, attn_type=attn_type))
|
268 |
+
down = nn.Module()
|
269 |
+
down.block = block
|
270 |
+
down.attn = attn
|
271 |
+
if i_level != self.num_resolutions - 1:
|
272 |
+
down.downsample = Downsample(block_in, resamp_with_conv)
|
273 |
+
curr_res = curr_res // 2
|
274 |
+
self.down.append(down)
|
275 |
+
|
276 |
+
# middle
|
277 |
+
self.mid = nn.Module()
|
278 |
+
self.mid.block_1 = ResnetBlock(
|
279 |
+
in_channels=block_in,
|
280 |
+
out_channels=block_in,
|
281 |
+
temb_channels=self.temb_ch,
|
282 |
+
dropout=dropout,
|
283 |
+
)
|
284 |
+
self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
|
285 |
+
self.mid.block_2 = ResnetBlock(
|
286 |
+
in_channels=block_in,
|
287 |
+
out_channels=block_in,
|
288 |
+
temb_channels=self.temb_ch,
|
289 |
+
dropout=dropout,
|
290 |
+
)
|
291 |
+
|
292 |
+
# upsampling
|
293 |
+
self.up = nn.ModuleList()
|
294 |
+
for i_level in reversed(range(self.num_resolutions)):
|
295 |
+
block = nn.ModuleList()
|
296 |
+
attn = nn.ModuleList()
|
297 |
+
block_out = ch * ch_mult[i_level]
|
298 |
+
skip_in = ch * ch_mult[i_level]
|
299 |
+
for i_block in range(self.num_res_blocks + 1):
|
300 |
+
if i_block == self.num_res_blocks:
|
301 |
+
skip_in = ch * in_ch_mult[i_level]
|
302 |
+
block.append(
|
303 |
+
ResnetBlock(
|
304 |
+
in_channels=block_in + skip_in,
|
305 |
+
out_channels=block_out,
|
306 |
+
temb_channels=self.temb_ch,
|
307 |
+
dropout=dropout,
|
308 |
+
)
|
309 |
+
)
|
310 |
+
block_in = block_out
|
311 |
+
if curr_res in attn_resolutions:
|
312 |
+
attn.append(make_attn(block_in, attn_type=attn_type))
|
313 |
+
up = nn.Module()
|
314 |
+
up.block = block
|
315 |
+
up.attn = attn
|
316 |
+
if i_level != 0:
|
317 |
+
up.upsample = Upsample(block_in, resamp_with_conv)
|
318 |
+
curr_res = curr_res * 2
|
319 |
+
self.up.insert(0, up) # prepend to get consistent order
|
320 |
+
|
321 |
+
# end
|
322 |
+
self.norm_out = Normalize(block_in)
|
323 |
+
self.conv_out = torch.nn.Conv2d(
|
324 |
+
block_in, out_ch, kernel_size=3, stride=1, padding=1
|
325 |
+
)
|
326 |
+
|
327 |
+
def forward(self, x, t=None, context=None):
|
328 |
+
# assert x.shape[2] == x.shape[3] == self.resolution
|
329 |
+
if context is not None:
|
330 |
+
# assume aligned context, cat along channel axis
|
331 |
+
x = torch.cat((x, context), dim=1)
|
332 |
+
if self.use_timestep:
|
333 |
+
# timestep embedding
|
334 |
+
assert t is not None
|
335 |
+
temb = get_timestep_embedding(t, self.ch)
|
336 |
+
temb = self.temb.dense[0](temb)
|
337 |
+
temb = nonlinearity(temb)
|
338 |
+
temb = self.temb.dense[1](temb)
|
339 |
+
else:
|
340 |
+
temb = None
|
341 |
+
|
342 |
+
# downsampling
|
343 |
+
hs = [self.conv_in(x)]
|
344 |
+
for i_level in range(self.num_resolutions):
|
345 |
+
for i_block in range(self.num_res_blocks):
|
346 |
+
h = self.down[i_level].block[i_block](hs[-1], temb)
|
347 |
+
if len(self.down[i_level].attn) > 0:
|
348 |
+
h = self.down[i_level].attn[i_block](h)
|
349 |
+
hs.append(h)
|
350 |
+
if i_level != self.num_resolutions - 1:
|
351 |
+
hs.append(self.down[i_level].downsample(hs[-1]))
|
352 |
+
|
353 |
+
# middle
|
354 |
+
h = hs[-1]
|
355 |
+
h = self.mid.block_1(h, temb)
|
356 |
+
h = self.mid.attn_1(h)
|
357 |
+
h = self.mid.block_2(h, temb)
|
358 |
+
|
359 |
+
# upsampling
|
360 |
+
for i_level in reversed(range(self.num_resolutions)):
|
361 |
+
for i_block in range(self.num_res_blocks + 1):
|
362 |
+
h = self.up[i_level].block[i_block](
|
363 |
+
torch.cat([h, hs.pop()], dim=1), temb
|
364 |
+
)
|
365 |
+
if len(self.up[i_level].attn) > 0:
|
366 |
+
h = self.up[i_level].attn[i_block](h)
|
367 |
+
if i_level != 0:
|
368 |
+
h = self.up[i_level].upsample(h)
|
369 |
+
|
370 |
+
# end
|
371 |
+
h = self.norm_out(h)
|
372 |
+
h = nonlinearity(h)
|
373 |
+
h = self.conv_out(h)
|
374 |
+
return h
|
375 |
+
|
376 |
+
def get_last_layer(self):
|
377 |
+
return self.conv_out.weight
|
378 |
+
|
379 |
+
|
380 |
+
class Encoder(nn.Module):
|
381 |
+
def __init__(
|
382 |
+
self,
|
383 |
+
*,
|
384 |
+
ch,
|
385 |
+
out_ch,
|
386 |
+
ch_mult=(1, 2, 4, 8),
|
387 |
+
num_res_blocks,
|
388 |
+
attn_resolutions,
|
389 |
+
dropout=0.0,
|
390 |
+
resamp_with_conv=True,
|
391 |
+
in_channels,
|
392 |
+
resolution,
|
393 |
+
z_channels,
|
394 |
+
double_z=True,
|
395 |
+
use_linear_attn=False,
|
396 |
+
attn_type="vanilla",
|
397 |
+
**ignore_kwargs,
|
398 |
+
):
|
399 |
+
super().__init__()
|
400 |
+
if use_linear_attn:
|
401 |
+
attn_type = "linear"
|
402 |
+
self.ch = ch
|
403 |
+
self.temb_ch = 0
|
404 |
+
self.num_resolutions = len(ch_mult)
|
405 |
+
self.num_res_blocks = num_res_blocks
|
406 |
+
self.resolution = resolution
|
407 |
+
self.in_channels = in_channels
|
408 |
+
|
409 |
+
# downsampling
|
410 |
+
self.conv_in = torch.nn.Conv2d(
|
411 |
+
in_channels, self.ch, kernel_size=3, stride=1, padding=1
|
412 |
+
)
|
413 |
+
|
414 |
+
curr_res = resolution
|
415 |
+
in_ch_mult = (1,) + tuple(ch_mult)
|
416 |
+
self.in_ch_mult = in_ch_mult
|
417 |
+
self.down = nn.ModuleList()
|
418 |
+
for i_level in range(self.num_resolutions):
|
419 |
+
block = nn.ModuleList()
|
420 |
+
attn = nn.ModuleList()
|
421 |
+
block_in = ch * in_ch_mult[i_level]
|
422 |
+
block_out = ch * ch_mult[i_level]
|
423 |
+
for i_block in range(self.num_res_blocks):
|
424 |
+
block.append(
|
425 |
+
ResnetBlock(
|
426 |
+
in_channels=block_in,
|
427 |
+
out_channels=block_out,
|
428 |
+
temb_channels=self.temb_ch,
|
429 |
+
dropout=dropout,
|
430 |
+
)
|
431 |
+
)
|
432 |
+
block_in = block_out
|
433 |
+
if curr_res in attn_resolutions:
|
434 |
+
attn.append(make_attn(block_in, attn_type=attn_type))
|
435 |
+
down = nn.Module()
|
436 |
+
down.block = block
|
437 |
+
down.attn = attn
|
438 |
+
if i_level != self.num_resolutions - 1:
|
439 |
+
down.downsample = Downsample(block_in, resamp_with_conv)
|
440 |
+
curr_res = curr_res // 2
|
441 |
+
self.down.append(down)
|
442 |
+
|
443 |
+
# middle
|
444 |
+
self.mid = nn.Module()
|
445 |
+
self.mid.block_1 = ResnetBlock(
|
446 |
+
in_channels=block_in,
|
447 |
+
out_channels=block_in,
|
448 |
+
temb_channels=self.temb_ch,
|
449 |
+
dropout=dropout,
|
450 |
+
)
|
451 |
+
self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
|
452 |
+
self.mid.block_2 = ResnetBlock(
|
453 |
+
in_channels=block_in,
|
454 |
+
out_channels=block_in,
|
455 |
+
temb_channels=self.temb_ch,
|
456 |
+
dropout=dropout,
|
457 |
+
)
|
458 |
+
|
459 |
+
# end
|
460 |
+
self.norm_out = Normalize(block_in)
|
461 |
+
self.conv_out = torch.nn.Conv2d(
|
462 |
+
block_in,
|
463 |
+
2 * z_channels if double_z else z_channels,
|
464 |
+
kernel_size=3,
|
465 |
+
stride=1,
|
466 |
+
padding=1,
|
467 |
+
)
|
468 |
+
|
469 |
+
def forward(self, x):
|
470 |
+
# timestep embedding
|
471 |
+
temb = None
|
472 |
+
|
473 |
+
# downsampling
|
474 |
+
hs = [self.conv_in(x)]
|
475 |
+
for i_level in range(self.num_resolutions):
|
476 |
+
for i_block in range(self.num_res_blocks):
|
477 |
+
h = self.down[i_level].block[i_block](hs[-1], temb)
|
478 |
+
if len(self.down[i_level].attn) > 0:
|
479 |
+
h = self.down[i_level].attn[i_block](h)
|
480 |
+
hs.append(h)
|
481 |
+
if i_level != self.num_resolutions - 1:
|
482 |
+
hs.append(self.down[i_level].downsample(hs[-1]))
|
483 |
+
|
484 |
+
# middle
|
485 |
+
h = hs[-1]
|
486 |
+
h = self.mid.block_1(h, temb)
|
487 |
+
h = self.mid.attn_1(h)
|
488 |
+
h = self.mid.block_2(h, temb)
|
489 |
+
|
490 |
+
# end
|
491 |
+
h = self.norm_out(h)
|
492 |
+
h = nonlinearity(h)
|
493 |
+
h = self.conv_out(h)
|
494 |
+
return h
|
495 |
+
|
496 |
+
|
497 |
+
class Decoder(nn.Module):
|
498 |
+
def __init__(
|
499 |
+
self,
|
500 |
+
*,
|
501 |
+
ch,
|
502 |
+
out_ch,
|
503 |
+
ch_mult=(1, 2, 4, 8),
|
504 |
+
num_res_blocks,
|
505 |
+
attn_resolutions,
|
506 |
+
dropout=0.0,
|
507 |
+
resamp_with_conv=True,
|
508 |
+
in_channels,
|
509 |
+
resolution,
|
510 |
+
z_channels,
|
511 |
+
give_pre_end=False,
|
512 |
+
tanh_out=False,
|
513 |
+
use_linear_attn=False,
|
514 |
+
attn_type="vanilla",
|
515 |
+
**ignorekwargs,
|
516 |
+
):
|
517 |
+
super().__init__()
|
518 |
+
if use_linear_attn:
|
519 |
+
attn_type = "linear"
|
520 |
+
self.ch = ch
|
521 |
+
self.temb_ch = 0
|
522 |
+
self.num_resolutions = len(ch_mult)
|
523 |
+
self.num_res_blocks = num_res_blocks
|
524 |
+
self.resolution = resolution
|
525 |
+
self.in_channels = in_channels
|
526 |
+
self.give_pre_end = give_pre_end
|
527 |
+
self.tanh_out = tanh_out
|
528 |
+
|
529 |
+
# compute in_ch_mult, block_in and curr_res at lowest res
|
530 |
+
in_ch_mult = (1,) + tuple(ch_mult)
|
531 |
+
block_in = ch * ch_mult[self.num_resolutions - 1]
|
532 |
+
curr_res = resolution // 2 ** (self.num_resolutions - 1)
|
533 |
+
self.z_shape = (1, z_channels, curr_res, curr_res)
|
534 |
+
print(
|
535 |
+
"Working with z of shape {} = {} dimensions.".format(
|
536 |
+
self.z_shape, np.prod(self.z_shape)
|
537 |
+
)
|
538 |
+
)
|
539 |
+
|
540 |
+
# z to block_in
|
541 |
+
self.conv_in = torch.nn.Conv2d(
|
542 |
+
z_channels, block_in, kernel_size=3, stride=1, padding=1
|
543 |
+
)
|
544 |
+
|
545 |
+
# middle
|
546 |
+
self.mid = nn.Module()
|
547 |
+
self.mid.block_1 = ResnetBlock(
|
548 |
+
in_channels=block_in,
|
549 |
+
out_channels=block_in,
|
550 |
+
temb_channels=self.temb_ch,
|
551 |
+
dropout=dropout,
|
552 |
+
)
|
553 |
+
self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
|
554 |
+
self.mid.block_2 = ResnetBlock(
|
555 |
+
in_channels=block_in,
|
556 |
+
out_channels=block_in,
|
557 |
+
temb_channels=self.temb_ch,
|
558 |
+
dropout=dropout,
|
559 |
+
)
|
560 |
+
|
561 |
+
# upsampling
|
562 |
+
self.up = nn.ModuleList()
|
563 |
+
for i_level in reversed(range(self.num_resolutions)):
|
564 |
+
block = nn.ModuleList()
|
565 |
+
attn = nn.ModuleList()
|
566 |
+
block_out = ch * ch_mult[i_level]
|
567 |
+
for i_block in range(self.num_res_blocks + 1):
|
568 |
+
block.append(
|
569 |
+
ResnetBlock(
|
570 |
+
in_channels=block_in,
|
571 |
+
out_channels=block_out,
|
572 |
+
temb_channels=self.temb_ch,
|
573 |
+
dropout=dropout,
|
574 |
+
)
|
575 |
+
)
|
576 |
+
block_in = block_out
|
577 |
+
if curr_res in attn_resolutions:
|
578 |
+
attn.append(make_attn(block_in, attn_type=attn_type))
|
579 |
+
up = nn.Module()
|
580 |
+
up.block = block
|
581 |
+
up.attn = attn
|
582 |
+
if i_level != 0:
|
583 |
+
up.upsample = Upsample(block_in, resamp_with_conv)
|
584 |
+
curr_res = curr_res * 2
|
585 |
+
self.up.insert(0, up) # prepend to get consistent order
|
586 |
+
|
587 |
+
# end
|
588 |
+
self.norm_out = Normalize(block_in)
|
589 |
+
self.conv_out = torch.nn.Conv2d(
|
590 |
+
block_in, out_ch, kernel_size=3, stride=1, padding=1
|
591 |
+
)
|
592 |
+
|
593 |
+
def forward(self, z):
|
594 |
+
# assert z.shape[1:] == self.z_shape[1:]
|
595 |
+
self.last_z_shape = z.shape
|
596 |
+
|
597 |
+
# timestep embedding
|
598 |
+
temb = None
|
599 |
+
|
600 |
+
# z to block_in
|
601 |
+
h = self.conv_in(z)
|
602 |
+
|
603 |
+
# middle
|
604 |
+
h = self.mid.block_1(h, temb)
|
605 |
+
h = self.mid.attn_1(h)
|
606 |
+
h = self.mid.block_2(h, temb)
|
607 |
+
|
608 |
+
# upsampling
|
609 |
+
for i_level in reversed(range(self.num_resolutions)):
|
610 |
+
for i_block in range(self.num_res_blocks + 1):
|
611 |
+
h = self.up[i_level].block[i_block](h, temb)
|
612 |
+
if len(self.up[i_level].attn) > 0:
|
613 |
+
h = self.up[i_level].attn[i_block](h)
|
614 |
+
if i_level != 0:
|
615 |
+
h = self.up[i_level].upsample(h)
|
616 |
+
|
617 |
+
# end
|
618 |
+
if self.give_pre_end:
|
619 |
+
return h
|
620 |
+
|
621 |
+
h = self.norm_out(h)
|
622 |
+
h = nonlinearity(h)
|
623 |
+
h = self.conv_out(h)
|
624 |
+
if self.tanh_out:
|
625 |
+
h = torch.tanh(h)
|
626 |
+
return h
|
627 |
+
|
628 |
+
|
629 |
+
class SimpleDecoder(nn.Module):
|
630 |
+
def __init__(self, in_channels, out_channels, *args, **kwargs):
|
631 |
+
super().__init__()
|
632 |
+
self.model = nn.ModuleList(
|
633 |
+
[
|
634 |
+
nn.Conv2d(in_channels, in_channels, 1),
|
635 |
+
ResnetBlock(
|
636 |
+
in_channels=in_channels,
|
637 |
+
out_channels=2 * in_channels,
|
638 |
+
temb_channels=0,
|
639 |
+
dropout=0.0,
|
640 |
+
),
|
641 |
+
ResnetBlock(
|
642 |
+
in_channels=2 * in_channels,
|
643 |
+
out_channels=4 * in_channels,
|
644 |
+
temb_channels=0,
|
645 |
+
dropout=0.0,
|
646 |
+
),
|
647 |
+
ResnetBlock(
|
648 |
+
in_channels=4 * in_channels,
|
649 |
+
out_channels=2 * in_channels,
|
650 |
+
temb_channels=0,
|
651 |
+
dropout=0.0,
|
652 |
+
),
|
653 |
+
nn.Conv2d(2 * in_channels, in_channels, 1),
|
654 |
+
Upsample(in_channels, with_conv=True),
|
655 |
+
]
|
656 |
+
)
|
657 |
+
# end
|
658 |
+
self.norm_out = Normalize(in_channels)
|
659 |
+
self.conv_out = torch.nn.Conv2d(
|
660 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
661 |
+
)
|
662 |
+
|
663 |
+
def forward(self, x):
|
664 |
+
for i, layer in enumerate(self.model):
|
665 |
+
if i in [1, 2, 3]:
|
666 |
+
x = layer(x, None)
|
667 |
+
else:
|
668 |
+
x = layer(x)
|
669 |
+
|
670 |
+
h = self.norm_out(x)
|
671 |
+
h = nonlinearity(h)
|
672 |
+
x = self.conv_out(h)
|
673 |
+
return x
|
674 |
+
|
675 |
+
|
676 |
+
class UpsampleDecoder(nn.Module):
|
677 |
+
def __init__(
|
678 |
+
self,
|
679 |
+
in_channels,
|
680 |
+
out_channels,
|
681 |
+
ch,
|
682 |
+
num_res_blocks,
|
683 |
+
resolution,
|
684 |
+
ch_mult=(2, 2),
|
685 |
+
dropout=0.0,
|
686 |
+
):
|
687 |
+
super().__init__()
|
688 |
+
# upsampling
|
689 |
+
self.temb_ch = 0
|
690 |
+
self.num_resolutions = len(ch_mult)
|
691 |
+
self.num_res_blocks = num_res_blocks
|
692 |
+
block_in = in_channels
|
693 |
+
curr_res = resolution // 2 ** (self.num_resolutions - 1)
|
694 |
+
self.res_blocks = nn.ModuleList()
|
695 |
+
self.upsample_blocks = nn.ModuleList()
|
696 |
+
for i_level in range(self.num_resolutions):
|
697 |
+
res_block = []
|
698 |
+
block_out = ch * ch_mult[i_level]
|
699 |
+
for i_block in range(self.num_res_blocks + 1):
|
700 |
+
res_block.append(
|
701 |
+
ResnetBlock(
|
702 |
+
in_channels=block_in,
|
703 |
+
out_channels=block_out,
|
704 |
+
temb_channels=self.temb_ch,
|
705 |
+
dropout=dropout,
|
706 |
+
)
|
707 |
+
)
|
708 |
+
block_in = block_out
|
709 |
+
self.res_blocks.append(nn.ModuleList(res_block))
|
710 |
+
if i_level != self.num_resolutions - 1:
|
711 |
+
self.upsample_blocks.append(Upsample(block_in, True))
|
712 |
+
curr_res = curr_res * 2
|
713 |
+
|
714 |
+
# end
|
715 |
+
self.norm_out = Normalize(block_in)
|
716 |
+
self.conv_out = torch.nn.Conv2d(
|
717 |
+
block_in, out_channels, kernel_size=3, stride=1, padding=1
|
718 |
+
)
|
719 |
+
|
720 |
+
def forward(self, x):
|
721 |
+
# upsampling
|
722 |
+
h = x
|
723 |
+
for k, i_level in enumerate(range(self.num_resolutions)):
|
724 |
+
for i_block in range(self.num_res_blocks + 1):
|
725 |
+
h = self.res_blocks[i_level][i_block](h, None)
|
726 |
+
if i_level != self.num_resolutions - 1:
|
727 |
+
h = self.upsample_blocks[k](h)
|
728 |
+
h = self.norm_out(h)
|
729 |
+
h = nonlinearity(h)
|
730 |
+
h = self.conv_out(h)
|
731 |
+
return h
|
732 |
+
|
733 |
+
|
734 |
+
class LatentRescaler(nn.Module):
|
735 |
+
def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
|
736 |
+
super().__init__()
|
737 |
+
# residual block, interpolate, residual block
|
738 |
+
self.factor = factor
|
739 |
+
self.conv_in = nn.Conv2d(
|
740 |
+
in_channels, mid_channels, kernel_size=3, stride=1, padding=1
|
741 |
+
)
|
742 |
+
self.res_block1 = nn.ModuleList(
|
743 |
+
[
|
744 |
+
ResnetBlock(
|
745 |
+
in_channels=mid_channels,
|
746 |
+
out_channels=mid_channels,
|
747 |
+
temb_channels=0,
|
748 |
+
dropout=0.0,
|
749 |
+
)
|
750 |
+
for _ in range(depth)
|
751 |
+
]
|
752 |
+
)
|
753 |
+
self.attn = AttnBlock(mid_channels)
|
754 |
+
self.res_block2 = nn.ModuleList(
|
755 |
+
[
|
756 |
+
ResnetBlock(
|
757 |
+
in_channels=mid_channels,
|
758 |
+
out_channels=mid_channels,
|
759 |
+
temb_channels=0,
|
760 |
+
dropout=0.0,
|
761 |
+
)
|
762 |
+
for _ in range(depth)
|
763 |
+
]
|
764 |
+
)
|
765 |
+
|
766 |
+
self.conv_out = nn.Conv2d(
|
767 |
+
mid_channels,
|
768 |
+
out_channels,
|
769 |
+
kernel_size=1,
|
770 |
+
)
|
771 |
+
|
772 |
+
def forward(self, x):
|
773 |
+
x = self.conv_in(x)
|
774 |
+
for block in self.res_block1:
|
775 |
+
x = block(x, None)
|
776 |
+
x = torch.nn.functional.interpolate(
|
777 |
+
x,
|
778 |
+
size=(
|
779 |
+
int(round(x.shape[2] * self.factor)),
|
780 |
+
int(round(x.shape[3] * self.factor)),
|
781 |
+
),
|
782 |
+
)
|
783 |
+
x = self.attn(x)
|
784 |
+
for block in self.res_block2:
|
785 |
+
x = block(x, None)
|
786 |
+
x = self.conv_out(x)
|
787 |
+
return x
|
788 |
+
|
789 |
+
|
790 |
+
class MergedRescaleEncoder(nn.Module):
|
791 |
+
def __init__(
|
792 |
+
self,
|
793 |
+
in_channels,
|
794 |
+
ch,
|
795 |
+
resolution,
|
796 |
+
out_ch,
|
797 |
+
num_res_blocks,
|
798 |
+
attn_resolutions,
|
799 |
+
dropout=0.0,
|
800 |
+
resamp_with_conv=True,
|
801 |
+
ch_mult=(1, 2, 4, 8),
|
802 |
+
rescale_factor=1.0,
|
803 |
+
rescale_module_depth=1,
|
804 |
+
):
|
805 |
+
super().__init__()
|
806 |
+
intermediate_chn = ch * ch_mult[-1]
|
807 |
+
self.encoder = Encoder(
|
808 |
+
in_channels=in_channels,
|
809 |
+
num_res_blocks=num_res_blocks,
|
810 |
+
ch=ch,
|
811 |
+
ch_mult=ch_mult,
|
812 |
+
z_channels=intermediate_chn,
|
813 |
+
double_z=False,
|
814 |
+
resolution=resolution,
|
815 |
+
attn_resolutions=attn_resolutions,
|
816 |
+
dropout=dropout,
|
817 |
+
resamp_with_conv=resamp_with_conv,
|
818 |
+
out_ch=None,
|
819 |
+
)
|
820 |
+
self.rescaler = LatentRescaler(
|
821 |
+
factor=rescale_factor,
|
822 |
+
in_channels=intermediate_chn,
|
823 |
+
mid_channels=intermediate_chn,
|
824 |
+
out_channels=out_ch,
|
825 |
+
depth=rescale_module_depth,
|
826 |
+
)
|
827 |
+
|
828 |
+
def forward(self, x):
|
829 |
+
x = self.encoder(x)
|
830 |
+
x = self.rescaler(x)
|
831 |
+
return x
|
832 |
+
|
833 |
+
|
834 |
+
class MergedRescaleDecoder(nn.Module):
|
835 |
+
def __init__(
|
836 |
+
self,
|
837 |
+
z_channels,
|
838 |
+
out_ch,
|
839 |
+
resolution,
|
840 |
+
num_res_blocks,
|
841 |
+
attn_resolutions,
|
842 |
+
ch,
|
843 |
+
ch_mult=(1, 2, 4, 8),
|
844 |
+
dropout=0.0,
|
845 |
+
resamp_with_conv=True,
|
846 |
+
rescale_factor=1.0,
|
847 |
+
rescale_module_depth=1,
|
848 |
+
):
|
849 |
+
super().__init__()
|
850 |
+
tmp_chn = z_channels * ch_mult[-1]
|
851 |
+
self.decoder = Decoder(
|
852 |
+
out_ch=out_ch,
|
853 |
+
z_channels=tmp_chn,
|
854 |
+
attn_resolutions=attn_resolutions,
|
855 |
+
dropout=dropout,
|
856 |
+
resamp_with_conv=resamp_with_conv,
|
857 |
+
in_channels=None,
|
858 |
+
num_res_blocks=num_res_blocks,
|
859 |
+
ch_mult=ch_mult,
|
860 |
+
resolution=resolution,
|
861 |
+
ch=ch,
|
862 |
+
)
|
863 |
+
self.rescaler = LatentRescaler(
|
864 |
+
factor=rescale_factor,
|
865 |
+
in_channels=z_channels,
|
866 |
+
mid_channels=tmp_chn,
|
867 |
+
out_channels=tmp_chn,
|
868 |
+
depth=rescale_module_depth,
|
869 |
+
)
|
870 |
+
|
871 |
+
def forward(self, x):
|
872 |
+
x = self.rescaler(x)
|
873 |
+
x = self.decoder(x)
|
874 |
+
return x
|
875 |
+
|
876 |
+
|
877 |
+
class Upsampler(nn.Module):
|
878 |
+
def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
|
879 |
+
super().__init__()
|
880 |
+
assert out_size >= in_size
|
881 |
+
num_blocks = int(np.log2(out_size // in_size)) + 1
|
882 |
+
factor_up = 1.0 + (out_size % in_size)
|
883 |
+
print(
|
884 |
+
f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}"
|
885 |
+
)
|
886 |
+
self.rescaler = LatentRescaler(
|
887 |
+
factor=factor_up,
|
888 |
+
in_channels=in_channels,
|
889 |
+
mid_channels=2 * in_channels,
|
890 |
+
out_channels=in_channels,
|
891 |
+
)
|
892 |
+
self.decoder = Decoder(
|
893 |
+
out_ch=out_channels,
|
894 |
+
resolution=out_size,
|
895 |
+
z_channels=in_channels,
|
896 |
+
num_res_blocks=2,
|
897 |
+
attn_resolutions=[],
|
898 |
+
in_channels=None,
|
899 |
+
ch=in_channels,
|
900 |
+
ch_mult=[ch_mult for _ in range(num_blocks)],
|
901 |
+
)
|
902 |
+
|
903 |
+
def forward(self, x):
|
904 |
+
x = self.rescaler(x)
|
905 |
+
x = self.decoder(x)
|
906 |
+
return x
|
907 |
+
|
908 |
+
|
909 |
+
class Resize(nn.Module):
|
910 |
+
def __init__(self, in_channels=None, learned=False, mode="bilinear"):
|
911 |
+
super().__init__()
|
912 |
+
self.with_conv = learned
|
913 |
+
self.mode = mode
|
914 |
+
if self.with_conv:
|
915 |
+
print(
|
916 |
+
f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode"
|
917 |
+
)
|
918 |
+
raise NotImplementedError()
|
919 |
+
assert in_channels is not None
|
920 |
+
# no asymmetric padding in torch conv, must do it ourselves
|
921 |
+
self.conv = torch.nn.Conv2d(
|
922 |
+
in_channels, in_channels, kernel_size=4, stride=2, padding=1
|
923 |
+
)
|
924 |
+
|
925 |
+
def forward(self, x, scale_factor=1.0):
|
926 |
+
if scale_factor == 1.0:
|
927 |
+
return x
|
928 |
+
else:
|
929 |
+
x = torch.nn.functional.interpolate(
|
930 |
+
x, mode=self.mode, align_corners=False, scale_factor=scale_factor
|
931 |
+
)
|
932 |
+
return x
|
933 |
+
|
934 |
+
|
935 |
+
class FirstStagePostProcessor(nn.Module):
|
936 |
+
|
937 |
+
def __init__(
|
938 |
+
self,
|
939 |
+
ch_mult: list,
|
940 |
+
in_channels,
|
941 |
+
pretrained_model: nn.Module = None,
|
942 |
+
reshape=False,
|
943 |
+
n_channels=None,
|
944 |
+
dropout=0.0,
|
945 |
+
pretrained_config=None,
|
946 |
+
):
|
947 |
+
super().__init__()
|
948 |
+
if pretrained_config is None:
|
949 |
+
assert (
|
950 |
+
pretrained_model is not None
|
951 |
+
), 'Either "pretrained_model" or "pretrained_config" must not be None'
|
952 |
+
self.pretrained_model = pretrained_model
|
953 |
+
else:
|
954 |
+
assert (
|
955 |
+
pretrained_config is not None
|
956 |
+
), 'Either "pretrained_model" or "pretrained_config" must not be None'
|
957 |
+
self.instantiate_pretrained(pretrained_config)
|
958 |
+
|
959 |
+
self.do_reshape = reshape
|
960 |
+
|
961 |
+
if n_channels is None:
|
962 |
+
n_channels = self.pretrained_model.encoder.ch
|
963 |
+
|
964 |
+
self.proj_norm = Normalize(in_channels, num_groups=in_channels // 2)
|
965 |
+
self.proj = nn.Conv2d(
|
966 |
+
in_channels, n_channels, kernel_size=3, stride=1, padding=1
|
967 |
+
)
|
968 |
+
|
969 |
+
blocks = []
|
970 |
+
downs = []
|
971 |
+
ch_in = n_channels
|
972 |
+
for m in ch_mult:
|
973 |
+
blocks.append(
|
974 |
+
ResnetBlock(
|
975 |
+
in_channels=ch_in, out_channels=m * n_channels, dropout=dropout
|
976 |
+
)
|
977 |
+
)
|
978 |
+
ch_in = m * n_channels
|
979 |
+
downs.append(Downsample(ch_in, with_conv=False))
|
980 |
+
|
981 |
+
self.model = nn.ModuleList(blocks)
|
982 |
+
self.downsampler = nn.ModuleList(downs)
|
983 |
+
|
984 |
+
def instantiate_pretrained(self, config):
|
985 |
+
model = instantiate_from_config(config)
|
986 |
+
self.pretrained_model = model.eval()
|
987 |
+
# self.pretrained_model.train = False
|
988 |
+
for param in self.pretrained_model.parameters():
|
989 |
+
param.requires_grad = False
|
990 |
+
|
991 |
+
@torch.no_grad()
|
992 |
+
def encode_with_pretrained(self, x):
|
993 |
+
c = self.pretrained_model.encode(x)
|
994 |
+
if isinstance(c, DiagonalGaussianDistribution):
|
995 |
+
c = c.mode()
|
996 |
+
return c
|
997 |
+
|
998 |
+
def forward(self, x):
|
999 |
+
z_fs = self.encode_with_pretrained(x)
|
1000 |
+
z = self.proj_norm(z_fs)
|
1001 |
+
z = self.proj(z)
|
1002 |
+
z = nonlinearity(z)
|
1003 |
+
|
1004 |
+
for submodel, downmodel in zip(self.model, self.downsampler):
|
1005 |
+
z = submodel(z, temb=None)
|
1006 |
+
z = downmodel(z)
|
1007 |
+
|
1008 |
+
if self.do_reshape:
|
1009 |
+
z = rearrange(z, "b c h w -> b (h w) c")
|
1010 |
+
return z
|
swim/modules/diffusionmodules/openaimodel.py
ADDED
@@ -0,0 +1,1012 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from abc import abstractmethod
|
2 |
+
from functools import partial
|
3 |
+
import math
|
4 |
+
from typing import Iterable
|
5 |
+
|
6 |
+
import numpy as np
|
7 |
+
import torch as th
|
8 |
+
import torch.nn as nn
|
9 |
+
import torch.nn.functional as F
|
10 |
+
|
11 |
+
from swim.modules.diffusionmodules.util import (
|
12 |
+
checkpoint,
|
13 |
+
conv_nd,
|
14 |
+
linear,
|
15 |
+
avg_pool_nd,
|
16 |
+
zero_module,
|
17 |
+
normalization,
|
18 |
+
timestep_embedding,
|
19 |
+
)
|
20 |
+
from swim.modules.attention import SpatialTransformer
|
21 |
+
|
22 |
+
|
23 |
+
# dummy replace
|
24 |
+
def convert_module_to_f16(x):
|
25 |
+
pass
|
26 |
+
|
27 |
+
|
28 |
+
def convert_module_to_f32(x):
|
29 |
+
pass
|
30 |
+
|
31 |
+
|
32 |
+
## go
|
33 |
+
class AttentionPool2d(nn.Module):
|
34 |
+
"""
|
35 |
+
Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
|
36 |
+
"""
|
37 |
+
|
38 |
+
def __init__(
|
39 |
+
self,
|
40 |
+
spacial_dim: int,
|
41 |
+
embed_dim: int,
|
42 |
+
num_heads_channels: int,
|
43 |
+
output_dim: int = None,
|
44 |
+
):
|
45 |
+
super().__init__()
|
46 |
+
self.positional_embedding = nn.Parameter(
|
47 |
+
th.randn(embed_dim, spacial_dim**2 + 1) / embed_dim**0.5
|
48 |
+
)
|
49 |
+
self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
|
50 |
+
self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
|
51 |
+
self.num_heads = embed_dim // num_heads_channels
|
52 |
+
self.attention = QKVAttention(self.num_heads)
|
53 |
+
|
54 |
+
def forward(self, x):
|
55 |
+
b, c, *_spatial = x.shape
|
56 |
+
x = x.reshape(b, c, -1) # NC(HW)
|
57 |
+
x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1)
|
58 |
+
x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1)
|
59 |
+
x = self.qkv_proj(x)
|
60 |
+
x = self.attention(x)
|
61 |
+
x = self.c_proj(x)
|
62 |
+
return x[:, :, 0]
|
63 |
+
|
64 |
+
|
65 |
+
class TimestepBlock(nn.Module):
|
66 |
+
"""
|
67 |
+
Any module where forward() takes timestep embeddings as a second argument.
|
68 |
+
"""
|
69 |
+
|
70 |
+
@abstractmethod
|
71 |
+
def forward(self, x, emb):
|
72 |
+
"""
|
73 |
+
Apply the module to `x` given `emb` timestep embeddings.
|
74 |
+
"""
|
75 |
+
|
76 |
+
|
77 |
+
class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
|
78 |
+
"""
|
79 |
+
A sequential module that passes timestep embeddings to the children that
|
80 |
+
support it as an extra input.
|
81 |
+
"""
|
82 |
+
|
83 |
+
def forward(self, x, emb, context=None):
|
84 |
+
for layer in self:
|
85 |
+
if isinstance(layer, TimestepBlock):
|
86 |
+
x = layer(x, emb)
|
87 |
+
elif isinstance(layer, SpatialTransformer):
|
88 |
+
x = layer(x, context)
|
89 |
+
else:
|
90 |
+
x = layer(x)
|
91 |
+
return x
|
92 |
+
|
93 |
+
|
94 |
+
class Upsample(nn.Module):
|
95 |
+
"""
|
96 |
+
An upsampling layer with an optional convolution.
|
97 |
+
:param channels: channels in the inputs and outputs.
|
98 |
+
:param use_conv: a bool determining if a convolution is applied.
|
99 |
+
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
|
100 |
+
upsampling occurs in the inner-two dimensions.
|
101 |
+
"""
|
102 |
+
|
103 |
+
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
|
104 |
+
super().__init__()
|
105 |
+
self.channels = channels
|
106 |
+
self.out_channels = out_channels or channels
|
107 |
+
self.use_conv = use_conv
|
108 |
+
self.dims = dims
|
109 |
+
if use_conv:
|
110 |
+
self.conv = conv_nd(
|
111 |
+
dims, self.channels, self.out_channels, 3, padding=padding
|
112 |
+
)
|
113 |
+
|
114 |
+
def forward(self, x):
|
115 |
+
assert x.shape[1] == self.channels
|
116 |
+
if self.dims == 3:
|
117 |
+
x = F.interpolate(
|
118 |
+
x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
|
119 |
+
)
|
120 |
+
else:
|
121 |
+
x = F.interpolate(x, scale_factor=2, mode="nearest")
|
122 |
+
if self.use_conv:
|
123 |
+
x = self.conv(x)
|
124 |
+
return x
|
125 |
+
|
126 |
+
|
127 |
+
class TransposedUpsample(nn.Module):
|
128 |
+
"Learned 2x upsampling without padding"
|
129 |
+
|
130 |
+
def __init__(self, channels, out_channels=None, ks=5):
|
131 |
+
super().__init__()
|
132 |
+
self.channels = channels
|
133 |
+
self.out_channels = out_channels or channels
|
134 |
+
|
135 |
+
self.up = nn.ConvTranspose2d(
|
136 |
+
self.channels, self.out_channels, kernel_size=ks, stride=2
|
137 |
+
)
|
138 |
+
|
139 |
+
def forward(self, x):
|
140 |
+
return self.up(x)
|
141 |
+
|
142 |
+
|
143 |
+
class Downsample(nn.Module):
|
144 |
+
"""
|
145 |
+
A downsampling layer with an optional convolution.
|
146 |
+
:param channels: channels in the inputs and outputs.
|
147 |
+
:param use_conv: a bool determining if a convolution is applied.
|
148 |
+
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
|
149 |
+
downsampling occurs in the inner-two dimensions.
|
150 |
+
"""
|
151 |
+
|
152 |
+
def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
|
153 |
+
super().__init__()
|
154 |
+
self.channels = channels
|
155 |
+
self.out_channels = out_channels or channels
|
156 |
+
self.use_conv = use_conv
|
157 |
+
self.dims = dims
|
158 |
+
stride = 2 if dims != 3 else (1, 2, 2)
|
159 |
+
if use_conv:
|
160 |
+
self.op = conv_nd(
|
161 |
+
dims,
|
162 |
+
self.channels,
|
163 |
+
self.out_channels,
|
164 |
+
3,
|
165 |
+
stride=stride,
|
166 |
+
padding=padding,
|
167 |
+
)
|
168 |
+
else:
|
169 |
+
assert self.channels == self.out_channels
|
170 |
+
self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
|
171 |
+
|
172 |
+
def forward(self, x):
|
173 |
+
assert x.shape[1] == self.channels
|
174 |
+
return self.op(x)
|
175 |
+
|
176 |
+
|
177 |
+
class ResBlock(TimestepBlock):
|
178 |
+
"""
|
179 |
+
A residual block that can optionally change the number of channels.
|
180 |
+
:param channels: the number of input channels.
|
181 |
+
:param emb_channels: the number of timestep embedding channels.
|
182 |
+
:param dropout: the rate of dropout.
|
183 |
+
:param out_channels: if specified, the number of out channels.
|
184 |
+
:param use_conv: if True and out_channels is specified, use a spatial
|
185 |
+
convolution instead of a smaller 1x1 convolution to change the
|
186 |
+
channels in the skip connection.
|
187 |
+
:param dims: determines if the signal is 1D, 2D, or 3D.
|
188 |
+
:param use_checkpoint: if True, use gradient checkpointing on this module.
|
189 |
+
:param up: if True, use this block for upsampling.
|
190 |
+
:param down: if True, use this block for downsampling.
|
191 |
+
"""
|
192 |
+
|
193 |
+
def __init__(
|
194 |
+
self,
|
195 |
+
channels,
|
196 |
+
emb_channels,
|
197 |
+
dropout,
|
198 |
+
out_channels=None,
|
199 |
+
use_conv=False,
|
200 |
+
use_scale_shift_norm=False,
|
201 |
+
dims=2,
|
202 |
+
use_checkpoint=False,
|
203 |
+
up=False,
|
204 |
+
down=False,
|
205 |
+
):
|
206 |
+
super().__init__()
|
207 |
+
self.channels = channels
|
208 |
+
self.emb_channels = emb_channels
|
209 |
+
self.dropout = dropout
|
210 |
+
self.out_channels = out_channels or channels
|
211 |
+
self.use_conv = use_conv
|
212 |
+
self.use_checkpoint = use_checkpoint
|
213 |
+
self.use_scale_shift_norm = use_scale_shift_norm
|
214 |
+
|
215 |
+
self.in_layers = nn.Sequential(
|
216 |
+
normalization(channels),
|
217 |
+
nn.SiLU(),
|
218 |
+
conv_nd(dims, channels, self.out_channels, 3, padding=1),
|
219 |
+
)
|
220 |
+
|
221 |
+
self.updown = up or down
|
222 |
+
|
223 |
+
if up:
|
224 |
+
self.h_upd = Upsample(channels, False, dims)
|
225 |
+
self.x_upd = Upsample(channels, False, dims)
|
226 |
+
elif down:
|
227 |
+
self.h_upd = Downsample(channels, False, dims)
|
228 |
+
self.x_upd = Downsample(channels, False, dims)
|
229 |
+
else:
|
230 |
+
self.h_upd = self.x_upd = nn.Identity()
|
231 |
+
|
232 |
+
self.emb_layers = nn.Sequential(
|
233 |
+
nn.SiLU(),
|
234 |
+
linear(
|
235 |
+
emb_channels,
|
236 |
+
2 * self.out_channels if use_scale_shift_norm else self.out_channels,
|
237 |
+
),
|
238 |
+
)
|
239 |
+
self.out_layers = nn.Sequential(
|
240 |
+
normalization(self.out_channels),
|
241 |
+
nn.SiLU(),
|
242 |
+
nn.Dropout(p=dropout),
|
243 |
+
zero_module(
|
244 |
+
conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
|
245 |
+
),
|
246 |
+
)
|
247 |
+
|
248 |
+
if self.out_channels == channels:
|
249 |
+
self.skip_connection = nn.Identity()
|
250 |
+
elif use_conv:
|
251 |
+
self.skip_connection = conv_nd(
|
252 |
+
dims, channels, self.out_channels, 3, padding=1
|
253 |
+
)
|
254 |
+
else:
|
255 |
+
self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
|
256 |
+
|
257 |
+
def forward(self, x, emb):
|
258 |
+
"""
|
259 |
+
Apply the block to a Tensor, conditioned on a timestep embedding.
|
260 |
+
:param x: an [N x C x ...] Tensor of features.
|
261 |
+
:param emb: an [N x emb_channels] Tensor of timestep embeddings.
|
262 |
+
:return: an [N x C x ...] Tensor of outputs.
|
263 |
+
"""
|
264 |
+
return checkpoint(
|
265 |
+
self._forward, (x, emb), self.parameters(), self.use_checkpoint
|
266 |
+
)
|
267 |
+
|
268 |
+
def _forward(self, x, emb):
|
269 |
+
if self.updown:
|
270 |
+
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
|
271 |
+
h = in_rest(x)
|
272 |
+
h = self.h_upd(h)
|
273 |
+
x = self.x_upd(x)
|
274 |
+
h = in_conv(h)
|
275 |
+
else:
|
276 |
+
h = self.in_layers(x)
|
277 |
+
emb_out = self.emb_layers(emb).type(h.dtype)
|
278 |
+
while len(emb_out.shape) < len(h.shape):
|
279 |
+
emb_out = emb_out[..., None]
|
280 |
+
if self.use_scale_shift_norm:
|
281 |
+
out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
|
282 |
+
scale, shift = th.chunk(emb_out, 2, dim=1)
|
283 |
+
h = out_norm(h) * (1 + scale) + shift
|
284 |
+
h = out_rest(h)
|
285 |
+
else:
|
286 |
+
h = h + emb_out
|
287 |
+
h = self.out_layers(h)
|
288 |
+
return self.skip_connection(x) + h
|
289 |
+
|
290 |
+
|
291 |
+
class AttentionBlock(nn.Module):
|
292 |
+
"""
|
293 |
+
An attention block that allows spatial positions to attend to each other.
|
294 |
+
Originally ported from here, but adapted to the N-d case.
|
295 |
+
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
|
296 |
+
"""
|
297 |
+
|
298 |
+
def __init__(
|
299 |
+
self,
|
300 |
+
channels,
|
301 |
+
num_heads=1,
|
302 |
+
num_head_channels=-1,
|
303 |
+
use_checkpoint=False,
|
304 |
+
use_new_attention_order=False,
|
305 |
+
):
|
306 |
+
super().__init__()
|
307 |
+
self.channels = channels
|
308 |
+
if num_head_channels == -1:
|
309 |
+
self.num_heads = num_heads
|
310 |
+
else:
|
311 |
+
assert (
|
312 |
+
channels % num_head_channels == 0
|
313 |
+
), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
|
314 |
+
self.num_heads = channels // num_head_channels
|
315 |
+
self.use_checkpoint = use_checkpoint
|
316 |
+
self.norm = normalization(channels)
|
317 |
+
self.qkv = conv_nd(1, channels, channels * 3, 1)
|
318 |
+
if use_new_attention_order:
|
319 |
+
# split qkv before split heads
|
320 |
+
self.attention = QKVAttention(self.num_heads)
|
321 |
+
else:
|
322 |
+
# split heads before split qkv
|
323 |
+
self.attention = QKVAttentionLegacy(self.num_heads)
|
324 |
+
|
325 |
+
self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
|
326 |
+
|
327 |
+
def forward(self, x):
|
328 |
+
return checkpoint(
|
329 |
+
self._forward, (x,), self.parameters(), True
|
330 |
+
) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
|
331 |
+
# return pt_checkpoint(self._forward, x) # pytorch
|
332 |
+
|
333 |
+
def _forward(self, x):
|
334 |
+
b, c, *spatial = x.shape
|
335 |
+
x = x.reshape(b, c, -1)
|
336 |
+
qkv = self.qkv(self.norm(x))
|
337 |
+
h = self.attention(qkv)
|
338 |
+
h = self.proj_out(h)
|
339 |
+
return (x + h).reshape(b, c, *spatial)
|
340 |
+
|
341 |
+
|
342 |
+
def count_flops_attn(model, _x, y):
|
343 |
+
"""
|
344 |
+
A counter for the `thop` package to count the operations in an
|
345 |
+
attention operation.
|
346 |
+
Meant to be used like:
|
347 |
+
macs, params = thop.profile(
|
348 |
+
model,
|
349 |
+
inputs=(inputs, timestamps),
|
350 |
+
custom_ops={QKVAttention: QKVAttention.count_flops},
|
351 |
+
)
|
352 |
+
"""
|
353 |
+
b, c, *spatial = y[0].shape
|
354 |
+
num_spatial = int(np.prod(spatial))
|
355 |
+
# We perform two matmuls with the same number of ops.
|
356 |
+
# The first computes the weight matrix, the second computes
|
357 |
+
# the combination of the value vectors.
|
358 |
+
matmul_ops = 2 * b * (num_spatial**2) * c
|
359 |
+
model.total_ops += th.DoubleTensor([matmul_ops])
|
360 |
+
|
361 |
+
|
362 |
+
class QKVAttentionLegacy(nn.Module):
|
363 |
+
"""
|
364 |
+
A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
|
365 |
+
"""
|
366 |
+
|
367 |
+
def __init__(self, n_heads):
|
368 |
+
super().__init__()
|
369 |
+
self.n_heads = n_heads
|
370 |
+
|
371 |
+
def forward(self, qkv):
|
372 |
+
"""
|
373 |
+
Apply QKV attention.
|
374 |
+
:param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
|
375 |
+
:return: an [N x (H * C) x T] tensor after attention.
|
376 |
+
"""
|
377 |
+
bs, width, length = qkv.shape
|
378 |
+
assert width % (3 * self.n_heads) == 0
|
379 |
+
ch = width // (3 * self.n_heads)
|
380 |
+
q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
|
381 |
+
scale = 1 / math.sqrt(math.sqrt(ch))
|
382 |
+
weight = th.einsum(
|
383 |
+
"bct,bcs->bts", q * scale, k * scale
|
384 |
+
) # More stable with f16 than dividing afterwards
|
385 |
+
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
|
386 |
+
a = th.einsum("bts,bcs->bct", weight, v)
|
387 |
+
return a.reshape(bs, -1, length)
|
388 |
+
|
389 |
+
@staticmethod
|
390 |
+
def count_flops(model, _x, y):
|
391 |
+
return count_flops_attn(model, _x, y)
|
392 |
+
|
393 |
+
|
394 |
+
class QKVAttention(nn.Module):
|
395 |
+
"""
|
396 |
+
A module which performs QKV attention and splits in a different order.
|
397 |
+
"""
|
398 |
+
|
399 |
+
def __init__(self, n_heads):
|
400 |
+
super().__init__()
|
401 |
+
self.n_heads = n_heads
|
402 |
+
|
403 |
+
def forward(self, qkv):
|
404 |
+
"""
|
405 |
+
Apply QKV attention.
|
406 |
+
:param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
|
407 |
+
:return: an [N x (H * C) x T] tensor after attention.
|
408 |
+
"""
|
409 |
+
bs, width, length = qkv.shape
|
410 |
+
assert width % (3 * self.n_heads) == 0
|
411 |
+
ch = width // (3 * self.n_heads)
|
412 |
+
q, k, v = qkv.chunk(3, dim=1)
|
413 |
+
scale = 1 / math.sqrt(math.sqrt(ch))
|
414 |
+
weight = th.einsum(
|
415 |
+
"bct,bcs->bts",
|
416 |
+
(q * scale).view(bs * self.n_heads, ch, length),
|
417 |
+
(k * scale).view(bs * self.n_heads, ch, length),
|
418 |
+
) # More stable with f16 than dividing afterwards
|
419 |
+
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
|
420 |
+
a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
|
421 |
+
return a.reshape(bs, -1, length)
|
422 |
+
|
423 |
+
@staticmethod
|
424 |
+
def count_flops(model, _x, y):
|
425 |
+
return count_flops_attn(model, _x, y)
|
426 |
+
|
427 |
+
|
428 |
+
class UNetModel(nn.Module):
|
429 |
+
"""
|
430 |
+
The full UNet model with attention and timestep embedding.
|
431 |
+
:param in_channels: channels in the input Tensor.
|
432 |
+
:param model_channels: base channel count for the model.
|
433 |
+
:param out_channels: channels in the output Tensor.
|
434 |
+
:param num_res_blocks: number of residual blocks per downsample.
|
435 |
+
:param attention_resolutions: a collection of downsample rates at which
|
436 |
+
attention will take place. May be a set, list, or tuple.
|
437 |
+
For example, if this contains 4, then at 4x downsampling, attention
|
438 |
+
will be used.
|
439 |
+
:param dropout: the dropout probability.
|
440 |
+
:param channel_mult: channel multiplier for each level of the UNet.
|
441 |
+
:param conv_resample: if True, use learned convolutions for upsampling and
|
442 |
+
downsampling.
|
443 |
+
:param dims: determines if the signal is 1D, 2D, or 3D.
|
444 |
+
:param num_classes: if specified (as an int), then this model will be
|
445 |
+
class-conditional with `num_classes` classes.
|
446 |
+
:param use_checkpoint: use gradient checkpointing to reduce memory usage.
|
447 |
+
:param num_heads: the number of attention heads in each attention layer.
|
448 |
+
:param num_heads_channels: if specified, ignore num_heads and instead use
|
449 |
+
a fixed channel width per attention head.
|
450 |
+
:param num_heads_upsample: works with num_heads to set a different number
|
451 |
+
of heads for upsampling. Deprecated.
|
452 |
+
:param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
|
453 |
+
:param resblock_updown: use residual blocks for up/downsampling.
|
454 |
+
:param use_new_attention_order: use a different attention pattern for potentially
|
455 |
+
increased efficiency.
|
456 |
+
"""
|
457 |
+
|
458 |
+
def __init__(
|
459 |
+
self,
|
460 |
+
image_size,
|
461 |
+
in_channels,
|
462 |
+
model_channels,
|
463 |
+
out_channels,
|
464 |
+
num_res_blocks,
|
465 |
+
attention_resolutions,
|
466 |
+
dropout=0,
|
467 |
+
channel_mult=(1, 2, 4, 8),
|
468 |
+
conv_resample=True,
|
469 |
+
dims=2,
|
470 |
+
num_classes=None,
|
471 |
+
use_checkpoint=False,
|
472 |
+
use_fp16=False,
|
473 |
+
num_heads=-1,
|
474 |
+
num_head_channels=-1,
|
475 |
+
num_heads_upsample=-1,
|
476 |
+
use_scale_shift_norm=False,
|
477 |
+
resblock_updown=False,
|
478 |
+
use_new_attention_order=False,
|
479 |
+
use_spatial_transformer=False, # custom transformer support
|
480 |
+
transformer_depth=1, # custom transformer support
|
481 |
+
context_dim=None, # custom transformer support
|
482 |
+
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
|
483 |
+
legacy=True,
|
484 |
+
):
|
485 |
+
super().__init__()
|
486 |
+
if use_spatial_transformer:
|
487 |
+
assert (
|
488 |
+
context_dim is not None
|
489 |
+
), "Fool!! You forgot to include the dimension of your cross-attention conditioning..."
|
490 |
+
|
491 |
+
if context_dim is not None:
|
492 |
+
assert (
|
493 |
+
use_spatial_transformer
|
494 |
+
), "Fool!! You forgot to use the spatial transformer for your cross-attention conditioning..."
|
495 |
+
from omegaconf.listconfig import ListConfig
|
496 |
+
|
497 |
+
if type(context_dim) == ListConfig:
|
498 |
+
context_dim = list(context_dim)
|
499 |
+
|
500 |
+
if num_heads_upsample == -1:
|
501 |
+
num_heads_upsample = num_heads
|
502 |
+
|
503 |
+
if num_heads == -1:
|
504 |
+
assert (
|
505 |
+
num_head_channels != -1
|
506 |
+
), "Either num_heads or num_head_channels has to be set"
|
507 |
+
|
508 |
+
if num_head_channels == -1:
|
509 |
+
assert (
|
510 |
+
num_heads != -1
|
511 |
+
), "Either num_heads or num_head_channels has to be set"
|
512 |
+
|
513 |
+
self.image_size = image_size
|
514 |
+
self.in_channels = in_channels
|
515 |
+
self.model_channels = model_channels
|
516 |
+
self.out_channels = out_channels
|
517 |
+
self.num_res_blocks = num_res_blocks
|
518 |
+
self.attention_resolutions = attention_resolutions
|
519 |
+
self.dropout = dropout
|
520 |
+
self.channel_mult = channel_mult
|
521 |
+
self.conv_resample = conv_resample
|
522 |
+
self.num_classes = num_classes
|
523 |
+
self.use_checkpoint = use_checkpoint
|
524 |
+
self.dtype = th.float16 if use_fp16 else th.float32
|
525 |
+
self.num_heads = num_heads
|
526 |
+
self.num_head_channels = num_head_channels
|
527 |
+
self.num_heads_upsample = num_heads_upsample
|
528 |
+
self.predict_codebook_ids = n_embed is not None
|
529 |
+
|
530 |
+
time_embed_dim = model_channels * 4
|
531 |
+
self.time_embed = nn.Sequential(
|
532 |
+
linear(model_channels, time_embed_dim),
|
533 |
+
nn.SiLU(),
|
534 |
+
linear(time_embed_dim, time_embed_dim),
|
535 |
+
)
|
536 |
+
|
537 |
+
if self.num_classes is not None:
|
538 |
+
self.label_emb = nn.Embedding(num_classes, time_embed_dim)
|
539 |
+
|
540 |
+
self.input_blocks = nn.ModuleList(
|
541 |
+
[
|
542 |
+
TimestepEmbedSequential(
|
543 |
+
conv_nd(dims, in_channels, model_channels, 3, padding=1)
|
544 |
+
)
|
545 |
+
]
|
546 |
+
)
|
547 |
+
self._feature_size = model_channels
|
548 |
+
input_block_chans = [model_channels]
|
549 |
+
ch = model_channels
|
550 |
+
ds = 1
|
551 |
+
for level, mult in enumerate(channel_mult):
|
552 |
+
for _ in range(num_res_blocks):
|
553 |
+
layers = [
|
554 |
+
ResBlock(
|
555 |
+
ch,
|
556 |
+
time_embed_dim,
|
557 |
+
dropout,
|
558 |
+
out_channels=mult * model_channels,
|
559 |
+
dims=dims,
|
560 |
+
use_checkpoint=use_checkpoint,
|
561 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
562 |
+
)
|
563 |
+
]
|
564 |
+
ch = mult * model_channels
|
565 |
+
if ds in attention_resolutions:
|
566 |
+
if num_head_channels == -1:
|
567 |
+
dim_head = ch // num_heads
|
568 |
+
else:
|
569 |
+
num_heads = ch // num_head_channels
|
570 |
+
dim_head = num_head_channels
|
571 |
+
if legacy:
|
572 |
+
# num_heads = 1
|
573 |
+
dim_head = (
|
574 |
+
ch // num_heads
|
575 |
+
if use_spatial_transformer
|
576 |
+
else num_head_channels
|
577 |
+
)
|
578 |
+
layers.append(
|
579 |
+
AttentionBlock(
|
580 |
+
ch,
|
581 |
+
use_checkpoint=use_checkpoint,
|
582 |
+
num_heads=num_heads,
|
583 |
+
num_head_channels=dim_head,
|
584 |
+
use_new_attention_order=use_new_attention_order,
|
585 |
+
)
|
586 |
+
if not use_spatial_transformer
|
587 |
+
else SpatialTransformer(
|
588 |
+
ch,
|
589 |
+
num_heads,
|
590 |
+
dim_head,
|
591 |
+
depth=transformer_depth,
|
592 |
+
context_dim=context_dim,
|
593 |
+
)
|
594 |
+
)
|
595 |
+
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
596 |
+
self._feature_size += ch
|
597 |
+
input_block_chans.append(ch)
|
598 |
+
if level != len(channel_mult) - 1:
|
599 |
+
out_ch = ch
|
600 |
+
self.input_blocks.append(
|
601 |
+
TimestepEmbedSequential(
|
602 |
+
ResBlock(
|
603 |
+
ch,
|
604 |
+
time_embed_dim,
|
605 |
+
dropout,
|
606 |
+
out_channels=out_ch,
|
607 |
+
dims=dims,
|
608 |
+
use_checkpoint=use_checkpoint,
|
609 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
610 |
+
down=True,
|
611 |
+
)
|
612 |
+
if resblock_updown
|
613 |
+
else Downsample(
|
614 |
+
ch, conv_resample, dims=dims, out_channels=out_ch
|
615 |
+
)
|
616 |
+
)
|
617 |
+
)
|
618 |
+
ch = out_ch
|
619 |
+
input_block_chans.append(ch)
|
620 |
+
ds *= 2
|
621 |
+
self._feature_size += ch
|
622 |
+
|
623 |
+
if num_head_channels == -1:
|
624 |
+
dim_head = ch // num_heads
|
625 |
+
else:
|
626 |
+
num_heads = ch // num_head_channels
|
627 |
+
dim_head = num_head_channels
|
628 |
+
if legacy:
|
629 |
+
# num_heads = 1
|
630 |
+
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
631 |
+
self.middle_block = TimestepEmbedSequential(
|
632 |
+
ResBlock(
|
633 |
+
ch,
|
634 |
+
time_embed_dim,
|
635 |
+
dropout,
|
636 |
+
dims=dims,
|
637 |
+
use_checkpoint=use_checkpoint,
|
638 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
639 |
+
),
|
640 |
+
(
|
641 |
+
AttentionBlock(
|
642 |
+
ch,
|
643 |
+
use_checkpoint=use_checkpoint,
|
644 |
+
num_heads=num_heads,
|
645 |
+
num_head_channels=dim_head,
|
646 |
+
use_new_attention_order=use_new_attention_order,
|
647 |
+
)
|
648 |
+
if not use_spatial_transformer
|
649 |
+
else SpatialTransformer(
|
650 |
+
ch,
|
651 |
+
num_heads,
|
652 |
+
dim_head,
|
653 |
+
depth=transformer_depth,
|
654 |
+
context_dim=context_dim,
|
655 |
+
)
|
656 |
+
),
|
657 |
+
ResBlock(
|
658 |
+
ch,
|
659 |
+
time_embed_dim,
|
660 |
+
dropout,
|
661 |
+
dims=dims,
|
662 |
+
use_checkpoint=use_checkpoint,
|
663 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
664 |
+
),
|
665 |
+
)
|
666 |
+
self._feature_size += ch
|
667 |
+
|
668 |
+
self.output_blocks = nn.ModuleList([])
|
669 |
+
for level, mult in list(enumerate(channel_mult))[::-1]:
|
670 |
+
for i in range(num_res_blocks + 1):
|
671 |
+
ich = input_block_chans.pop()
|
672 |
+
layers = [
|
673 |
+
ResBlock(
|
674 |
+
ch + ich,
|
675 |
+
time_embed_dim,
|
676 |
+
dropout,
|
677 |
+
out_channels=model_channels * mult,
|
678 |
+
dims=dims,
|
679 |
+
use_checkpoint=use_checkpoint,
|
680 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
681 |
+
)
|
682 |
+
]
|
683 |
+
ch = model_channels * mult
|
684 |
+
if ds in attention_resolutions:
|
685 |
+
if num_head_channels == -1:
|
686 |
+
dim_head = ch // num_heads
|
687 |
+
else:
|
688 |
+
num_heads = ch // num_head_channels
|
689 |
+
dim_head = num_head_channels
|
690 |
+
if legacy:
|
691 |
+
# num_heads = 1
|
692 |
+
dim_head = (
|
693 |
+
ch // num_heads
|
694 |
+
if use_spatial_transformer
|
695 |
+
else num_head_channels
|
696 |
+
)
|
697 |
+
layers.append(
|
698 |
+
AttentionBlock(
|
699 |
+
ch,
|
700 |
+
use_checkpoint=use_checkpoint,
|
701 |
+
num_heads=num_heads_upsample,
|
702 |
+
num_head_channels=dim_head,
|
703 |
+
use_new_attention_order=use_new_attention_order,
|
704 |
+
)
|
705 |
+
if not use_spatial_transformer
|
706 |
+
else SpatialTransformer(
|
707 |
+
ch,
|
708 |
+
num_heads,
|
709 |
+
dim_head,
|
710 |
+
depth=transformer_depth,
|
711 |
+
context_dim=context_dim,
|
712 |
+
)
|
713 |
+
)
|
714 |
+
if level and i == num_res_blocks:
|
715 |
+
out_ch = ch
|
716 |
+
layers.append(
|
717 |
+
ResBlock(
|
718 |
+
ch,
|
719 |
+
time_embed_dim,
|
720 |
+
dropout,
|
721 |
+
out_channels=out_ch,
|
722 |
+
dims=dims,
|
723 |
+
use_checkpoint=use_checkpoint,
|
724 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
725 |
+
up=True,
|
726 |
+
)
|
727 |
+
if resblock_updown
|
728 |
+
else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
|
729 |
+
)
|
730 |
+
ds //= 2
|
731 |
+
self.output_blocks.append(TimestepEmbedSequential(*layers))
|
732 |
+
self._feature_size += ch
|
733 |
+
|
734 |
+
self.out = nn.Sequential(
|
735 |
+
normalization(ch),
|
736 |
+
nn.SiLU(),
|
737 |
+
zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
|
738 |
+
)
|
739 |
+
if self.predict_codebook_ids:
|
740 |
+
self.id_predictor = nn.Sequential(
|
741 |
+
normalization(ch),
|
742 |
+
conv_nd(dims, model_channels, n_embed, 1),
|
743 |
+
# nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits
|
744 |
+
)
|
745 |
+
|
746 |
+
def convert_to_fp16(self):
|
747 |
+
"""
|
748 |
+
Convert the torso of the model to float16.
|
749 |
+
"""
|
750 |
+
self.input_blocks.apply(convert_module_to_f16)
|
751 |
+
self.middle_block.apply(convert_module_to_f16)
|
752 |
+
self.output_blocks.apply(convert_module_to_f16)
|
753 |
+
|
754 |
+
def convert_to_fp32(self):
|
755 |
+
"""
|
756 |
+
Convert the torso of the model to float32.
|
757 |
+
"""
|
758 |
+
self.input_blocks.apply(convert_module_to_f32)
|
759 |
+
self.middle_block.apply(convert_module_to_f32)
|
760 |
+
self.output_blocks.apply(convert_module_to_f32)
|
761 |
+
|
762 |
+
def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
|
763 |
+
"""
|
764 |
+
Apply the model to an input batch.
|
765 |
+
:param x: an [N x C x ...] Tensor of inputs.
|
766 |
+
:param timesteps: a 1-D batch of timesteps.
|
767 |
+
:param context: conditioning plugged in via crossattn
|
768 |
+
:param y: an [N] Tensor of labels, if class-conditional.
|
769 |
+
:return: an [N x C x ...] Tensor of outputs.
|
770 |
+
"""
|
771 |
+
assert (y is not None) == (
|
772 |
+
self.num_classes is not None
|
773 |
+
), "must specify y if and only if the model is class-conditional"
|
774 |
+
hs = []
|
775 |
+
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
|
776 |
+
emb = self.time_embed(t_emb)
|
777 |
+
|
778 |
+
if self.num_classes is not None:
|
779 |
+
assert y.shape == (x.shape[0],)
|
780 |
+
emb = emb + self.label_emb(y)
|
781 |
+
|
782 |
+
h = x.type(self.dtype)
|
783 |
+
for module in self.input_blocks:
|
784 |
+
h = module(h, emb, context)
|
785 |
+
hs.append(h)
|
786 |
+
h = self.middle_block(h, emb, context)
|
787 |
+
for module in self.output_blocks:
|
788 |
+
h = th.cat([h, hs.pop()], dim=1)
|
789 |
+
h = module(h, emb, context)
|
790 |
+
h = h.type(x.dtype)
|
791 |
+
if self.predict_codebook_ids:
|
792 |
+
return self.id_predictor(h)
|
793 |
+
else:
|
794 |
+
return self.out(h)
|
795 |
+
|
796 |
+
|
797 |
+
class EncoderUNetModel(nn.Module):
|
798 |
+
"""
|
799 |
+
The half UNet model with attention and timestep embedding.
|
800 |
+
For usage, see UNet.
|
801 |
+
"""
|
802 |
+
|
803 |
+
def __init__(
|
804 |
+
self,
|
805 |
+
image_size,
|
806 |
+
in_channels,
|
807 |
+
model_channels,
|
808 |
+
out_channels,
|
809 |
+
num_res_blocks,
|
810 |
+
attention_resolutions,
|
811 |
+
dropout=0,
|
812 |
+
channel_mult=(1, 2, 4, 8),
|
813 |
+
conv_resample=True,
|
814 |
+
dims=2,
|
815 |
+
use_checkpoint=False,
|
816 |
+
use_fp16=False,
|
817 |
+
num_heads=1,
|
818 |
+
num_head_channels=-1,
|
819 |
+
num_heads_upsample=-1,
|
820 |
+
use_scale_shift_norm=False,
|
821 |
+
resblock_updown=False,
|
822 |
+
use_new_attention_order=False,
|
823 |
+
pool="adaptive",
|
824 |
+
*args,
|
825 |
+
**kwargs,
|
826 |
+
):
|
827 |
+
super().__init__()
|
828 |
+
|
829 |
+
if num_heads_upsample == -1:
|
830 |
+
num_heads_upsample = num_heads
|
831 |
+
|
832 |
+
self.in_channels = in_channels
|
833 |
+
self.model_channels = model_channels
|
834 |
+
self.out_channels = out_channels
|
835 |
+
self.num_res_blocks = num_res_blocks
|
836 |
+
self.attention_resolutions = attention_resolutions
|
837 |
+
self.dropout = dropout
|
838 |
+
self.channel_mult = channel_mult
|
839 |
+
self.conv_resample = conv_resample
|
840 |
+
self.use_checkpoint = use_checkpoint
|
841 |
+
self.dtype = th.float16 if use_fp16 else th.float32
|
842 |
+
self.num_heads = num_heads
|
843 |
+
self.num_head_channels = num_head_channels
|
844 |
+
self.num_heads_upsample = num_heads_upsample
|
845 |
+
|
846 |
+
time_embed_dim = model_channels * 4
|
847 |
+
self.time_embed = nn.Sequential(
|
848 |
+
linear(model_channels, time_embed_dim),
|
849 |
+
nn.SiLU(),
|
850 |
+
linear(time_embed_dim, time_embed_dim),
|
851 |
+
)
|
852 |
+
|
853 |
+
self.input_blocks = nn.ModuleList(
|
854 |
+
[
|
855 |
+
TimestepEmbedSequential(
|
856 |
+
conv_nd(dims, in_channels, model_channels, 3, padding=1)
|
857 |
+
)
|
858 |
+
]
|
859 |
+
)
|
860 |
+
self._feature_size = model_channels
|
861 |
+
input_block_chans = [model_channels]
|
862 |
+
ch = model_channels
|
863 |
+
ds = 1
|
864 |
+
for level, mult in enumerate(channel_mult):
|
865 |
+
for _ in range(num_res_blocks):
|
866 |
+
layers = [
|
867 |
+
ResBlock(
|
868 |
+
ch,
|
869 |
+
time_embed_dim,
|
870 |
+
dropout,
|
871 |
+
out_channels=mult * model_channels,
|
872 |
+
dims=dims,
|
873 |
+
use_checkpoint=use_checkpoint,
|
874 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
875 |
+
)
|
876 |
+
]
|
877 |
+
ch = mult * model_channels
|
878 |
+
if ds in attention_resolutions:
|
879 |
+
layers.append(
|
880 |
+
AttentionBlock(
|
881 |
+
ch,
|
882 |
+
use_checkpoint=use_checkpoint,
|
883 |
+
num_heads=num_heads,
|
884 |
+
num_head_channels=num_head_channels,
|
885 |
+
use_new_attention_order=use_new_attention_order,
|
886 |
+
)
|
887 |
+
)
|
888 |
+
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
889 |
+
self._feature_size += ch
|
890 |
+
input_block_chans.append(ch)
|
891 |
+
if level != len(channel_mult) - 1:
|
892 |
+
out_ch = ch
|
893 |
+
self.input_blocks.append(
|
894 |
+
TimestepEmbedSequential(
|
895 |
+
ResBlock(
|
896 |
+
ch,
|
897 |
+
time_embed_dim,
|
898 |
+
dropout,
|
899 |
+
out_channels=out_ch,
|
900 |
+
dims=dims,
|
901 |
+
use_checkpoint=use_checkpoint,
|
902 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
903 |
+
down=True,
|
904 |
+
)
|
905 |
+
if resblock_updown
|
906 |
+
else Downsample(
|
907 |
+
ch, conv_resample, dims=dims, out_channels=out_ch
|
908 |
+
)
|
909 |
+
)
|
910 |
+
)
|
911 |
+
ch = out_ch
|
912 |
+
input_block_chans.append(ch)
|
913 |
+
ds *= 2
|
914 |
+
self._feature_size += ch
|
915 |
+
|
916 |
+
self.middle_block = TimestepEmbedSequential(
|
917 |
+
ResBlock(
|
918 |
+
ch,
|
919 |
+
time_embed_dim,
|
920 |
+
dropout,
|
921 |
+
dims=dims,
|
922 |
+
use_checkpoint=use_checkpoint,
|
923 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
924 |
+
),
|
925 |
+
AttentionBlock(
|
926 |
+
ch,
|
927 |
+
use_checkpoint=use_checkpoint,
|
928 |
+
num_heads=num_heads,
|
929 |
+
num_head_channels=num_head_channels,
|
930 |
+
use_new_attention_order=use_new_attention_order,
|
931 |
+
),
|
932 |
+
ResBlock(
|
933 |
+
ch,
|
934 |
+
time_embed_dim,
|
935 |
+
dropout,
|
936 |
+
dims=dims,
|
937 |
+
use_checkpoint=use_checkpoint,
|
938 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
939 |
+
),
|
940 |
+
)
|
941 |
+
self._feature_size += ch
|
942 |
+
self.pool = pool
|
943 |
+
if pool == "adaptive":
|
944 |
+
self.out = nn.Sequential(
|
945 |
+
normalization(ch),
|
946 |
+
nn.SiLU(),
|
947 |
+
nn.AdaptiveAvgPool2d((1, 1)),
|
948 |
+
zero_module(conv_nd(dims, ch, out_channels, 1)),
|
949 |
+
nn.Flatten(),
|
950 |
+
)
|
951 |
+
elif pool == "attention":
|
952 |
+
assert num_head_channels != -1
|
953 |
+
self.out = nn.Sequential(
|
954 |
+
normalization(ch),
|
955 |
+
nn.SiLU(),
|
956 |
+
AttentionPool2d(
|
957 |
+
(image_size // ds), ch, num_head_channels, out_channels
|
958 |
+
),
|
959 |
+
)
|
960 |
+
elif pool == "spatial":
|
961 |
+
self.out = nn.Sequential(
|
962 |
+
nn.Linear(self._feature_size, 2048),
|
963 |
+
nn.ReLU(),
|
964 |
+
nn.Linear(2048, self.out_channels),
|
965 |
+
)
|
966 |
+
elif pool == "spatial_v2":
|
967 |
+
self.out = nn.Sequential(
|
968 |
+
nn.Linear(self._feature_size, 2048),
|
969 |
+
normalization(2048),
|
970 |
+
nn.SiLU(),
|
971 |
+
nn.Linear(2048, self.out_channels),
|
972 |
+
)
|
973 |
+
else:
|
974 |
+
raise NotImplementedError(f"Unexpected {pool} pooling")
|
975 |
+
|
976 |
+
def convert_to_fp16(self):
|
977 |
+
"""
|
978 |
+
Convert the torso of the model to float16.
|
979 |
+
"""
|
980 |
+
self.input_blocks.apply(convert_module_to_f16)
|
981 |
+
self.middle_block.apply(convert_module_to_f16)
|
982 |
+
|
983 |
+
def convert_to_fp32(self):
|
984 |
+
"""
|
985 |
+
Convert the torso of the model to float32.
|
986 |
+
"""
|
987 |
+
self.input_blocks.apply(convert_module_to_f32)
|
988 |
+
self.middle_block.apply(convert_module_to_f32)
|
989 |
+
|
990 |
+
def forward(self, x, timesteps):
|
991 |
+
"""
|
992 |
+
Apply the model to an input batch.
|
993 |
+
:param x: an [N x C x ...] Tensor of inputs.
|
994 |
+
:param timesteps: a 1-D batch of timesteps.
|
995 |
+
:return: an [N x K] Tensor of outputs.
|
996 |
+
"""
|
997 |
+
emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
|
998 |
+
|
999 |
+
results = []
|
1000 |
+
h = x.type(self.dtype)
|
1001 |
+
for module in self.input_blocks:
|
1002 |
+
h = module(h, emb)
|
1003 |
+
if self.pool.startswith("spatial"):
|
1004 |
+
results.append(h.type(x.dtype).mean(dim=(2, 3)))
|
1005 |
+
h = self.middle_block(h, emb)
|
1006 |
+
if self.pool.startswith("spatial"):
|
1007 |
+
results.append(h.type(x.dtype).mean(dim=(2, 3)))
|
1008 |
+
h = th.cat(results, axis=-1)
|
1009 |
+
return self.out(h)
|
1010 |
+
else:
|
1011 |
+
h = h.type(x.dtype)
|
1012 |
+
return self.out(h)
|
swim/modules/diffusionmodules/util.py
ADDED
@@ -0,0 +1,296 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# adopted from
|
2 |
+
# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
|
3 |
+
# and
|
4 |
+
# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
|
5 |
+
# and
|
6 |
+
# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
|
7 |
+
#
|
8 |
+
# thanks!
|
9 |
+
|
10 |
+
|
11 |
+
import os
|
12 |
+
import math
|
13 |
+
import torch
|
14 |
+
import torch.nn as nn
|
15 |
+
import numpy as np
|
16 |
+
from einops import repeat
|
17 |
+
|
18 |
+
from swim.utils import instantiate_from_config
|
19 |
+
|
20 |
+
|
21 |
+
def make_beta_schedule(
|
22 |
+
schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3
|
23 |
+
):
|
24 |
+
if schedule == "linear":
|
25 |
+
betas = (
|
26 |
+
torch.linspace(
|
27 |
+
linear_start**0.5, linear_end**0.5, n_timestep, dtype=torch.float64
|
28 |
+
)
|
29 |
+
** 2
|
30 |
+
)
|
31 |
+
|
32 |
+
elif schedule == "cosine":
|
33 |
+
timesteps = (
|
34 |
+
torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
|
35 |
+
)
|
36 |
+
alphas = timesteps / (1 + cosine_s) * np.pi / 2
|
37 |
+
alphas = torch.cos(alphas).pow(2)
|
38 |
+
alphas = alphas / alphas[0]
|
39 |
+
betas = 1 - alphas[1:] / alphas[:-1]
|
40 |
+
betas = np.clip(betas, a_min=0, a_max=0.999)
|
41 |
+
|
42 |
+
elif schedule == "sqrt_linear":
|
43 |
+
betas = torch.linspace(
|
44 |
+
linear_start, linear_end, n_timestep, dtype=torch.float64
|
45 |
+
)
|
46 |
+
elif schedule == "sqrt":
|
47 |
+
betas = (
|
48 |
+
torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
|
49 |
+
** 0.5
|
50 |
+
)
|
51 |
+
else:
|
52 |
+
raise ValueError(f"schedule '{schedule}' unknown.")
|
53 |
+
return betas.numpy()
|
54 |
+
|
55 |
+
|
56 |
+
def make_ddim_timesteps(
|
57 |
+
ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True
|
58 |
+
):
|
59 |
+
if ddim_discr_method == "uniform":
|
60 |
+
c = num_ddpm_timesteps // num_ddim_timesteps
|
61 |
+
ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
|
62 |
+
elif ddim_discr_method == "quad":
|
63 |
+
ddim_timesteps = (
|
64 |
+
(np.linspace(0, np.sqrt(num_ddpm_timesteps * 0.8), num_ddim_timesteps)) ** 2
|
65 |
+
).astype(int)
|
66 |
+
else:
|
67 |
+
raise NotImplementedError(
|
68 |
+
f'There is no ddim discretization method called "{ddim_discr_method}"'
|
69 |
+
)
|
70 |
+
|
71 |
+
# assert ddim_timesteps.shape[0] == num_ddim_timesteps
|
72 |
+
# add one to get the final alpha values right (the ones from first scale to data during sampling)
|
73 |
+
steps_out = ddim_timesteps + 1
|
74 |
+
if verbose:
|
75 |
+
print(f"Selected timesteps for ddim sampler: {steps_out}")
|
76 |
+
return steps_out
|
77 |
+
|
78 |
+
|
79 |
+
def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
|
80 |
+
# select alphas for computing the variance schedule
|
81 |
+
alphas = alphacums[ddim_timesteps]
|
82 |
+
alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
|
83 |
+
|
84 |
+
# according the the formula provided in https://arxiv.org/abs/2010.02502
|
85 |
+
sigmas = eta * np.sqrt(
|
86 |
+
(1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)
|
87 |
+
)
|
88 |
+
if verbose:
|
89 |
+
print(
|
90 |
+
f"Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}"
|
91 |
+
)
|
92 |
+
print(
|
93 |
+
f"For the chosen value of eta, which is {eta}, "
|
94 |
+
f"this results in the following sigma_t schedule for ddim sampler {sigmas}"
|
95 |
+
)
|
96 |
+
return sigmas, alphas, alphas_prev
|
97 |
+
|
98 |
+
|
99 |
+
def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
|
100 |
+
"""
|
101 |
+
Create a beta schedule that discretizes the given alpha_t_bar function,
|
102 |
+
which defines the cumulative product of (1-beta) over time from t = [0,1].
|
103 |
+
:param num_diffusion_timesteps: the number of betas to produce.
|
104 |
+
:param alpha_bar: a lambda that takes an argument t from 0 to 1 and
|
105 |
+
produces the cumulative product of (1-beta) up to that
|
106 |
+
part of the diffusion process.
|
107 |
+
:param max_beta: the maximum beta to use; use values lower than 1 to
|
108 |
+
prevent singularities.
|
109 |
+
"""
|
110 |
+
betas = []
|
111 |
+
for i in range(num_diffusion_timesteps):
|
112 |
+
t1 = i / num_diffusion_timesteps
|
113 |
+
t2 = (i + 1) / num_diffusion_timesteps
|
114 |
+
betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
|
115 |
+
return np.array(betas)
|
116 |
+
|
117 |
+
|
118 |
+
def extract_into_tensor(a, t, x_shape):
|
119 |
+
b, *_ = t.shape
|
120 |
+
out = a.gather(-1, t)
|
121 |
+
return out.reshape(b, *((1,) * (len(x_shape) - 1)))
|
122 |
+
|
123 |
+
|
124 |
+
def checkpoint(func, inputs, params, flag):
|
125 |
+
"""
|
126 |
+
Evaluate a function without caching intermediate activations, allowing for
|
127 |
+
reduced memory at the expense of extra compute in the backward pass.
|
128 |
+
:param func: the function to evaluate.
|
129 |
+
:param inputs: the argument sequence to pass to `func`.
|
130 |
+
:param params: a sequence of parameters `func` depends on but does not
|
131 |
+
explicitly take as arguments.
|
132 |
+
:param flag: if False, disable gradient checkpointing.
|
133 |
+
"""
|
134 |
+
if flag:
|
135 |
+
args = tuple(inputs) + tuple(params)
|
136 |
+
return CheckpointFunction.apply(func, len(inputs), *args)
|
137 |
+
else:
|
138 |
+
return func(*inputs)
|
139 |
+
|
140 |
+
|
141 |
+
class CheckpointFunction(torch.autograd.Function):
|
142 |
+
@staticmethod
|
143 |
+
def forward(ctx, run_function, length, *args):
|
144 |
+
ctx.run_function = run_function
|
145 |
+
ctx.input_tensors = list(args[:length])
|
146 |
+
ctx.input_params = list(args[length:])
|
147 |
+
|
148 |
+
with torch.no_grad():
|
149 |
+
output_tensors = ctx.run_function(*ctx.input_tensors)
|
150 |
+
return output_tensors
|
151 |
+
|
152 |
+
@staticmethod
|
153 |
+
def backward(ctx, *output_grads):
|
154 |
+
ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
|
155 |
+
with torch.enable_grad():
|
156 |
+
# Fixes a bug where the first op in run_function modifies the
|
157 |
+
# Tensor storage in place, which is not allowed for detach()'d
|
158 |
+
# Tensors.
|
159 |
+
shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
|
160 |
+
output_tensors = ctx.run_function(*shallow_copies)
|
161 |
+
input_grads = torch.autograd.grad(
|
162 |
+
output_tensors,
|
163 |
+
ctx.input_tensors + ctx.input_params,
|
164 |
+
output_grads,
|
165 |
+
allow_unused=True,
|
166 |
+
)
|
167 |
+
del ctx.input_tensors
|
168 |
+
del ctx.input_params
|
169 |
+
del output_tensors
|
170 |
+
return (None, None) + input_grads
|
171 |
+
|
172 |
+
|
173 |
+
def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
|
174 |
+
"""
|
175 |
+
Create sinusoidal timestep embeddings.
|
176 |
+
:param timesteps: a 1-D Tensor of N indices, one per batch element.
|
177 |
+
These may be fractional.
|
178 |
+
:param dim: the dimension of the output.
|
179 |
+
:param max_period: controls the minimum frequency of the embeddings.
|
180 |
+
:return: an [N x dim] Tensor of positional embeddings.
|
181 |
+
"""
|
182 |
+
if not repeat_only:
|
183 |
+
half = dim // 2
|
184 |
+
freqs = torch.exp(
|
185 |
+
-math.log(max_period)
|
186 |
+
* torch.arange(start=0, end=half, dtype=torch.float32)
|
187 |
+
/ half
|
188 |
+
).to(device=timesteps.device)
|
189 |
+
args = timesteps[:, None].float() * freqs[None]
|
190 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
191 |
+
if dim % 2:
|
192 |
+
embedding = torch.cat(
|
193 |
+
[embedding, torch.zeros_like(embedding[:, :1])], dim=-1
|
194 |
+
)
|
195 |
+
else:
|
196 |
+
embedding = repeat(timesteps, "b -> b d", d=dim)
|
197 |
+
return embedding
|
198 |
+
|
199 |
+
|
200 |
+
def zero_module(module):
|
201 |
+
"""
|
202 |
+
Zero out the parameters of a module and return it.
|
203 |
+
"""
|
204 |
+
for p in module.parameters():
|
205 |
+
p.detach().zero_()
|
206 |
+
return module
|
207 |
+
|
208 |
+
|
209 |
+
def scale_module(module, scale):
|
210 |
+
"""
|
211 |
+
Scale the parameters of a module and return it.
|
212 |
+
"""
|
213 |
+
for p in module.parameters():
|
214 |
+
p.detach().mul_(scale)
|
215 |
+
return module
|
216 |
+
|
217 |
+
|
218 |
+
def mean_flat(tensor):
|
219 |
+
"""
|
220 |
+
Take the mean over all non-batch dimensions.
|
221 |
+
"""
|
222 |
+
return tensor.mean(dim=list(range(1, len(tensor.shape))))
|
223 |
+
|
224 |
+
|
225 |
+
def normalization(channels):
|
226 |
+
"""
|
227 |
+
Make a standard normalization layer.
|
228 |
+
:param channels: number of input channels.
|
229 |
+
:return: an nn.Module for normalization.
|
230 |
+
"""
|
231 |
+
return GroupNorm32(32, channels)
|
232 |
+
|
233 |
+
|
234 |
+
# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
|
235 |
+
class SiLU(nn.Module):
|
236 |
+
def forward(self, x):
|
237 |
+
return x * torch.sigmoid(x)
|
238 |
+
|
239 |
+
|
240 |
+
class GroupNorm32(nn.GroupNorm):
|
241 |
+
def forward(self, x):
|
242 |
+
return super().forward(x.float()).type(x.dtype)
|
243 |
+
|
244 |
+
|
245 |
+
def conv_nd(dims, *args, **kwargs):
|
246 |
+
"""
|
247 |
+
Create a 1D, 2D, or 3D convolution module.
|
248 |
+
"""
|
249 |
+
if dims == 1:
|
250 |
+
return nn.Conv1d(*args, **kwargs)
|
251 |
+
elif dims == 2:
|
252 |
+
return nn.Conv2d(*args, **kwargs)
|
253 |
+
elif dims == 3:
|
254 |
+
return nn.Conv3d(*args, **kwargs)
|
255 |
+
raise ValueError(f"unsupported dimensions: {dims}")
|
256 |
+
|
257 |
+
|
258 |
+
def linear(*args, **kwargs):
|
259 |
+
"""
|
260 |
+
Create a linear module.
|
261 |
+
"""
|
262 |
+
return nn.Linear(*args, **kwargs)
|
263 |
+
|
264 |
+
|
265 |
+
def avg_pool_nd(dims, *args, **kwargs):
|
266 |
+
"""
|
267 |
+
Create a 1D, 2D, or 3D average pooling module.
|
268 |
+
"""
|
269 |
+
if dims == 1:
|
270 |
+
return nn.AvgPool1d(*args, **kwargs)
|
271 |
+
elif dims == 2:
|
272 |
+
return nn.AvgPool2d(*args, **kwargs)
|
273 |
+
elif dims == 3:
|
274 |
+
return nn.AvgPool3d(*args, **kwargs)
|
275 |
+
raise ValueError(f"unsupported dimensions: {dims}")
|
276 |
+
|
277 |
+
|
278 |
+
class HybridConditioner(nn.Module):
|
279 |
+
|
280 |
+
def __init__(self, c_concat_config, c_crossattn_config):
|
281 |
+
super().__init__()
|
282 |
+
self.concat_conditioner = instantiate_from_config(c_concat_config)
|
283 |
+
self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
|
284 |
+
|
285 |
+
def forward(self, c_concat, c_crossattn):
|
286 |
+
c_concat = self.concat_conditioner(c_concat)
|
287 |
+
c_crossattn = self.crossattn_conditioner(c_crossattn)
|
288 |
+
return {"c_concat": [c_concat], "c_crossattn": [c_crossattn]}
|
289 |
+
|
290 |
+
|
291 |
+
def noise_like(shape, device, repeat=False):
|
292 |
+
repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(
|
293 |
+
shape[0], *((1,) * (len(shape) - 1))
|
294 |
+
)
|
295 |
+
noise = lambda: torch.randn(shape, device=device)
|
296 |
+
return repeat_noise() if repeat else noise()
|
swim/modules/discriminators/n_layer_discriminator.py
ADDED
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import functools
|
2 |
+
import torch.nn as nn
|
3 |
+
|
4 |
+
|
5 |
+
from swim.utils import ActNorm
|
6 |
+
|
7 |
+
|
8 |
+
def weights_init(m):
|
9 |
+
classname = m.__class__.__name__
|
10 |
+
if classname.find("Conv") != -1:
|
11 |
+
nn.init.normal_(m.weight.data, 0.0, 0.02)
|
12 |
+
elif classname.find("BatchNorm") != -1:
|
13 |
+
nn.init.normal_(m.weight.data, 1.0, 0.02)
|
14 |
+
nn.init.constant_(m.bias.data, 0)
|
15 |
+
|
16 |
+
|
17 |
+
class NLayerDiscriminator(nn.Module):
|
18 |
+
"""Defines a PatchGAN discriminator as in Pix2Pix
|
19 |
+
--> see https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/networks.py
|
20 |
+
"""
|
21 |
+
|
22 |
+
def __init__(self, input_nc=3, ndf=64, n_layers=3, use_actnorm=False):
|
23 |
+
"""Construct a PatchGAN discriminator
|
24 |
+
Parameters:
|
25 |
+
input_nc (int) -- the number of channels in input images
|
26 |
+
ndf (int) -- the number of filters in the last conv layer
|
27 |
+
n_layers (int) -- the number of conv layers in the discriminator
|
28 |
+
norm_layer -- normalization layer
|
29 |
+
"""
|
30 |
+
super(NLayerDiscriminator, self).__init__()
|
31 |
+
if not use_actnorm:
|
32 |
+
norm_layer = nn.BatchNorm2d
|
33 |
+
else:
|
34 |
+
norm_layer = ActNorm
|
35 |
+
if (
|
36 |
+
type(norm_layer) == functools.partial
|
37 |
+
): # no need to use bias as BatchNorm2d has affine parameters
|
38 |
+
use_bias = norm_layer.func != nn.BatchNorm2d
|
39 |
+
else:
|
40 |
+
use_bias = norm_layer != nn.BatchNorm2d
|
41 |
+
|
42 |
+
kw = 4
|
43 |
+
padw = 1
|
44 |
+
sequence = [
|
45 |
+
nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw),
|
46 |
+
nn.LeakyReLU(0.2, True),
|
47 |
+
]
|
48 |
+
nf_mult = 1
|
49 |
+
nf_mult_prev = 1
|
50 |
+
for n in range(1, n_layers): # gradually increase the number of filters
|
51 |
+
nf_mult_prev = nf_mult
|
52 |
+
nf_mult = min(2**n, 8)
|
53 |
+
sequence += [
|
54 |
+
nn.Conv2d(
|
55 |
+
ndf * nf_mult_prev,
|
56 |
+
ndf * nf_mult,
|
57 |
+
kernel_size=kw,
|
58 |
+
stride=2,
|
59 |
+
padding=padw,
|
60 |
+
bias=use_bias,
|
61 |
+
),
|
62 |
+
norm_layer(ndf * nf_mult),
|
63 |
+
nn.LeakyReLU(0.2, True),
|
64 |
+
]
|
65 |
+
|
66 |
+
nf_mult_prev = nf_mult
|
67 |
+
nf_mult = min(2**n_layers, 8)
|
68 |
+
sequence += [
|
69 |
+
nn.Conv2d(
|
70 |
+
ndf * nf_mult_prev,
|
71 |
+
ndf * nf_mult,
|
72 |
+
kernel_size=kw,
|
73 |
+
stride=1,
|
74 |
+
padding=padw,
|
75 |
+
bias=use_bias,
|
76 |
+
),
|
77 |
+
norm_layer(ndf * nf_mult),
|
78 |
+
nn.LeakyReLU(0.2, True),
|
79 |
+
]
|
80 |
+
|
81 |
+
sequence += [
|
82 |
+
nn.Conv2d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw)
|
83 |
+
] # output 1 channel prediction map
|
84 |
+
self.main = nn.Sequential(*sequence)
|
85 |
+
|
86 |
+
def forward(self, input):
|
87 |
+
"""Standard forward."""
|
88 |
+
return self.main(input)
|
swim/modules/distributions/__init__.py
ADDED
File without changes
|
swim/modules/distributions/distributions.py
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import numpy as np
|
3 |
+
|
4 |
+
|
5 |
+
class AbstractDistribution:
|
6 |
+
def sample(self):
|
7 |
+
raise NotImplementedError()
|
8 |
+
|
9 |
+
def mode(self):
|
10 |
+
raise NotImplementedError()
|
11 |
+
|
12 |
+
|
13 |
+
class DiracDistribution(AbstractDistribution):
|
14 |
+
def __init__(self, value):
|
15 |
+
self.value = value
|
16 |
+
|
17 |
+
def sample(self):
|
18 |
+
return self.value
|
19 |
+
|
20 |
+
def mode(self):
|
21 |
+
return self.value
|
22 |
+
|
23 |
+
|
24 |
+
class DiagonalGaussianDistribution(object):
|
25 |
+
def __init__(self, parameters, deterministic=False):
|
26 |
+
self.parameters = parameters
|
27 |
+
self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
|
28 |
+
self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
|
29 |
+
self.deterministic = deterministic
|
30 |
+
self.std = torch.exp(0.5 * self.logvar)
|
31 |
+
self.var = torch.exp(self.logvar)
|
32 |
+
if self.deterministic:
|
33 |
+
self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
|
34 |
+
|
35 |
+
def sample(self):
|
36 |
+
x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
|
37 |
+
return x
|
38 |
+
|
39 |
+
def kl(self, other=None):
|
40 |
+
if self.deterministic:
|
41 |
+
return torch.Tensor([0.])
|
42 |
+
else:
|
43 |
+
if other is None:
|
44 |
+
return 0.5 * torch.sum(torch.pow(self.mean, 2)
|
45 |
+
+ self.var - 1.0 - self.logvar,
|
46 |
+
dim=[1, 2, 3])
|
47 |
+
else:
|
48 |
+
return 0.5 * torch.sum(
|
49 |
+
torch.pow(self.mean - other.mean, 2) / other.var
|
50 |
+
+ self.var / other.var - 1.0 - self.logvar + other.logvar,
|
51 |
+
dim=[1, 2, 3])
|
52 |
+
|
53 |
+
def nll(self, sample, dims=[1,2,3]):
|
54 |
+
if self.deterministic:
|
55 |
+
return torch.Tensor([0.])
|
56 |
+
logtwopi = np.log(2.0 * np.pi)
|
57 |
+
return 0.5 * torch.sum(
|
58 |
+
logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
|
59 |
+
dim=dims)
|
60 |
+
|
61 |
+
def mode(self):
|
62 |
+
return self.mean
|
63 |
+
|
64 |
+
|
65 |
+
def normal_kl(mean1, logvar1, mean2, logvar2):
|
66 |
+
"""
|
67 |
+
source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
|
68 |
+
Compute the KL divergence between two gaussians.
|
69 |
+
Shapes are automatically broadcasted, so batches can be compared to
|
70 |
+
scalars, among other use cases.
|
71 |
+
"""
|
72 |
+
tensor = None
|
73 |
+
for obj in (mean1, logvar1, mean2, logvar2):
|
74 |
+
if isinstance(obj, torch.Tensor):
|
75 |
+
tensor = obj
|
76 |
+
break
|
77 |
+
assert tensor is not None, "at least one argument must be a Tensor"
|
78 |
+
|
79 |
+
# Force variances to be Tensors. Broadcasting helps convert scalars to
|
80 |
+
# Tensors, but it does not work for torch.exp().
|
81 |
+
logvar1, logvar2 = [
|
82 |
+
x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
|
83 |
+
for x in (logvar1, logvar2)
|
84 |
+
]
|
85 |
+
|
86 |
+
return 0.5 * (
|
87 |
+
-1.0
|
88 |
+
+ logvar2
|
89 |
+
- logvar1
|
90 |
+
+ torch.exp(logvar1 - logvar2)
|
91 |
+
+ ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
|
92 |
+
)
|
swim/modules/ema.py
ADDED
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn
|
3 |
+
|
4 |
+
|
5 |
+
class LitEma(nn.Module):
|
6 |
+
def __init__(self, model, decay=0.9999, use_num_upates=True):
|
7 |
+
super().__init__()
|
8 |
+
if decay < 0.0 or decay > 1.0:
|
9 |
+
raise ValueError('Decay must be between 0 and 1')
|
10 |
+
|
11 |
+
self.m_name2s_name = {}
|
12 |
+
self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
|
13 |
+
self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates
|
14 |
+
else torch.tensor(-1,dtype=torch.int))
|
15 |
+
|
16 |
+
for name, p in model.named_parameters():
|
17 |
+
if p.requires_grad:
|
18 |
+
#remove as '.'-character is not allowed in buffers
|
19 |
+
s_name = name.replace('.','')
|
20 |
+
self.m_name2s_name.update({name:s_name})
|
21 |
+
self.register_buffer(s_name,p.clone().detach().data)
|
22 |
+
|
23 |
+
self.collected_params = []
|
24 |
+
|
25 |
+
def forward(self,model):
|
26 |
+
decay = self.decay
|
27 |
+
|
28 |
+
if self.num_updates >= 0:
|
29 |
+
self.num_updates += 1
|
30 |
+
decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates))
|
31 |
+
|
32 |
+
one_minus_decay = 1.0 - decay
|
33 |
+
|
34 |
+
with torch.no_grad():
|
35 |
+
m_param = dict(model.named_parameters())
|
36 |
+
shadow_params = dict(self.named_buffers())
|
37 |
+
|
38 |
+
for key in m_param:
|
39 |
+
if m_param[key].requires_grad:
|
40 |
+
sname = self.m_name2s_name[key]
|
41 |
+
shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
|
42 |
+
shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
|
43 |
+
else:
|
44 |
+
assert not key in self.m_name2s_name
|
45 |
+
|
46 |
+
def copy_to(self, model):
|
47 |
+
m_param = dict(model.named_parameters())
|
48 |
+
shadow_params = dict(self.named_buffers())
|
49 |
+
for key in m_param:
|
50 |
+
if m_param[key].requires_grad:
|
51 |
+
m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
|
52 |
+
else:
|
53 |
+
assert not key in self.m_name2s_name
|
54 |
+
|
55 |
+
def store(self, parameters):
|
56 |
+
"""
|
57 |
+
Save the current parameters for restoring later.
|
58 |
+
Args:
|
59 |
+
parameters: Iterable of `torch.nn.Parameter`; the parameters to be
|
60 |
+
temporarily stored.
|
61 |
+
"""
|
62 |
+
self.collected_params = [param.clone() for param in parameters]
|
63 |
+
|
64 |
+
def restore(self, parameters):
|
65 |
+
"""
|
66 |
+
Restore the parameters stored with the `store` method.
|
67 |
+
Useful to validate the model with EMA parameters without affecting the
|
68 |
+
original optimization process. Store the parameters before the
|
69 |
+
`copy_to` method. After validation (or model saving), use this to
|
70 |
+
restore the former parameters.
|
71 |
+
Args:
|
72 |
+
parameters: Iterable of `torch.nn.Parameter`; the parameters to be
|
73 |
+
updated with the stored parameters.
|
74 |
+
"""
|
75 |
+
for c_param, param in zip(self.collected_params, parameters):
|
76 |
+
param.data.copy_(c_param.data)
|
swim/modules/losses/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
from swim.modules.losses.contperceptual import LPIPSWithDiscriminator
|
swim/modules/losses/contperceptual.py
ADDED
@@ -0,0 +1,181 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from lpips import LPIPS
|
5 |
+
|
6 |
+
from swim.modules.discriminators.n_layer_discriminator import (
|
7 |
+
NLayerDiscriminator,
|
8 |
+
weights_init,
|
9 |
+
)
|
10 |
+
|
11 |
+
|
12 |
+
def adopt_weight(weight, global_step, threshold=0, value=0.0):
|
13 |
+
if global_step < threshold:
|
14 |
+
weight = value
|
15 |
+
return weight
|
16 |
+
|
17 |
+
|
18 |
+
def hinge_d_loss(logits_real, logits_fake):
|
19 |
+
loss_real = torch.mean(F.relu(1.0 - logits_real))
|
20 |
+
loss_fake = torch.mean(F.relu(1.0 + logits_fake))
|
21 |
+
d_loss = 0.5 * (loss_real + loss_fake)
|
22 |
+
return d_loss
|
23 |
+
|
24 |
+
|
25 |
+
def vanilla_d_loss(logits_real, logits_fake):
|
26 |
+
d_loss = 0.5 * (
|
27 |
+
torch.mean(torch.nn.functional.softplus(-logits_real))
|
28 |
+
+ torch.mean(torch.nn.functional.softplus(logits_fake))
|
29 |
+
)
|
30 |
+
return d_loss
|
31 |
+
|
32 |
+
|
33 |
+
class LPIPSWithDiscriminator(nn.Module):
|
34 |
+
def __init__(
|
35 |
+
self,
|
36 |
+
disc_start,
|
37 |
+
logvar_init=0.0,
|
38 |
+
kl_weight=1.0,
|
39 |
+
pixelloss_weight=1.0,
|
40 |
+
disc_num_layers=3,
|
41 |
+
disc_in_channels=3,
|
42 |
+
disc_factor=1.0,
|
43 |
+
disc_weight=1.0,
|
44 |
+
perceptual_weight=1.0,
|
45 |
+
use_actnorm=False,
|
46 |
+
disc_conditional=False,
|
47 |
+
disc_loss="hinge",
|
48 |
+
):
|
49 |
+
|
50 |
+
super().__init__()
|
51 |
+
assert disc_loss in ["hinge", "vanilla"]
|
52 |
+
self.kl_weight = kl_weight
|
53 |
+
self.pixel_weight = pixelloss_weight
|
54 |
+
self.perceptual_loss = LPIPS(net="vgg").eval()
|
55 |
+
self.perceptual_weight = perceptual_weight
|
56 |
+
# output log variance
|
57 |
+
self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init)
|
58 |
+
|
59 |
+
self.discriminator = NLayerDiscriminator(
|
60 |
+
input_nc=disc_in_channels, n_layers=disc_num_layers, use_actnorm=use_actnorm
|
61 |
+
).apply(weights_init)
|
62 |
+
self.discriminator_iter_start = disc_start
|
63 |
+
self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss
|
64 |
+
self.disc_factor = disc_factor
|
65 |
+
self.discriminator_weight = disc_weight
|
66 |
+
self.disc_conditional = disc_conditional
|
67 |
+
|
68 |
+
def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
|
69 |
+
if last_layer is not None:
|
70 |
+
nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
|
71 |
+
g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
|
72 |
+
else:
|
73 |
+
nll_grads = torch.autograd.grad(
|
74 |
+
nll_loss, self.last_layer[0], retain_graph=True
|
75 |
+
)[0]
|
76 |
+
g_grads = torch.autograd.grad(
|
77 |
+
g_loss, self.last_layer[0], retain_graph=True
|
78 |
+
)[0]
|
79 |
+
|
80 |
+
d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
|
81 |
+
d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
|
82 |
+
d_weight = d_weight * self.discriminator_weight
|
83 |
+
return d_weight
|
84 |
+
|
85 |
+
def forward(
|
86 |
+
self,
|
87 |
+
inputs,
|
88 |
+
reconstructions,
|
89 |
+
posteriors,
|
90 |
+
optimizer_idx,
|
91 |
+
global_step,
|
92 |
+
last_layer=None,
|
93 |
+
cond=None,
|
94 |
+
split="train",
|
95 |
+
weights=None,
|
96 |
+
):
|
97 |
+
rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
|
98 |
+
if self.perceptual_weight > 0:
|
99 |
+
p_loss = self.perceptual_loss(
|
100 |
+
inputs.contiguous(), reconstructions.contiguous()
|
101 |
+
)
|
102 |
+
rec_loss = rec_loss + self.perceptual_weight * p_loss
|
103 |
+
|
104 |
+
nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar
|
105 |
+
weighted_nll_loss = nll_loss
|
106 |
+
if weights is not None:
|
107 |
+
weighted_nll_loss = weights * nll_loss
|
108 |
+
weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0]
|
109 |
+
nll_loss = torch.sum(nll_loss) / nll_loss.shape[0]
|
110 |
+
kl_loss = posteriors.kl()
|
111 |
+
kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
|
112 |
+
|
113 |
+
# now the GAN part
|
114 |
+
if optimizer_idx == 0:
|
115 |
+
# generator update
|
116 |
+
if cond is None:
|
117 |
+
assert not self.disc_conditional
|
118 |
+
logits_fake = self.discriminator(reconstructions.contiguous())
|
119 |
+
else:
|
120 |
+
assert self.disc_conditional
|
121 |
+
logits_fake = self.discriminator(
|
122 |
+
torch.cat((reconstructions.contiguous(), cond), dim=1)
|
123 |
+
)
|
124 |
+
g_loss = -torch.mean(logits_fake)
|
125 |
+
|
126 |
+
if self.disc_factor > 0.0:
|
127 |
+
try:
|
128 |
+
d_weight = self.calculate_adaptive_weight(
|
129 |
+
nll_loss, g_loss, last_layer=last_layer
|
130 |
+
)
|
131 |
+
except RuntimeError:
|
132 |
+
assert not self.training
|
133 |
+
d_weight = torch.tensor(0.0)
|
134 |
+
else:
|
135 |
+
d_weight = torch.tensor(0.0)
|
136 |
+
|
137 |
+
disc_factor = adopt_weight(
|
138 |
+
self.disc_factor, global_step, threshold=self.discriminator_iter_start
|
139 |
+
)
|
140 |
+
loss = (
|
141 |
+
weighted_nll_loss
|
142 |
+
+ self.kl_weight * kl_loss
|
143 |
+
+ d_weight * disc_factor * g_loss
|
144 |
+
)
|
145 |
+
|
146 |
+
log = {
|
147 |
+
"{}/total_loss".format(split): loss.clone().detach().mean(),
|
148 |
+
"{}/logvar".format(split): self.logvar.detach(),
|
149 |
+
"{}/kl_loss".format(split): kl_loss.detach().mean(),
|
150 |
+
"{}/nll_loss".format(split): nll_loss.detach().mean(),
|
151 |
+
"{}/rec_loss".format(split): rec_loss.detach().mean(),
|
152 |
+
"{}/d_weight".format(split): d_weight.detach(),
|
153 |
+
"{}/disc_factor".format(split): torch.tensor(disc_factor),
|
154 |
+
"{}/g_loss".format(split): g_loss.detach().mean(),
|
155 |
+
}
|
156 |
+
return loss, log
|
157 |
+
|
158 |
+
if optimizer_idx == 1:
|
159 |
+
# second pass for discriminator update
|
160 |
+
if cond is None:
|
161 |
+
logits_real = self.discriminator(inputs.contiguous().detach())
|
162 |
+
logits_fake = self.discriminator(reconstructions.contiguous().detach())
|
163 |
+
else:
|
164 |
+
logits_real = self.discriminator(
|
165 |
+
torch.cat((inputs.contiguous().detach(), cond), dim=1)
|
166 |
+
)
|
167 |
+
logits_fake = self.discriminator(
|
168 |
+
torch.cat((reconstructions.contiguous().detach(), cond), dim=1)
|
169 |
+
)
|
170 |
+
|
171 |
+
disc_factor = adopt_weight(
|
172 |
+
self.disc_factor, global_step, threshold=self.discriminator_iter_start
|
173 |
+
)
|
174 |
+
d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
|
175 |
+
|
176 |
+
log = {
|
177 |
+
"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
|
178 |
+
"{}/logits_real".format(split): logits_real.detach().mean(),
|
179 |
+
"{}/logits_fake".format(split): logits_fake.detach().mean(),
|
180 |
+
}
|
181 |
+
return d_loss, log
|
swim/modules/x_transformer.py
ADDED
@@ -0,0 +1,641 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers"""
|
2 |
+
import torch
|
3 |
+
from torch import nn, einsum
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from functools import partial
|
6 |
+
from inspect import isfunction
|
7 |
+
from collections import namedtuple
|
8 |
+
from einops import rearrange, repeat, reduce
|
9 |
+
|
10 |
+
# constants
|
11 |
+
|
12 |
+
DEFAULT_DIM_HEAD = 64
|
13 |
+
|
14 |
+
Intermediates = namedtuple('Intermediates', [
|
15 |
+
'pre_softmax_attn',
|
16 |
+
'post_softmax_attn'
|
17 |
+
])
|
18 |
+
|
19 |
+
LayerIntermediates = namedtuple('Intermediates', [
|
20 |
+
'hiddens',
|
21 |
+
'attn_intermediates'
|
22 |
+
])
|
23 |
+
|
24 |
+
|
25 |
+
class AbsolutePositionalEmbedding(nn.Module):
|
26 |
+
def __init__(self, dim, max_seq_len):
|
27 |
+
super().__init__()
|
28 |
+
self.emb = nn.Embedding(max_seq_len, dim)
|
29 |
+
self.init_()
|
30 |
+
|
31 |
+
def init_(self):
|
32 |
+
nn.init.normal_(self.emb.weight, std=0.02)
|
33 |
+
|
34 |
+
def forward(self, x):
|
35 |
+
n = torch.arange(x.shape[1], device=x.device)
|
36 |
+
return self.emb(n)[None, :, :]
|
37 |
+
|
38 |
+
|
39 |
+
class FixedPositionalEmbedding(nn.Module):
|
40 |
+
def __init__(self, dim):
|
41 |
+
super().__init__()
|
42 |
+
inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
|
43 |
+
self.register_buffer('inv_freq', inv_freq)
|
44 |
+
|
45 |
+
def forward(self, x, seq_dim=1, offset=0):
|
46 |
+
t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset
|
47 |
+
sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq)
|
48 |
+
emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
|
49 |
+
return emb[None, :, :]
|
50 |
+
|
51 |
+
|
52 |
+
# helpers
|
53 |
+
|
54 |
+
def exists(val):
|
55 |
+
return val is not None
|
56 |
+
|
57 |
+
|
58 |
+
def default(val, d):
|
59 |
+
if exists(val):
|
60 |
+
return val
|
61 |
+
return d() if isfunction(d) else d
|
62 |
+
|
63 |
+
|
64 |
+
def always(val):
|
65 |
+
def inner(*args, **kwargs):
|
66 |
+
return val
|
67 |
+
return inner
|
68 |
+
|
69 |
+
|
70 |
+
def not_equals(val):
|
71 |
+
def inner(x):
|
72 |
+
return x != val
|
73 |
+
return inner
|
74 |
+
|
75 |
+
|
76 |
+
def equals(val):
|
77 |
+
def inner(x):
|
78 |
+
return x == val
|
79 |
+
return inner
|
80 |
+
|
81 |
+
|
82 |
+
def max_neg_value(tensor):
|
83 |
+
return -torch.finfo(tensor.dtype).max
|
84 |
+
|
85 |
+
|
86 |
+
# keyword argument helpers
|
87 |
+
|
88 |
+
def pick_and_pop(keys, d):
|
89 |
+
values = list(map(lambda key: d.pop(key), keys))
|
90 |
+
return dict(zip(keys, values))
|
91 |
+
|
92 |
+
|
93 |
+
def group_dict_by_key(cond, d):
|
94 |
+
return_val = [dict(), dict()]
|
95 |
+
for key in d.keys():
|
96 |
+
match = bool(cond(key))
|
97 |
+
ind = int(not match)
|
98 |
+
return_val[ind][key] = d[key]
|
99 |
+
return (*return_val,)
|
100 |
+
|
101 |
+
|
102 |
+
def string_begins_with(prefix, str):
|
103 |
+
return str.startswith(prefix)
|
104 |
+
|
105 |
+
|
106 |
+
def group_by_key_prefix(prefix, d):
|
107 |
+
return group_dict_by_key(partial(string_begins_with, prefix), d)
|
108 |
+
|
109 |
+
|
110 |
+
def groupby_prefix_and_trim(prefix, d):
|
111 |
+
kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
|
112 |
+
kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
|
113 |
+
return kwargs_without_prefix, kwargs
|
114 |
+
|
115 |
+
|
116 |
+
# classes
|
117 |
+
class Scale(nn.Module):
|
118 |
+
def __init__(self, value, fn):
|
119 |
+
super().__init__()
|
120 |
+
self.value = value
|
121 |
+
self.fn = fn
|
122 |
+
|
123 |
+
def forward(self, x, **kwargs):
|
124 |
+
x, *rest = self.fn(x, **kwargs)
|
125 |
+
return (x * self.value, *rest)
|
126 |
+
|
127 |
+
|
128 |
+
class Rezero(nn.Module):
|
129 |
+
def __init__(self, fn):
|
130 |
+
super().__init__()
|
131 |
+
self.fn = fn
|
132 |
+
self.g = nn.Parameter(torch.zeros(1))
|
133 |
+
|
134 |
+
def forward(self, x, **kwargs):
|
135 |
+
x, *rest = self.fn(x, **kwargs)
|
136 |
+
return (x * self.g, *rest)
|
137 |
+
|
138 |
+
|
139 |
+
class ScaleNorm(nn.Module):
|
140 |
+
def __init__(self, dim, eps=1e-5):
|
141 |
+
super().__init__()
|
142 |
+
self.scale = dim ** -0.5
|
143 |
+
self.eps = eps
|
144 |
+
self.g = nn.Parameter(torch.ones(1))
|
145 |
+
|
146 |
+
def forward(self, x):
|
147 |
+
norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
|
148 |
+
return x / norm.clamp(min=self.eps) * self.g
|
149 |
+
|
150 |
+
|
151 |
+
class RMSNorm(nn.Module):
|
152 |
+
def __init__(self, dim, eps=1e-8):
|
153 |
+
super().__init__()
|
154 |
+
self.scale = dim ** -0.5
|
155 |
+
self.eps = eps
|
156 |
+
self.g = nn.Parameter(torch.ones(dim))
|
157 |
+
|
158 |
+
def forward(self, x):
|
159 |
+
norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
|
160 |
+
return x / norm.clamp(min=self.eps) * self.g
|
161 |
+
|
162 |
+
|
163 |
+
class Residual(nn.Module):
|
164 |
+
def forward(self, x, residual):
|
165 |
+
return x + residual
|
166 |
+
|
167 |
+
|
168 |
+
class GRUGating(nn.Module):
|
169 |
+
def __init__(self, dim):
|
170 |
+
super().__init__()
|
171 |
+
self.gru = nn.GRUCell(dim, dim)
|
172 |
+
|
173 |
+
def forward(self, x, residual):
|
174 |
+
gated_output = self.gru(
|
175 |
+
rearrange(x, 'b n d -> (b n) d'),
|
176 |
+
rearrange(residual, 'b n d -> (b n) d')
|
177 |
+
)
|
178 |
+
|
179 |
+
return gated_output.reshape_as(x)
|
180 |
+
|
181 |
+
|
182 |
+
# feedforward
|
183 |
+
|
184 |
+
class GEGLU(nn.Module):
|
185 |
+
def __init__(self, dim_in, dim_out):
|
186 |
+
super().__init__()
|
187 |
+
self.proj = nn.Linear(dim_in, dim_out * 2)
|
188 |
+
|
189 |
+
def forward(self, x):
|
190 |
+
x, gate = self.proj(x).chunk(2, dim=-1)
|
191 |
+
return x * F.gelu(gate)
|
192 |
+
|
193 |
+
|
194 |
+
class FeedForward(nn.Module):
|
195 |
+
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
|
196 |
+
super().__init__()
|
197 |
+
inner_dim = int(dim * mult)
|
198 |
+
dim_out = default(dim_out, dim)
|
199 |
+
project_in = nn.Sequential(
|
200 |
+
nn.Linear(dim, inner_dim),
|
201 |
+
nn.GELU()
|
202 |
+
) if not glu else GEGLU(dim, inner_dim)
|
203 |
+
|
204 |
+
self.net = nn.Sequential(
|
205 |
+
project_in,
|
206 |
+
nn.Dropout(dropout),
|
207 |
+
nn.Linear(inner_dim, dim_out)
|
208 |
+
)
|
209 |
+
|
210 |
+
def forward(self, x):
|
211 |
+
return self.net(x)
|
212 |
+
|
213 |
+
|
214 |
+
# attention.
|
215 |
+
class Attention(nn.Module):
|
216 |
+
def __init__(
|
217 |
+
self,
|
218 |
+
dim,
|
219 |
+
dim_head=DEFAULT_DIM_HEAD,
|
220 |
+
heads=8,
|
221 |
+
causal=False,
|
222 |
+
mask=None,
|
223 |
+
talking_heads=False,
|
224 |
+
sparse_topk=None,
|
225 |
+
use_entmax15=False,
|
226 |
+
num_mem_kv=0,
|
227 |
+
dropout=0.,
|
228 |
+
on_attn=False
|
229 |
+
):
|
230 |
+
super().__init__()
|
231 |
+
if use_entmax15:
|
232 |
+
raise NotImplementedError("Check out entmax activation instead of softmax activation!")
|
233 |
+
self.scale = dim_head ** -0.5
|
234 |
+
self.heads = heads
|
235 |
+
self.causal = causal
|
236 |
+
self.mask = mask
|
237 |
+
|
238 |
+
inner_dim = dim_head * heads
|
239 |
+
|
240 |
+
self.to_q = nn.Linear(dim, inner_dim, bias=False)
|
241 |
+
self.to_k = nn.Linear(dim, inner_dim, bias=False)
|
242 |
+
self.to_v = nn.Linear(dim, inner_dim, bias=False)
|
243 |
+
self.dropout = nn.Dropout(dropout)
|
244 |
+
|
245 |
+
# talking heads
|
246 |
+
self.talking_heads = talking_heads
|
247 |
+
if talking_heads:
|
248 |
+
self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads))
|
249 |
+
self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads))
|
250 |
+
|
251 |
+
# explicit topk sparse attention
|
252 |
+
self.sparse_topk = sparse_topk
|
253 |
+
|
254 |
+
# entmax
|
255 |
+
#self.attn_fn = entmax15 if use_entmax15 else F.softmax
|
256 |
+
self.attn_fn = F.softmax
|
257 |
+
|
258 |
+
# add memory key / values
|
259 |
+
self.num_mem_kv = num_mem_kv
|
260 |
+
if num_mem_kv > 0:
|
261 |
+
self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
|
262 |
+
self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
|
263 |
+
|
264 |
+
# attention on attention
|
265 |
+
self.attn_on_attn = on_attn
|
266 |
+
self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim)
|
267 |
+
|
268 |
+
def forward(
|
269 |
+
self,
|
270 |
+
x,
|
271 |
+
context=None,
|
272 |
+
mask=None,
|
273 |
+
context_mask=None,
|
274 |
+
rel_pos=None,
|
275 |
+
sinusoidal_emb=None,
|
276 |
+
prev_attn=None,
|
277 |
+
mem=None
|
278 |
+
):
|
279 |
+
b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device
|
280 |
+
kv_input = default(context, x)
|
281 |
+
|
282 |
+
q_input = x
|
283 |
+
k_input = kv_input
|
284 |
+
v_input = kv_input
|
285 |
+
|
286 |
+
if exists(mem):
|
287 |
+
k_input = torch.cat((mem, k_input), dim=-2)
|
288 |
+
v_input = torch.cat((mem, v_input), dim=-2)
|
289 |
+
|
290 |
+
if exists(sinusoidal_emb):
|
291 |
+
# in shortformer, the query would start at a position offset depending on the past cached memory
|
292 |
+
offset = k_input.shape[-2] - q_input.shape[-2]
|
293 |
+
q_input = q_input + sinusoidal_emb(q_input, offset=offset)
|
294 |
+
k_input = k_input + sinusoidal_emb(k_input)
|
295 |
+
|
296 |
+
q = self.to_q(q_input)
|
297 |
+
k = self.to_k(k_input)
|
298 |
+
v = self.to_v(v_input)
|
299 |
+
|
300 |
+
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))
|
301 |
+
|
302 |
+
input_mask = None
|
303 |
+
if any(map(exists, (mask, context_mask))):
|
304 |
+
q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool())
|
305 |
+
k_mask = q_mask if not exists(context) else context_mask
|
306 |
+
k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool())
|
307 |
+
q_mask = rearrange(q_mask, 'b i -> b () i ()')
|
308 |
+
k_mask = rearrange(k_mask, 'b j -> b () () j')
|
309 |
+
input_mask = q_mask * k_mask
|
310 |
+
|
311 |
+
if self.num_mem_kv > 0:
|
312 |
+
mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v))
|
313 |
+
k = torch.cat((mem_k, k), dim=-2)
|
314 |
+
v = torch.cat((mem_v, v), dim=-2)
|
315 |
+
if exists(input_mask):
|
316 |
+
input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True)
|
317 |
+
|
318 |
+
dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
|
319 |
+
mask_value = max_neg_value(dots)
|
320 |
+
|
321 |
+
if exists(prev_attn):
|
322 |
+
dots = dots + prev_attn
|
323 |
+
|
324 |
+
pre_softmax_attn = dots
|
325 |
+
|
326 |
+
if talking_heads:
|
327 |
+
dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous()
|
328 |
+
|
329 |
+
if exists(rel_pos):
|
330 |
+
dots = rel_pos(dots)
|
331 |
+
|
332 |
+
if exists(input_mask):
|
333 |
+
dots.masked_fill_(~input_mask, mask_value)
|
334 |
+
del input_mask
|
335 |
+
|
336 |
+
if self.causal:
|
337 |
+
i, j = dots.shape[-2:]
|
338 |
+
r = torch.arange(i, device=device)
|
339 |
+
mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j')
|
340 |
+
mask = F.pad(mask, (j - i, 0), value=False)
|
341 |
+
dots.masked_fill_(mask, mask_value)
|
342 |
+
del mask
|
343 |
+
|
344 |
+
if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]:
|
345 |
+
top, _ = dots.topk(self.sparse_topk, dim=-1)
|
346 |
+
vk = top[..., -1].unsqueeze(-1).expand_as(dots)
|
347 |
+
mask = dots < vk
|
348 |
+
dots.masked_fill_(mask, mask_value)
|
349 |
+
del mask
|
350 |
+
|
351 |
+
attn = self.attn_fn(dots, dim=-1)
|
352 |
+
post_softmax_attn = attn
|
353 |
+
|
354 |
+
attn = self.dropout(attn)
|
355 |
+
|
356 |
+
if talking_heads:
|
357 |
+
attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous()
|
358 |
+
|
359 |
+
out = einsum('b h i j, b h j d -> b h i d', attn, v)
|
360 |
+
out = rearrange(out, 'b h n d -> b n (h d)')
|
361 |
+
|
362 |
+
intermediates = Intermediates(
|
363 |
+
pre_softmax_attn=pre_softmax_attn,
|
364 |
+
post_softmax_attn=post_softmax_attn
|
365 |
+
)
|
366 |
+
|
367 |
+
return self.to_out(out), intermediates
|
368 |
+
|
369 |
+
|
370 |
+
class AttentionLayers(nn.Module):
|
371 |
+
def __init__(
|
372 |
+
self,
|
373 |
+
dim,
|
374 |
+
depth,
|
375 |
+
heads=8,
|
376 |
+
causal=False,
|
377 |
+
cross_attend=False,
|
378 |
+
only_cross=False,
|
379 |
+
use_scalenorm=False,
|
380 |
+
use_rmsnorm=False,
|
381 |
+
use_rezero=False,
|
382 |
+
rel_pos_num_buckets=32,
|
383 |
+
rel_pos_max_distance=128,
|
384 |
+
position_infused_attn=False,
|
385 |
+
custom_layers=None,
|
386 |
+
sandwich_coef=None,
|
387 |
+
par_ratio=None,
|
388 |
+
residual_attn=False,
|
389 |
+
cross_residual_attn=False,
|
390 |
+
macaron=False,
|
391 |
+
pre_norm=True,
|
392 |
+
gate_residual=False,
|
393 |
+
**kwargs
|
394 |
+
):
|
395 |
+
super().__init__()
|
396 |
+
ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs)
|
397 |
+
attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs)
|
398 |
+
|
399 |
+
dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
|
400 |
+
|
401 |
+
self.dim = dim
|
402 |
+
self.depth = depth
|
403 |
+
self.layers = nn.ModuleList([])
|
404 |
+
|
405 |
+
self.has_pos_emb = position_infused_attn
|
406 |
+
self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None
|
407 |
+
self.rotary_pos_emb = always(None)
|
408 |
+
|
409 |
+
assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
|
410 |
+
self.rel_pos = None
|
411 |
+
|
412 |
+
self.pre_norm = pre_norm
|
413 |
+
|
414 |
+
self.residual_attn = residual_attn
|
415 |
+
self.cross_residual_attn = cross_residual_attn
|
416 |
+
|
417 |
+
norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm
|
418 |
+
norm_class = RMSNorm if use_rmsnorm else norm_class
|
419 |
+
norm_fn = partial(norm_class, dim)
|
420 |
+
|
421 |
+
norm_fn = nn.Identity if use_rezero else norm_fn
|
422 |
+
branch_fn = Rezero if use_rezero else None
|
423 |
+
|
424 |
+
if cross_attend and not only_cross:
|
425 |
+
default_block = ('a', 'c', 'f')
|
426 |
+
elif cross_attend and only_cross:
|
427 |
+
default_block = ('c', 'f')
|
428 |
+
else:
|
429 |
+
default_block = ('a', 'f')
|
430 |
+
|
431 |
+
if macaron:
|
432 |
+
default_block = ('f',) + default_block
|
433 |
+
|
434 |
+
if exists(custom_layers):
|
435 |
+
layer_types = custom_layers
|
436 |
+
elif exists(par_ratio):
|
437 |
+
par_depth = depth * len(default_block)
|
438 |
+
assert 1 < par_ratio <= par_depth, 'par ratio out of range'
|
439 |
+
default_block = tuple(filter(not_equals('f'), default_block))
|
440 |
+
par_attn = par_depth // par_ratio
|
441 |
+
depth_cut = par_depth * 2 // 3 # 2 / 3 attention layer cutoff suggested by PAR paper
|
442 |
+
par_width = (depth_cut + depth_cut // par_attn) // par_attn
|
443 |
+
assert len(default_block) <= par_width, 'default block is too large for par_ratio'
|
444 |
+
par_block = default_block + ('f',) * (par_width - len(default_block))
|
445 |
+
par_head = par_block * par_attn
|
446 |
+
layer_types = par_head + ('f',) * (par_depth - len(par_head))
|
447 |
+
elif exists(sandwich_coef):
|
448 |
+
assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
|
449 |
+
layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
|
450 |
+
else:
|
451 |
+
layer_types = default_block * depth
|
452 |
+
|
453 |
+
self.layer_types = layer_types
|
454 |
+
self.num_attn_layers = len(list(filter(equals('a'), layer_types)))
|
455 |
+
|
456 |
+
for layer_type in self.layer_types:
|
457 |
+
if layer_type == 'a':
|
458 |
+
layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs)
|
459 |
+
elif layer_type == 'c':
|
460 |
+
layer = Attention(dim, heads=heads, **attn_kwargs)
|
461 |
+
elif layer_type == 'f':
|
462 |
+
layer = FeedForward(dim, **ff_kwargs)
|
463 |
+
layer = layer if not macaron else Scale(0.5, layer)
|
464 |
+
else:
|
465 |
+
raise Exception(f'invalid layer type {layer_type}')
|
466 |
+
|
467 |
+
if isinstance(layer, Attention) and exists(branch_fn):
|
468 |
+
layer = branch_fn(layer)
|
469 |
+
|
470 |
+
if gate_residual:
|
471 |
+
residual_fn = GRUGating(dim)
|
472 |
+
else:
|
473 |
+
residual_fn = Residual()
|
474 |
+
|
475 |
+
self.layers.append(nn.ModuleList([
|
476 |
+
norm_fn(),
|
477 |
+
layer,
|
478 |
+
residual_fn
|
479 |
+
]))
|
480 |
+
|
481 |
+
def forward(
|
482 |
+
self,
|
483 |
+
x,
|
484 |
+
context=None,
|
485 |
+
mask=None,
|
486 |
+
context_mask=None,
|
487 |
+
mems=None,
|
488 |
+
return_hiddens=False
|
489 |
+
):
|
490 |
+
hiddens = []
|
491 |
+
intermediates = []
|
492 |
+
prev_attn = None
|
493 |
+
prev_cross_attn = None
|
494 |
+
|
495 |
+
mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers
|
496 |
+
|
497 |
+
for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)):
|
498 |
+
is_last = ind == (len(self.layers) - 1)
|
499 |
+
|
500 |
+
if layer_type == 'a':
|
501 |
+
hiddens.append(x)
|
502 |
+
layer_mem = mems.pop(0)
|
503 |
+
|
504 |
+
residual = x
|
505 |
+
|
506 |
+
if self.pre_norm:
|
507 |
+
x = norm(x)
|
508 |
+
|
509 |
+
if layer_type == 'a':
|
510 |
+
out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos,
|
511 |
+
prev_attn=prev_attn, mem=layer_mem)
|
512 |
+
elif layer_type == 'c':
|
513 |
+
out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn)
|
514 |
+
elif layer_type == 'f':
|
515 |
+
out = block(x)
|
516 |
+
|
517 |
+
x = residual_fn(out, residual)
|
518 |
+
|
519 |
+
if layer_type in ('a', 'c'):
|
520 |
+
intermediates.append(inter)
|
521 |
+
|
522 |
+
if layer_type == 'a' and self.residual_attn:
|
523 |
+
prev_attn = inter.pre_softmax_attn
|
524 |
+
elif layer_type == 'c' and self.cross_residual_attn:
|
525 |
+
prev_cross_attn = inter.pre_softmax_attn
|
526 |
+
|
527 |
+
if not self.pre_norm and not is_last:
|
528 |
+
x = norm(x)
|
529 |
+
|
530 |
+
if return_hiddens:
|
531 |
+
intermediates = LayerIntermediates(
|
532 |
+
hiddens=hiddens,
|
533 |
+
attn_intermediates=intermediates
|
534 |
+
)
|
535 |
+
|
536 |
+
return x, intermediates
|
537 |
+
|
538 |
+
return x
|
539 |
+
|
540 |
+
|
541 |
+
class Encoder(AttentionLayers):
|
542 |
+
def __init__(self, **kwargs):
|
543 |
+
assert 'causal' not in kwargs, 'cannot set causality on encoder'
|
544 |
+
super().__init__(causal=False, **kwargs)
|
545 |
+
|
546 |
+
|
547 |
+
|
548 |
+
class TransformerWrapper(nn.Module):
|
549 |
+
def __init__(
|
550 |
+
self,
|
551 |
+
*,
|
552 |
+
num_tokens,
|
553 |
+
max_seq_len,
|
554 |
+
attn_layers,
|
555 |
+
emb_dim=None,
|
556 |
+
max_mem_len=0.,
|
557 |
+
emb_dropout=0.,
|
558 |
+
num_memory_tokens=None,
|
559 |
+
tie_embedding=False,
|
560 |
+
use_pos_emb=True
|
561 |
+
):
|
562 |
+
super().__init__()
|
563 |
+
assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'
|
564 |
+
|
565 |
+
dim = attn_layers.dim
|
566 |
+
emb_dim = default(emb_dim, dim)
|
567 |
+
|
568 |
+
self.max_seq_len = max_seq_len
|
569 |
+
self.max_mem_len = max_mem_len
|
570 |
+
self.num_tokens = num_tokens
|
571 |
+
|
572 |
+
self.token_emb = nn.Embedding(num_tokens, emb_dim)
|
573 |
+
self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if (
|
574 |
+
use_pos_emb and not attn_layers.has_pos_emb) else always(0)
|
575 |
+
self.emb_dropout = nn.Dropout(emb_dropout)
|
576 |
+
|
577 |
+
self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity()
|
578 |
+
self.attn_layers = attn_layers
|
579 |
+
self.norm = nn.LayerNorm(dim)
|
580 |
+
|
581 |
+
self.init_()
|
582 |
+
|
583 |
+
self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t()
|
584 |
+
|
585 |
+
# memory tokens (like [cls]) from Memory Transformers paper
|
586 |
+
num_memory_tokens = default(num_memory_tokens, 0)
|
587 |
+
self.num_memory_tokens = num_memory_tokens
|
588 |
+
if num_memory_tokens > 0:
|
589 |
+
self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim))
|
590 |
+
|
591 |
+
# let funnel encoder know number of memory tokens, if specified
|
592 |
+
if hasattr(attn_layers, 'num_memory_tokens'):
|
593 |
+
attn_layers.num_memory_tokens = num_memory_tokens
|
594 |
+
|
595 |
+
def init_(self):
|
596 |
+
nn.init.normal_(self.token_emb.weight, std=0.02)
|
597 |
+
|
598 |
+
def forward(
|
599 |
+
self,
|
600 |
+
x,
|
601 |
+
return_embeddings=False,
|
602 |
+
mask=None,
|
603 |
+
return_mems=False,
|
604 |
+
return_attn=False,
|
605 |
+
mems=None,
|
606 |
+
**kwargs
|
607 |
+
):
|
608 |
+
b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens
|
609 |
+
x = self.token_emb(x)
|
610 |
+
x += self.pos_emb(x)
|
611 |
+
x = self.emb_dropout(x)
|
612 |
+
|
613 |
+
x = self.project_emb(x)
|
614 |
+
|
615 |
+
if num_mem > 0:
|
616 |
+
mem = repeat(self.memory_tokens, 'n d -> b n d', b=b)
|
617 |
+
x = torch.cat((mem, x), dim=1)
|
618 |
+
|
619 |
+
# auto-handle masking after appending memory tokens
|
620 |
+
if exists(mask):
|
621 |
+
mask = F.pad(mask, (num_mem, 0), value=True)
|
622 |
+
|
623 |
+
x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
|
624 |
+
x = self.norm(x)
|
625 |
+
|
626 |
+
mem, x = x[:, :num_mem], x[:, num_mem:]
|
627 |
+
|
628 |
+
out = self.to_logits(x) if not return_embeddings else x
|
629 |
+
|
630 |
+
if return_mems:
|
631 |
+
hiddens = intermediates.hiddens
|
632 |
+
new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens
|
633 |
+
new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems))
|
634 |
+
return out, new_mems
|
635 |
+
|
636 |
+
if return_attn:
|
637 |
+
attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
|
638 |
+
return out, attn_maps
|
639 |
+
|
640 |
+
return out
|
641 |
+
|
swim/unet.py
DELETED
@@ -1,169 +0,0 @@
|
|
1 |
-
import math
|
2 |
-
from typing import List
|
3 |
-
|
4 |
-
import torch
|
5 |
-
import torch.nn as nn
|
6 |
-
import torch.nn.functional as F
|
7 |
-
|
8 |
-
from .attention_blocks import SpatialTransformer
|
9 |
-
from .blocks import (
|
10 |
-
DownSample,
|
11 |
-
ResnetBlock,
|
12 |
-
TimestepEmbedSequential,
|
13 |
-
UpSample,
|
14 |
-
Normalization,
|
15 |
-
get_timestep_embedding,
|
16 |
-
)
|
17 |
-
|
18 |
-
|
19 |
-
class UNet(nn.Module):
|
20 |
-
"""
|
21 |
-
## U-Net model
|
22 |
-
"""
|
23 |
-
|
24 |
-
def __init__(
|
25 |
-
self,
|
26 |
-
*,
|
27 |
-
in_channels: int,
|
28 |
-
out_channels: int,
|
29 |
-
channels: int,
|
30 |
-
n_res_blocks: int,
|
31 |
-
attention_levels: List[int],
|
32 |
-
channel_multipliers: List[int],
|
33 |
-
n_heads: int,
|
34 |
-
tf_layers: int = 1,
|
35 |
-
d_cond: int = 768
|
36 |
-
):
|
37 |
-
"""
|
38 |
-
:param in_channels: is the number of channels in the input feature map
|
39 |
-
:param out_channels: is the number of channels in the output feature map
|
40 |
-
:param channels: is the base channel count for the model
|
41 |
-
:param n_res_blocks: number of residual blocks at each level
|
42 |
-
:param attention_levels: are the levels at which attention should be performed
|
43 |
-
:param channel_multipliers: are the multiplicative factors for number of channels for each level
|
44 |
-
:param n_heads: is the number of attention heads in the transformers
|
45 |
-
:param tf_layers: is the number of transformer layers in the transformers
|
46 |
-
:param d_cond: is the size of the conditional embedding in the transformers
|
47 |
-
"""
|
48 |
-
super().__init__()
|
49 |
-
self.channels = channels
|
50 |
-
|
51 |
-
# Number of levels
|
52 |
-
levels = len(channel_multipliers)
|
53 |
-
# Size time embeddings
|
54 |
-
d_time_emb = channels * 4
|
55 |
-
self.time_embed = nn.Sequential(
|
56 |
-
nn.Linear(channels, d_time_emb),
|
57 |
-
nn.SiLU(),
|
58 |
-
nn.Linear(d_time_emb, d_time_emb),
|
59 |
-
)
|
60 |
-
|
61 |
-
# Input half of the U-Net
|
62 |
-
self.input_blocks = nn.ModuleList()
|
63 |
-
# Initial $3 \times 3$ convolution that maps the input to `channels`.
|
64 |
-
# The blocks are wrapped in `TimestepEmbedSequential` module because
|
65 |
-
# different modules have different forward function signatures;
|
66 |
-
# for example, convolution only accepts the feature map and
|
67 |
-
# residual blocks accept the feature map and time embedding.
|
68 |
-
# `TimestepEmbedSequential` calls them accordingly.
|
69 |
-
self.input_blocks.append(
|
70 |
-
TimestepEmbedSequential(nn.Conv2d(in_channels, channels, 3, padding=1))
|
71 |
-
)
|
72 |
-
# Number of channels at each block in the input half of U-Net
|
73 |
-
input_block_channels = [channels]
|
74 |
-
# Number of channels at each level
|
75 |
-
channels_list = [channels * m for m in channel_multipliers]
|
76 |
-
# Prepare levels
|
77 |
-
for i in range(levels):
|
78 |
-
# Add the residual blocks and attentions
|
79 |
-
for _ in range(n_res_blocks):
|
80 |
-
# Residual block maps from previous number of channels to the number of
|
81 |
-
# channels in the current level
|
82 |
-
layers = [
|
83 |
-
ResnetBlock(channels, d_time_emb, out_channels=channels_list[i])
|
84 |
-
]
|
85 |
-
channels = channels_list[i]
|
86 |
-
# Add transformer
|
87 |
-
if i in attention_levels:
|
88 |
-
layers.append(
|
89 |
-
SpatialTransformer(channels, n_heads, tf_layers, d_cond)
|
90 |
-
)
|
91 |
-
# Add them to the input half of the U-Net and keep track of the number of channels of
|
92 |
-
# its output
|
93 |
-
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
94 |
-
input_block_channels.append(channels)
|
95 |
-
# Down sample at all levels except last
|
96 |
-
if i != levels - 1:
|
97 |
-
self.input_blocks.append(TimestepEmbedSequential(DownSample(channels)))
|
98 |
-
input_block_channels.append(channels)
|
99 |
-
|
100 |
-
# The middle of the U-Net
|
101 |
-
self.middle_block = TimestepEmbedSequential(
|
102 |
-
ResnetBlock(channels, d_time_emb),
|
103 |
-
SpatialTransformer(channels, n_heads, tf_layers, d_cond),
|
104 |
-
ResnetBlock(channels, d_time_emb),
|
105 |
-
)
|
106 |
-
|
107 |
-
# Second half of the U-Net
|
108 |
-
self.output_blocks = nn.ModuleList([])
|
109 |
-
# Prepare levels in reverse order
|
110 |
-
for i in reversed(range(levels)):
|
111 |
-
# Add the residual blocks and attentions
|
112 |
-
for j in range(n_res_blocks + 1):
|
113 |
-
# Residual block maps from previous number of channels plus the
|
114 |
-
# skip connections from the input half of U-Net to the number of
|
115 |
-
# channels in the current level.
|
116 |
-
layers = [
|
117 |
-
ResnetBlock(
|
118 |
-
channels + input_block_channels.pop(),
|
119 |
-
d_time_emb,
|
120 |
-
out_channels=channels_list[i],
|
121 |
-
)
|
122 |
-
]
|
123 |
-
channels = channels_list[i]
|
124 |
-
# Add transformer
|
125 |
-
if i in attention_levels:
|
126 |
-
layers.append(
|
127 |
-
SpatialTransformer(channels, n_heads, tf_layers, d_cond)
|
128 |
-
)
|
129 |
-
# Up-sample at every level after last residual block
|
130 |
-
# except the last one.
|
131 |
-
# Note that we are iterating in reverse; i.e. `i == 0` is the last.
|
132 |
-
if i != 0 and j == n_res_blocks:
|
133 |
-
layers.append(UpSample(channels))
|
134 |
-
# Add to the output half of the U-Net
|
135 |
-
self.output_blocks.append(TimestepEmbedSequential(*layers))
|
136 |
-
|
137 |
-
# Final normalization and $3 \times 3$ convolution
|
138 |
-
self.out = nn.Sequential(
|
139 |
-
Normalization(channels),
|
140 |
-
nn.SiLU(),
|
141 |
-
nn.Conv2d(channels, out_channels, 3, padding=1),
|
142 |
-
)
|
143 |
-
|
144 |
-
def forward(self, x: torch.Tensor, timesteps: torch.Tensor, cond: torch.Tensor):
|
145 |
-
"""
|
146 |
-
:param x: is the input feature map of shape `[batch_size, channels, width, height]`
|
147 |
-
:param timesteps: are the time steps of shape `[batch_size]`
|
148 |
-
:param cond: conditioning of shape `[batch_size, n_cond, d_cond]`
|
149 |
-
"""
|
150 |
-
# To store the input half outputs for skip connections
|
151 |
-
x_input_block = []
|
152 |
-
|
153 |
-
# Get time step embeddings
|
154 |
-
t_emb = get_timestep_embedding(timesteps, self.channels * 2)
|
155 |
-
t_emb = self.time_embed(t_emb)
|
156 |
-
|
157 |
-
# Input half of the U-Net
|
158 |
-
for module in self.input_blocks:
|
159 |
-
x = module(x, t_emb, cond)
|
160 |
-
x_input_block.append(x)
|
161 |
-
# Middle of the U-Net
|
162 |
-
x = self.middle_block(x, t_emb, cond)
|
163 |
-
# Output half of the U-Net
|
164 |
-
for module in self.output_blocks:
|
165 |
-
x = torch.cat([x, x_input_block.pop()], dim=1)
|
166 |
-
x = module(x, t_emb, cond)
|
167 |
-
|
168 |
-
# Final normalization and $3 \times 3$ convolution
|
169 |
-
return self.out(x)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
swim/utils.py
ADDED
@@ -0,0 +1,297 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import importlib
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
import numpy as np
|
6 |
+
from collections import abc
|
7 |
+
from einops import rearrange
|
8 |
+
from functools import partial
|
9 |
+
|
10 |
+
import multiprocessing as mp
|
11 |
+
from threading import Thread
|
12 |
+
from queue import Queue
|
13 |
+
|
14 |
+
from inspect import isfunction
|
15 |
+
from PIL import Image, ImageDraw, ImageFont
|
16 |
+
|
17 |
+
|
18 |
+
def log_txt_as_img(wh, xc, size=10):
|
19 |
+
# wh a tuple of (width, height)
|
20 |
+
# xc a list of captions to plot
|
21 |
+
b = len(xc)
|
22 |
+
txts = list()
|
23 |
+
for bi in range(b):
|
24 |
+
txt = Image.new("RGB", wh, color="white")
|
25 |
+
draw = ImageDraw.Draw(txt)
|
26 |
+
font = ImageFont.truetype("data/DejaVuSans.ttf", size=size)
|
27 |
+
nc = int(40 * (wh[0] / 256))
|
28 |
+
lines = "\n".join(
|
29 |
+
xc[bi][start : start + nc] for start in range(0, len(xc[bi]), nc)
|
30 |
+
)
|
31 |
+
|
32 |
+
try:
|
33 |
+
draw.text((0, 0), lines, fill="black", font=font)
|
34 |
+
except UnicodeEncodeError:
|
35 |
+
print("Cant encode string for logging. Skipping.")
|
36 |
+
|
37 |
+
txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0
|
38 |
+
txts.append(txt)
|
39 |
+
txts = np.stack(txts)
|
40 |
+
txts = torch.tensor(txts)
|
41 |
+
return txts
|
42 |
+
|
43 |
+
|
44 |
+
def ismap(x):
|
45 |
+
if not isinstance(x, torch.Tensor):
|
46 |
+
return False
|
47 |
+
return (len(x.shape) == 4) and (x.shape[1] > 3)
|
48 |
+
|
49 |
+
|
50 |
+
def isimage(x):
|
51 |
+
if not isinstance(x, torch.Tensor):
|
52 |
+
return False
|
53 |
+
return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
|
54 |
+
|
55 |
+
|
56 |
+
def exists(x):
|
57 |
+
return x is not None
|
58 |
+
|
59 |
+
|
60 |
+
def default(val, d):
|
61 |
+
if exists(val):
|
62 |
+
return val
|
63 |
+
return d() if isfunction(d) else d
|
64 |
+
|
65 |
+
|
66 |
+
def mean_flat(tensor):
|
67 |
+
"""
|
68 |
+
https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86
|
69 |
+
Take the mean over all non-batch dimensions.
|
70 |
+
"""
|
71 |
+
return tensor.mean(dim=list(range(1, len(tensor.shape))))
|
72 |
+
|
73 |
+
|
74 |
+
def count_params(model, verbose=False):
|
75 |
+
total_params = sum(p.numel() for p in model.parameters())
|
76 |
+
if verbose:
|
77 |
+
print(f"{model.__class__.__name__} has {total_params * 1.e-6:.2f} M params.")
|
78 |
+
return total_params
|
79 |
+
|
80 |
+
|
81 |
+
def instantiate_from_config(config):
|
82 |
+
if not "target" in config:
|
83 |
+
if config == "__is_first_stage__":
|
84 |
+
return None
|
85 |
+
elif config == "__is_unconditional__":
|
86 |
+
return None
|
87 |
+
raise KeyError("Expected key `target` to instantiate.")
|
88 |
+
return get_obj_from_str(config["target"])(**config.get("params", dict()))
|
89 |
+
|
90 |
+
|
91 |
+
def get_obj_from_str(string, reload=False):
|
92 |
+
module, cls = string.rsplit(".", 1)
|
93 |
+
if reload:
|
94 |
+
module_imp = importlib.import_module(module)
|
95 |
+
importlib.reload(module_imp)
|
96 |
+
return getattr(importlib.import_module(module, package=None), cls)
|
97 |
+
|
98 |
+
|
99 |
+
def _do_parallel_data_prefetch(func, Q, data, idx, idx_to_fn=False):
|
100 |
+
# create dummy dataset instance
|
101 |
+
|
102 |
+
# run prefetching
|
103 |
+
if idx_to_fn:
|
104 |
+
res = func(data, worker_id=idx)
|
105 |
+
else:
|
106 |
+
res = func(data)
|
107 |
+
Q.put([idx, res])
|
108 |
+
Q.put("Done")
|
109 |
+
|
110 |
+
|
111 |
+
def parallel_data_prefetch(
|
112 |
+
func: callable,
|
113 |
+
data,
|
114 |
+
n_proc,
|
115 |
+
target_data_type="ndarray",
|
116 |
+
cpu_intensive=True,
|
117 |
+
use_worker_id=False,
|
118 |
+
):
|
119 |
+
# if target_data_type not in ["ndarray", "list"]:
|
120 |
+
# raise ValueError(
|
121 |
+
# "Data, which is passed to parallel_data_prefetch has to be either of type list or ndarray."
|
122 |
+
# )
|
123 |
+
if isinstance(data, np.ndarray) and target_data_type == "list":
|
124 |
+
raise ValueError("list expected but function got ndarray.")
|
125 |
+
elif isinstance(data, abc.Iterable):
|
126 |
+
if isinstance(data, dict):
|
127 |
+
print(
|
128 |
+
f'WARNING:"data" argument passed to parallel_data_prefetch is a dict: Using only its values and disregarding keys.'
|
129 |
+
)
|
130 |
+
data = list(data.values())
|
131 |
+
if target_data_type == "ndarray":
|
132 |
+
data = np.asarray(data)
|
133 |
+
else:
|
134 |
+
data = list(data)
|
135 |
+
else:
|
136 |
+
raise TypeError(
|
137 |
+
f"The data, that shall be processed parallel has to be either an np.ndarray or an Iterable, but is actually {type(data)}."
|
138 |
+
)
|
139 |
+
|
140 |
+
if cpu_intensive:
|
141 |
+
Q = mp.Queue(1000)
|
142 |
+
proc = mp.Process
|
143 |
+
else:
|
144 |
+
Q = Queue(1000)
|
145 |
+
proc = Thread
|
146 |
+
# spawn processes
|
147 |
+
if target_data_type == "ndarray":
|
148 |
+
arguments = [
|
149 |
+
[func, Q, part, i, use_worker_id]
|
150 |
+
for i, part in enumerate(np.array_split(data, n_proc))
|
151 |
+
]
|
152 |
+
else:
|
153 |
+
step = (
|
154 |
+
int(len(data) / n_proc + 1)
|
155 |
+
if len(data) % n_proc != 0
|
156 |
+
else int(len(data) / n_proc)
|
157 |
+
)
|
158 |
+
arguments = [
|
159 |
+
[func, Q, part, i, use_worker_id]
|
160 |
+
for i, part in enumerate(
|
161 |
+
[data[i : i + step] for i in range(0, len(data), step)]
|
162 |
+
)
|
163 |
+
]
|
164 |
+
processes = []
|
165 |
+
for i in range(n_proc):
|
166 |
+
p = proc(target=_do_parallel_data_prefetch, args=arguments[i])
|
167 |
+
processes += [p]
|
168 |
+
|
169 |
+
# start processes
|
170 |
+
print(f"Start prefetching...")
|
171 |
+
import time
|
172 |
+
|
173 |
+
start = time.time()
|
174 |
+
gather_res = [[] for _ in range(n_proc)]
|
175 |
+
try:
|
176 |
+
for p in processes:
|
177 |
+
p.start()
|
178 |
+
|
179 |
+
k = 0
|
180 |
+
while k < n_proc:
|
181 |
+
# get result
|
182 |
+
res = Q.get()
|
183 |
+
if res == "Done":
|
184 |
+
k += 1
|
185 |
+
else:
|
186 |
+
gather_res[res[0]] = res[1]
|
187 |
+
|
188 |
+
except Exception as e:
|
189 |
+
print("Exception: ", e)
|
190 |
+
for p in processes:
|
191 |
+
p.terminate()
|
192 |
+
|
193 |
+
raise e
|
194 |
+
finally:
|
195 |
+
for p in processes:
|
196 |
+
p.join()
|
197 |
+
print(f"Prefetching complete. [{time.time() - start} sec.]")
|
198 |
+
|
199 |
+
if target_data_type == "ndarray":
|
200 |
+
if not isinstance(gather_res[0], np.ndarray):
|
201 |
+
return np.concatenate([np.asarray(r) for r in gather_res], axis=0)
|
202 |
+
|
203 |
+
# order outputs
|
204 |
+
return np.concatenate(gather_res, axis=0)
|
205 |
+
elif target_data_type == "list":
|
206 |
+
out = []
|
207 |
+
for r in gather_res:
|
208 |
+
out.extend(r)
|
209 |
+
return out
|
210 |
+
else:
|
211 |
+
return gather_res
|
212 |
+
|
213 |
+
|
214 |
+
class ActNorm(nn.Module):
|
215 |
+
def __init__(
|
216 |
+
self, num_features, logdet=False, affine=True, allow_reverse_init=False
|
217 |
+
):
|
218 |
+
assert affine
|
219 |
+
super().__init__()
|
220 |
+
self.logdet = logdet
|
221 |
+
self.loc = nn.Parameter(torch.zeros(1, num_features, 1, 1))
|
222 |
+
self.scale = nn.Parameter(torch.ones(1, num_features, 1, 1))
|
223 |
+
self.allow_reverse_init = allow_reverse_init
|
224 |
+
|
225 |
+
self.register_buffer("initialized", torch.tensor(0, dtype=torch.uint8))
|
226 |
+
|
227 |
+
def initialize(self, input):
|
228 |
+
with torch.no_grad():
|
229 |
+
flatten = input.permute(1, 0, 2, 3).contiguous().view(input.shape[1], -1)
|
230 |
+
mean = (
|
231 |
+
flatten.mean(1)
|
232 |
+
.unsqueeze(1)
|
233 |
+
.unsqueeze(2)
|
234 |
+
.unsqueeze(3)
|
235 |
+
.permute(1, 0, 2, 3)
|
236 |
+
)
|
237 |
+
std = (
|
238 |
+
flatten.std(1)
|
239 |
+
.unsqueeze(1)
|
240 |
+
.unsqueeze(2)
|
241 |
+
.unsqueeze(3)
|
242 |
+
.permute(1, 0, 2, 3)
|
243 |
+
)
|
244 |
+
|
245 |
+
self.loc.data.copy_(-mean)
|
246 |
+
self.scale.data.copy_(1 / (std + 1e-6))
|
247 |
+
|
248 |
+
def forward(self, input, reverse=False):
|
249 |
+
if reverse:
|
250 |
+
return self.reverse(input)
|
251 |
+
if len(input.shape) == 2:
|
252 |
+
input = input[:, :, None, None]
|
253 |
+
squeeze = True
|
254 |
+
else:
|
255 |
+
squeeze = False
|
256 |
+
|
257 |
+
_, _, height, width = input.shape
|
258 |
+
|
259 |
+
if self.training and self.initialized.item() == 0:
|
260 |
+
self.initialize(input)
|
261 |
+
self.initialized.fill_(1)
|
262 |
+
|
263 |
+
h = self.scale * (input + self.loc)
|
264 |
+
|
265 |
+
if squeeze:
|
266 |
+
h = h.squeeze(-1).squeeze(-1)
|
267 |
+
|
268 |
+
if self.logdet:
|
269 |
+
log_abs = torch.log(torch.abs(self.scale))
|
270 |
+
logdet = height * width * torch.sum(log_abs)
|
271 |
+
logdet = logdet * torch.ones(input.shape[0]).to(input)
|
272 |
+
return h, logdet
|
273 |
+
|
274 |
+
return h
|
275 |
+
|
276 |
+
def reverse(self, output):
|
277 |
+
if self.training and self.initialized.item() == 0:
|
278 |
+
if not self.allow_reverse_init:
|
279 |
+
raise RuntimeError(
|
280 |
+
"Initializing ActNorm in reverse direction is "
|
281 |
+
"disabled by default. Use allow_reverse_init=True to enable."
|
282 |
+
)
|
283 |
+
else:
|
284 |
+
self.initialize(output)
|
285 |
+
self.initialized.fill_(1)
|
286 |
+
|
287 |
+
if len(output.shape) == 2:
|
288 |
+
output = output[:, :, None, None]
|
289 |
+
squeeze = True
|
290 |
+
else:
|
291 |
+
squeeze = False
|
292 |
+
|
293 |
+
h = output / self.scale - self.loc
|
294 |
+
|
295 |
+
if squeeze:
|
296 |
+
h = h.squeeze(-1).squeeze(-1)
|
297 |
+
return h
|
train.py
DELETED
@@ -1,72 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
from torchinfo import summary
|
3 |
-
|
4 |
-
from swim.autoencoder import Autoencoder
|
5 |
-
from diffusers import AutoencoderKL, UNet2DModel
|
6 |
-
|
7 |
-
# vae = Autoencoder(
|
8 |
-
# z_channels=4,
|
9 |
-
# in_channels=3,
|
10 |
-
# channels=128,
|
11 |
-
# channel_multipliers=[1, 2, 4, 4],
|
12 |
-
# n_resnet_blocks=2,
|
13 |
-
# emb_channels=4,
|
14 |
-
# ).to("meta")
|
15 |
-
# lol_vae = AutoencoderKL.from_pretrained(
|
16 |
-
# "stabilityai/stable-diffusion-2-1", subfolder="vae"
|
17 |
-
# ).to("meta")
|
18 |
-
|
19 |
-
# # copy weights from lol_vae to vae
|
20 |
-
# import json
|
21 |
-
|
22 |
-
# with open("lolvae.json", "w") as f:
|
23 |
-
# json.dump(list(lol_vae.state_dict().keys()), f, indent=4)
|
24 |
-
|
25 |
-
# with open("vae.json", "w") as f:
|
26 |
-
# json.dump(list(vae.state_dict().keys()), f, indent=4)
|
27 |
-
|
28 |
-
# sample = torch.randn(1, 3, 512, 512).to("meta")
|
29 |
-
# # lantent = vae.encoder(sample)
|
30 |
-
|
31 |
-
from diffusers import UNet2DModel
|
32 |
-
|
33 |
-
model = UNet2DModel(
|
34 |
-
sample_size=512, # the target image resolution
|
35 |
-
in_channels=3, # the number of input channels, 3 for RGB images
|
36 |
-
out_channels=3, # the number of output channels
|
37 |
-
layers_per_block=2, # how many ResNet layers to use per UNet block
|
38 |
-
block_out_channels=(
|
39 |
-
128,
|
40 |
-
128,
|
41 |
-
256,
|
42 |
-
256,
|
43 |
-
512,
|
44 |
-
512,
|
45 |
-
), # the number of output channels for each UNet block
|
46 |
-
down_block_types=(
|
47 |
-
"DownBlock2D", # a regular ResNet downsampling block
|
48 |
-
"DownBlock2D",
|
49 |
-
"DownBlock2D",
|
50 |
-
"DownBlock2D",
|
51 |
-
"AttnDownBlock2D", # a ResNet downsampling block with spatial self-attention
|
52 |
-
"DownBlock2D",
|
53 |
-
),
|
54 |
-
up_block_types=(
|
55 |
-
"UpBlock2D", # a regular ResNet upsampling block
|
56 |
-
"AttnUpBlock2D", # a ResNet upsampling block with spatial self-attention
|
57 |
-
"UpBlock2D",
|
58 |
-
"UpBlock2D",
|
59 |
-
"UpBlock2D",
|
60 |
-
"UpBlock2D",
|
61 |
-
),
|
62 |
-
).to("meta")
|
63 |
-
|
64 |
-
sample = torch.randn(1, 3, 512, 512).to("meta")
|
65 |
-
|
66 |
-
summary(
|
67 |
-
model,
|
68 |
-
input_data=(
|
69 |
-
sample,
|
70 |
-
0,
|
71 |
-
),
|
72 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|