Spaces:
Running
Running
Eduard-Sebastian Zamfir
commited on
Commit
•
9080570
1
Parent(s):
022d36d
add gradio app
Browse files- .gitignore +2 -0
- README.md +6 -6
- app.py +153 -0
- assets/arch.svg +0 -0
- configs/eval_seemore_t_x4.yml +14 -0
- images/img002x4.png +0 -0
- images/img003x4.png +0 -0
- images/img004x4.png +0 -0
- images/img035x4.png +0 -0
- images/img053x4.png +0 -0
- images/img064x4.png +0 -0
- images/img083x4.png +0 -0
- images/img092x4.png +0 -0
- models/seemore.py +416 -0
- requirements.txt +6 -0
.gitignore
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
__pycache__/
|
2 |
+
flagged/
|
README.md
CHANGED
@@ -1,13 +1,13 @@
|
|
1 |
---
|
2 |
-
title:
|
3 |
-
emoji:
|
4 |
-
colorFrom:
|
5 |
-
colorTo:
|
6 |
sdk: gradio
|
7 |
-
sdk_version: 4.
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
-
license:
|
11 |
---
|
12 |
|
13 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
1 |
---
|
2 |
+
title: SeemoRe
|
3 |
+
emoji: 💻
|
4 |
+
colorFrom: purple
|
5 |
+
colorTo: blue
|
6 |
sdk: gradio
|
7 |
+
sdk_version: 4.16.0
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
+
license: mit
|
11 |
---
|
12 |
|
13 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.py
ADDED
@@ -0,0 +1,153 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import yaml
|
3 |
+
import torch
|
4 |
+
import argparse
|
5 |
+
import numpy as np
|
6 |
+
import gradio as gr
|
7 |
+
|
8 |
+
from PIL import Image
|
9 |
+
from copy import deepcopy
|
10 |
+
from torch.nn.parallel import DataParallel, DistributedDataParallel
|
11 |
+
|
12 |
+
from huggingface_hub import hf_hub_download
|
13 |
+
from gradio_imageslider import ImageSlider
|
14 |
+
|
15 |
+
## local code
|
16 |
+
from models import seemore
|
17 |
+
|
18 |
+
|
19 |
+
def dict2namespace(config):
|
20 |
+
namespace = argparse.Namespace()
|
21 |
+
for key, value in config.items():
|
22 |
+
if isinstance(value, dict):
|
23 |
+
new_value = dict2namespace(value)
|
24 |
+
else:
|
25 |
+
new_value = value
|
26 |
+
setattr(namespace, key, new_value)
|
27 |
+
return namespace
|
28 |
+
|
29 |
+
def load_img (filename, norm=True,):
|
30 |
+
img = np.array(Image.open(filename).convert("RGB"))
|
31 |
+
if norm:
|
32 |
+
img = img / 255.
|
33 |
+
img = img.astype(np.float32)
|
34 |
+
return img
|
35 |
+
|
36 |
+
def process_img (image):
|
37 |
+
img = np.array(image)
|
38 |
+
img = img / 255.
|
39 |
+
img = img.astype(np.float32)
|
40 |
+
y = torch.tensor(img).permute(2,0,1).unsqueeze(0).to(device)
|
41 |
+
|
42 |
+
with torch.no_grad():
|
43 |
+
x_hat = model(y)
|
44 |
+
|
45 |
+
restored_img = x_hat.squeeze().permute(1,2,0).clamp_(0, 1).cpu().detach().numpy()
|
46 |
+
restored_img = np.clip(restored_img, 0. , 1.)
|
47 |
+
|
48 |
+
restored_img = (restored_img * 255.0).round().astype(np.uint8) # float32 to uint8
|
49 |
+
#return Image.fromarray(restored_img) #
|
50 |
+
return (image, Image.fromarray(restored_img))
|
51 |
+
|
52 |
+
def load_network(net, load_path, strict=True, param_key='params'):
|
53 |
+
if isinstance(net, (DataParallel, DistributedDataParallel)):
|
54 |
+
net = net.module
|
55 |
+
load_net = torch.load(load_path, map_location=lambda storage, loc: storage)
|
56 |
+
if param_key is not None:
|
57 |
+
if param_key not in load_net and 'params' in load_net:
|
58 |
+
param_key = 'params'
|
59 |
+
load_net = load_net[param_key]
|
60 |
+
# remove unnecessary 'module.'
|
61 |
+
for k, v in deepcopy(load_net).items():
|
62 |
+
if k.startswith('module.'):
|
63 |
+
load_net[k[7:]] = v
|
64 |
+
load_net.pop(k)
|
65 |
+
net.load_state_dict(load_net, strict=strict)
|
66 |
+
|
67 |
+
CONFIG = "configs/eval_seemore_t_x4.yml"
|
68 |
+
MODEL_NAME = "checkpoints/SeemoRe_T/X4/net_g_latest.pth"
|
69 |
+
|
70 |
+
# parse config file
|
71 |
+
with open(os.path.join(CONFIG), "r") as f:
|
72 |
+
config = yaml.safe_load(f)
|
73 |
+
|
74 |
+
cfg = dict2namespace(config)
|
75 |
+
|
76 |
+
device = torch.device("cpu")
|
77 |
+
model = seemore.SeemoRe(scale=cfg.model.scale, in_chans=cfg.model.in_chans,
|
78 |
+
num_experts=cfg.model.num_experts, num_layers=cfg.model.num_layers, embedding_dim=cfg.model.embedding_dim,
|
79 |
+
img_range=cfg.model.img_range, use_shuffle=cfg.model.use_shuffle, global_kernel_size=cfg.model.global_kernel_size,
|
80 |
+
recursive=cfg.model.recursive, lr_space=cfg.model.lr_space, topk=cfg.model.topk)
|
81 |
+
|
82 |
+
model = model.to(device)
|
83 |
+
print ("IMAGE MODEL CKPT:", MODEL_NAME)
|
84 |
+
load_network(model, MODEL_NAME, strict=True, param_key='params')
|
85 |
+
|
86 |
+
|
87 |
+
|
88 |
+
|
89 |
+
title = "See More Details"
|
90 |
+
description = ''' ### See More Details: Efficient Image Super-Resolution by Experts Mining
|
91 |
+
|
92 |
+
#### [Eduard Zamfir<sup>1</sup>](https://eduardzamfir.github.io), [Zongwei Wu<sup>1*</sup>](https://sites.google.com/view/zwwu/accueil), [Nancy Mehta<sup>1</sup>](https://scholar.google.com/citations?user=WwdYdlUAAAAJ&hl=en&oi=ao), [Yulun Zhang<sup>2,3*</sup>](http://yulunzhang.com/) and [Radu Timofte<sup>1</sup>](https://www.informatik.uni-wuerzburg.de/computervision/)
|
93 |
+
|
94 |
+
#### **<sup>1</sup> University of Würzburg, Germany - <sup>2</sup> Shanghai Jiao Tong University, China - <sup>3</sup> ETH Zürich, Switzerland**
|
95 |
+
#### **<sup>*</sup> Corresponding authors**
|
96 |
+
|
97 |
+
<details>
|
98 |
+
<summary> <b> Abstract</b> (click me to read)</summary>
|
99 |
+
<p>
|
100 |
+
Reconstructing high-resolution (HR) images from low-resolution (LR) inputs poses a significant challenge in image super-resolution (SR). While recent approaches have demonstrated the efficacy of intricate operations customized for various objectives, the straightforward stacking of these disparate operations can result in a substantial computational burden, hampering their practical utility. In response, we introduce **S**eemo**R**e, an efficient SR model employing expert mining. Our approach strategically incorporates experts at different levels, adopting a collaborative methodology. At the macro scale, our experts address rank-wise and spatial-wise informative features, providing a holistic understanding. Subsequently, the model delves into the subtleties of rank choice by leveraging a mixture of low-rank experts. By tapping into experts specialized in distinct key factors crucial for accurate SR, our model excels in uncovering intricate intra-feature details. This collaborative approach is reminiscent of the concept of **see more**, allowing our model to achieve an optimal performance with minimal computational costs in efficient settings
|
101 |
+
</p>
|
102 |
+
</details>
|
103 |
+
|
104 |
+
|
105 |
+
<br>
|
106 |
+
|
107 |
+
<code>
|
108 |
+
@inproceedings{zamfir2024details,
|
109 |
+
title={See More Details: Efficient Image Super-Resolution by Experts Mining},
|
110 |
+
author={Eduard Zamfir and Zongwei Wu and Nancy Mehta and Yulun Zhang and Radu Timofte},
|
111 |
+
booktitle={International Conference on Machine Learning},
|
112 |
+
year={2024},
|
113 |
+
organization={PMLR}
|
114 |
+
}
|
115 |
+
</code>
|
116 |
+
<br>
|
117 |
+
'''
|
118 |
+
|
119 |
+
|
120 |
+
article = "<p style='text-align: center'><a href='https://eduardzamfir.github.io/seemore' target='_blank'>See More Details: Efficient Image Super-Resolution by Experts Mining</a></p>"
|
121 |
+
|
122 |
+
#### Image,Prompts examples
|
123 |
+
examples = [['images/img002x4.png'],
|
124 |
+
['images/img003x4.png'],
|
125 |
+
['images/img004x4.png'],
|
126 |
+
['images/img035x4.png'],
|
127 |
+
['images/img053x4.png'],
|
128 |
+
['images/img064x4.png'],
|
129 |
+
['images/img083x4.png'],
|
130 |
+
['images/img092x4.png'],
|
131 |
+
]
|
132 |
+
|
133 |
+
css = """
|
134 |
+
.image-frame img, .image-container img {
|
135 |
+
width: auto;
|
136 |
+
height: auto;
|
137 |
+
max-width: none;
|
138 |
+
}
|
139 |
+
"""
|
140 |
+
|
141 |
+
demo = gr.Interface(
|
142 |
+
fn=process_img,
|
143 |
+
inputs=[gr.Image(type="pil", label="Input", value="images/img002x4.png"),],
|
144 |
+
outputs=ImageSlider(label="Super-Resolved Image", type="pil"), #[gr.Image(type="pil", label="Ouput", min_width=500)],
|
145 |
+
title=title,
|
146 |
+
description=description,
|
147 |
+
article=article,
|
148 |
+
examples=examples,
|
149 |
+
css=css,
|
150 |
+
)
|
151 |
+
|
152 |
+
if __name__ == "__main__":
|
153 |
+
demo.launch()
|
assets/arch.svg
ADDED
configs/eval_seemore_t_x4.yml
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
arch: "SeemoRe"
|
3 |
+
scale: 4
|
4 |
+
in_chans: 3
|
5 |
+
num_experts: 3
|
6 |
+
img_range: 1.0
|
7 |
+
num_layers: 6
|
8 |
+
embedding_dim: 36
|
9 |
+
use_shuffle: True
|
10 |
+
lr_space: exp
|
11 |
+
topk: 1
|
12 |
+
recursive: 2
|
13 |
+
global_kernel_size: 11
|
14 |
+
|
images/img002x4.png
ADDED
images/img003x4.png
ADDED
images/img004x4.png
ADDED
images/img035x4.png
ADDED
images/img053x4.png
ADDED
images/img064x4.png
ADDED
images/img083x4.png
ADDED
images/img092x4.png
ADDED
models/seemore.py
ADDED
@@ -0,0 +1,416 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple, List
|
2 |
+
from torch import Tensor
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
import torch.nn.functional as F
|
7 |
+
from einops.layers.torch import Rearrange
|
8 |
+
|
9 |
+
|
10 |
+
######################
|
11 |
+
# Meta Architecture
|
12 |
+
######################
|
13 |
+
class SeemoRe(nn.Module):
|
14 |
+
def __init__(self,
|
15 |
+
scale: int = 4,
|
16 |
+
in_chans: int = 3,
|
17 |
+
num_experts: int = 6,
|
18 |
+
num_layers: int = 6,
|
19 |
+
embedding_dim: int = 64,
|
20 |
+
img_range: float = 1.0,
|
21 |
+
use_shuffle: bool = False,
|
22 |
+
global_kernel_size: int = 11,
|
23 |
+
recursive: int = 2,
|
24 |
+
lr_space: int = 1,
|
25 |
+
topk: int = 2,):
|
26 |
+
super().__init__()
|
27 |
+
self.scale = scale
|
28 |
+
self.num_in_channels = in_chans
|
29 |
+
self.num_out_channels = in_chans
|
30 |
+
self.img_range = img_range
|
31 |
+
|
32 |
+
rgb_mean = (0.4488, 0.4371, 0.4040)
|
33 |
+
self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
|
34 |
+
|
35 |
+
|
36 |
+
# -- SHALLOW FEATURES --
|
37 |
+
self.conv_1 = nn.Conv2d(self.num_in_channels, embedding_dim, kernel_size=3, padding=1)
|
38 |
+
|
39 |
+
# -- DEEP FEATURES --
|
40 |
+
self.body = nn.ModuleList(
|
41 |
+
[ResGroup(in_ch=embedding_dim,
|
42 |
+
num_experts=num_experts,
|
43 |
+
use_shuffle=use_shuffle,
|
44 |
+
topk=topk,
|
45 |
+
lr_space=lr_space,
|
46 |
+
recursive=recursive,
|
47 |
+
global_kernel_size=global_kernel_size) for i in range(num_layers)]
|
48 |
+
)
|
49 |
+
|
50 |
+
# -- UPSCALE --
|
51 |
+
self.norm = LayerNorm(embedding_dim, data_format='channels_first')
|
52 |
+
self.conv_2 = nn.Conv2d(embedding_dim, embedding_dim, kernel_size=3, padding=1)
|
53 |
+
self.upsampler = nn.Sequential(
|
54 |
+
nn.Conv2d(embedding_dim, (scale**2) * self.num_out_channels, kernel_size=3, padding=1),
|
55 |
+
nn.PixelShuffle(scale)
|
56 |
+
)
|
57 |
+
|
58 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
59 |
+
self.mean = self.mean.type_as(x)
|
60 |
+
x = (x - self.mean) * self.img_range
|
61 |
+
|
62 |
+
# -- SHALLOW FEATURES --
|
63 |
+
x = self.conv_1(x)
|
64 |
+
res = x
|
65 |
+
|
66 |
+
# -- DEEP FEATURES --
|
67 |
+
for idx, layer in enumerate(self.body):
|
68 |
+
x = layer(x)
|
69 |
+
|
70 |
+
x = self.norm(x)
|
71 |
+
|
72 |
+
# -- HR IMAGE RECONSTRUCTION --
|
73 |
+
x = self.conv_2(x) + res
|
74 |
+
x = self.upsampler(x)
|
75 |
+
|
76 |
+
x = x / self.img_range + self.mean
|
77 |
+
return x
|
78 |
+
|
79 |
+
|
80 |
+
|
81 |
+
#############################
|
82 |
+
# Components
|
83 |
+
#############################
|
84 |
+
class ResGroup(nn.Module):
|
85 |
+
def __init__(self,
|
86 |
+
in_ch: int,
|
87 |
+
num_experts: int,
|
88 |
+
global_kernel_size: int = 11,
|
89 |
+
lr_space: int = 1,
|
90 |
+
topk: int = 2,
|
91 |
+
recursive: int = 2,
|
92 |
+
use_shuffle: bool = False):
|
93 |
+
super().__init__()
|
94 |
+
|
95 |
+
self.local_block = RME(in_ch=in_ch,
|
96 |
+
num_experts=num_experts,
|
97 |
+
use_shuffle=use_shuffle,
|
98 |
+
lr_space=lr_space,
|
99 |
+
topk=topk,
|
100 |
+
recursive=recursive)
|
101 |
+
self.global_block = SME(in_ch=in_ch,
|
102 |
+
kernel_size=global_kernel_size)
|
103 |
+
|
104 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
105 |
+
x = self.local_block(x)
|
106 |
+
x = self.global_block(x)
|
107 |
+
return x
|
108 |
+
|
109 |
+
|
110 |
+
|
111 |
+
#############################
|
112 |
+
# Global Block
|
113 |
+
#############################
|
114 |
+
class SME(nn.Module):
|
115 |
+
def __init__(self,
|
116 |
+
in_ch: int,
|
117 |
+
kernel_size: int = 11):
|
118 |
+
super().__init__()
|
119 |
+
|
120 |
+
self.norm_1 = LayerNorm(in_ch, data_format='channels_first')
|
121 |
+
self.block = StripedConvFormer(in_ch=in_ch, kernel_size=kernel_size)
|
122 |
+
|
123 |
+
self.norm_2 = LayerNorm(in_ch, data_format='channels_first')
|
124 |
+
self.ffn = GatedFFN(in_ch, mlp_ratio=2, kernel_size=3, act_layer=nn.GELU())
|
125 |
+
|
126 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
127 |
+
x = self.block(self.norm_1(x)) + x
|
128 |
+
x = self.ffn(self.norm_2(x)) + x
|
129 |
+
return x
|
130 |
+
|
131 |
+
|
132 |
+
|
133 |
+
|
134 |
+
class StripedConvFormer(nn.Module):
|
135 |
+
def __init__(self,
|
136 |
+
in_ch: int,
|
137 |
+
kernel_size: int):
|
138 |
+
super().__init__()
|
139 |
+
self.in_ch = in_ch
|
140 |
+
self.kernel_size = kernel_size
|
141 |
+
self.padding = kernel_size // 2
|
142 |
+
|
143 |
+
self.proj = nn.Conv2d(in_ch, in_ch, kernel_size=1, padding=0)
|
144 |
+
self.to_qv = nn.Sequential(
|
145 |
+
nn.Conv2d(in_ch, in_ch * 2, kernel_size=1, padding=0),
|
146 |
+
nn.GELU(),
|
147 |
+
)
|
148 |
+
|
149 |
+
self.attn = StripedConv2d(in_ch, kernel_size=kernel_size, depthwise=True)
|
150 |
+
|
151 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
152 |
+
q, v = self.to_qv(x).chunk(2, dim=1)
|
153 |
+
q = self.attn(q)
|
154 |
+
x = self.proj(q * v)
|
155 |
+
return x
|
156 |
+
|
157 |
+
|
158 |
+
|
159 |
+
#############################
|
160 |
+
# Local Blocks
|
161 |
+
#############################
|
162 |
+
class RME(nn.Module):
|
163 |
+
def __init__(self,
|
164 |
+
in_ch: int,
|
165 |
+
num_experts: int,
|
166 |
+
topk: int,
|
167 |
+
lr_space: int = 1,
|
168 |
+
recursive: int = 2,
|
169 |
+
use_shuffle: bool = False,):
|
170 |
+
super().__init__()
|
171 |
+
|
172 |
+
self.norm_1 = LayerNorm(in_ch, data_format='channels_first')
|
173 |
+
self.block = MoEBlock(in_ch=in_ch, num_experts=num_experts, topk=topk, use_shuffle=use_shuffle, recursive=recursive, lr_space=lr_space,)
|
174 |
+
|
175 |
+
self.norm_2 = LayerNorm(in_ch, data_format='channels_first')
|
176 |
+
self.ffn = GatedFFN(in_ch, mlp_ratio=2, kernel_size=3, act_layer=nn.GELU())
|
177 |
+
|
178 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
179 |
+
x = self.block(self.norm_1(x)) + x
|
180 |
+
x = self.ffn(self.norm_2(x)) + x
|
181 |
+
return x
|
182 |
+
|
183 |
+
|
184 |
+
|
185 |
+
#################
|
186 |
+
# MoE Layer
|
187 |
+
#################
|
188 |
+
class MoEBlock(nn.Module):
|
189 |
+
def __init__(self,
|
190 |
+
in_ch: int,
|
191 |
+
num_experts: int,
|
192 |
+
topk: int,
|
193 |
+
use_shuffle: bool = False,
|
194 |
+
lr_space: str = "linear",
|
195 |
+
recursive: int = 2):
|
196 |
+
super().__init__()
|
197 |
+
self.use_shuffle = use_shuffle
|
198 |
+
self.recursive = recursive
|
199 |
+
|
200 |
+
self.conv_1 = nn.Sequential(
|
201 |
+
nn.Conv2d(in_ch, in_ch, kernel_size=3, padding=1),
|
202 |
+
nn.GELU(),
|
203 |
+
nn.Conv2d(in_ch, 2*in_ch, kernel_size=1, padding=0)
|
204 |
+
)
|
205 |
+
|
206 |
+
self.agg_conv = nn.Sequential(
|
207 |
+
nn.Conv2d(in_ch, in_ch, kernel_size=4, stride=4, groups=in_ch),
|
208 |
+
nn.GELU())
|
209 |
+
|
210 |
+
self.conv = nn.Sequential(
|
211 |
+
nn.Conv2d(in_ch, in_ch, kernel_size=3, stride=1, padding=1, groups=in_ch),
|
212 |
+
nn.Conv2d(in_ch, in_ch, kernel_size=1, padding=0)
|
213 |
+
)
|
214 |
+
|
215 |
+
self.conv_2 = nn.Sequential(
|
216 |
+
StripedConv2d(in_ch, kernel_size=3, depthwise=True),
|
217 |
+
nn.GELU())
|
218 |
+
|
219 |
+
if lr_space == "linear":
|
220 |
+
grow_func = lambda i: i+2
|
221 |
+
elif lr_space == "exp":
|
222 |
+
grow_func = lambda i: 2**(i+1)
|
223 |
+
elif lr_space == "double":
|
224 |
+
grow_func = lambda i: 2*i+2
|
225 |
+
else:
|
226 |
+
raise NotImplementedError(f"lr_space {lr_space} not implemented")
|
227 |
+
|
228 |
+
self.moe_layer = MoELayer(
|
229 |
+
experts=[Expert(in_ch=in_ch, low_dim=grow_func(i)) for i in range(num_experts)], # add here multiple of 2 as low_dim
|
230 |
+
gate=Router(in_ch=in_ch, num_experts=num_experts),
|
231 |
+
num_expert=topk,
|
232 |
+
)
|
233 |
+
|
234 |
+
self.proj = nn.Conv2d(in_ch, in_ch, kernel_size=1, padding=0)
|
235 |
+
|
236 |
+
def calibrate(self, x: torch.Tensor) -> torch.Tensor:
|
237 |
+
b, c, h, w = x.shape
|
238 |
+
res = x
|
239 |
+
|
240 |
+
for _ in range(self.recursive):
|
241 |
+
x = self.agg_conv(x)
|
242 |
+
x = self.conv(x)
|
243 |
+
x = F.interpolate(x, size=(h, w), mode="bilinear", align_corners=False)
|
244 |
+
return res + x
|
245 |
+
|
246 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
247 |
+
x = self.conv_1(x)
|
248 |
+
|
249 |
+
if self.use_shuffle:
|
250 |
+
x = channel_shuffle(x, groups=2)
|
251 |
+
x, k = torch.chunk(x, chunks=2, dim=1)
|
252 |
+
|
253 |
+
x = self.conv_2(x)
|
254 |
+
k = self.calibrate(k)
|
255 |
+
|
256 |
+
x = self.moe_layer(x, k)
|
257 |
+
x = self.proj(x)
|
258 |
+
return x
|
259 |
+
|
260 |
+
|
261 |
+
class MoELayer(nn.Module):
|
262 |
+
def __init__(self, experts: List[nn.Module], gate: nn.Module, num_expert: int = 1):
|
263 |
+
super().__init__()
|
264 |
+
assert len(experts) > 0
|
265 |
+
self.experts = nn.ModuleList(experts)
|
266 |
+
self.gate = gate
|
267 |
+
self.num_expert = num_expert
|
268 |
+
|
269 |
+
def forward(self, inputs: torch.Tensor, k: torch.Tensor):
|
270 |
+
out = self.gate(inputs)
|
271 |
+
weights = F.softmax(out, dim=1, dtype=torch.float).to(inputs.dtype)
|
272 |
+
topk_weights, topk_experts = torch.topk(weights, self.num_expert)
|
273 |
+
out = inputs.clone()
|
274 |
+
|
275 |
+
if self.training:
|
276 |
+
exp_weights = torch.zeros_like(weights)
|
277 |
+
exp_weights.scatter_(1, topk_experts, weights.gather(1, topk_experts))
|
278 |
+
for i, expert in enumerate(self.experts):
|
279 |
+
out += expert(inputs, k) * exp_weights[:, i:i+1, None, None]
|
280 |
+
else:
|
281 |
+
selected_experts = [self.experts[i] for i in topk_experts.squeeze(dim=0)]
|
282 |
+
for i, expert in enumerate(selected_experts):
|
283 |
+
out += expert(inputs, k) * topk_weights[:, i:i+1, None, None]
|
284 |
+
|
285 |
+
return out
|
286 |
+
|
287 |
+
|
288 |
+
|
289 |
+
class Expert(nn.Module):
|
290 |
+
def __init__(self,
|
291 |
+
in_ch: int,
|
292 |
+
low_dim: int,):
|
293 |
+
super().__init__()
|
294 |
+
self.conv_1 = nn.Conv2d(in_ch, low_dim, kernel_size=1, padding=0)
|
295 |
+
self.conv_2 = nn.Conv2d(in_ch, low_dim, kernel_size=1, padding=0)
|
296 |
+
self.conv_3 = nn.Conv2d(low_dim, in_ch, kernel_size=1, padding=0)
|
297 |
+
|
298 |
+
def forward(self, x: torch.Tensor, k: torch.Tensor) -> torch.Tensor:
|
299 |
+
x = self.conv_1(x)
|
300 |
+
x = self.conv_2(k) * x # here no more sigmoid
|
301 |
+
x = self.conv_3(x)
|
302 |
+
return x
|
303 |
+
|
304 |
+
|
305 |
+
class Router(nn.Module):
|
306 |
+
def __init__(self,
|
307 |
+
in_ch: int,
|
308 |
+
num_experts: int):
|
309 |
+
super().__init__()
|
310 |
+
|
311 |
+
self.body = nn.Sequential(
|
312 |
+
nn.AdaptiveAvgPool2d(1),
|
313 |
+
Rearrange('b c 1 1 -> b c'),
|
314 |
+
nn.Linear(in_ch, num_experts, bias=False),
|
315 |
+
)
|
316 |
+
|
317 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
318 |
+
return self.body(x)
|
319 |
+
|
320 |
+
|
321 |
+
|
322 |
+
#################
|
323 |
+
# Utilities
|
324 |
+
#################
|
325 |
+
class StripedConv2d(nn.Module):
|
326 |
+
def __init__(self,
|
327 |
+
in_ch: int,
|
328 |
+
kernel_size: int,
|
329 |
+
depthwise: bool = False):
|
330 |
+
super().__init__()
|
331 |
+
self.in_ch = in_ch
|
332 |
+
self.kernel_size = kernel_size
|
333 |
+
self.padding = kernel_size // 2
|
334 |
+
|
335 |
+
self.conv = nn.Sequential(
|
336 |
+
nn.Conv2d(in_ch, in_ch, kernel_size=(1, self.kernel_size), padding=(0, self.padding), groups=in_ch if depthwise else 1),
|
337 |
+
nn.Conv2d(in_ch, in_ch, kernel_size=(self.kernel_size, 1), padding=(self.padding, 0), groups=in_ch if depthwise else 1),
|
338 |
+
)
|
339 |
+
|
340 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
341 |
+
return self.conv(x)
|
342 |
+
|
343 |
+
|
344 |
+
|
345 |
+
def channel_shuffle(x, groups=2):
|
346 |
+
bat_size, channels, w, h = x.shape
|
347 |
+
group_c = channels // groups
|
348 |
+
x = x.view(bat_size, groups, group_c, w, h)
|
349 |
+
x = torch.transpose(x, 1, 2).contiguous()
|
350 |
+
x = x.view(bat_size, -1, w, h)
|
351 |
+
return x
|
352 |
+
|
353 |
+
|
354 |
+
class GatedFFN(nn.Module):
|
355 |
+
def __init__(self,
|
356 |
+
in_ch,
|
357 |
+
mlp_ratio,
|
358 |
+
kernel_size,
|
359 |
+
act_layer,):
|
360 |
+
super().__init__()
|
361 |
+
mlp_ch = in_ch * mlp_ratio
|
362 |
+
|
363 |
+
self.fn_1 = nn.Sequential(
|
364 |
+
nn.Conv2d(in_ch, mlp_ch, kernel_size=1, padding=0),
|
365 |
+
act_layer,
|
366 |
+
)
|
367 |
+
self.fn_2 = nn.Sequential(
|
368 |
+
nn.Conv2d(in_ch, in_ch, kernel_size=1, padding=0),
|
369 |
+
act_layer,
|
370 |
+
)
|
371 |
+
|
372 |
+
self.gate = nn.Conv2d(mlp_ch // 2, mlp_ch // 2,
|
373 |
+
kernel_size=kernel_size, padding=kernel_size // 2, groups=mlp_ch // 2)
|
374 |
+
|
375 |
+
def feat_decompose(self, x):
|
376 |
+
s = x - self.gate(x)
|
377 |
+
x = x + self.sigma * s
|
378 |
+
return x
|
379 |
+
|
380 |
+
def forward(self, x: torch.Tensor):
|
381 |
+
x = self.fn_1(x)
|
382 |
+
x, gate = torch.chunk(x, 2, dim=1)
|
383 |
+
|
384 |
+
gate = self.gate(gate)
|
385 |
+
x = x * gate
|
386 |
+
|
387 |
+
x = self.fn_2(x)
|
388 |
+
return x
|
389 |
+
|
390 |
+
|
391 |
+
|
392 |
+
class LayerNorm(nn.Module):
|
393 |
+
r""" LayerNorm that supports two data formats: channels_last (default) or channels_first.
|
394 |
+
The ordering of the dimensions in the inputs. channels_last corresponds to inputs with
|
395 |
+
shape (batch_size, height, width, channels) while channels_first corresponds to inputs
|
396 |
+
with shape (batch_size, channels, height, width).
|
397 |
+
"""
|
398 |
+
def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):
|
399 |
+
super().__init__()
|
400 |
+
self.weight = nn.Parameter(torch.ones(normalized_shape))
|
401 |
+
self.bias = nn.Parameter(torch.zeros(normalized_shape))
|
402 |
+
self.eps = eps
|
403 |
+
self.data_format = data_format
|
404 |
+
if self.data_format not in ["channels_last", "channels_first"]:
|
405 |
+
raise NotImplementedError
|
406 |
+
self.normalized_shape = (normalized_shape, )
|
407 |
+
|
408 |
+
def forward(self, x):
|
409 |
+
if self.data_format == "channels_last":
|
410 |
+
return F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
|
411 |
+
elif self.data_format == "channels_first":
|
412 |
+
u = x.mean(1, keepdim=True)
|
413 |
+
s = (x - u).pow(2).mean(1, keepdim=True)
|
414 |
+
x = (x - u) / torch.sqrt(s + self.eps)
|
415 |
+
x = self.weight[:, None, None] * x + self.bias[:, None, None]
|
416 |
+
return x
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch
|
2 |
+
numpy
|
3 |
+
PyYAML
|
4 |
+
Pillow>=6.2.2
|
5 |
+
gradio==4.16.0
|
6 |
+
gradio_imageslider==0.0.18
|