Upload 2 files
Browse files- configuration_mle.py +46 -0
- modeling_mle.py +413 -0
configuration_mle.py
ADDED
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from collections import OrderedDict
|
2 |
+
from typing import Any, List, Mapping, Optional
|
3 |
+
|
4 |
+
from transformers.configuration_utils import PretrainedConfig
|
5 |
+
from transformers.onnx import OnnxConfigWithPast, PatchingSpec
|
6 |
+
from transformers.utils import logging
|
7 |
+
|
8 |
+
|
9 |
+
logger = logging.get_logger(__name__)
|
10 |
+
|
11 |
+
|
12 |
+
class MLEConfig(PretrainedConfig):
|
13 |
+
model_type = "mle"
|
14 |
+
|
15 |
+
def __init__(
|
16 |
+
self,
|
17 |
+
in_channels=1,
|
18 |
+
num_encoder_layers=[2, 3, 5, 7, 12],
|
19 |
+
num_decoder_layers=[7, 5, 3, 2, 2],
|
20 |
+
last_hidden_channels=16,
|
21 |
+
block_stride_size=4,
|
22 |
+
block_kernel_size=3,
|
23 |
+
block_patch_size=24,
|
24 |
+
upsample_ratio=2,
|
25 |
+
batch_norm_eps=1e-3,
|
26 |
+
hidden_act="leaky_relu",
|
27 |
+
negative_slope=0.2,
|
28 |
+
**kwargs,
|
29 |
+
):
|
30 |
+
self.in_channels = in_channels
|
31 |
+
self.num_encoder_layers = num_encoder_layers
|
32 |
+
self.num_decoder_layers = num_decoder_layers
|
33 |
+
self.last_hidden_channels = last_hidden_channels
|
34 |
+
|
35 |
+
self.block_stride_size = block_stride_size
|
36 |
+
# if isinstance(block_kernel_size, int):
|
37 |
+
# self.block_kernel_size = (block_kernel_size, block_kernel_size)
|
38 |
+
self.block_kernel_size = block_kernel_size
|
39 |
+
self.block_patch_size = block_patch_size
|
40 |
+
|
41 |
+
self.upsample_ratio = upsample_ratio
|
42 |
+
self.batch_norm_eps = batch_norm_eps
|
43 |
+
self.hidden_act = hidden_act
|
44 |
+
self.negative_slope = negative_slope
|
45 |
+
|
46 |
+
super().__init__(**kwargs)
|
modeling_mle.py
ADDED
@@ -0,0 +1,413 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""PyTorch MLE (Mnaga Line Extraction) model"""
|
2 |
+
|
3 |
+
from dataclasses import dataclass
|
4 |
+
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
|
8 |
+
from transformers import PreTrainedModel
|
9 |
+
from transformers.modeling_outputs import ModelOutput, BaseModelOutput
|
10 |
+
from transformers.activations import ACT2FN
|
11 |
+
|
12 |
+
from .configuration_mle import MLEConfig
|
13 |
+
|
14 |
+
|
15 |
+
@dataclass
|
16 |
+
class MLEModelOutput(ModelOutput):
|
17 |
+
last_hidden_state: torch.FloatTensor | None = None
|
18 |
+
|
19 |
+
|
20 |
+
@dataclass
|
21 |
+
class MLEForAnimeLineExtractionOutput(ModelOutput):
|
22 |
+
last_hidden_state: torch.FloatTensor | None = None
|
23 |
+
pixel_values: torch.Tensor | None = None
|
24 |
+
|
25 |
+
|
26 |
+
class MLEBatchNorm(nn.Module):
|
27 |
+
def __init__(
|
28 |
+
self,
|
29 |
+
config: MLEConfig,
|
30 |
+
in_features: int,
|
31 |
+
):
|
32 |
+
super().__init__()
|
33 |
+
|
34 |
+
self.norm = nn.BatchNorm2d(in_features, eps=config.batch_norm_eps)
|
35 |
+
# the original model uses leaky_relu
|
36 |
+
if config.hidden_act == "leaky_relu":
|
37 |
+
self.act_fn = nn.LeakyReLU(negative_slope=config.negative_slope)
|
38 |
+
else:
|
39 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
40 |
+
|
41 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
42 |
+
hidden_states = self.norm(hidden_states)
|
43 |
+
hidden_states = self.act_fn(hidden_states)
|
44 |
+
|
45 |
+
return hidden_states
|
46 |
+
|
47 |
+
|
48 |
+
class MLEResBlock(nn.Module):
|
49 |
+
def __init__(
|
50 |
+
self,
|
51 |
+
config: MLEConfig,
|
52 |
+
in_channels: int,
|
53 |
+
out_channels: int,
|
54 |
+
stride_size: int,
|
55 |
+
):
|
56 |
+
super().__init__()
|
57 |
+
|
58 |
+
self.norm1 = MLEBatchNorm(config, in_channels)
|
59 |
+
self.conv1 = nn.Conv2d(
|
60 |
+
in_channels,
|
61 |
+
out_channels,
|
62 |
+
config.block_kernel_size,
|
63 |
+
stride=stride_size,
|
64 |
+
padding=config.block_kernel_size // 2,
|
65 |
+
)
|
66 |
+
|
67 |
+
self.norm2 = MLEBatchNorm(config, out_channels)
|
68 |
+
self.conv2 = nn.Conv2d(
|
69 |
+
out_channels,
|
70 |
+
out_channels,
|
71 |
+
config.block_kernel_size,
|
72 |
+
stride=1,
|
73 |
+
padding=config.block_kernel_size // 2,
|
74 |
+
)
|
75 |
+
|
76 |
+
if in_channels != out_channels or stride_size != 1:
|
77 |
+
self.resize = nn.Conv2d(
|
78 |
+
in_channels,
|
79 |
+
out_channels,
|
80 |
+
kernel_size=1,
|
81 |
+
stride=stride_size,
|
82 |
+
)
|
83 |
+
else:
|
84 |
+
self.resize = None
|
85 |
+
|
86 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
87 |
+
output = self.norm1(hidden_states)
|
88 |
+
output = self.conv1(output)
|
89 |
+
output = self.norm2(output)
|
90 |
+
output = self.conv2(output)
|
91 |
+
|
92 |
+
if self.resize is not None:
|
93 |
+
resized_input = self.resize(hidden_states)
|
94 |
+
output += resized_input
|
95 |
+
else:
|
96 |
+
output += hidden_states
|
97 |
+
|
98 |
+
return output
|
99 |
+
|
100 |
+
|
101 |
+
class MLEEncoderLayer(nn.Module):
|
102 |
+
def __init__(
|
103 |
+
self,
|
104 |
+
config: MLEConfig,
|
105 |
+
in_features: int,
|
106 |
+
out_features: int,
|
107 |
+
num_layers: int,
|
108 |
+
stride_sizes: list[int],
|
109 |
+
):
|
110 |
+
super().__init__()
|
111 |
+
|
112 |
+
self.blocks = nn.ModuleList(
|
113 |
+
[
|
114 |
+
MLEResBlock(
|
115 |
+
config,
|
116 |
+
in_channels=in_features if i == 0 else out_features,
|
117 |
+
out_channels=out_features,
|
118 |
+
stride_size=stride_sizes[i],
|
119 |
+
)
|
120 |
+
for i in range(num_layers)
|
121 |
+
]
|
122 |
+
)
|
123 |
+
|
124 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
125 |
+
for block in self.blocks:
|
126 |
+
hidden_states = block(hidden_states)
|
127 |
+
return hidden_states
|
128 |
+
|
129 |
+
|
130 |
+
class MLEEncoder(nn.Module):
|
131 |
+
def __init__(
|
132 |
+
self,
|
133 |
+
config: MLEConfig,
|
134 |
+
):
|
135 |
+
super().__init__()
|
136 |
+
|
137 |
+
self.layers = nn.ModuleList(
|
138 |
+
[
|
139 |
+
MLEEncoderLayer(
|
140 |
+
config,
|
141 |
+
in_features=(
|
142 |
+
config.in_channels
|
143 |
+
if i == 0
|
144 |
+
else config.in_channels
|
145 |
+
* config.block_patch_size
|
146 |
+
* (config.upsample_ratio ** (i - 1))
|
147 |
+
),
|
148 |
+
out_features=config.in_channels
|
149 |
+
* config.block_patch_size
|
150 |
+
* (config.upsample_ratio**i),
|
151 |
+
num_layers=num_layers,
|
152 |
+
stride_sizes=(
|
153 |
+
[
|
154 |
+
1 if i_layer < num_layers - 1 else 2
|
155 |
+
for i_layer in range(num_layers)
|
156 |
+
]
|
157 |
+
if i > 0
|
158 |
+
else [1 for _ in range(num_layers)]
|
159 |
+
),
|
160 |
+
)
|
161 |
+
for i, num_layers in enumerate(config.num_encoder_layers)
|
162 |
+
]
|
163 |
+
)
|
164 |
+
|
165 |
+
def forward(
|
166 |
+
self, hidden_states: torch.Tensor
|
167 |
+
) -> tuple[torch.Tensor, tuple[torch.Tensor, ...]]:
|
168 |
+
all_hidden_states: tuple[torch.Tensor, ...] = ()
|
169 |
+
for layer in self.layers:
|
170 |
+
hidden_states = layer(hidden_states)
|
171 |
+
all_hidden_states += (hidden_states,)
|
172 |
+
return hidden_states, all_hidden_states
|
173 |
+
|
174 |
+
|
175 |
+
class MLEUpsampleBlock(nn.Module):
|
176 |
+
def __init__(self, config: MLEConfig, in_features: int, out_features: int):
|
177 |
+
super().__init__()
|
178 |
+
|
179 |
+
self.norm = MLEBatchNorm(config, in_features=in_features)
|
180 |
+
self.conv = nn.Conv2d(
|
181 |
+
in_features,
|
182 |
+
out_features,
|
183 |
+
config.block_kernel_size,
|
184 |
+
stride=1,
|
185 |
+
padding=config.block_kernel_size // 2,
|
186 |
+
)
|
187 |
+
self.upsample = nn.Upsample(scale_factor=config.upsample_ratio)
|
188 |
+
|
189 |
+
def forward(self, hidden_states: torch.Tensor):
|
190 |
+
output = self.norm(hidden_states)
|
191 |
+
output = self.conv(output)
|
192 |
+
output = self.upsample(output)
|
193 |
+
|
194 |
+
return output
|
195 |
+
|
196 |
+
|
197 |
+
class MLEUpsampleResBlock(nn.Module):
|
198 |
+
def __init__(self, config: MLEConfig, in_features: int, out_features: int):
|
199 |
+
super().__init__()
|
200 |
+
|
201 |
+
self.upsample = MLEUpsampleBlock(
|
202 |
+
config, in_features=in_features, out_features=out_features
|
203 |
+
)
|
204 |
+
|
205 |
+
self.norm = MLEBatchNorm(config, in_features=out_features)
|
206 |
+
self.conv = nn.Conv2d(
|
207 |
+
out_features,
|
208 |
+
out_features,
|
209 |
+
config.block_kernel_size,
|
210 |
+
stride=1,
|
211 |
+
padding=config.block_kernel_size // 2,
|
212 |
+
)
|
213 |
+
|
214 |
+
if in_features != out_features:
|
215 |
+
self.resize = nn.Sequential(
|
216 |
+
nn.Conv2d(
|
217 |
+
in_features,
|
218 |
+
out_features,
|
219 |
+
kernel_size=1,
|
220 |
+
stride=1,
|
221 |
+
),
|
222 |
+
nn.Upsample(scale_factor=config.upsample_ratio),
|
223 |
+
)
|
224 |
+
else:
|
225 |
+
self.resize = None
|
226 |
+
|
227 |
+
def forward(self, hidden_states: torch.Tensor):
|
228 |
+
output = self.upsample(hidden_states)
|
229 |
+
output = self.norm(output)
|
230 |
+
output = self.conv(output)
|
231 |
+
|
232 |
+
if self.resize is not None:
|
233 |
+
output += self.resize(hidden_states)
|
234 |
+
|
235 |
+
return output
|
236 |
+
|
237 |
+
|
238 |
+
class MLEDecoderLayer(nn.Module):
|
239 |
+
def __init__(
|
240 |
+
self,
|
241 |
+
config: MLEConfig,
|
242 |
+
in_features: int,
|
243 |
+
out_features: int,
|
244 |
+
num_layers: int,
|
245 |
+
):
|
246 |
+
super().__init__()
|
247 |
+
|
248 |
+
self.blocks = nn.ModuleList(
|
249 |
+
[
|
250 |
+
(
|
251 |
+
MLEResBlock(
|
252 |
+
config,
|
253 |
+
in_channels=out_features,
|
254 |
+
out_channels=out_features,
|
255 |
+
stride_size=1,
|
256 |
+
)
|
257 |
+
if i > 0
|
258 |
+
else MLEUpsampleResBlock(
|
259 |
+
config,
|
260 |
+
in_features=in_features,
|
261 |
+
out_features=out_features,
|
262 |
+
)
|
263 |
+
)
|
264 |
+
for i in range(num_layers)
|
265 |
+
]
|
266 |
+
)
|
267 |
+
|
268 |
+
def forward(
|
269 |
+
self, hidden_states: torch.Tensor, shortcut_states: torch.Tensor
|
270 |
+
) -> torch.Tensor:
|
271 |
+
for block in self.blocks:
|
272 |
+
hidden_states = block(hidden_states)
|
273 |
+
|
274 |
+
hidden_states += shortcut_states
|
275 |
+
|
276 |
+
return hidden_states
|
277 |
+
|
278 |
+
|
279 |
+
class MLEDecoderHead(nn.Module):
|
280 |
+
def __init__(self, config: MLEConfig, num_layers: int):
|
281 |
+
super().__init__()
|
282 |
+
|
283 |
+
self.layer = MLEEncoderLayer(
|
284 |
+
config,
|
285 |
+
in_features=config.block_patch_size,
|
286 |
+
out_features=config.last_hidden_channels,
|
287 |
+
stride_sizes=[1 for _ in range(num_layers)],
|
288 |
+
num_layers=num_layers,
|
289 |
+
)
|
290 |
+
self.norm = MLEBatchNorm(config, in_features=config.last_hidden_channels)
|
291 |
+
self.conv = nn.Conv2d(
|
292 |
+
config.last_hidden_channels,
|
293 |
+
out_channels=1,
|
294 |
+
kernel_size=1,
|
295 |
+
stride=1,
|
296 |
+
)
|
297 |
+
|
298 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
299 |
+
hidden_states = self.layer(hidden_states)
|
300 |
+
hidden_states = self.norm(hidden_states)
|
301 |
+
pixel_values = self.conv(hidden_states)
|
302 |
+
return pixel_values
|
303 |
+
|
304 |
+
|
305 |
+
class MLEDecoder(nn.Module):
|
306 |
+
def __init__(
|
307 |
+
self,
|
308 |
+
config: MLEConfig,
|
309 |
+
):
|
310 |
+
super().__init__()
|
311 |
+
|
312 |
+
encoder_output_channels = (
|
313 |
+
config.in_channels
|
314 |
+
* config.block_patch_size
|
315 |
+
* (config.upsample_ratio ** (len(config.num_encoder_layers) - 1))
|
316 |
+
)
|
317 |
+
upsample_ratio = config.upsample_ratio
|
318 |
+
num_decoder_layers = config.num_decoder_layers
|
319 |
+
|
320 |
+
self.layers = nn.ModuleList(
|
321 |
+
[
|
322 |
+
(
|
323 |
+
MLEDecoderLayer(
|
324 |
+
config,
|
325 |
+
in_features=encoder_output_channels // (upsample_ratio**i),
|
326 |
+
out_features=encoder_output_channels
|
327 |
+
// (upsample_ratio ** (i + 1)),
|
328 |
+
num_layers=num_layers,
|
329 |
+
)
|
330 |
+
if i < len(num_decoder_layers) - 1
|
331 |
+
else MLEDecoderHead(
|
332 |
+
config,
|
333 |
+
num_layers=num_layers,
|
334 |
+
)
|
335 |
+
)
|
336 |
+
for i, num_layers in enumerate(num_decoder_layers)
|
337 |
+
]
|
338 |
+
)
|
339 |
+
|
340 |
+
def forward(
|
341 |
+
self,
|
342 |
+
last_hidden_states: torch.Tensor,
|
343 |
+
encoder_hidden_states: tuple[torch.Tensor, ...],
|
344 |
+
) -> torch.Tensor:
|
345 |
+
hidden_states = last_hidden_states
|
346 |
+
num_encoder_hidden_states = len(encoder_hidden_states) # 5
|
347 |
+
|
348 |
+
for i, layer in enumerate(self.layers):
|
349 |
+
if i < len(self.layers) - 1:
|
350 |
+
hidden_states = layer(
|
351 |
+
hidden_states,
|
352 |
+
# 0, 1, 2, 3, 4
|
353 |
+
# ↓ ↓ ↓ ↓ ↓
|
354 |
+
# 8, 7, 6, 5, 5
|
355 |
+
encoder_hidden_states[num_encoder_hidden_states - 2 - i],
|
356 |
+
)
|
357 |
+
else:
|
358 |
+
# decoder head
|
359 |
+
hidden_states = layer(hidden_states)
|
360 |
+
|
361 |
+
return hidden_states
|
362 |
+
|
363 |
+
|
364 |
+
class MLEPretrainedModel(PreTrainedModel):
|
365 |
+
config_class = MLEConfig
|
366 |
+
base_model_prefix = "model"
|
367 |
+
supports_gradient_checkpointing = True
|
368 |
+
|
369 |
+
|
370 |
+
class MLEModel(MLEPretrainedModel):
|
371 |
+
def __init__(self, config: MLEConfig):
|
372 |
+
super().__init__(config)
|
373 |
+
self.config = config
|
374 |
+
|
375 |
+
self.encoder = MLEEncoder(config)
|
376 |
+
self.decoder = MLEDecoder(config)
|
377 |
+
|
378 |
+
# Initialize weights and apply final processing
|
379 |
+
self.post_init()
|
380 |
+
|
381 |
+
def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
|
382 |
+
encoder_output, all_hidden_states = self.encoder(pixel_values)
|
383 |
+
decoder_output = self.decoder(encoder_output, all_hidden_states)
|
384 |
+
|
385 |
+
return decoder_output
|
386 |
+
|
387 |
+
|
388 |
+
class MLEForAnimeLineExtraction(MLEPretrainedModel):
|
389 |
+
def __init__(self, config: MLEConfig):
|
390 |
+
super().__init__(config)
|
391 |
+
|
392 |
+
self.model = MLEModel(config)
|
393 |
+
|
394 |
+
def postprocess(self, output_tensor: torch.Tensor, input_shape: torch.Size):
|
395 |
+
pixel_values = output_tensor[0, 0, :, :]
|
396 |
+
pixel_values = torch.clip(pixel_values, 0, 255)
|
397 |
+
|
398 |
+
pixel_values = pixel_values[0 : input_shape[2], 0 : input_shape[3]]
|
399 |
+
return pixel_values
|
400 |
+
|
401 |
+
def forward(
|
402 |
+
self, pixel_values: torch.Tensor, return_dict: bool = True
|
403 |
+
) -> tuple[torch.Tensor, ...] | MLEForAnimeLineExtractionOutput:
|
404 |
+
model_output = self.model(pixel_values)
|
405 |
+
|
406 |
+
if not return_dict:
|
407 |
+
return (model_output, self.postprocess(model_output, pixel_values.shape))
|
408 |
+
|
409 |
+
else:
|
410 |
+
return MLEForAnimeLineExtractionOutput(
|
411 |
+
last_hidden_state=model_output,
|
412 |
+
pixel_values=self.postprocess(model_output, pixel_values.shape),
|
413 |
+
)
|