-
Notifications
You must be signed in to change notification settings - Fork 0
/
segformer_cityscapes_run.py
446 lines (350 loc) · 124 KB
/
segformer_cityscapes_run.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
# -*- coding: utf-8 -*-
"""Segformer Cityscapes Run.ipynb
Automatically generated by Colab.
Original file is located at
https://colab.research.google.com/gist/Jeremy26/13f71c273f0a1a93f758d02b2b77802e/segformer-cityscapes-run.ipynb
[Image reference](https://arxiv.org/abs/2105.15203)
## Imports & Download
"""
# basic imports
import numpy as np
# DL library imports
import torch
import torch.nn as nn
import torch.nn.functional as F
# libraries for loading image, plotting
import cv2
import matplotlib.pyplot as plt
try:
from einops import rearrange
import segmentation_models_pytorch as smp
from timm.models.layers import drop_path, trunc_normal_
except:
!pip install timm
!pip install einops
!pip install segmentation-models-pytorch
from einops import rearrange
import segmentation_models_pytorch as smp
from timm.models.layers import drop_path, trunc_normal_
"""## 1. Dataset : Download and use Cityscapes dataset"""
targetWidth = 1024
targetHeight = 512
# utility functions to get Cityscapes Pytorch dataset and dataloaders
from utils import get_dataloaders
from cityScapes_utils import get_cs_datasets
train_set, val_set, test_set= get_cs_datasets(rootDir='')
sample_image, sample_label = train_set[0]
print(f"There are {len(train_set)} train images, {len(val_set)} validation images, {len(test_set)} test Images")
print(f"Input shape = {sample_image.shape}, output label shape = {sample_label.shape}")
train_dataloader, val_dataloader, test_dataloader = get_dataloaders(train_set, val_set, test_set)#, batch_size=4)
"""### Show Sample images from dataset"""
from utils import inverse_transform
from cityScapes_utils import train_id_to_color as cs_train_id_to_color
rgb_image, label = train_set[np.random.choice(len(train_set))]
rgb_image = inverse_transform(rgb_image).permute(1, 2, 0).cpu().detach().numpy()
label = label.cpu().detach().numpy()
# plot sample image
fig, axes = plt.subplots(1,2, figsize=(20,10))
axes[0].imshow(rgb_image);
axes[0].set_title("Image");
axes[0].axis('off');
axes[1].imshow(cs_train_id_to_color[label]);
axes[1].set_title("Label");
axes[1].axis('off');
"""## 2 — Segformer Network
The Segformer is made of 2 parts: An Encoder and a Decoder. Let's begin with the Encoder.
### Mix Transformer / Encoder
* Overlap Patch Embedding
* Efficient Self-Attention
* Mix FFNs
**The 3 elements form a single Transformer Block.**
"""
class overlap_patch_embed(nn.Module):
def __init__(self, patch_size, stride, in_chans, embed_dim):
super().__init__()
self.patch_size = patch_size
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=stride,
padding=(patch_size // 2, patch_size // 2))
self.norm = nn.LayerNorm(embed_dim)
def forward(self, x):
x = self.proj(x)
_, _, h, w = x.shape
x = rearrange(x, 'b c h w -> b (h w) c')
x = self.norm(x)
return x, h, w
class efficient_self_attention(nn.Module):
def __init__(self, attn_dim, num_heads, dropout_p, sr_ratio):
super().__init__()
assert attn_dim % num_heads == 0, f'expected attn_dim {attn_dim} to be a multiple of num_heads {num_heads}'
self.attn_dim = attn_dim
self.num_heads = num_heads
self.dropout_p = dropout_p
self.sr_ratio = sr_ratio
if sr_ratio > 1:
self.sr = nn.Conv2d(attn_dim, attn_dim, kernel_size=sr_ratio, stride=sr_ratio)
self.norm = nn.LayerNorm(attn_dim)
# Multi-head Self-Attention using dot product
# Query - Key Dot product is scaled by root of head_dim
self.q = nn.Linear(attn_dim, attn_dim, bias=True)
self.kv = nn.Linear(attn_dim, attn_dim * 2, bias=True)
self.scale = (attn_dim // num_heads) ** -0.5
# Projecting concatenated outputs from
# multiple heads to single `attn_dim` size
self.proj = nn.Linear(attn_dim, attn_dim)
def forward(self, x, h, w):
q = self.q(x)
q = rearrange(q, ('b hw (m c) -> b m hw c'), m=self.num_heads)
if self.sr_ratio > 1:
x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
x = self.sr(x)
x = rearrange(x, 'b c h w -> b (h w) c')
x = self.norm(x)
x = self.kv(x)
x = rearrange(x, 'b d (a m c) -> a b m d c', a=2, m=self.num_heads)
k, v = x[0], x[1] # x.unbind(0)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
x = attn @ v
x = rearrange(x, 'b m hw c -> b hw (m c)')
x = self.proj(x)
x = F.dropout(x, p=self.dropout_p, training=self.training)
return x
class mix_feedforward(nn.Module):
def __init__(self, in_features, out_features, hidden_features, dropout_p = 0.0):
super().__init__()
self.fc1 = nn.Linear(in_features, hidden_features)
self.fc2 = nn.Linear(hidden_features, out_features)
# Depth-wise separable convolution
self.conv = nn.Conv2d(hidden_features, hidden_features, (3, 3), padding=(1, 1),
bias=True, groups=hidden_features)
self.dropout_p = dropout_p
def forward(self, x, h, w):
x = self.fc1(x)
x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
x = self.conv(x)
x = rearrange(x, 'b c h w -> b (h w) c')
x = F.gelu(x)
x = F.dropout(x, p=self.dropout_p, training=self.training)
x = self.fc2(x)
x = F.dropout(x, p=self.dropout_p, training=self.training)
return x
class transformer_block(nn.Module):
def __init__(self, dim, num_heads, dropout_p, drop_path_p, sr_ratio):
super().__init__()
# One transformer block is defined as :
# Norm -> self-attention -> Norm -> FeedForward
# skip-connections are added after attention and FF layers
self.attn = efficient_self_attention(attn_dim=dim, num_heads=num_heads,
dropout_p=dropout_p, sr_ratio=sr_ratio)
self.ffn = mix_feedforward( dim, dim, hidden_features=dim * 4, dropout_p=dropout_p)
self.drop_path_p = drop_path_p
self.norm1 = nn.LayerNorm(dim, eps=1e-6)
self.norm2 = nn.LayerNorm(dim, eps=1e-6)
def forward(self, x, h, w):
# Norm -> self-attention
skip = x
x = self.norm1(x)
x = self.attn(x, h, w)
x = drop_path(x, drop_prob=self.drop_path_p, training=self.training)
x = x + skip
# Norm -> FeedForward
skip = x
x = self.norm2(x)
x = self.ffn(x, h, w)
x = drop_path(x, drop_prob=self.drop_path_p, training=self.training)
x = x + skip
return x
"""Then, we'll build a set of transformer stages: each stage is Nx Transformer Blocks:"""
class mix_transformer_stage(nn.Module):
def __init__(self, patch_embed, blocks, norm):
super().__init__()
self.patch_embed = patch_embed
self.blocks = nn.ModuleList(blocks)
self.norm = norm
def forward(self, x):
x, h, w = self.patch_embed(x)
for block in self.blocks:
x = block(x, h, w)
x = self.norm(x)
x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
return x
"""Now we create the full decoder"""
class mix_transformer(nn.Module):
def __init__(self, in_chans, embed_dims, num_heads, depths,
sr_ratios, dropout_p, drop_path_p):
super().__init__()
self.stages = nn.ModuleList()
for stage_i in range(len(depths)):
# Each Stage consists of following blocks :
# Overlap patch embedding -> mix_transformer_block -> norm
blocks = []
for i in range(depths[stage_i]):
blocks.append(transformer_block(dim = embed_dims[stage_i],
num_heads= num_heads[stage_i], dropout_p=dropout_p,
drop_path_p = drop_path_p * (sum(depths[:stage_i])+i) / (sum(depths)-1),
sr_ratio = sr_ratios[stage_i] ))
if(stage_i == 0):
patch_size = 7
stride = 4
in_chans = in_chans
else:
patch_size = 3
stride = 2
in_chans = embed_dims[stage_i -1]
patch_embed = overlap_patch_embed(patch_size, stride=stride, in_chans=in_chans,
embed_dim= embed_dims[stage_i])
norm = nn.LayerNorm(embed_dims[stage_i], eps=1e-6)
self.stages.append(mix_transformer_stage(patch_embed, blocks, norm))
def forward(self, x):
outputs = []
for stage in self.stages:
x = stage(x)
outputs.append(x)
return outputs
"""### Decoder Head"""
class segformer_head(nn.Module):
def __init__(self, in_channels, num_classes, embed_dim, dropout_p=0.1):
super().__init__()
self.in_channels = in_channels
self.num_classes = num_classes
self.embed_dim = embed_dim
self.dropout_p = dropout_p
# 1x1 conv to fuse multi-scale output from encoder
self.layers = nn.ModuleList([nn.Conv2d(chans, embed_dim, (1, 1))
for chans in reversed(in_channels)])
self.linear_fuse = nn.Conv2d(embed_dim * len(self.layers), embed_dim, (1, 1), bias=False)
self.bn = nn.BatchNorm2d(embed_dim, eps=1e-5)
# 1x1 conv to get num_class channel predictions
self.linear_pred = nn.Conv2d(self.embed_dim, num_classes, kernel_size=(1, 1))
self.init_weights()
def init_weights(self):
nn.init.kaiming_normal_(self.linear_fuse.weight, mode='fan_out', nonlinearity='relu')
nn.init.constant_(self.bn.weight, 1)
nn.init.constant_(self.bn.bias, 0)
def forward(self, x):
feature_size = x[0].shape[2:]
# project each encoder stage output to H/4, W/4
x = [layer(xi) for layer, xi in zip(self.layers, reversed(x))]
x = [F.interpolate(xi, size=feature_size, mode='bilinear', align_corners=False)
for xi in x[:-1]] + [x[-1]]
# concatenate project output and use 1x1
# convs to get num_class channel output
x = self.linear_fuse(torch.cat(x, dim=1))
x = self.bn(x)
x = F.relu(x, inplace=True)
x = F.dropout(x, p=self.dropout_p, training=self.training)
x = self.linear_pred(x)
return x
"""### Full Segformer"""
class segformer_mit_b3(nn.Module):
def __init__(self, in_channels, num_classes):
super().__init__()
# Encoder block
self.backbone = mix_transformer(in_chans=in_channels, embed_dims=(64, 128, 320, 512),
num_heads=(1, 2, 5, 8), depths=(3, 4, 18, 3),
sr_ratios=(8, 4, 2, 1), dropout_p=0.0, drop_path_p=0.1)
# decoder block
self.decoder_head = segformer_head(in_channels=(64, 128, 320, 512),
num_classes=num_classes, embed_dim=256)
# init weights
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
elif isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
def forward(self, x):
image_hw = x.shape[2:]
x = self.backbone(x)
x = self.decoder_head(x)
x = F.interpolate(x, size=image_hw, mode='bilinear', align_corners=False)
return x
"""## 3. Training : Train and validate model on the custom dataset
We will reuse the utility functions we defined in FCN notebook
"""
from utils import meanIoU # metric class
from utils import plot_training_results # function to plot training curves
from utils import evaluate_model # evaluation function
from utils import train_validate_model # train validate function
"""### Model Training"""
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# MODEL HYPERPARAMETERS
N_EPOCHS = 12
NUM_CLASSES = 19
MAX_LR = 1e-3
MODEL_NAME = f'segformer_mit_b3_cs_pretrain_19CLS_{targetHeight}_{targetWidth}_CE_loss'
import os
import torch.optim as optim
from torch.optim.lr_scheduler import OneCycleLR
# criterion = smp.losses.DiceLoss('multiclass', classes=np.arange(20).tolist(), log_loss = True, smooth=1.0)
criterion = nn.CrossEntropyLoss(ignore_index=19)
# create model, load imagenet pretrained weights
model = segformer_mit_b3(in_channels=3, num_classes=NUM_CLASSES).to(device)
model.backbone.load_state_dict(torch.load('segformer_mit_b3_imagenet_weights.pt', map_location=device))
# create optimizer, lr_scheduler and pass to training function
optimizer = optim.Adam(model.parameters(), lr=MAX_LR)
scheduler = OneCycleLR(optimizer, max_lr= MAX_LR, epochs = N_EPOCHS,steps_per_epoch = len(train_dataloader),
pct_start=0.3, div_factor=10, anneal_strategy='cos')
_ = train_validate_model(model, N_EPOCHS, MODEL_NAME, criterion, optimizer,
device, train_dataloader, val_dataloader, meanIoU, 'meanIoU',
NUM_CLASSES, lr_scheduler = scheduler, output_path = "")
"""## 4. Evaluate : Evaluate the model on Test Data and visualize results"""
model.load_state_dict(torch.load(f'{MODEL_NAME}.pt', map_location=device))
_, test_metric = evaluate_model(model, test_dataloader, criterion, meanIoU, NUM_CLASSES, device)
print(f"\nModel has {test_metric} mean IoU in test set")
from utils import visualize_predictions
num_test_samples = 2
_, axes = plt.subplots(num_test_samples, 3, figsize=(3*6, num_test_samples * 4))
visualize_predictions(model, test_set, axes, device, numTestSamples=num_test_samples,
id_to_color = cs_train_id_to_color)
"""## Test on sample video"""
import os
from tqdm import tqdm
from utils import preprocess
def predict_cs_video(model, model_name, demo_video_path, id_to_color,
output_dir, target_width, target_height, device,
fps : int = 20, alpha : float = 0.3):
test_images = [os.path.join(demo_video_path, *[x]) for x in sorted(os.listdir(demo_video_path))]
output_filename = f'{model_name}_cs_part_overlay_demo_video.avi'
output_video_path = os.path.join(output_dir, *[output_filename])
# handles for input output videos
output_handle = cv2.VideoWriter(output_video_path, cv2.VideoWriter_fourcc(*'DIVX'), \
20, (target_width, target_height))
# create progress bar
num_frames = int(len(test_images))
pbar = tqdm(total = num_frames, position=0, leave=True)
for i in range(num_frames):
frame = cv2.imread(test_images[i])
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
# create torch tensor to give as input to model
pt_image = preprocess(frame)
pt_image = pt_image.to(device)
# get model prediction and remap certain labels to showcase
# only certain colors. class index 19 has color map (0,0,0),
# so remap unwanted classes to 19
y_pred = torch.argmax(model(pt_image.unsqueeze(0)), dim=1).squeeze(0)
predicted_labels = y_pred.cpu().detach().numpy()
predicted_labels[(predicted_labels < 11) & (predicted_labels != 0) & (predicted_labels != 6) & (predicted_labels != 7)] = 19
# convert to corresponding color
cm_labels = (id_to_color[predicted_labels]).astype(np.uint8)
# overlay prediction over input frame
overlay_image = cv2.addWeighted(frame, 1, cm_labels, alpha, 0)
overlay_image = cv2.cvtColor(overlay_image, cv2.COLOR_RGB2BGR)
# write output result and update progress
output_handle.write(overlay_image)
pbar.update(1)
output_handle.release()
predict_cs_video(model, MODEL_NAME,
demo_video_path = 'demoVideo/stuttgart_00',
id_to_color = cs_train_id_to_color,
output_dir = '.',
target_width = targetWidth,
target_height = targetHeight,
device = device,
alpha=0.7)