Spaces:
Runtime error
Runtime error
Upload train.py
Browse files
train.py
ADDED
@@ -0,0 +1,362 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import datetime
|
3 |
+
import os
|
4 |
+
import traceback
|
5 |
+
|
6 |
+
import numpy as np
|
7 |
+
import torch
|
8 |
+
from tensorboardX import SummaryWriter
|
9 |
+
from torch import nn
|
10 |
+
from torchvision import transforms
|
11 |
+
from tqdm.autonotebook import tqdm
|
12 |
+
|
13 |
+
from val import val
|
14 |
+
from backbone import HybridNetsBackbone
|
15 |
+
from hybridnets.loss import FocalLoss
|
16 |
+
from utils.sync_batchnorm import patch_replication_callback
|
17 |
+
from utils.utils import replace_w_sync_bn, CustomDataParallel, get_last_weights, init_weights, boolean_string, \
|
18 |
+
save_checkpoint, DataLoaderX, Params
|
19 |
+
from hybridnets.dataset import BddDataset
|
20 |
+
from hybridnets.loss import FocalLossSeg, TverskyLoss
|
21 |
+
from hybridnets.autoanchor import run_anchor
|
22 |
+
|
23 |
+
|
24 |
+
def get_args():
|
25 |
+
parser = argparse.ArgumentParser('HybridNets: End-to-End Perception Network - DatVu')
|
26 |
+
parser.add_argument('-p', '--project', type=str, default='bdd100k', help='Project file that contains parameters')
|
27 |
+
parser.add_argument('-c', '--compound_coef', type=int, default=3, help='Coefficient of efficientnet backbone')
|
28 |
+
parser.add_argument('-n', '--num_workers', type=int, default=12, help='Num_workers of dataloader')
|
29 |
+
parser.add_argument('-b', '--batch_size', type=int, default=12, help='Number of images per batch among all devices')
|
30 |
+
parser.add_argument('--freeze_backbone', type=boolean_string, default=False,
|
31 |
+
help='Freeze encoder and neck (effnet and bifpn)')
|
32 |
+
parser.add_argument('--freeze_det', type=boolean_string, default=False,
|
33 |
+
help='Freeze detection head')
|
34 |
+
parser.add_argument('--freeze_seg', type=boolean_string, default=False,
|
35 |
+
help='Freeze segmentation head')
|
36 |
+
parser.add_argument('--lr', type=float, default=1e-4)
|
37 |
+
parser.add_argument('--optim', type=str, default='adamw', help='Select optimizer for training, '
|
38 |
+
'suggest using \'admaw\' until the'
|
39 |
+
' very final stage then switch to \'sgd\'')
|
40 |
+
parser.add_argument('--num_epochs', type=int, default=500)
|
41 |
+
parser.add_argument('--val_interval', type=int, default=1, help='Number of epoches between valing phases')
|
42 |
+
parser.add_argument('--save_interval', type=int, default=500, help='Number of steps between saving')
|
43 |
+
parser.add_argument('--es_min_delta', type=float, default=0.0,
|
44 |
+
help='Early stopping\'s parameter: minimum change loss to qualify as an improvement')
|
45 |
+
parser.add_argument('--es_patience', type=int, default=0,
|
46 |
+
help='Early stopping\'s parameter: number of epochs with no improvement after which '
|
47 |
+
'training will be stopped. Set to 0 to disable this technique')
|
48 |
+
parser.add_argument('--data_path', type=str, default='datasets/', help='The root folder of dataset')
|
49 |
+
parser.add_argument('--log_path', type=str, default='checkpoints/')
|
50 |
+
parser.add_argument('-w', '--load_weights', type=str, default=None,
|
51 |
+
help='Whether to load weights from a checkpoint, set None to initialize,'
|
52 |
+
'set \'last\' to load last checkpoint')
|
53 |
+
parser.add_argument('--saved_path', type=str, default='checkpoints/')
|
54 |
+
parser.add_argument('--debug', type=boolean_string, default=False,
|
55 |
+
help='Whether visualize the predicted boxes of training, '
|
56 |
+
'the output images will be in test/')
|
57 |
+
parser.add_argument('--cal_map', type=boolean_string, default=True,
|
58 |
+
help='Calculate mAP in validation')
|
59 |
+
parser.add_argument('-v', '--verbose', type=boolean_string, default=True,
|
60 |
+
help='Whether to print results per class when valing')
|
61 |
+
parser.add_argument('--plots', type=boolean_string, default=True,
|
62 |
+
help='Whether to plot confusion matrix when valing')
|
63 |
+
parser.add_argument('--num_gpus', type=int, default=1,
|
64 |
+
help='Number of GPUs to be used (0 to use CPU)')
|
65 |
+
|
66 |
+
args = parser.parse_args()
|
67 |
+
return args
|
68 |
+
|
69 |
+
|
70 |
+
class ModelWithLoss(nn.Module):
|
71 |
+
def __init__(self, model, debug=False):
|
72 |
+
super().__init__()
|
73 |
+
self.criterion = FocalLoss()
|
74 |
+
self.seg_criterion1 = TverskyLoss(mode='multilabel', alpha=0.7, beta=0.3, gamma=4.0 / 3, from_logits=False)
|
75 |
+
self.seg_criterion2 = FocalLossSeg(mode='multilabel', alpha=0.25)
|
76 |
+
self.model = model
|
77 |
+
self.debug = debug
|
78 |
+
|
79 |
+
def forward(self, imgs, annotations, seg_annot, obj_list=None):
|
80 |
+
_, regression, classification, anchors, segmentation = self.model(imgs)
|
81 |
+
|
82 |
+
if self.debug:
|
83 |
+
cls_loss, reg_loss = self.criterion(classification, regression, anchors, annotations,
|
84 |
+
imgs=imgs, obj_list=obj_list)
|
85 |
+
tversky_loss = self.seg_criterion1(segmentation, seg_annot)
|
86 |
+
focal_loss = self.seg_criterion2(segmentation, seg_annot)
|
87 |
+
else:
|
88 |
+
cls_loss, reg_loss = self.criterion(classification, regression, anchors, annotations)
|
89 |
+
tversky_loss = self.seg_criterion1(segmentation, seg_annot)
|
90 |
+
focal_loss = self.seg_criterion2(segmentation, seg_annot)
|
91 |
+
|
92 |
+
# Visualization
|
93 |
+
# seg_0 = seg_annot[0]
|
94 |
+
# # print('bbb', seg_0.shape)
|
95 |
+
# seg_0 = torch.argmax(seg_0, dim = 0)
|
96 |
+
# # print('before', seg_0.shape)
|
97 |
+
# seg_0 = seg_0.cpu().numpy()
|
98 |
+
# #.transpose(1, 2, 0)
|
99 |
+
# print(seg_0.shape)
|
100 |
+
#
|
101 |
+
# anh = np.zeros((384,640,3))
|
102 |
+
#
|
103 |
+
# anh[seg_0 == 0] = (255,0,0)
|
104 |
+
# anh[seg_0 == 1] = (0,255,0)
|
105 |
+
# anh[seg_0 == 2] = (0,0,255)
|
106 |
+
#
|
107 |
+
# anh = np.uint8(anh)
|
108 |
+
#
|
109 |
+
# cv2.imwrite('anh.jpg',anh)
|
110 |
+
|
111 |
+
seg_loss = tversky_loss + 1 * focal_loss
|
112 |
+
# print("TVERSKY", tversky_loss)
|
113 |
+
# print("FOCAL", focal_loss)
|
114 |
+
|
115 |
+
return cls_loss, reg_loss, seg_loss, regression, classification, anchors, segmentation
|
116 |
+
|
117 |
+
|
118 |
+
def train(opt):
|
119 |
+
params = Params(f'projects/{opt.project}.yml')
|
120 |
+
|
121 |
+
if opt.num_gpus == 0:
|
122 |
+
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
|
123 |
+
|
124 |
+
if torch.cuda.is_available():
|
125 |
+
torch.cuda.manual_seed(42)
|
126 |
+
else:
|
127 |
+
torch.manual_seed(42)
|
128 |
+
|
129 |
+
opt.saved_path = opt.saved_path + f'/{params.project_name}/'
|
130 |
+
opt.log_path = opt.log_path + f'/{params.project_name}/tensorboard/'
|
131 |
+
os.makedirs(opt.log_path, exist_ok=True)
|
132 |
+
os.makedirs(opt.saved_path, exist_ok=True)
|
133 |
+
|
134 |
+
train_dataset = BddDataset(
|
135 |
+
params=params,
|
136 |
+
is_train=True,
|
137 |
+
inputsize=params.model['image_size'],
|
138 |
+
transform=transforms.Compose([
|
139 |
+
transforms.ToTensor(),
|
140 |
+
transforms.Normalize(
|
141 |
+
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
|
142 |
+
)
|
143 |
+
])
|
144 |
+
)
|
145 |
+
|
146 |
+
training_generator = DataLoaderX(
|
147 |
+
train_dataset,
|
148 |
+
batch_size=opt.batch_size,
|
149 |
+
shuffle=True,
|
150 |
+
num_workers=opt.num_workers,
|
151 |
+
pin_memory=params.pin_memory,
|
152 |
+
collate_fn=BddDataset.collate_fn
|
153 |
+
)
|
154 |
+
|
155 |
+
valid_dataset = BddDataset(
|
156 |
+
params=params,
|
157 |
+
is_train=False,
|
158 |
+
inputsize=params.model['image_size'],
|
159 |
+
transform=transforms.Compose([
|
160 |
+
transforms.ToTensor(),
|
161 |
+
transforms.Normalize(
|
162 |
+
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
|
163 |
+
)
|
164 |
+
])
|
165 |
+
)
|
166 |
+
|
167 |
+
val_generator = DataLoaderX(
|
168 |
+
valid_dataset,
|
169 |
+
batch_size=opt.batch_size,
|
170 |
+
shuffle=False,
|
171 |
+
num_workers=opt.num_workers,
|
172 |
+
pin_memory=params.pin_memory,
|
173 |
+
collate_fn=BddDataset.collate_fn
|
174 |
+
)
|
175 |
+
|
176 |
+
if params.need_autoanchor:
|
177 |
+
params.anchors_scales, params.anchors_ratios = run_anchor(None, train_dataset)
|
178 |
+
|
179 |
+
model = HybridNetsBackbone(num_classes=len(params.obj_list), compound_coef=opt.compound_coef,
|
180 |
+
ratios=eval(params.anchors_ratios), scales=eval(params.anchors_scales),
|
181 |
+
seg_classes=len(params.seg_list))
|
182 |
+
|
183 |
+
# load last weights
|
184 |
+
ckpt = {}
|
185 |
+
# last_step = None
|
186 |
+
if opt.load_weights:
|
187 |
+
if opt.load_weights.endswith('.pth'):
|
188 |
+
weights_path = opt.load_weights
|
189 |
+
else:
|
190 |
+
weights_path = get_last_weights(opt.saved_path)
|
191 |
+
# try:
|
192 |
+
# last_step = int(os.path.basename(weights_path).split('_')[-1].split('.')[0])
|
193 |
+
# except:
|
194 |
+
# last_step = 0
|
195 |
+
|
196 |
+
try:
|
197 |
+
ckpt = torch.load(weights_path)
|
198 |
+
model.load_state_dict(ckpt.get('model', ckpt), strict=False)
|
199 |
+
except RuntimeError as e:
|
200 |
+
print(f'[Warning] Ignoring {e}')
|
201 |
+
print(
|
202 |
+
'[Warning] Don\'t panic if you see this, this might be because you load a pretrained weights with different number of classes. The rest of the weights should be loaded already.')
|
203 |
+
else:
|
204 |
+
print('[Info] initializing weights...')
|
205 |
+
init_weights(model)
|
206 |
+
|
207 |
+
print('[Info] Successfully!!!')
|
208 |
+
|
209 |
+
if opt.freeze_backbone:
|
210 |
+
def freeze_backbone(m):
|
211 |
+
classname = m.__class__.__name__
|
212 |
+
if classname in ['EfficientNetEncoder', 'BiFPN']: # replace backbone classname when using another backbone
|
213 |
+
print("[Info] freezing {}".format(classname))
|
214 |
+
for param in m.parameters():
|
215 |
+
param.requires_grad = False
|
216 |
+
model.apply(freeze_backbone)
|
217 |
+
print('[Info] freezed backbone')
|
218 |
+
|
219 |
+
if opt.freeze_det:
|
220 |
+
def freeze_det(m):
|
221 |
+
classname = m.__class__.__name__
|
222 |
+
if classname in ['Regressor', 'Classifier', 'Anchors']:
|
223 |
+
print("[Info] freezing {}".format(classname))
|
224 |
+
for param in m.parameters():
|
225 |
+
param.requires_grad = False
|
226 |
+
model.apply(freeze_det)
|
227 |
+
print('[Info] freezed detection head')
|
228 |
+
|
229 |
+
if opt.freeze_seg:
|
230 |
+
def freeze_seg(m):
|
231 |
+
classname = m.__class__.__name__
|
232 |
+
if classname in ['BiFPNDecoder', 'SegmentationHead']:
|
233 |
+
print("[Info] freezing {}".format(classname))
|
234 |
+
for param in m.parameters():
|
235 |
+
param.requires_grad = False
|
236 |
+
model.apply(freeze_seg)
|
237 |
+
print('[Info] freezed segmentation head')
|
238 |
+
|
239 |
+
# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
|
240 |
+
# apply sync_bn when using multiple gpu and batch_size per gpu is lower than 4
|
241 |
+
# useful when gpu memory is limited.
|
242 |
+
# because when bn is disable, the training will be very unstable or slow to converge,
|
243 |
+
# apply sync_bn can solve it,
|
244 |
+
# by packing all mini-batch across all gpus as one batch and normalize, then send it back to all gpus.
|
245 |
+
# but it would also slow down the training by a little bit.
|
246 |
+
if opt.num_gpus > 1 and opt.batch_size // opt.num_gpus < 4:
|
247 |
+
model.apply(replace_w_sync_bn)
|
248 |
+
use_sync_bn = True
|
249 |
+
else:
|
250 |
+
use_sync_bn = False
|
251 |
+
|
252 |
+
writer = SummaryWriter(opt.log_path + f'/{datetime.datetime.now().strftime("%Y%m%d-%H%M%S")}/')
|
253 |
+
|
254 |
+
# wrap the model with loss function, to reduce the memory usage on gpu0 and speedup
|
255 |
+
model = ModelWithLoss(model, debug=opt.debug)
|
256 |
+
|
257 |
+
if opt.num_gpus > 0:
|
258 |
+
model = model.cuda()
|
259 |
+
if opt.num_gpus > 1:
|
260 |
+
model = CustomDataParallel(model, opt.num_gpus)
|
261 |
+
if use_sync_bn:
|
262 |
+
patch_replication_callback(model)
|
263 |
+
|
264 |
+
if opt.optim == 'adamw':
|
265 |
+
optimizer = torch.optim.AdamW(model.parameters(), opt.lr)
|
266 |
+
else:
|
267 |
+
optimizer = torch.optim.SGD(model.parameters(), opt.lr, momentum=0.9, nesterov=True)
|
268 |
+
# print(ckpt)
|
269 |
+
if opt.load_weights is not None and ckpt.get('optimizer', None):
|
270 |
+
optimizer.load_state_dict(ckpt['optimizer'])
|
271 |
+
|
272 |
+
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=3, verbose=True)
|
273 |
+
|
274 |
+
epoch = 0
|
275 |
+
best_loss = 1e5
|
276 |
+
best_epoch = 0
|
277 |
+
last_step = ckpt['step'] if opt.load_weights is not None and ckpt.get('step', None) else 0
|
278 |
+
best_fitness = ckpt['best_fitness'] if opt.load_weights is not None and ckpt.get('best_fitness', None) else 0
|
279 |
+
step = max(0, last_step)
|
280 |
+
model.train()
|
281 |
+
|
282 |
+
num_iter_per_epoch = len(training_generator)
|
283 |
+
try:
|
284 |
+
for epoch in range(opt.num_epochs):
|
285 |
+
last_epoch = step // num_iter_per_epoch
|
286 |
+
if epoch < last_epoch:
|
287 |
+
continue
|
288 |
+
|
289 |
+
epoch_loss = []
|
290 |
+
progress_bar = tqdm(training_generator)
|
291 |
+
for iter, data in enumerate(progress_bar):
|
292 |
+
if iter < step - last_epoch * num_iter_per_epoch:
|
293 |
+
progress_bar.update()
|
294 |
+
continue
|
295 |
+
try:
|
296 |
+
imgs = data['img']
|
297 |
+
annot = data['annot']
|
298 |
+
seg_annot = data['segmentation']
|
299 |
+
|
300 |
+
if opt.num_gpus == 1:
|
301 |
+
# if only one gpu, just send it to cuda:0
|
302 |
+
# elif multiple gpus, send it to multiple gpus in CustomDataParallel, not here
|
303 |
+
imgs = imgs.cuda()
|
304 |
+
annot = annot.cuda()
|
305 |
+
seg_annot = seg_annot.cuda().long()
|
306 |
+
|
307 |
+
optimizer.zero_grad()
|
308 |
+
cls_loss, reg_loss, seg_loss, regression, classification, anchors, segmentation = model(imgs, annot,
|
309 |
+
seg_annot,
|
310 |
+
obj_list=params.obj_list)
|
311 |
+
cls_loss = cls_loss.mean()
|
312 |
+
reg_loss = reg_loss.mean()
|
313 |
+
seg_loss = seg_loss.mean()
|
314 |
+
|
315 |
+
loss = cls_loss + reg_loss + seg_loss
|
316 |
+
if loss == 0 or not torch.isfinite(loss):
|
317 |
+
continue
|
318 |
+
|
319 |
+
loss.backward()
|
320 |
+
# torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
|
321 |
+
optimizer.step()
|
322 |
+
|
323 |
+
epoch_loss.append(float(loss))
|
324 |
+
|
325 |
+
progress_bar.set_description(
|
326 |
+
'Step: {}. Epoch: {}/{}. Iteration: {}/{}. Cls loss: {:.5f}. Reg loss: {:.5f}. Seg loss: {:.5f}. Total loss: {:.5f}'.format(
|
327 |
+
step, epoch, opt.num_epochs, iter + 1, num_iter_per_epoch, cls_loss.item(),
|
328 |
+
reg_loss.item(), seg_loss.item(), loss.item()))
|
329 |
+
writer.add_scalars('Loss', {'train': loss}, step)
|
330 |
+
writer.add_scalars('Regression_loss', {'train': reg_loss}, step)
|
331 |
+
writer.add_scalars('Classfication_loss', {'train': cls_loss}, step)
|
332 |
+
writer.add_scalars('Segmentation_loss', {'train': seg_loss}, step)
|
333 |
+
|
334 |
+
# log learning_rate
|
335 |
+
current_lr = optimizer.param_groups[0]['lr']
|
336 |
+
writer.add_scalar('learning_rate', current_lr, step)
|
337 |
+
|
338 |
+
step += 1
|
339 |
+
|
340 |
+
if step % opt.save_interval == 0 and step > 0:
|
341 |
+
save_checkpoint(model, opt.saved_path, f'hybridnets-d{opt.compound_coef}_{epoch}_{step}.pth')
|
342 |
+
print('checkpoint...')
|
343 |
+
|
344 |
+
except Exception as e:
|
345 |
+
print('[Error]', traceback.format_exc())
|
346 |
+
print(e)
|
347 |
+
continue
|
348 |
+
|
349 |
+
scheduler.step(np.mean(epoch_loss))
|
350 |
+
|
351 |
+
if epoch % opt.val_interval == 0:
|
352 |
+
best_fitness, best_loss, best_epoch = val(model, optimizer, val_generator, params, opt, writer, epoch,
|
353 |
+
step, best_fitness, best_loss, best_epoch)
|
354 |
+
except KeyboardInterrupt:
|
355 |
+
save_checkpoint(model, opt.saved_path, f'hybridnets-d{opt.compound_coef}_{epoch}_{step}.pth')
|
356 |
+
finally:
|
357 |
+
writer.close()
|
358 |
+
|
359 |
+
|
360 |
+
if __name__ == '__main__':
|
361 |
+
opt = get_args()
|
362 |
+
train(opt)
|