-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_avs.py
283 lines (239 loc) · 11.1 KB
/
train_avs.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
import argparse
import os
import time
import ipdb
import pdb
import yaml
import numpy as np
from tqdm import tqdm
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import CosineAnnealingLR
import warnings
warnings.simplefilter("ignore", UserWarning)
import datasets
import models
import utils
from statistics import mean
import torch
import torch.distributed as dist
torch.distributed.init_process_group(backend='nccl')
# torch.distributed.init_process_group(backend='nccl', init_method='env://', timeout=datetime.timedelta(seconds=5400))
local_rank = torch.distributed.get_rank()
torch.cuda.set_device(local_rank)
device = torch.device("cuda", local_rank)
from collections import defaultdict
from utility import mask_iou, Eval_Fmeasure, AverageMeter, MetricLogger
def make_data_loader(spec, tag='', args=None):
if args.subset == 'ms3':
if tag =='train':
from datasets.avsb_dataloader_vggish_ms3_train import S4Dataset
dataset = S4Dataset(split=tag, args=args)
else:
from datasets.avsb_dataloader_vggish_ms3_eval import S4Dataset
dataset = S4Dataset(split=tag, args=args)
elif args.subset == 's4':
from datasets.avsb_dataloader_vggish import S4Dataset
dataset = S4Dataset(split=tag, args=args)
elif args.subset == 'synthetic':
from datasets.AVSSynthetic_dataloader import SyntheticDataset
dataset = SyntheticDataset(split=tag, args=args)
else:
raise NotImplementedError("To be implemented")
if local_rank == 0:
log('{} dataset: size={}'.format(tag, len(dataset)))
sampler = torch.utils.data.distributed.DistributedSampler(dataset)
loader = DataLoader(dataset, batch_size=spec['batch_size'],
shuffle=False, num_workers=args.n_threads, pin_memory=True, sampler=sampler)
return loader
def make_data_loaders(args=None):
train_loader = make_data_loader(config.get('train_dataset'), tag='train', args=args)
val_loader = make_data_loader(config.get('val_dataset'), tag='val', args=args)
return train_loader, val_loader
@torch.no_grad()
def validate(loader, model):
model.eval()
device = model.device
avg_meter_miou = AverageMeter('miou')
avg_meter_F = AverageMeter('F_score')
metric_logger = MetricLogger(delimiter=" ")
header = 'Test:'
for batch in tqdm(metric_logger.log_every(loader, 10, header)):
img, spec, mask, category, video_name = batch
# img: Bx5xCxHxW, spec: Bx5x1xHxW, mask: BxTx1xHxW -> BxTxHxW
bs, T = img.size()[:2]
mask = mask.squeeze(dim=2).to(device)
bs, T, H, W = mask.size()
all_pred_masks = []
for idx in range(bs):
img_i = img[idx].to(device)
spec_i = spec[idx].to(device)
mask_i = mask[idx].to(device)
with torch.no_grad():
mask_pred = model.infer(img_i, spec_i)
all_pred_masks.append(mask_pred)
all_pred_masks = torch.stack(all_pred_masks, dim=0) # BxTxHxW
gt_masks = mask.reshape(bs*T, H, W)
pred_masks = all_pred_masks.reshape(bs*T, H, W)
miou = mask_iou(pred_masks, gt_masks)
avg_meter_miou.add({'miou': miou})
F_score = Eval_Fmeasure(pred_masks, gt_masks)
avg_meter_F.add({'F_score': F_score})
miou = (avg_meter_miou.pop('miou'))
F_score = (avg_meter_F.pop('F_score'))
eval_metrics = {'miou': miou.item(),
'F_score': F_score
}
return eval_metrics
def prepare_training():
if config.get('resume') is not None:
model = models.make(config['model']).cuda()
optimizer = utils.make_optimizer(
model.parameters(), config['optimizer'])
epoch_start = config.get('resume') + 1
else:
model = models.make(config['model']).cuda()
optimizer = utils.make_optimizer(
model.parameters(), config['optimizer'])
epoch_start = 1
max_epoch = config.get('epoch_max')
lr_scheduler = CosineAnnealingLR(optimizer, max_epoch, eta_min=config.get('lr_min'))
if local_rank == 0:
log('model: #params={}'.format(utils.compute_num_params(model, text=True)))
return model, optimizer, epoch_start, lr_scheduler
def train(train_loader, model):
model.train()
if local_rank == 0:
pbar = tqdm(total=len(train_loader), leave=False, desc='train')
else:
pbar = None
loss_list = []
for batch in train_loader:
img, spec, mask, category, video_name = batch
model.set_input(img, spec, mask)
model.optimize_parameters()
batch_loss = [torch.zeros_like(model.loss_G) for _ in range(dist.get_world_size())]
dist.all_gather(batch_loss, model.loss_G)
loss_list.extend(batch_loss)
if pbar is not None:
pbar.update(1)
if pbar is not None:
pbar.close()
loss = [i.item() for i in loss_list]
return mean(loss)
def main(config_, save_path, args):
global config, log, writer, log_info
config = config_
if args.local_rank == 0:
log, writer = utils.set_save_path(save_path, remove=False)
with open(os.path.join(save_path, 'config.yaml'), 'w') as f:
yaml.dump(config, f, sort_keys=False)
train_loader, val_loader = make_data_loaders(args)
if config.get('data_norm') is None:
config['data_norm'] = {
'inp': {'sub': [0], 'div': [1]},
'gt': {'sub': [0], 'div': [1]}
}
model, optimizer, epoch_start, lr_scheduler = prepare_training()
model.optimizer = optimizer
lr_scheduler = CosineAnnealingLR(model.optimizer, config['epoch_max'], eta_min=config.get('lr_min'))
model = model.cuda()
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True,
broadcast_buffers=False
)
model = model.module
sam_checkpoint = torch.load(config['sam_checkpoint'])
model.load_state_dict(sam_checkpoint, strict=False)
if os.path.isfile(args.pretrained_weights):
ckpt = torch.load(args.pretrained_weights, map_location=torch.device('cpu'))
model.load_state_dict(ckpt, strict=False)
print(f"Successfully load pretrained model weights from {args.pretrained_weights}!")
for name, para in model.named_parameters():
if "image_encoder" in name and "prompt_generator" not in name:
para.requires_grad_(False)
if local_rank == 0:
model_total_params = sum(p.numel() for p in model.parameters())
model_grad_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('model_grad_params:' + str(model_grad_params), '\nmodel_total_params:' + str(model_total_params))
epoch_max = config['epoch_max']
epoch_val = config.get('epoch_val')
max_val_v = 1e-8
timer = utils.Timer()
for epoch in range(epoch_start, epoch_max + 1):
train_loader.sampler.set_epoch(epoch)
t_epoch_start = timer.t()
train_loss_G = train(train_loader, model)
lr_scheduler.step()
if local_rank == 0:
log_info = ['epoch {}/{}'.format(epoch, epoch_max)]
writer.add_scalar('lr', optimizer.param_groups[0]['lr'], epoch)
log_info.append('train G: loss={:.4f}'.format(train_loss_G))
writer.add_scalars('loss', {'train G': train_loss_G}, epoch)
model_spec = config['model']
model_spec['sd'] = model.state_dict()
optimizer_spec = config['optimizer']
optimizer_spec['sd'] = optimizer.state_dict()
save(config, model, save_path, 'last')
if (epoch_val is not None) and (epoch % epoch_val == 0):
eval_results = validate(val_loader, model)
metric1 = 'miou'
result1 = eval_results[metric1]
metric2 = 'F_score'
result2 = eval_results[metric2]
if local_rank == 0:
log_info.append('val: {}={:.4f}'.format(metric1, result1))
writer.add_scalars(metric1, {'val': result1}, epoch)
log_info.append('val: {}={:.4f}'.format(metric2, result2))
writer.add_scalars(metric2, {'val': result2}, epoch)
if result1 > max_val_v:
max_val_v = result1
save(config, model, save_path, 'best')
t = timer.t()
prog = (epoch - epoch_start + 1) / (epoch_max - epoch_start + 1)
t_epoch = utils.time_text(t - t_epoch_start)
t_elapsed, t_all = utils.time_text(t), utils.time_text(t / prog)
log_info.append('{} {}/{}'.format(t_epoch, t_elapsed, t_all))
log(', '.join(log_info))
writer.flush()
def save(config, model, save_path, name):
if config['model']['name'] == 'segformer' or config['model']['name'] == 'setr':
if config['model']['args']['encoder_mode']['name'] == 'evp':
prompt_generator = model.encoder.backbone.prompt_generator.state_dict()
decode_head = model.encoder.decode_head.state_dict()
torch.save({"prompt": prompt_generator, "decode_head": decode_head},
os.path.join(save_path, f"prompt_epoch_{name}.pth"))
else:
torch.save(model.state_dict(), os.path.join(save_path, f"model_epoch_{name}.pth"))
else:
torch.save(model.state_dict(), os.path.join(save_path, f"model_epoch_{name}.pth"))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config', default="configs/sam_avs_adapter.yaml")
parser.add_argument('--name', type=str, default=None)
parser.add_argument('--tag', type=str, default=None)
parser.add_argument("--n_threads", type=int, default=8, help="")
parser.add_argument("--local_rank", type=int, default=-1, help="")
parser.add_argument("--inp_size", type=int, default=1024, help="")
parser.add_argument('--dir_prefix', type=str, default="/home/", help="")
parser.add_argument('--subset', type=str, default="s4", help="which subset of avsbench: s4 | ms3 | synthetic")
parser.add_argument('--pretrained_weights', type=str, default="", help="Load pretrained weights")
parser.add_argument('--trainset_shuffle', default=False, action='store_true', help=' ')
parser.add_argument("--trainset_ratio", type=float, default=1, help="Use the ratio of S4 subset")
parser.add_argument('--openset', default=False, action='store_true', help='Open set traing and evaluation of S4 subset ')
args = parser.parse_args()
with open(args.config, 'r') as f:
config = yaml.load(f, Loader=yaml.FullLoader)
if local_rank == 0:
print('config loaded.')
current_time = time.strftime("%Y_%m_%d_%H_%M_%S", time.localtime())
save_name = current_time + '_' + args.name
if save_name is None:
save_name = '_' + args.config.split('/')[-1][:-len('.yaml')]
if args.tag is not None:
save_name += '_' + args.tag
save_path = os.path.join('./ckpts', save_name)
args.config = config
main(config, save_path, args=args)