-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
247 lines (226 loc) · 9.76 KB
/
train.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
import os
import random
from collections import defaultdict
import argparse
import sys
sys.path.insert(0, 'lib/mmdetection')
import pandas as pd
import numpy as np
import mmcv
from mmcv.runner import load_checkpoint
from mmcv.utils import Config
import torch
import torch.distributed as dist
from torch.utils.data import DataLoader
from sklearn.model_selection import KFold, StratifiedKFold, train_test_split
from sklearn.metrics import f1_score, confusion_matrix
from mmcls.models import build_classifier
from mmcls.datasets import build_dataset
from lib.datasets import (ImageSequenceDataset, ClassBalancedSubsetSampler,
DistributedClassBalancedSubsetSampler,
DistributedSubsetSampler, CombinedSampler,
DistributedReversedSubsetSampler, DASampler)
from lib.models import Classifier, BBNLoss
from lib.utils.dist_utils import collect_results_cpu
def eval(model, dataloader, **kwargs):
rank = dist.get_rank()
self = model
self.eval()
all_preds = np.empty((0,))
all_labels = np.empty((0,))
if rank == 0:
progress_bar = mmcv.ProgressBar(len(dataloader))
for data in dataloader:
imgs = data.pop('imgs')
labels = data['labels']
labels = labels.to(next(self.parameters()).device)
with torch.no_grad():
preds = self(imgs, **data)
all_labels = np.hstack([all_labels, labels.cpu().numpy()])
all_preds = np.hstack([all_preds,
preds.argmax(1).detach().cpu().numpy()])
if rank == 0:
progress_bar.update()
all_preds = collect_results_cpu(
all_preds, len(val_loader.dataset))
all_labels = collect_results_cpu(
all_labels, len(val_loader.dataset))
if rank == 0:
cm = confusion_matrix(all_labels, all_preds)
f1_scores = f1_score(all_labels, all_preds, average=None)
class_weights = kwargs.get('class_weights', [0.25] * 4)
f1 = (np.array(class_weights) * f1_scores).sum()
return dict(f1_scores=f1_scores,
f1=f1,
cm=cm)
else:
return None
def train(self, dataloader, **kwargs):
rank = dist.get_rank()
world_size = dist.get_world_size()
optimizer = torch.optim.SGD(
list(self.parameters()), lr=kwargs.get('lr', 1e-3),
momentum=kwargs.get('momentum', 0.9),
weight_decay=kwargs.get('weight_decay', 0.0005))
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer, kwargs.get('milestones', [3, ]),
kwargs.get('gamma', 0.1))
#lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
# optimizer, kwargs.get('max_epoch', 5) // 3)
max_epoch = kwargs.get('max_epoch', 5)
save_dir = kwargs.get('save_dir', 'checkpoints')
if not os.path.exists(save_dir):
os.makedirs(save_dir)
logs = defaultdict(list)
for epoch in range(max_epoch):
cur_lr = optimizer.param_groups[0]['lr']
self.train()
#all_preds = np.empty((0,))
#all_labels = np.empty((0,))
dataloader.sampler.set_epoch(epoch)
for i, data in enumerate(dataloader):
imgs = data.pop('imgs')
labels = data['labels']
labels = labels.to(next(self.parameters()).device)
losses = self(imgs, **data)
loss, log_vars = self.module._parse_losses(losses)
'''
#acc = (preds.argmax(1) == labels).sum().item() / len(labels)
all_labels = np.hstack([all_labels, labels.cpu().numpy()])
all_preds = np.hstack([all_preds,
preds.argmax(1).detach().cpu().numpy()])
'''
if rank == 0 and i % kwargs.get('log_iters', 5) == 0:
log_str = f' '.join(f'{k}: {v:.4f}'
for k, v in log_vars.items())
print(f'Epoch {epoch+1}/{max_epoch} '
f'Iter: {i+1}/{len(dataloader)} lr: {cur_lr} {log_str}')
optimizer.zero_grad()
loss.backward()
optimizer.step()
'''
all_preds = collect_results_cpu(all_preds, 10000) # collect all
all_labels = collect_results_cpu(all_labels, 10000)
if rank == 0:
f1_scores = f1_score(all_labels, all_preds, average=None)
class_weights = kwargs.get('class_weights', [0.25] * 4)
f1 = (np.array(class_weights) * f1_scores).sum()
for cat_id in range(len(f1_scores)):
logs['score_train_%d'%cat_id] += [f1_scores[cat_id]]
logs['score_train'] += [f1]
'''
if kwargs.get('val_dataloader'):
result = eval(self, kwargs.get('val_dataloader'), **kwargs)
if rank == 0:
logs['score'] += [result['f1']]
logs['cm'] = result['cm']
f1_scores = result['f1_scores']
for cat_id in range(len(f1_scores)):
logs['score_%d'%cat_id] += [f1_scores[cat_id]]
torch.save(self.module.state_dict(), os.path.join(
save_dir, f'classifier_epoch{epoch+1}.pth'))
if kwargs.get('test_dataloader'):
result = eval(self, kwargs.get('test_dataloader'), **kwargs)
if rank == 0:
logs['score_test'] += [result['f1']]
f1_scores = result['f1_scores']
for cat_id in range(len(f1_scores)):
logs['score_%d_test'%cat_id] += [f1_scores[cat_id]]
lr_scheduler.step()
if rank == 0:
print(f'\nEpoch {epoch+1}/{max_epoch} lr: {cur_lr}')
print('confusion matrix:')
print(logs.pop('cm'))
print(pd.DataFrame(logs).transpose())
if rank == 0:
best_epoch = np.argmax(logs['score'])
outputs = {key: value[best_epoch] for key, value in logs.items()}
outputs['model_name'] = f'classifier_epoch{best_epoch+1}.pth'
return outputs
else:
return None
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config', type=str, default='')
parser.add_argument('--img_root', type=str, default='')
parser.add_argument('--ann_file', type=str, default='')
parser.add_argument('--key_frame_only', action='store_true',
default=False)
parser.add_argument('--samples_per_gpu', type=int, default=32)
parser.add_argument('--local_rank', type=int, default=0)
parser.add_argument('--max_epoch', type=int, default=10)
parser.add_argument('--lr', type=float, default=0.001)
parser.add_argument('--milestones', nargs='+', type=int,
default=[8, ])
args = parser.parse_args()
cfg = Config.fromfile(args.config)
if args.local_rank == 0:
print(cfg.model)
torch.distributed.init_process_group('nccl')
torch.cuda.set_device(args.local_rank)
training_set = build_dataset(cfg.data.train)
val_set = build_dataset(cfg.data.val)
outputs = []
if cfg.get('domain_adaption', False):
indices = np.arange(len(training_set.datasets[0]))
else:
indices = np.arange(len(training_set))
labels = np.array([training_set.get_cat_ids(_) for _ in indices])
# split train/test
train_indices, test_indices = next(StratifiedKFold(
6, shuffle=True, random_state=666).split(indices, labels))
k = 5
# cross validation
kf = StratifiedKFold(k, shuffle=True, random_state=666)
for idx, (train_inds, val_inds) in enumerate(
kf.split(train_indices, labels[train_indices])
):
train_inds = train_indices[train_inds] # map to sample index
val_inds = train_indices[val_inds] # map to sample index
torch.manual_seed(666)
torch.cuda.manual_seed_all(666)
np.random.seed(666)
random.seed(666)
bs = args.samples_per_gpu
samplers = [
DistributedSubsetSampler(train_inds),
DistributedReversedSubsetSampler(training_set, train_inds),
DistributedClassBalancedSubsetSampler(training_set, train_inds),
#DistributedSubsetSampler(
# np.arange(len(training_set.datasets[1]))
# + len(training_set.datasets[0])),
]
# specify sampler here to use different long-tail distribution handling
train_loader = DataLoader(training_set, batch_size=bs, num_workers=4,
sampler=samplers[2])
#DASampler(samplers[2:4]))
# samplers[2])
#CombinedSampler(samplers[:2])) # for BBN
val_loader = DataLoader(val_set, batch_size=bs, num_workers=4,
sampler=DistributedSubsetSampler(val_inds, shuffle=False))
test_loader = DataLoader(val_set, batch_size=bs, num_workers=4,
sampler=DistributedSubsetSampler(test_indices, shuffle=False))
model = build_classifier(cfg.model).to(args.local_rank)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], find_unused_parameters=True)
if cfg.load_from:
load_checkpoint(model, cfg.load_from, map_location='cpu')
output = train(model, train_loader,
val_dataloader=val_loader,
test_dataloader=test_loader,
log_iters=10,
max_epoch=args.max_epoch,
milestones=args.milestones,
lr=args.lr,
class_weights=[0.1, 0.2, 0.3, 0.4],
save_dir=f'work_dirs/classification/fold{idx + 1}')
outputs.append(output)
if dist.get_rank() == 0:
print(f'{k}-fold cross validation: {idx + 1}.')
#print output of the last fold
print(pd.DataFrame(outputs)[-1:].transpose())
if dist.get_rank() == 0:
print(pd.DataFrame(outputs)
.describe()
.transpose()
[['mean','std','min','max']])