-
Notifications
You must be signed in to change notification settings - Fork 9
/
train.py
252 lines (213 loc) · 9.96 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
248
249
250
251
252
""" Defines the Trainer class which handles train/validation/validation_video
"""
import torch
import itertools
import numpy as np
import pdb
from utils import map as meanap
from utils.utils import dump_gpumem, AverageMeter, submission_file, Timer
def adjust_learning_rate(startlr, decay_rate, optimizer, epoch):
"""Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
lr = startlr * (0.1 ** (epoch // decay_rate))
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def parse(x):
return x[0], x[1], x[2] if len(x) > 2 else {'id': [1] * x[1].shape[0]}
def accuracy(output, target, topk=(1,)):
"""Computes the precision@k for the specified values of k"""
maxk = max(topk)
batch_size = target.size(0)
_, pred = output.topk(maxk, 1, True, True)
pred = pred.t()
correct = pred.eq(target.view(1, -1).expand_as(pred))
res = []
for k in topk:
correct_k = correct[:k].view(-1).float().sum(0, keepdim=True)
res.append(correct_k.mul_(100.0 / batch_size))
return res
def triplet_accuracy(output, target, weights=None):
"""
if target>0 then first output should be smaller than right output
optional weighted average
"""
if type(output) is not list:
output = [(x.data[0], y.data[0]) for x, y in zip(*output)]
correct = [x < y if t > 0 else y < x for (x, y), t in zip(output, target)]
if weights is None:
return np.mean(correct)
else:
weights = weights.numpy()
weights = weights / (1e-5 + np.sum(weights))
return np.sum(np.array(correct).astype(float) * weights)
def triplet_topk(output, target, weights, topk=5):
weights = np.array(weights)
n = weights.shape[0]
topkn = int(np.ceil(.01 * topk * n))
ind = np.argsort(weights)
ind = ind[-topkn:].tolist()
return triplet_accuracy([output[x] for x in ind], [target[x] for x in ind])
def triplet_allk(output, target, weights):
out = {}
for k in (1, 2, 5, 10, 50):
out['topk{}'.format(k)] = triplet_topk(output, target, weights, topk=k)
return out
def forward(inputs, target, model, criterion, ids, train=True):
target = target.float().cuda(async=True)
input_vars = [torch.autograd.Variable(inp.cuda(), volatile=not train)
for inp in inputs]
target_var = torch.autograd.Variable(target, volatile=not train)
output = model(*input_vars)
loss, weights = criterion(*(list(output) + [target_var, ids]))
return output[:2], loss, weights
class Trainer():
def train(self, loader, model, criterion, optimizer, epoch, args):
adjust_learning_rate(args.lr, args.lr_decay_rate, optimizer, epoch)
timer = Timer()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
wtop1 = AverageMeter()
metrics = {}
# switch to train mode
model.train()
optimizer.zero_grad()
def part(x):
return itertools.islice(x, int(len(x) * args.train_size))
for i, x in enumerate(part(loader)):
inputs, target, meta = parse(x)
data_time.update(timer.thetime() - timer.end)
output, loss, weights = forward(inputs, target, model, criterion, meta['id'])
prec1 = triplet_accuracy(output, target)
wprec1 = triplet_accuracy(output, target, weights)
losses.update(loss.data[0], inputs[0].size(0))
top1.update(prec1, inputs[0].size(0))
wtop1.update(wprec1, inputs[0].size(0))
loss.backward()
if i % args.accum_grad == args.accum_grad - 1:
print('updating parameters')
optimizer.step()
optimizer.zero_grad()
timer.tic()
if i % args.print_freq == 0:
print('[{name}] Epoch: [{0}][{1}/{2}({3})]\t'
'Time {timer.val:.3f} ({timer.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'WAcc@1 {wtop1.val:.3f} ({wtop1.avg:.3f})\t'.format(
epoch, i, int(len(loader) * args.train_size), len(loader), name=args.name,
timer=timer, data_time=data_time, loss=losses, top1=top1, wtop1=wtop1))
metrics.update({'top1': top1.avg, 'wtop1': wtop1.avg})
return metrics
def validate(self, loader, model, criterion, epoch, args):
timer = Timer()
losses = AverageMeter()
top1 = AverageMeter()
wtop1 = AverageMeter()
alloutputs = []
metrics = {}
# switch to evaluate mode
model.eval()
def part(x):
return itertools.islice(x, int(len(x) * args.val_size))
for i, x in enumerate(part(loader)):
inputs, target, meta = parse(x)
output, loss, weights = forward(inputs, target, model, criterion, meta['id'], train=False)
prec1 = triplet_accuracy(output, target)
wprec1 = triplet_accuracy(output, target, weights)
losses.update(loss.data[0], inputs[0].size(0))
top1.update(prec1, inputs[0].size(0))
wtop1.update(wprec1, inputs[0].size(0))
alloutputs.extend(zip([(x.data[0], y.data[0]) for x, y in zip(*output)], target, weights))
timer.tic()
if i % args.print_freq == 0:
print('[{name}] Test [{epoch}]: [{0}/{1} ({2})]\t'
'Time {timer.val:.3f} ({timer.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'WAcc@1 {wtop1.val:.3f} ({wtop1.avg:.3f})\t'.format(
i, int(len(loader) * args.val_size), len(loader), name=args.name,
timer=timer, loss=losses, top1=top1, epoch=epoch, wtop1=wtop1))
metrics.update(triplet_allk(*zip(*alloutputs)))
metrics.update({'top1val': top1.avg, 'wtop1val': wtop1.avg})
print(' * Acc@1 {top1val:.3f} \t WAcc@1 {wtop1val:.3f}'
'\n topk1: {topk1:.3f} \t topk2: {topk2:.3f} \t '
'topk5: {topk5:.3f} \t topk10: {topk10:.3f} \t topk50: {topk50:.3f}'
.format(**metrics))
return metrics
def validate_video(self, loader, model, epoch, args):
""" Run video-level validation on the Charades test set"""
timer = Timer()
outputs, gts, ids = [], [], []
metrics = {}
# switch to evaluate mode
model.eval()
for i, x in enumerate(loader):
inputs, target, meta = parse(x)
target = target.long().cuda(async=True)
assert target[0, :].eq(target[1, :]).all(), "val_video not synced"
input_vars = [torch.autograd.Variable(inp.cuda(), volatile=True)
for inp in inputs]
output = model(*input_vars)[-1] # classification should be last output
output = torch.nn.Softmax(dim=1)(output)
# store predictions
output_video = output.mean(dim=0)
outputs.append(output_video.data.cpu().numpy())
gts.append(target[0, :])
ids.append(meta['id'][0])
timer.tic()
if i % args.print_freq == 0:
print('Test2: [{0}/{1}]\t'
'Time {timer.val:.3f} ({timer.avg:.3f})'.format(
i, len(loader), timer=timer))
# mAP, _, ap = meanap.map(np.vstack(outputs), np.vstack(gts))
mAP, _, ap = meanap.charades_map(np.vstack(outputs), np.vstack(gts))
metrics['mAP'] = mAP
print(ap)
print(' * mAP {:.3f}'.format(mAP))
submission_file(
ids, outputs, '{}/epoch_{:03d}.txt'.format(args.cache, epoch + 1))
return metrics
def validate_egovideo(self, loader, model, epoch, args):
""" Run video-level validation on the Charades ego test set"""
timer = Timer()
outputs, gts, ids = [], [], []
outputsw = []
metrics = {}
# switch to evaluate mode
model.eval()
for i, x in enumerate(loader):
inp, target, meta = parse(x)
target = target.long().cuda(async=True)
assert target[0, :].eq(target[1, :]).all(), "val_video not synced"
input_var = torch.autograd.Variable(inp.cuda(), volatile=True)
output, w_x, w_z = model(input_var)
output = torch.nn.Softmax(dim=1)(output)
sw_x = torch.nn.Softmax(dim=0)(w_x) * w_x.shape[0]
sw_x = (sw_x - sw_x.mean()) / sw_x.std()
scale = torch.clamp(1 + (sw_x - 1) * 0.05, 0, 100)
print('scale min: {}\t max: {}\t std: {}'.format(scale.min().data[0], scale.max().data[0], scale.std().data[0]))
scale = torch.clamp(scale, 0, 100)
scale *= scale.shape[0] / scale.sum()
outputw = output * scale.unsqueeze(1)
# store predictions
output_video = output.mean(dim=0)
outputs.append(output_video.data.cpu().numpy())
outputsw.append(outputw.mean(dim=0).data.cpu().numpy())
gts.append(target[0, :])
ids.append(meta['id'][0])
timer.tic()
if i % args.print_freq == 0:
print('Test2: [{0}/{1}]\t'
'Time {timer.val:.3f} ({timer.avg:.3f})'.format(
i, len(loader), timer=timer))
# mAP, _, ap = meanap.map(np.vstack(outputs), np.vstack(gts))
mAP, _, ap = meanap.charades_nanmap(np.vstack(outputs), np.vstack(gts))
mAPw, _, _ = meanap.charades_nanmap(np.vstack(outputsw), np.vstack(gts))
metrics['mAPego'] = mAP
metrics['mAPegow'] = mAPw
print(ap)
print(' * mAPego {mAPego:.3f} \t mAPegow {mAPegow:.3f}'.format(**metrics))
submission_file(
ids, outputs, '{}/egoepoch_{:03d}.txt'.format(args.cache, epoch + 1))
return metrics