def __init__(self,
num_class,
num_segments,
modality,
base_model='resnet101',
new_length=None,
consensus_type='avg',
before_softmax=True,
dropout=0.8,
crop_num=1,
partial_bn=True):
super(TSN, self).__init__()
self.modality = modality
self.num_segments = num_segments
self.reshape = True
self.before_softmax = before_softmax
self.dropout = dropout
self.crop_num = crop_num
self.consensus_type = consensus_type
if not before_softmax and consensus_type != 'avg':
raise ValueError("Only avg consensus can be used after Softmax")
if new_length is None:
self.new_length = 1 if modality == "RGB" else 5
else:
self.new_length = new_length
log(("""
Initializing TSN with base model: {}.
TSN Configurations:
input_modality: {}
num_segments: {}
new_length: {}
consensus_module: {}
dropout_ratio: {}
""".format(base_model, self.modality, self.num_segments,
self.new_length, consensus_type, self.dropout)))
self._prepare_base_model(base_model)
feature_dim = self._prepare_tsn(num_class)
if self.modality == 'Flow':
log("Converting the ImageNet model to a flow init model")
self.base_model = self._construct_flow_model(self.base_model)
log("Done. Flow model ready...")
elif self.modality == 'RGBDiff':
log("Converting the ImageNet model to RGB+Diff init model")
self.base_model = self._construct_diff_model(self.base_model)
log("Done. RGBDiff model ready.")
self.consensus = ConsensusModule(consensus_type)
if not self.before_softmax:
self.softmax = nn.Softmax()
self._enable_pbn = partial_bn
if partial_bn:
self.partialBN(True)
def validate(val_loader, model, criterion, iter, logger=None):
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to evaluate mode
model.eval()
end = time.time()
for i, (input, target) in enumerate(val_loader):
target = target.cuda(async=True)
input_var = torch.autograd.Variable(input, volatile=True)
target_var = torch.autograd.Variable(target, volatile=True)
# compute output
output = model(input_var)
loss = criterion(output, target_var)
# measure accuracy and record loss
prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
losses.update(loss.data[0], input.size(0))
top1.update(prec1[0], input.size(0))
top5.update(prec5[0], input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % args.log_freq == 0:
log(('Test: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
i,
len(val_loader),
batch_time=batch_time,
loss=losses,
top1=top1,
top5=top5)))
log((
'Testing Results: Prec@1 {top1.avg:.3f} Prec@5 {top5.avg:.3f} Loss {loss.avg:.5f}'
.format(top1=top1, top5=top5, loss=losses)))
return top1.avg
def train(self, mode=True):
"""
Override the default train() to freeze the BN parameters 通过重写train方法实现PBN
:return:
"""
super(TSN, self).train(mode)
count = 0
if self._enable_pbn:
log("Freezing BatchNorm2D except the first one.")
for m in self.base_model.modules():
if isinstance(m, nn.BatchNorm2d):
count += 1
if count >= (2 if self._enable_pbn else 1):
m.eval()
# shutdown update in frozen mode
m.weight.requires_grad = False
m.bias.requires_grad = False
def _load_image(self, directory, record, idx):
if self.modality == 'RGB' or self.modality == 'RGBDiff':
directory += record.path
return [
Image.open(os.path.join(
directory, self.image_tmpl.format(idx))).convert('RGB')
]
elif self.modality == 'Flow':
log('directory:', directory)
log('idx:', idx)
#每一次返回两张光流图分为x,y方向各一张。
x_img = Image.open(
os.path.join(directory + 'u/' + record.path,
self.image_tmpl.format(idx))).convert('L')
y_img = Image.open(
os.path.join(directory + 'v/' + record.path,
self.image_tmpl.format(idx))).convert('L')
return [x_img, y_img]
def blank_loglikes(n):
a = ones((NUM_CHARS, n)) * 0.1
a[0, :] = 0.9
a /= sqrt(square(a).sum(axis=0))
return log(a)
def uniform_loglikes(n):
return log(ones((NUM_CHARS, n)) / float(NUM_CHARS))
def cost_and_grad(self, data, labels, back=True, prev_h0=None):
hps = self.hps
T = data.shape[1]
bsize = data.shape[2]
# FIXME gnumpy reallocates if try and use same parameters?
#us = self.us[:, 0:T, 0:bsize]
#dus = self.dus[:, 0:T, 0:bsize]
#hs = self.hs[:, 0:T, 0:bsize]
#dhs = self.dhs[:, 0:T, 0:bsize]
#probs = self.probs[:, 0:T, 0:bsize]
#dprobs = self.dprobs[:, 0:T, 0:bsize]
#costs = self.costs[0:T, 0:bsize]
us = list()
dus = list()
hs = list()
dhs = list()
h0 = list()
for k in xrange(hps.hidden_layers):
us.append(list())
dus.append(list())
hs.append(list())
dhs.append(list())
h0.append(empty((hps.hidden_size, bsize)))
for t in xrange(T):
us[k].append(zeros((hps.hidden_size, bsize)))
dus[k].append(zeros((hps.hidden_size, bsize)))
hs[k].append(zeros((hps.hidden_size, bsize)))
dhs[k].append(zeros((hps.hidden_size, bsize)))
probs = list()
for t in xrange(T):
probs.append(zeros((hps.output_size, bsize)))
costs = np.zeros((T, bsize))
if prev_h0 is not None:
h0 = prev_h0
else:
for k in xrange(hps.hidden_layers):
h0[k] = tile(self.params['h0'][:, k].reshape(-1, 1), bsize)
bih = self.params['bih']
Wih = self.params['Wih']
Whh = self.params['Whh']
bhh = self.params['bhh']
Who = self.params['Who']
bho = self.params['bho']
# Forward prop
for t in xrange(T):
for k in xrange(hps.hidden_layers):
if t == 0:
hprev = h0[k]
else:
hprev = hs[k][t-1]
if k == 0:
us[k][t] = mult(Wih, data[:, t, :]) + bih
else:
us[k][t] = mult(self.params['Wh%d' % k], hs[k-1][t])
if k == hps.recurrent_layer - 1:
us[k][t] += mult(Whh, hprev) + bhh
# Clip maximum activation
mask = us[k][t] < hps.max_act
us[k][t] = us[k][t] * mask + hps.max_act * (1 - mask)
elif k != 0:
us[k][t] += self.params['bh%d' % k]
hs[k][t] = self.nl(us[k][t])
probs[t] = softmax(mult(Who, hs[-1][t]) + bho)
self.last_h = list()
for k in xrange(hps.hidden_layers):
self.last_h.append(hs[k][-1])
if labels is None:
return None, probs
probs_neg_log = list()
dprobs = list()
for t in xrange(T):
probs_neg_log.append(as_np(-1 * log(probs[t])))
dprobs.append(as_np(probs[t].copy()))
for k in xrange(bsize):
for t in xrange(len(labels[k])):
costs[t, k] = probs_neg_log[t][labels[k][t], k]
dprobs[t][labels[k][t], k] -= 1
for t in xrange(T):
dprobs[t] = array(dprobs[t])
# NOTE Summing costs over time
# NOTE FIXME Dividing by T to get better sense if objective
# is decreasing, remove for grad checking
cost = costs.sum() / bsize / float(T)
if not back:
return cost, probs
# Backprop
for k in self.grads:
self.grads[k][:] = 0
for t in reversed(xrange(T)):
self.grads['bho'] += dprobs[t][:, :].sum(axis=-1).reshape((-1, 1)) / bsize
self.grads['Who'] += mult(dprobs[t], hs[-1][t].T) / bsize
for k in reversed(xrange(hps.hidden_layers)):
if k == hps.hidden_layers - 1:
dhs[k][t] += mult(Who.T, dprobs[t])
else:
dhs[k][t] += mult(self.params['Wh%d' % (k+1)].T, dhs[k+1][t])
dus[k][t] += get_nl_grad(self.hps.nl, us[k][t]) * dhs[k][t]
if k > 0:
self.grads['Wh%d' % k] += mult(dus[k][t], hs[k-1][t].T) / bsize
self.grads['bh%d' % k] += dus[k][t].sum(axis=-1).reshape((-1, 1)) / bsize
if k == hps.recurrent_layer - 1:
if t == 0:
hprev = h0[k]
self.grads['h0'][:, k] = mult(Whh.T, dus[k][t]).sum(axis=-1) / bsize
else:
hprev = hs[k][t-1]
dhs[k][t-1] = mult(Whh.T, dus[k][t])
self.grads['Whh'] += mult(dus[k][t], hprev.T) / bsize
self.grads['bhh'] += dus[k][t].sum(axis=-1).reshape((-1, 1)) / bsize
self.grads['Wih'] += mult(dus[0][t], data[:, t, :].T) / bsize
self.grads['bih'] += dus[0][t].sum(axis=-1).reshape((-1, 1)) / bsize
return cost, self.grads
def cost_and_grad(self, data, labels, back=True, prev_h0=None):
hps = self.hps
T = data.shape[1]
bsize = data.shape[2]
# FIXME gnumpy reallocates if try and use same parameters?
#us = self.us[:, 0:T, 0:bsize]
#dus = self.dus[:, 0:T, 0:bsize]
#hs = self.hs[:, 0:T, 0:bsize]
#dhs = self.dhs[:, 0:T, 0:bsize]
#probs = self.probs[:, 0:T, 0:bsize]
#dprobs = self.dprobs[:, 0:T, 0:bsize]
#costs = self.costs[0:T, 0:bsize]
us = list()
dus = list()
hs = list()
dhs = list()
h0 = list()
for k in xrange(hps.hidden_layers):
us.append(list())
dus.append(list())
hs.append(list())
dhs.append(list())
h0.append(empty((hps.hidden_size, bsize)))
for t in xrange(T):
us[k].append(zeros((hps.hidden_size, bsize)))
dus[k].append(zeros((hps.hidden_size, bsize)))
hs[k].append(zeros((hps.hidden_size, bsize)))
dhs[k].append(zeros((hps.hidden_size, bsize)))
probs = list()
for t in xrange(T):
probs.append(zeros((hps.output_size, bsize)))
costs = np.zeros((T, bsize))
if prev_h0 is not None:
h0 = prev_h0
else:
for k in xrange(hps.hidden_layers):
h0[k] = tile(self.params['h0'][:, k].reshape(-1, 1), bsize)
bih = self.params['bih']
Wih = self.params['Wih']
Whh = self.params['Whh']
bhh = self.params['bhh']
Who = self.params['Who']
bho = self.params['bho']
# Forward prop
for t in xrange(T):
for k in xrange(hps.hidden_layers):
if t == 0:
hprev = h0[k]
else:
hprev = hs[k][t - 1]
if k == 0:
us[k][t] = mult(Wih, data[:, t, :]) + bih
else:
us[k][t] = mult(self.params['Wh%d' % k], hs[k - 1][t])
if k == hps.recurrent_layer - 1:
us[k][t] += mult(Whh, hprev) + bhh
# Clip maximum activation
mask = us[k][t] < hps.max_act
us[k][t] = us[k][t] * mask + hps.max_act * (1 - mask)
elif k != 0:
us[k][t] += self.params['bh%d' % k]
hs[k][t] = self.nl(us[k][t])
probs[t] = softmax(mult(Who, hs[-1][t]) + bho)
self.last_h = list()
for k in xrange(hps.hidden_layers):
self.last_h.append(hs[k][-1])
if labels is None:
return None, probs
probs_neg_log = list()
dprobs = list()
for t in xrange(T):
probs_neg_log.append(as_np(-1 * log(probs[t])))
dprobs.append(as_np(probs[t].copy()))
for k in xrange(bsize):
for t in xrange(len(labels[k])):
costs[t, k] = probs_neg_log[t][labels[k][t], k]
dprobs[t][labels[k][t], k] -= 1
for t in xrange(T):
dprobs[t] = array(dprobs[t])
# NOTE Summing costs over time
# NOTE FIXME Dividing by T to get better sense if objective
# is decreasing, remove for grad checking
cost = costs.sum() / bsize / float(T)
if not back:
return cost, probs
# Backprop
for k in self.grads:
self.grads[k][:] = 0
for t in reversed(xrange(T)):
self.grads['bho'] += dprobs[t][:, :].sum(axis=-1).reshape(
(-1, 1)) / bsize
self.grads['Who'] += mult(dprobs[t], hs[-1][t].T) / bsize
for k in reversed(xrange(hps.hidden_layers)):
if k == hps.hidden_layers - 1:
dhs[k][t] += mult(Who.T, dprobs[t])
else:
dhs[k][t] += mult(self.params['Wh%d' % (k + 1)].T,
dhs[k + 1][t])
dus[k][t] += get_nl_grad(self.hps.nl, us[k][t]) * dhs[k][t]
if k > 0:
self.grads['Wh%d' %
k] += mult(dus[k][t], hs[k - 1][t].T) / bsize
self.grads['bh%d' % k] += dus[k][t].sum(axis=-1).reshape(
(-1, 1)) / bsize
if k == hps.recurrent_layer - 1:
if t == 0:
hprev = h0[k]
self.grads['h0'][:, k] = mult(
Whh.T, dus[k][t]).sum(axis=-1) / bsize
else:
hprev = hs[k][t - 1]
dhs[k][t - 1] = mult(Whh.T, dus[k][t])
self.grads['Whh'] += mult(dus[k][t], hprev.T) / bsize
self.grads['bhh'] += dus[k][t].sum(axis=-1).reshape(
(-1, 1)) / bsize
self.grads['Wih'] += mult(dus[0][t], data[:, t, :].T) / bsize
self.grads['bih'] += dus[0][t].sum(axis=-1).reshape(
(-1, 1)) / bsize
return cost, self.grads
elif args.dataset == 'hmdb51':
num_class = 51
elif args.dataset == 'kinetics':
num_class = 400
else:
raise ValueError('Unknown dataset ' + args.dataset)
net = TSN(num_class,
1,
args.modality,
base_model=args.arch,
consensus_type=args.crop_fusion_type,
dropout=args.dropout)
checkpoint = torch.load(args.weights)
log("model epoch {} best prec@1: {}".format(checkpoint['epoch'],
checkpoint['best_prec1']))
base_dict = {
'.'.join(k.split('.')[1:]): v
for k, v in list(checkpoint['state_dict'].items())
}
net.load_state_dict(base_dict)
if args.test_crops == 1:
cropping = torchvision.transforms.Compose([
GroupScale(net.scale_size),
GroupCenterCrop(net.input_size),
])
elif args.test_crops == 10:
cropping = torchvision.transforms.Compose(
[GroupOverSample(net.input_size, net.scale_size)])
def blank_loglikes(n):
a = ones((NUM_CHARS, n)) * 0.1
a[0, :] = 0.9
a /= sqrt(square(a).sum(axis=0))
return log(a)
def uniform_loglikes(n):
return log(ones((NUM_CHARS, n)) / float(NUM_CHARS))
def main():
global args, best_prec1
args = parser.parse_args()
if args.dataset == 'ucf101':
num_class = 101
elif args.dataset == 'hmdb51':
num_class = 51
elif args.dataset == 'kinetics':
num_class = 400
else:
raise ValueError('Unknown dataset ' + args.dataset)
model = TSN(num_class,
args.num_segments,
args.modality,
base_model=args.arch,
consensus_type=args.consensus_type,
dropout=args.dropout,
partial_bn=not args.no_partialbn) #只返回网络结构
# state_dict = model_zoo.load_url(model_urls[args.arch],model_dir='./')
# pretrained_dict = {}
# for key,value in state_dict.items():
# pretrained_dict['base_model.'+key] = value
#
# model_dict = model.state_dict()
#
# #filter weight
# pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
# # log(pretrained_dict.keys())
#
# model_dict.update(pretrained_dict)
#
# model.load_state_dict(model_dict) #模型加载预训练权重
crop_size = model.crop_size
scale_size = model.scale_size
input_mean = model.input_mean
input_std = model.input_std
policies = model.get_optim_policies()
train_augmentation = model.get_augmentation()
model = torch.nn.DataParallel(model,
device_ids=args.gpus).cuda() #实现单机多卡的分布式训练。
if args.resume:
if os.path.isfile(args.resume):
log(("=> loading checkpoint '{}'".format(args.resume)))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
log(("=> loaded checkpoint '{}' (epoch {})".format(
args.evaluate, checkpoint['epoch'])))
else:
log(("=> no checkpoint found at '{}'".format(args.resume)))
cudnn.benchmark = True
# Data loading code
if args.modality != 'RGBDiff':
normalize = GroupNormalize(input_mean, input_std)
else:
normalize = IdentityTransform()
if args.modality == 'RGB':
data_length = 1 #每一个segment采集的帧的数量。
elif args.modality in ['Flow', 'RGBDiff']:
data_length = 5
train_loader = torch.utils.data.DataLoader(
TSNDataSet(
root_path="./jpegs_256/"
if args.modality == 'RGB' else '../tvl1_flow/',
list_file=args.train_list,
num_segments=args.num_segments, #默认为3
new_length=data_length,
modality=args.modality,
# image_tmpl="frame{:06d}.jpg" if args.modality in ["RGB", "RGBDiff"] else args.flow_prefix+"{}_{:05d}.jpg",
image_tmpl="frame{:06}.jpg",
transform=torchvision.transforms.Compose([
train_augmentation,
Stack(roll=args.arch == 'BNInception'),
ToTorchFormatTensor(div=args.arch != 'BNInception'),
normalize,
])),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(
TSNDataSet(
root_path="./jpegs_256/"
if args.modality == 'RGB' else '../tvl1_flow/',
list_file=args.val_list,
num_segments=args.num_segments,
new_length=data_length,
modality=args.modality,
# image_tmpl="frame{:06d}.jpg" if args.modality in ["RGB", "RGBDiff"] else args.flow_prefix+"{}_{:05d}.jpg",
image_tmpl="frame{:06}.jpg",
random_shift=False,
transform=torchvision.transforms.Compose([
GroupScale(int(scale_size)),
GroupCenterCrop(crop_size),
Stack(roll=args.arch == 'BNInception'),
ToTorchFormatTensor(div=args.arch != 'BNInception'),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
# define loss function (criterion) and optimizer
if args.loss_type == 'nll':
criterion = torch.nn.CrossEntropyLoss().cuda()
else:
raise ValueError("Unknown loss type")
for group in policies:
log(('group: {} has {} params, lr_mult: {}, decay_mult: {}'.format(
group['name'], len(group['params']), group['lr_mult'],
group['decay_mult'])))
optimizer = torch.optim.SGD(policies,
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.evaluate:
validate(val_loader, model, criterion, 0)
return
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch, args.lr_steps)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# break
# evaluate on validation set
if (epoch + 1) % args.eval_freq == 0 or epoch == args.epochs - 1:
prec1 = validate(val_loader, model, criterion,
(epoch + 1) * len(train_loader))
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
}, is_best)
def train(train_loader, model, criterion, optimizer, epoch):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
if args.no_partialbn:
model.module.partialBN(False)
else:
model.module.partialBN(True)
# switch to train mode
model.train() #转换到train模式，此模式也可以自己重写自己定义哪些层需要训练。
end = time.time()
for i, (input, target) in enumerate(train_loader):
# measure data loading time
#RGB 模态
# input 每一个sample应该是三张图片。 Input.size() ==> [16,9,224,224]具体请看model.py文件
# taeget.size() == > [16]
#FLOW 模态
# Input 每一个sample应该是三个堆叠的光流图堆，Input.size() ==> [16,30,224,224]
# taeget.size() ==> 16 batch_size=16，一个batch0总共有16段视频。故总共16个target.
# log('input:',input.size())
# log('target:',target)
data_time.update(time.time() - end)
target = target.cuda(async=True)
input_var = torch.autograd.Variable(input)
target_var = torch.autograd.Variable(target)
# compute output
output = model(input_var)
# break
loss = criterion(output, target_var)
# measure accuracy and record loss
prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
losses.update(loss.data[0], input.size(0))
top1.update(prec1[0], input.size(0))
top5.update(prec5[0], input.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
if args.clip_gradient is not None:
total_norm = clip_grad_norm(model.parameters(), args.clip_gradient)
if total_norm > args.clip_gradient:
log("clipping gradient: {} with coef {}".format(
total_norm, args.clip_gradient / total_norm))
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % args.log_freq == 0:
log(('Epoch: [{0}][{1}/{2}], lr: {lr:.5f}\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
epoch,
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
top1=top1,
top5=top5,
lr=optimizer.param_groups[-1]['lr'])))
