def __init__(self, cfg: DictConfig):
super().__init__()
self.train_gulp_dir = Path(cfg.data.train_gulp_dir)
self.val_gulp_dir = Path(cfg.data.val_gulp_dir)
self.test_gulp_dir = Path(cfg.data.test_gulp_dir)
self.cfg = cfg
channel_count = (3 if self.cfg.modality == "RGB" else 2 *
self.cfg.data.segment_length)
common_transform = Compose([
Stack(bgr=self.cfg.modality == "RGB"
and self.cfg.data.preprocessing.get("bgr", False)),
ToTorchFormatTensor(div=self.cfg.data.preprocessing.rescale),
GroupNormalize(
mean=list(self.cfg.data.preprocessing.mean),
std=list(self.cfg.data.preprocessing.std),
),
ExtractTimeFromChannel(channel_count),
])
self.train_transform = Compose([
GroupMultiScaleCrop(
self.cfg.data.preprocessing.input_size,
self.cfg.data.train_augmentation.multiscale_crop_scales,
),
GroupRandomHorizontalFlip(is_flow=self.cfg.modality == "Flow"),
common_transform,
])
self.test_transform = Compose([
GroupScale(self.cfg.data.test_augmentation.rescale_size),
GroupCenterCrop(self.cfg.data.preprocessing.input_size),
common_transform,
])
def main():
global args
global best_prec1
args = parser.parse_args()
print('Training arguments:')
for k, v in vars(args).items():
print('\t{}: {}'.format(k, v))
if args.data_name == 'ucf101':
num_class = 101
elif args.data_name == 'hmdb51':
num_class = 51
else:
raise ValueError('Unknown dataset ' + args.data_name)
model = Model(num_class,
args.num_segments,
args.representation,
base_model=args.arch)
print(model)
train_loader = torch.utils.data.DataLoader(CoviarDataSet(
args.data_root,
args.data_name,
video_list=args.train_list,
num_segments=args.num_segments,
representation=args.representation,
transform=model.get_augmentation(),
is_train=True,
accumulate=(not args.no_accumulation),
),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CoviarDataSet(
args.data_root,
args.data_name,
video_list=args.test_list,
num_segments=args.num_segments,
representation=args.representation,
transform=torchvision.transforms.Compose([
GroupScale(int(model.scale_size)),
GroupCenterCrop(model.crop_size),
]),
is_train=False,
accumulate=(not args.no_accumulation),
),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
model = torch.nn.DataParallel(model, device_ids=args.gpus).cuda()
cudnn.benchmark = True
params_dict = dict(model.named_parameters())
params = []
for key, value in params_dict.items():
decay_mult = 0.0 if 'bias' in key else 1.0
if ('module.base_model.conv1' in key or 'module.base_model.bn1' in key
or 'data_bn'
in key) and args.representation in ['mv', 'residual']:
lr_mult = 0.1
elif '.fc.' in key:
lr_mult = 1.0
else:
lr_mult = 0.01
params += [{
'params': value,
'lr': args.lr,
'lr_mult': lr_mult,
'decay_mult': decay_mult
}]
optimizer = torch.optim.Adam(params,
weight_decay=args.weight_decay,
eps=0.001)
criterion = torch.nn.CrossEntropyLoss().cuda()
for epoch in range(args.epochs):
cur_lr = adjust_learning_rate(optimizer, epoch, args.lr_steps,
args.lr_decay)
train(train_loader, model, criterion, optimizer, epoch, cur_lr)
if epoch % args.eval_freq == 0 or epoch == args.epochs - 1:
prec1 = validate(val_loader, model, criterion)
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if is_best or epoch % SAVE_FREQ == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
},
is_best,
filename='checkpoint.pth.tar')
def main(conf, test_set, test_part=-1):
gulp_path = os.path.join(conf.gulp_test_dir, conf.modality.lower(), 'test',
test_set)
gulp_path = os.path.realpath(gulp_path)
gulp_path = Path(gulp_path)
classes_map = pickle.load(open(conf.classes_map, "rb"))
conf.num_classes = count_num_classes(classes_map)
net = TSN(conf.num_classes,
1,
conf.modality,
base_model=conf.arch,
consensus_type=conf.crop_fusion_type,
dropout=conf.dropout)
checkpoint = torch.load(conf.weights)
print("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 conf.test_crops == 1:
cropping = torchvision.transforms.Compose([
GroupScale(net.scale_size),
GroupCenterCrop(net.input_size),
])
elif conf.test_crops == 10:
cropping = torchvision.transforms.Compose(
[GroupOverSample(net.input_size, net.scale_size)])
else:
raise ValueError(
"Only 1 and 10 crops are supported while we got {}".format(
conf.test_crops))
class_type = 'verb+noun' if conf.class_type == 'action' else conf.class_type
if conf.modality == 'Flow':
dataset = EpicVideoFlowDataset(gulp_path=gulp_path,
class_type=class_type)
else:
dataset = EpicVideoDataset(gulp_path=gulp_path, class_type=class_type)
data_loader = torch.utils.data.DataLoader(EpicTSNTestDataset(
dataset,
classes_map,
num_segments=conf.test_segments,
new_length=1 if conf.modality == "RGB" else 5,
modality=conf.modality,
transform=torchvision.transforms.Compose([
cropping,
Stack(roll=conf.arch == 'BNInception'),
ToTorchFormatTensor(div=conf.arch != 'BNInception'),
GroupNormalize(net.input_mean, net.input_std),
]),
part=test_part),
batch_size=1,
shuffle=False,
num_workers=conf.workers * 2,
pin_memory=True)
net = torch.nn.DataParallel(net, device_ids=conf.gpus).cuda()
net.eval()
total_num = len(data_loader.dataset)
output = []
proc_start_time = time.time()
for i, (keys, input_) in enumerate(data_loader):
rst = eval_video(conf, (i, keys, input_), net)
output.append(rst[1:])
cnt_time = time.time() - proc_start_time
print('video {} done, total {}/{}, average {} sec/video'.format(
i, i + 1, total_num,
float(cnt_time) / (i + 1)))
video_index = [x[0] for x in output]
scores = [x[1] for x in output]
save_scores = './{}/tsn_{}_{}_testset_{}_{}_lr_{}_model_{:03d}.npz'.format(
conf.checkpoint, conf.class_type, conf.modality.lower(), test_set,
conf.arch, conf.lr, checkpoint['epoch'])
if test_part > 0:
save_scores = save_scores.replace('.npz',
'_part-{}.npz'.format(test_part))
np.savez(save_scores, segment_indices=video_index, scores=scores)
def main():
global args
global best_prec1
args = parser.parse_args()
print('Training arguments:')
for k, v in vars(args).items():
print('\t{}: {}'.format(k, v))
if args.data_name == 'ucf101':
num_class = 101
elif args.data_name == 'hmdb51':
num_class = 51
else:
raise ValueError('Unknown dataset ' + args.data_name)
# num_class: total number of classes
# num_segments: number of TSN segments, default=3
# representation: iframe, mv, residual
# base_model: base architecture
model = Model(num_class,
args.num_segments,
args.representation,
base_model=args.arch,
mv_stack_size=args.mv_stack_size)
print(model)
# dataset (Dataset) – dataset from which to load the data.
# batch_size – how many samples per batch to load (default: 1).
# shuffle – set to True to have the data reshuffled at every epoch.
# num_workers – how many subprocesses to use for data loading. 0 means that the data will be loaded in the main process. (default: 0)
# pin_memory – If True, the data loader will copy tensors into CUDA pinned memory before returning them.
train_loader = torch.utils.data.DataLoader(
CoviarDataSet(
args.data_root,
args.data_name,
video_list=args.train_list,
num_segments=args.num_segments,
representation=args.representation,
transform=model.get_augmentation(),
# get_augmentation() =
# GroupMultiScaleCrop + GroupRandomHorizontalFlip
# GroupMultiScaleCrop contains stack mv
# seems np.stack in resize_mv() called in GroupMultiScaleCrop
# has the same effects as Stack() in TSN
# -----------------------
# TSN:
# transform=torchvision.transforms.Compose([
# train_augmentation, # train_augmentation = model.get_augmentation(), same
# Stack(roll=args.arch == 'BNInception'), # this line seems important
# ToTorchFormatTensor(div=args.arch != 'BNInception'),
# normalize, # used for RGBDiff
# ])),
# ----------------------
is_train=True,
accumulate=(not args.no_accumulation),
mv_stack_size=args.mv_stack_size),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(
CoviarDataSet(
args.data_root,
args.data_name,
video_list=args.test_list,
num_segments=args.num_segments,
representation=args.representation,
transform=torchvision.transforms.
Compose([ # seems important to stacking
GroupScale(int(model.scale_size)),
GroupCenterCrop(
model.crop_size
), # here they both use model.crop_size (instead of TSN's net.input_size in test_model.py)
]), # this function contains stack
# seems np.stack in resize_mv() called in GroupCenterCrop
# has the same effects as Stack() in TSN
# -----------------------
# TSN:
# transform=torchvision.transforms.Compose([
# GroupScale(int(scale_size)),
# GroupCenterCrop(crop_size),
# Stack(roll=args.arch == 'BNInception'), # this line seems important
# ToTorchFormatTensor(div=args.arch != 'BNInception'),
# normalize,
# ])),
# -----------------------
is_train=False,
accumulate=(not args.no_accumulation),
mv_stack_size=args.mv_stack_size),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
# parallel gpu setting
model = torch.nn.DataParallel(model, device_ids=args.gpus).cuda()
cudnn.benchmark = True
params_dict = dict(model.named_parameters())
params = []
for key, value in params_dict.items():
decay_mult = 0.0 if 'bias' in key else 1.0
if ('module.base_model.conv1' in key or 'module.base_model.bn1' in key
or 'data_bn'
in key) and args.representation in ['mv', 'residual']:
lr_mult = 0.1
elif '.fc.' in key:
lr_mult = 1.0
else:
lr_mult = 0.01
params += [{
'params': value,
'lr': args.lr,
'lr_mult': lr_mult,
'decay_mult': decay_mult
}]
optimizer = torch.optim.Adam(params,
weight_decay=args.weight_decay,
eps=0.001)
criterion = torch.nn.CrossEntropyLoss().cuda()
for epoch in range(args.epochs):
cur_lr = adjust_learning_rate(optimizer, epoch, args.lr_steps,
args.lr_decay)
train(train_loader, model, criterion, optimizer, epoch, cur_lr)
if epoch % args.eval_freq == 0 or epoch == args.epochs - 1:
prec1 = validate(val_loader, model, criterion)
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if is_best or epoch % SAVE_FREQ == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
},
is_best,
filename='checkpoint.pth.tar')
def main():
global args
global best_prec1
args = parser.parse_args()
print('Training arguments:')
for k, v in vars(args).items():
print('\t{}: {}'.format(k, v))
if args.data_name == 'ucf101':
num_class = 101
elif args.data_name == 'hmdb51':
num_class = 51
elif args.data_name == 'mine':
num_class = 2
else:
raise ValueError('Unknown dataset ' + args.data_name)
model = Model(num_class,
args.num_segments,
args.representation,
base_model=args.arch)
print(model)
if 'resnet3D' in args.arch:
train_crop_min_ratio = 0.75
train_crop_min_scale = 0.25
mean = [0.4345, 0.4051, 0.3775]
std = [0.2768, 0.2713, 0.2737]
value_scale = 1
train_transform = Compose([
RandomResizedCrop(
model.crop_size, (train_crop_min_scale, 1.0),
(train_crop_min_ratio, 1.0 / train_crop_min_ratio)),
RandomHorizontalFlip(),
ToTensor(),
ScaleValue(value_scale),
Normalize(mean, std)
])
test_trainsform = Compose([
Resize(model.crop_size),
CenterCrop(model.crop_size),
ToTensor(), # range [0, 255] -> [0.0,1.0]
ScaleValue(1),
Normalize(mean, std)
])
train_loader = torch.utils.data.DataLoader(
CoviarDataSet(
args.data_root,
args.data_name,
video_list=args.train_list,
num_segments=args.num_segments,
representation=args.representation,
transform=model.get_augmentation(), #train_transform,
is_train=True,
accumulate=(not args.no_accumulation),
model_name=args.arch),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True,
worker_init_fn=worker_init_fn)
val_loader = torch.utils.data.DataLoader(
CoviarDataSet(
args.data_root,
args.data_name,
video_list=args.test_list,
num_segments=args.num_segments,
representation=args.representation,
transform=torchvision.transforms.Compose([
GroupScale(int(model.scale_size)),
GroupCenterCrop(model.crop_size)
]), #test_trainsform,
is_train=True,
accumulate=(not args.no_accumulation),
model_name=args.arch),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
worker_init_fn=worker_init_fn)
model = torch.nn.DataParallel(model, device_ids=args.gpus).cuda()
cudnn.benchmark = True
params_dict = dict(model.named_parameters())
params = []
for key, value in params_dict.items():
decay_mult = 0.0 if 'bias' in key else 1.0
if ('module.base_model.conv1' in key or 'module.base_model.bn1' in key
or 'data_bn'
in key) and args.representation in ['mv', 'residual']:
lr_mult = 0.1
elif '.fc.' in key:
lr_mult = 1.0
else:
lr_mult = 0.01
params += [{
'params': value,
'lr': args.lr,
'lr_mult': lr_mult,
'decay_mult': decay_mult
}]
#optimizer = torch.optim.SGD(params, weight_decay=0.001, momentum=0.9, nesterov=False)
#scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=10)
optimizer = torch.optim.Adam(params,
weight_decay=args.weight_decay,
eps=0.001)
criterion = torch.nn.CrossEntropyLoss().cuda()
for epoch in range(args.epochs):
cur_lr = adjust_learning_rate(optimizer, epoch, args.lr_steps,
args.lr_decay)
#cur_lr = get_lr(optimizer)
train(train_loader, model, criterion, optimizer, epoch, cur_lr)
#prec1, prev_val_loss = validate(val_loader, model, criterion)
#scheduler.step(prev_val_loss)
if epoch % args.eval_freq == 0 or epoch == args.epochs - 1:
prec1, _ = validate(val_loader, model, criterion)
# 紀錄訓練歷程
np.savez("train_history/train_history.npz",
loss=np.array(train_loss),
top1=np.array(train_prec),
lr=np.array(train_lr))
np.savez("train_history/valid_history.npz",
loss=np.array(valid_loss),
top1=np.array(valid_prec))
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if is_best or epoch % SAVE_FREQ == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
},
is_best,
filename='checkpoint.pth.tar')
def main():
net = Model(num_class,
args.test_segments,
args.representation,
base_model=args.arch)
checkpoint = torch.load(args.weights)
print("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.crop_size),
])
elif args.test_crops == 10:
cropping = torchvision.transforms.Compose([
GroupOverSample(net.crop_size,
net.scale_size,
is_mv=(args.representation == 'mv'))
])
else:
raise ValueError(
"Only 1 and 10 crops are supported, but got {}.".format(
args.test_crops))
data_loader = torch.utils.data.DataLoader(CoviarDataSet(
args.data_root,
args.data_name,
video_list=args.test_list,
num_segments=args.test_segments,
representation=args.representation,
transform=cropping,
is_train=False,
accumulate=(not args.no_accumulation),
),
batch_size=1,
shuffle=False,
num_workers=args.workers * 2,
pin_memory=True)
if args.gpus is not None:
devices = [args.gpus[i] for i in range(args.workers)]
else:
devices = list(range(args.workers))
net = torch.nn.DataParallel(net.cuda(devices[0]), device_ids=devices)
net.eval()
data_gen = enumerate(data_loader)
total_num = len(data_loader.dataset)
output = []
def forward_video(data):
input_var = torch.autograd.Variable(data, volatile=True)
scores = net(input_var)
scores = scores.view((-1, args.test_segments * args.test_crops) +
scores.size()[1:])
scores = torch.mean(scores, dim=1)
return scores.data.cpu().numpy().copy()
proc_start_time = time.time()
for i, (data, label) in data_gen:
video_scores = forward_video(data)
output.append((video_scores, label[0]))
cnt_time = time.time() - proc_start_time
if (i + 1) % 100 == 0:
print('video {} done, total {}/{}, average {} sec/video'.format(
i, i + 1, total_num,
float(cnt_time) / (i + 1)))
video_pred = [np.argmax(x[0]) for x in output]
video_labels = [x[1] for x in output]
print('Accuracy {:.02f}% ({})'.format(
float(np.sum(np.array(video_pred) == np.array(video_labels))) /
len(video_pred) * 100.0, len(video_pred)))
if args.save_scores is not None:
name_list = [x.strip().split()[0] for x in open(args.test_list)]
order_dict = {e: i for i, e in enumerate(sorted(name_list))}
reorder_output = [None] * len(output)
reorder_label = [None] * len(output)
reorder_name = [None] * len(output)
for i in range(len(output)):
idx = order_dict[name_list[i]]
reorder_output[idx] = output[i]
reorder_label[idx] = video_labels[i]
reorder_name[idx] = name_list[i]
np.savez(args.save_scores,
scores=reorder_output,
labels=reorder_label,
names=reorder_name)
# Move to GPU if available and set to evaluation
model.eval()
model.to(device)
# Define the transform
batch_size = 1
snippet_length = 1 # Number of frames composing the snippet, 1 for RGB, 5 for optical flow
snippet_channels = 3 # Number of channels in a frame, 3 for RGB, 2 for optical flow
height, width = 224, 224
crop_count = 10
if crop_count == 1:
cropping = Compose([
GroupScale(model.scale_size),
GroupCenterCrop(model.input_size),
])
elif crop_count == 10:
cropping = GroupOverSample(model.input_size, model.scale_size)
else:
raise ValueError("Only 1 and 10 crop_count are supported while we got {}".format(crop_count))
transform = Compose([
cropping,
Stack(roll=base_model == base_model),
ToTorchFormatTensor(div=base_model != base_model),
GroupNormalize(model.input_mean, model.input_std),
])
pred_verb_indices = []
pred_noun_indices = []
def main():
# define the model
net = Model(num_class,
args.test_segments,
args.representation,
base_model=args.arch,
new_length=args.new_length,
use_databn=args.use_databn,
gen_flow_or_delta=args.gen_flow_or_delta,
gen_flow_ds_factor=args.gen_flow_ds_factor,
arch_estimator=args.arch_estimator,
att=args.att)
# load the trained model
checkpoint = torch.load(args.weights,
map_location=lambda storage, loc: storage)
print("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, strict=False)
# setup the data loader
if args.test_crops == 1:
cropping = torchvision.transforms.Compose([
GroupScale(net.scale_size),
GroupCenterCrop(net.crop_size),
])
elif args.test_crops == 10:
cropping = torchvision.transforms.Compose(
[GroupOverSample(net.crop_size, net.scale_size)])
else:
raise ValueError(
"Only 1 and 10 crops are supported, but got {}.".format(
args.test_crops))
data_loader = torch.utils.data.DataLoader(CoviarDataSet(
args.data_root,
args.flow_root,
args.data_name,
video_list=args.test_list,
num_segments=args.test_segments,
representation=args.representation,
new_length=args.new_length,
flow_ds_factor=args.flow_ds_factor,
upsample_interp=args.upsample_interp,
transform=cropping,
is_train=False,
accumulate=(not args.no_accumulation),
gop=args.gop,
flow_folder=args.data_flow,
viz=args.viz),
batch_size=1,
shuffle=False,
num_workers=args.workers * 2,
pin_memory=True)
# deploy model on gpu
if args.gpus is not None:
devices = [args.gpus[i] for i in range(args.workers)]
else:
devices = list(range(args.workers))
net.cuda(devices[0])
#net.base_model.cuda(devices[-1])
net = torch.nn.DataParallel(net, device_ids=devices)
# switch to inference model and start to iterate over the test set
net.eval()
total_num = len(data_loader.dataset)
output = []
# process each video to obtain its predictions
def forward_video(input_mv, input_residual, att=0):
input_mv_var = torch.autograd.Variable(input_mv, volatile=True)
input_residual_var = torch.autograd.Variable(input_residual,
volatile=True)
if att == 0:
scores, gen_flow = net(input_mv_var, input_residual_var)
if att == 1:
scores, gen_flow, att_flow = net(input_mv_var, input_residual_var)
scores = scores.view((-1, args.test_segments * args.test_crops) +
scores.size()[1:])
scores = torch.mean(scores, dim=1)
if att == 0:
return scores.data.cpu().numpy().copy(), gen_flow
if att == 1:
return scores.data.cpu().numpy().copy(), gen_flow, att_flow
proc_start_time = time.time()
# iterate over the whole test set
for i, (input_flow, input_mv, input_residual,
label) in enumerate(data_loader):
input_mv = input_mv.cuda(args.gpus[-1], async=True)
input_residual = input_residual.cuda(args.gpus[0], async=True)
input_flow = input_flow.cuda(args.gpus[-1], async=True)
# print("input_flow shape:")
# print(input_flow.shape) # torch.Size([batch_size, num_crops*num_segments, 2, 224, 224])
# print("input_flow type:") # print(input_flow.type()) # torch.cuda.FloatTensor
if args.att == 0:
video_scores, gen_flow = forward_video(input_mv, input_residual)
if args.att == 1:
video_scores, gen_flow, att_flow = forward_video(
input_mv, input_residual, args.att)
output.append((video_scores, label[0]))
cnt_time = time.time() - proc_start_time
if (i + 1) % 100 == 0:
print('video {} done, total {}/{}, average {} sec/video'.format(
i, i + 1, total_num,
float(cnt_time) / (i + 1)))
video_pred = [np.argmax(x[0]) for x in output]
video_labels = [x[1] for x in output]
print('Accuracy {:.02f}% ({})'.format(
float(np.sum(np.array(video_pred) == np.array(video_labels))) /
len(video_pred) * 100.0, len(video_pred)))
if args.save_scores is not None:
name_list = [x.strip().split()[0] for x in open(args.test_list)]
order_dict = {e: i for i, e in enumerate(sorted(name_list))}
reorder_output = [None] * len(output)
reorder_label = [None] * len(output)
reorder_name = [None] * len(output)
for i in range(len(output)):
idx = order_dict[name_list[i]]
reorder_output[idx] = output[i]
reorder_label[idx] = video_labels[i]
reorder_name[idx] = name_list[i]
np.savez(args.save_scores,
scores=reorder_output,
labels=reorder_label,
names=reorder_name)
def main():
# load trained model
'''
@Param
num_class: total number of classes
num_segments: number of TSN segments, test default = 25
representation: iframe, mv, residual
base_model: base architecture
'''
net = Model(num_class,
args.test_segments,
args.representation,
base_model=args.arch,
mv_stack_size=args.mv_stack_size)
# -----------------------------MODIFIED_CODE_START-------------------------------
# print(net)
# -----------------------------MODIFIED_CODE_END---------------------------------
# checkpoint trained model ? (not best model
checkpoint = torch.load(args.weights)
print("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)
# -----------------------
# CLASS torchvision.transforms.Compose(transforms)[SOURCE]
# Composes several transforms together.
# Parameters: transforms (list of Transform objects) – list of transforms to compose.
# -----------------------
# -----------------------
# TSN:
# if args.test_crops == 1:
# cropping = torchvision.transforms.Compose([
# GroupScale(net.scale_size),
# GroupCenterCrop(net.input_size),
# ])
# -----------------------
if args.test_crops == 1:
cropping = torchvision.transforms.Compose([
GroupScale(net.scale_size),
GroupCenterCrop(net.crop_size),
])
# ??? what's difference between net.input_size and net.crop_size
# line 70 in model.py
# def crop_size(self):
# return self._input_size
# seems they are same here
# -----------------------
# TSN:
# elif args.test_crops == 10:
# cropping = torchvision.transforms.Compose([
# GroupOverSample(net.input_size, net.scale_size)
# ])
# -----------------------
# is_mv=(args.representation == 'mv') seems quite important
elif args.test_crops == 10:
cropping = torchvision.transforms.Compose([
GroupOverSample(net.crop_size,
net.scale_size,
is_mv=(args.representation == 'mv'))
])
# --test-crops specifies how many crops per segment.
# The value should be 1 or 10.
# 1 means using only one center crop.
# 10 means using 5 crops for both (horizontal) flips.
else:
raise ValueError(
"Only 1 and 10 crops are supported, but got {}.".format(
args.test_crops))
data_loader = torch.utils.data.DataLoader(
CoviarDataSet(
args.data_root,
args.data_name,
video_list=args.test_list,
num_segments=args.test_segments,
representation=args.representation,
transform=cropping, # seems important to stacking
# test_crops == 1: GroupScale + GroupCenterCrop
# the same as val_data_loader in train.py
# seems np.stack in resize_mv() called in GroupCenterCrop
# has the same effects as Stack() in TSN
# test_crops == 10: GroupOverSample
# -----------------------
# TSN:
# transform=torchvision.transforms.Compose([
# cropping,
# Stack(roll=args.arch == 'BNInception'), # this line seems important
# ToTorchFormatTensor(div=args.arch != 'BNInception'),
# GroupNormalize(net.input_mean, net.input_std),
# ])),
# -----------------------
is_train=False,
accumulate=(not args.no_accumulation),
mv_stack_size=args.mv_stack_size),
batch_size=1,
shuffle=False,
# -----------------------------ORIGINAL_CODE_START-----------------------------
# num_workers=args.workers * 2, pin_memory=True)
# -----------------------------ORIGINAL_CODE_END-------------------------------
# -----------------------------MODIFIED_CODE_START-----------------------------
num_workers=args.workers,
pin_memory=True)
# -----------------------------MODIFIED_CODE_END-------------------------------
if args.gpus is not None:
devices = [args.gpus[i] for i in range(args.workers)]
else:
devices = list(range(args.workers))
net = torch.nn.DataParallel(net.cuda(devices[0]), device_ids=devices)
net.eval()
data_gen = enumerate(data_loader)
total_num = len(data_loader.dataset)
output = []
def forward_video(data):
# torch.Size([batch_size, num_segment, 2*MV_STACK_SIZE, height, width])
# -----------------------------MODIFIED_CODE_START-------------------------------
# print("data.shape"+str(data.shape)) # testing: torch.Size([1, 25, 10, 224, 224])
# training: torch.Size([40, 3, 10, 224, 224])
# original:data.shape:torch.Size([1, 250, 2, 224, 224])
# so it seems that the format of input data in this function is not correct
# -----------------------------MODIFIED_CODE_END---------------------------------
input_var = torch.autograd.Variable(data, volatile=True)
# -----------------------------MODIFIED_CODE_START-------------------------------
# print("input_var:"+str(input_var.shape)) # input_var:torch.Size([1, 25, 10, 224, 224])
# original: input_var.shape:torch.Size([1, 250, 2, 224, 224])
# -----------------------------MODIFIED_CODE_END---------------------------------
# compute output
scores = net(input_var)
# -----------------------------MODIFIED_CODE_START-------------------------------
# torch.Size([batch_size*num_segment, num_class])
# print("scores: "+str(scores.shape)) # testing: torch.Size([25, 101])
# training: torch.Size([120, 101])
# print("scores.size()")
# print(scores.size()) # torch.Size([25, 101])
# -----------------------------MODIFIED_CODE_END---------------------------------
# what does args.test_segments * args.test_crops mean??
# view(*shape) → Tensor: Returns a new tensor with the same data as the self tensor but of a different shape.
# Parameters shape (torch.Size or int...) – the desired size
scores = scores.view((-1, args.test_segments * args.test_crops) +
scores.size()[1:])
scores = torch.mean(scores, dim=1)
return scores.data.cpu().numpy().copy()
proc_start_time = time.time()
for i, (data, label) in data_gen:
video_scores = forward_video(data)
output.append((video_scores, label[0]))
cnt_time = time.time() - proc_start_time
if (i + 1) % 100 == 0:
print('video {} done, total {}/{}, average {} sec/video'.format(
i, i + 1, total_num,
float(cnt_time) / (i + 1)))
video_pred = [np.argmax(x[0]) for x in output]
video_labels = [x[1] for x in output]
print('Accuracy {:.02f}% ({})'.format(
float(np.sum(np.array(video_pred) == np.array(video_labels))) /
len(video_pred) * 100.0, len(video_pred)))
if args.save_scores is not None:
name_list = [x.strip().split()[0] for x in open(args.test_list)]
order_dict = {e: i for i, e in enumerate(sorted(name_list))}
reorder_output = [None] * len(output)
reorder_label = [None] * len(output)
reorder_name = [None] * len(output)
for i in range(len(output)):
idx = order_dict[name_list[i]]
reorder_output[idx] = output[i]
reorder_label[idx] = video_labels[i]
reorder_name[idx] = name_list[i]
np.savez(args.save_scores,
scores=reorder_output,
labels=reorder_label,
names=reorder_name)
def main():
writter = SummaryWriter('./log/test', comment='')
net = Model(2, args.num_segments, args.representation,
base_model=args.arch)
checkpoint = torch.load(args.weights)
# print("model epoch {} best prec@1: {}".format(checkpoint['epoch'], checkpoint['best_prec1']))
print("model epoch {} lowest loss {}".format(checkpoint['epoch'], checkpoint['loss_min']))
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.crop_size),
])
elif args.test_crops == 10:
cropping = torchvision.transforms.Compose([
GroupOverSample(net.crop_size, net.scale_size, is_mv=(args.representation == 'mv'))
])
else:
raise ValueError("Only 1 and 10 crops are supported, but got {}.".format(args.test_crops))
data_loader = torch.utils.data.DataLoader(
CoviarDataSet(
args.data_root,
video_list=args.test_list,
num_segments=args.num_segments,
representation=args.representation,
transform=cropping,
is_train=False,
accumulate=(not args.no_accumulation),
),
batch_size=1, shuffle=False,
num_workers=args.workers * 2, pin_memory=True)
devices = [torch.device("cuda:%d" % device) for device in args.gpus]
net = torch.nn.DataParallel(net.cuda(devices[0]), device_ids=devices)
net.eval()
total_num = len(data_loader.dataset)
scores = []
labels = []
proc_start_time = time.time()
correct_nums = 0
for i, (input_pairs, label) in enumerate(data_loader):
with torch.no_grad:
input_pairs[0] = input_pairs[0].float().to(devices[0])
input_pairs[1] = input_pairs[1].float().to(devices[0])
label = label.float().to(devices[0])
outputs, y = net(input_pairs)
_, predicts = torch.max(y, 1)
scores.append(y.detach().cpu().numpy())
labels.append(label.detach().cpu().numpy())
correct_nums += (predicts == label.clone().long()).sum()
cnt_time = time.time() - proc_start_time
if (i + 1) % 100 == 0:
print('video {} done, total {}/{}, average {} sec/video'.format(i, i + 1,
total_num,
float(cnt_time) / (i + 1)))
predits = np.argmax(scores, 1)
labels = np.around(labels).astype(np.long).ravel()
acc = 100 * correct_nums / len(data_loader.dataset)
target_names = ['Copy', 'Not Copy']
# writter.add_pr_curve('Precision/Recall', labels, predits)
writter.add_text('Accuracy', '%.3f%%' % acc)
writter.add_text(classification_report(labels, predits, target_names=target_names))
print(('Validating Results: accuracy: {accuracy:.3f}%'.format(accuracy=acc)))
if args.save_scores is not None:
with open(args.save_scores + '_scores.pkl', 'wb') as fp:
pickle.dump(scores, fp)
with open(args.save_scores + '_labels.pkl', 'wb') as fp:
pickle.dump(labels, fp)
def main():
# loading input arguments for training
global args
global best_prec1
global start_epoch
start_epoch = 0
args = parser.parse_args()
print('Training arguments:')
for k, v in vars(args).items():
print('\t{}: {}'.format(k, v))
if args.data_name == 'ucf101':
num_class = 101
elif args.data_name == 'hmdb51':
num_class = 51
elif args.data_name == 'kinetics400':
num_class = 400
else:
raise ValueError('Unknown dataset ' + args.data_name)
# define the model architecture
model = Model(num_class, args.num_segments, args.representation,
base_model=args.arch,
new_length=args.new_length,
use_databn=args.use_databn,
gen_flow_or_delta=args.gen_flow_or_delta,
gen_flow_ds_factor=args.gen_flow_ds_factor,
arch_estimator=args.arch_estimator,
arch_d=args.arch_d,
att=args.att)
print(model)
# load the pre-trained model
if args.weights is not None:
checkpoint = torch.load(args.weights, map_location=lambda storage, loc: storage)
print("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())}
model.load_state_dict(base_dict, strict=False)
# define the data loader for reading training data
train_loader = torch.utils.data.DataLoader(
CoviarDataSet(
args.data_root,
args.flow_root,
args.data_name,
video_list=args.train_list,
num_segments=args.num_segments,
representation=args.representation,
new_length=args.new_length,
flow_ds_factor=args.flow_ds_factor,
upsample_interp=args.upsample_interp,
transform=model.get_augmentation(),
is_train=True,
accumulate=(not args.no_accumulation),
gop=args.gop,
flow_folder=args.data_flow,
mv_minmaxnorm=args.mv_minmaxnorm,
),
batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
# define the data loader for reading val data
val_loader = torch.utils.data.DataLoader(
CoviarDataSet(
args.data_root,
args.flow_root,
args.data_name,
video_list=args.test_list,
num_segments=args.num_segments,
representation=args.representation,
new_length=args.new_length,
flow_ds_factor=args.flow_ds_factor,
upsample_interp=args.upsample_interp,
transform=torchvision.transforms.Compose([
GroupScale(int(model.scale_size)),
GroupCenterCrop(model.crop_size),
]),
is_train=False,
accumulate=(not args.no_accumulation),
gop=args.gop,
flow_folder=args.data_flow,
mv_minmaxnorm=args.mv_minmaxnorm,
),
batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
model = torch.nn.DataParallel(model, device_ids=args.gpus).cuda(args.gpus[0])
cudnn.benchmark = True
# define optimizer and specify the corresponding parameters
params_dict = dict(model.named_parameters())
params_cls = []
params_gf = []
params_d = []
for key, value in params_dict.items():
if 'base_model' in key:
decay_mult = 0.0 if 'bias' in key else 1.0
lr_mult = args.lr_cls_mult # for cls, just finetune. if '.fc.' in key: lr_mult = 1.0
params_cls += [{'params': value, 'lr': args.lr, 'lr_mult': lr_mult, 'decay_mult': decay_mult}]
if 'gen_flow_model' in key:
decay_mult = 0.0 if 'bias' in key else 1.0
lr_mult = args.lr_mse_mult
params_gf += [{'params': value, 'lr': args.lr, 'lr_mult': lr_mult, 'decay_mult': decay_mult}]
if 'discriminator' in key:
decay_mult = 0.0 if 'bias' in key else 1.0
lr_mult = args.lr_d_mult
params_d += [{'params': value, 'lr': args.lr, 'lr_mult': lr_mult, 'decay_mult': decay_mult}]
optimizer_cls = torch.optim.Adam(
params_cls,
weight_decay=args.weight_decay,
eps=0.001)
optimizer_gf = torch.optim.Adam(
params_gf,
weight_decay=args.weight_decay,
eps=0.001)
optimizer_d = torch.optim.Adam(
params_d,
weight_decay=args.weight_decay,
eps=0.001)
# resume training from previous checkpoint
if args.resume is not None:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume, map_location=lambda storage, loc: storage)
start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
if 'optimizer_cls' in checkpoint.keys():
optimizer_cls.load_state_dict(checkpoint['optimizer_cls'])
optimizer_gf.load_state_dict(checkpoint['optimizer_gf'])
optimizer_d.load_state_dict(checkpoint['optimizer_d'])
def load_opt_update_cuda(optimizer, cuda_id):
for state in optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda(cuda_id)
load_opt_update_cuda(optimizer_cls, args.gpus[0])
load_opt_update_cuda(optimizer_gf, args.gpus[0])
load_opt_update_cuda(optimizer_d, args.gpus[0])
print("=> loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
# define several loss functions
criterion = torch.nn.CrossEntropyLoss().cuda(args.gpus[0])
if args.loss_mse == 'MSELoss':
criterion_mse = torch.nn.MSELoss().cuda(args.gpus[0])
elif args.loss_mse == 'SmoothL1Loss':
criterion_mse = torch.nn.SmoothL1Loss().cuda(args.gpus[0])
elif args.loss_mse == 'L1':
criterion_mse = torch.nn.L1Loss().cuda(args.gpus[0])
# finally done with setup and start to train model
for epoch in range(start_epoch, args.epochs):
# determine the learning rate for the current epoch
cur_lr_cls = adjust_learning_rate(optimizer_cls, epoch, args.lr_steps, args.lr_decay) #, freeze=True, epoch_thre=args.epoch_thre)
cur_lr_gf = adjust_learning_rate(optimizer_gf, epoch, args.lr_steps, args.lr_decay)
cur_lr_d = adjust_learning_rate(optimizer_d, epoch, args.lr_steps, args.lr_decay)
# perform training
train(train_loader, model, criterion, criterion_mse, optimizer_cls,
optimizer_gf, optimizer_d, epoch, cur_lr_cls, cur_lr_gf, cur_lr_d, args.lr_cls, args.lr_adv_g, args.lr_adv_d, args.lr_mse, args.att)
# perform validation if needed
if epoch % args.eval_freq == 0 or epoch == args.epochs - 1:
prec1 = validate(val_loader, model, criterion, criterion_mse, args.lr_cls, args.lr_adv_g, args.lr_adv_d, args.lr_mse, args.att)
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if is_best or epoch % SAVE_FREQ == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer_cls': optimizer_cls.state_dict(),
'optimizer_gf': optimizer_gf.state_dict(),
'optimizer_d': optimizer_d.state_dict(),
},
is_best,
filename='checkpoint.pth.tar')
def main():
net = Model(num_class, base_model=args.arch)
checkpoint = torch.load(args.weights)
print("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.crop_size),
])
elif args.test_crops == 10:
cropping = torchvision.transforms.Compose([
GroupOverSample(net.crop_size,
net.scale_size,
is_mv=(args.representation == 'mv'))
])
else:
raise ValueError(
"Only 1 and 10 crops are supported, but got {}.".format(
args.test_crops))
data_loader = torch.utils.data.DataLoader(FoodDataSet(
args.data_root,
img_list=args.test_list,
transform=cropping,
is_train=False,
),
batch_size=1,
shuffle=False,
num_workers=args.workers * 2,
pin_memory=True)
if args.gpus is not None:
devices = [args.gpus[i] for i in range(args.workers)]
else:
devices = list(range(args.workers))
net = torch.nn.DataParallel(net.cuda(devices[0]), device_ids=devices)
net.eval()
data_gen = enumerate(data_loader)
total_num = len(data_loader.dataset)
output = []
def forward_img(data):
"""
Args:
data (Tensor): size [batch_size, c, h, w]
Returns:
scores (Tensor) : size [batch_size, num_class]
"""
with torch.no_grad():
input_var = torch.autograd.Variable(data, volatile=True)
scores = net(input_var)
scores = scores.view((-1, args.test_crops) + scores.size()[1:])
scores = torch.mean(scores, dim=1)
return scores.data.cpu().numpy().copy()
proc_start_time = time.time()
for i, (data, label) in data_gen:
# data = [1, c, h ,w], label = [1]
img_scores = forward_img(data)
output.append((img_scores[0], label[0]))
cnt_time = time.time() - proc_start_time
if (i + 1) % 100 == 0:
print('image {} done, total {}/{}, average {} sec/image'.format(
i, i + 1, total_num,
float(cnt_time) / (i + 1)))
img_pred = [np.argmax(x[0]) for x in output]
img_labels = [x[1] for x in output]
print('Accuracy {:.02f}% ({})'.format(
float(np.sum(np.array(img_pred) == np.array(img_labels))) /
len(img_pred) * 100.0, len(img_pred)))