def video_transform(input_config, clip_input):
normalize = transforms.Normalize(mean=input_config['mean'], std=input_config['std'])
video_transform = transforms.Compose([
transforms.Resize((256, 256)),
transforms.RandomCrop((224, 224)),
# transforms.CenterCrop((224, 224)), # we did not use center crop in our paper
# transforms.RandomHorizontalFlip(), # we did not use mirror in our paper
transforms.ToTensor(),
normalize,
])
return video_transform(clip_input)
def prepare_data():
# split the video list
vnames = get_video_names(cfg.DATASET.VIDEO_LIST)
train_split = vnames
test_split = vnames[:int(0.2 * len(train_split))]
train_transforms = transforms.Compose([
video_transforms.Resize(cfg.DATA_TRANSFORM.LOADSIZE),
video_transforms.RandomCrop(cfg.DATA_TRANSFORM.FINESIZE)
])
#video_transforms.RandomHorizontalFlip()])
test_transforms = transforms.Compose([
video_transforms.Resize(cfg.DATA_TRANSFORM.LOADSIZE),
video_transforms.CenterCrop(cfg.DATA_TRANSFORM.FINESIZE)
])
video_root = cfg.DATASET.VIDEO_ROOT
annot_path = cfg.DATASET.ANNOT_JSON_SAVE_PATH
dataset = Dataset(train_split, video_root, annot_path, train_transforms)
train_dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=cfg.TRAIN.BATCH_SIZE,
shuffle=True,
num_workers=cfg.NUM_WORKERS,
drop_last=True,
pin_memory=True)
video_root = cfg.DATASET.VIDEO_ROOT
annot_path = cfg.DATASET.ANNOT_JSON_SAVE_PATH
test_dataset = Dataset(test_split, video_root, annot_path, test_transforms)
test_dataloader = torch.utils.data.DataLoader(
test_dataset,
batch_size=cfg.TEST.BATCH_SIZE,
shuffle=False,
num_workers=cfg.NUM_WORKERS,
drop_last=False,
pin_memory=True)
dataloaders = {'train': train_dataloader, 'test': test_dataloader}
return dataloaders
normalize = transforms.Normalize(mean=input_config['mean'],
std=input_config['std'])
val_sampler = sampler.RandomSampling(num=args.clip_length,
interval=args.frame_interval,
speed=[1.0, 1.0])
val_loader = VideoIter(
video_prefix=os.path.join(data_root, 'raw',
'data'), # change this part accordingly
txt_list=os.path.join(
data_root, 'raw', 'list_cvt',
'testlist01.txt'), # change this part accordingly
sampler=val_sampler,
force_color=True,
video_transform=transforms.Compose([
transforms.Resize((256, 256)),
transforms.RandomCrop((224, 224)),
# transforms.CenterCrop((224, 224)), # we did not use center crop in our paper
# transforms.RandomHorizontalFlip(), # we did not use mirror in our paper
transforms.ToTensor(),
normalize,
]),
name='test',
return_item_subpath=True,
)
eval_iter = torch.utils.data.DataLoader(
val_loader,
batch_size=args.batch_size,
shuffle=True,
num_workers=4, # change this part accordingly
pin_memory=True)
else:
sym_net = torch.nn.DataParallel(sym_net)
criterion = torch.nn.CrossEntropyLoss()
net = static_model(net=sym_net, criterion=criterion, model_prefix=args.model_prefix)
net.load_checkpoint(epoch=args.load_epoch)
# data iterator:
data_root = "../dataset/{}".format(args.dataset)
video_location = os.path.join(data_root, 'raw', 'test_data')
normalize = transforms.Normalize(mean=input_config['mean'], std=input_config['std'])
val_sampler = sampler.RandomSampling(num=args.clip_length, interval=args.frame_interval, speed=[1.0, 1.0], seed=1)
val_loader = VideoIter(video_prefix=video_location, csv_list=os.path.join(data_root, 'raw', 'list_cvt', args.list_file), sampler=val_sampler,
force_color=True, video_transform=transforms.Compose([transforms.Resize((256,256)), transforms.RandomCrop((224,224)), transforms.ToTensor(), normalize]),
name='predict', return_item_subpath=True,)
eval_iter = torch.utils.data.DataLoader(val_loader, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, pin_memory=True)
# main loop
net.net.eval()
sum_batch_elapse = 0.
softmax = torch.nn.Softmax(dim=1)
field_names = ['VideoID', 'Video', 'ClassID']
pred_rows = []
pred_file = 'track1_pred.csv'
i_batch = 0
for datas, targets, video_subpaths in eval_iter:
