def __init__(self, dataset_folder, spt_path, boxes_file, vid2idx, mode='train',get_loader=get_default_video_loader,
sample_size=112, classes_idx=None):
self.dataset_folder = dataset_folder
self.sample_size = sample_size
self.boxes_file = boxes_file
self.vid2idx = vid2idx
self.mode = mode
self.data, self.max_frames, self.max_actions = make_dataset_names( dataset_folder, spt_path, boxes_file, mode)
self.loader = get_loader()
self.classes_idx = classes_idx
# mean = [112.07945832, 112.87372333, 106.90993363] # ucf-101 24 classes
mean = [103.29825354, 104.63845484, 90.79830328] # jhmdb from .png
spatial_transform = Compose([Scale(sample_size), # [Resize(sample_size),
ToTensor(),
Normalize(mean, [1, 1, 1])])
self.spatial_transform=spatial_transform
split_txt_path = os.path.abspath(
os.path.join(root_path, dataset_cfg.dataset.split_txt_path))
### get videos id
actions = dataset_cfg.dataset.classes
cls2idx = {actions[i]: i for i in range(0, len(actions))}
vid2idx, vid_names = get_vid_dict(dataset_frames)
# # get mean
# mean = [112.07945832, 112.87372333, 106.90993363] # ucf-101 24 classes
mean = [0.5, 0.5, 0.5]
std = [0.5, 0.5, 0.5]
spatial_transform = Compose([
Scale(sample_size), # [Resize(sample_size),
ToTensor(),
Normalize(mean, std)
])
temporal_transform = LoopPadding(sample_duration)
n_classes = len(actions)
#######################################################
# Part 1-1 - train nTPN - without reg #
#######################################################
print(' -----------------------------------------------------')
print('| Part 1-1 - train TPN - without reg |')
print(' -----------------------------------------------------')
## Define Dataloaders
cls2idx = { classes[i] : i for i in range(0, len(classes)) }
dataset_folder = '/gpu-data2/sgal/JHMDB-act-detector-frames'
splt_txt_path = '/gpu-data2/sgal/splits'
boxes_file = '/gpu-data2/sgal/poses.json'
sample_size = 112
sample_duration = 16 #len(images)
batch_size = 1
n_threads = 0
mean = [103.29825354, 104.63845484, 90.79830328] # jhmdb from .png
spatial_transform = Compose([Scale(sample_size), # [Resize(sample_size),
# CenterCrop(sample_size),
ToTensor(),
Normalize(mean, [1, 1, 1])])
temporal_transform = LoopPadding(sample_duration)
data = Video(dataset_folder, frames_dur=sample_duration, spatial_transform=spatial_transform,
temporal_transform=temporal_transform, json_file = boxes_file,
split_txt_path=splt_txt_path, mode='train', classes_idx=cls2idx)
data_loader = torch.utils.data.DataLoader(data, batch_size=batch_size,
shuffle=False, num_workers=n_threads, pin_memory=True)
# clips, (h,w), gt_tubes, gt_bboxes, n_actions, n_frames = next(data_loader.__iter__())
# for i in data:
# clips, (h,w), gt_tubes, gt_bboxes, n_actions, n_frames = i
# # print('gt_bboxes.shape :',gt_bboxes)
# # print('gt_bboxes.shape :',gt_bboxes.shape)
# # print('gt_tubes :',gt_tubes)
# # print('clips.shape :',clips.shape)
sample_size = 112
sample_duration = 16 # len(images)
batch_size = 1
n_threads = 0
# # get mean
# mean = [103.75581543 104.79421473 91.16894564] # jhmdb
mean = [103.29825354, 104.63845484, 90.79830328] # jhmdb from .png
# generate model
last_fc = False
scale_size = [sample_size, sample_size]
spatial_transform = Compose([
Scale(sample_size), # [Resize(sample_size),
ToTensor()
])
temporal_transform = LoopPadding(sample_duration)
spatial_transform2 = Compose([ # [Resize(sample_size),
ToTensor()
])
## UCF code
dataset_folder = '/gpu-data/sgal/UCF-101-frames'
boxes_file = './pyannot.pkl'
actions = [
'Basketball', 'BasketballDunk', 'Biking', 'CliffDiving',
'CricketBowling', 'Diving', 'Fencing', 'FloorGymnastics', 'GolfSwing',
'HorseRiding', 'IceDancing', 'LongJump', 'PoleVault', 'RopeClimbing',
'SalsaSpin', 'SkateBoarding', 'Skiing', 'Skijet', 'SoccerJuggling',