def get_dataloader(type='segment'):
COMP_PATH = './'
split = 'test'
test_split = path.join(COMP_PATH, 'splits/test.split1.bundle')
mapping_loc = path.join(COMP_PATH, 'splits/mapping_bf.txt')
GT_folder = path.join(
COMP_PATH,
'groundTruth/') #Ground Truth Labels for each training video
DATA_folder = path.join(COMP_PATH,
'Data/') #Frame I3D features for all videos
actions_dict = read_mapping_dict(mapping_loc)
test_feat = load_data(test_split,
actions_dict,
GT_folder,
DATA_folder,
datatype=split,
target=type) #
if type == 'segment':
test_dataset = TestDataset(test_feat, test_seg_txt, seed=2)
else:
# todo
pass
test_dataloader = tud.DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=8,
pin_memory=True)
return test_dataloader
def get_dataloader(type='segment'):
valid_feat, valid_labels = load_data(val_split,
actions_dict,
GT_folder,
DATA_folder,
datatype=split,
target=type) #
if type == 'segment':
valid_dataset = SegmentDataset(valid_feat, valid_labels)
else:
valid_dataset = VideoDataset(valid_feat, valid_labels)
valid_dataloader = tud.DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=8,
pin_memory=True)
return valid_dataloader
train_split = os.path.join(COMP_PATH,
'splits/train.exclude_val.bundle') #Train Split
# train_split = os.path.join(COMP_PATH, 'splits/dev_train.split1.bundle') #Train Split
val_split = os.path.join(COMP_PATH, 'splits/val.split1.bundle')
test_split = os.path.join(COMP_PATH, 'splits/test.split1.bundle') #Test Split
GT_folder = os.path.join(
COMP_PATH, 'groundTruth/') #Ground Truth Labels for each training video
DATA_folder = os.path.join(COMP_PATH,
'Data/') #Frame I3D features for all videos
mapping_loc = os.path.join(COMP_PATH, 'splits/mapping_bf.txt')
actions_dict = read_mapping_dict(mapping_loc)
data_feat, data_labels = load_data(train_split,
actions_dict,
GT_folder,
DATA_folder,
datatype=split,
target='file') #
valid_feat, valid_labels = load_data(val_split,
actions_dict,
GT_folder,
DATA_folder,
datatype=split,
target='file') #
# balance data
# label_dict = b.generate_label_dictionary(data_feat, data_labels)
# total_segments = len(data_labels)
# data_feat, data_labels = b.balance_data(label_dict, total_segments)
# # model part
train_split = os.path.join(COMP_PATH,
'splits/train.exclude_val.bundle') #Train Split
# train_split = os.path.join(COMP_PATH, 'splits/dev_train.split1.bundle') #Train Split
val_split = os.path.join(COMP_PATH, 'splits/val.split1.bundle')
test_split = os.path.join(COMP_PATH, 'splits/test.split1.bundle') #Test Split
GT_folder = os.path.join(
COMP_PATH, 'groundTruth/') #Ground Truth Labels for each training video
DATA_folder = os.path.join(COMP_PATH,
'Data/') #Frame I3D features for all videos
mapping_loc = os.path.join(COMP_PATH, 'splits/mapping_bf.txt')
actions_dict = read_mapping_dict(mapping_loc)
data_feat, data_labels = load_data(train_split,
actions_dict,
GT_folder,
DATA_folder,
datatype=split,
target='frame') #
x = torch.cat(data_feat).reshape(-1, 400)
print(len(data_labels))
labels = torch.cat([torch.Tensor(labels)
for labels in data_labels]).long().flatten()
print(x.shape, labels.shape)
# filter 'silent' data
mask = labels == 0
labels = labels[~mask]
x = x[~mask]
'''
training to load train set
test to load test set
'''
split = 'training'
#split = 'test'
train_split = os.path.join(COMP_PATH, 'splits/train.split1.bundle') #Train Split
test_split = os.path.join(COMP_PATH, 'splits/test.split1.bundle') #Test Split
GT_folder = os.path.join(COMP_PATH, 'groundTruth/') #Ground Truth Labels for each training video
DATA_folder = os.path.join(COMP_PATH, 'Data/') #Frame I3D features for all videos
mapping_loc = os.path.join(COMP_PATH, 'splits/mapping_bf.txt')
actions_dict = read_mapping_dict(mapping_loc)
data_feat, data_labels = load_data( train_split, actions_dict, GT_folder, DATA_folder, datatype = split) #Get features and labels
validation_split = os.path.join(COMP_PATH, 'splits/val.split1.bundle') #Validation split
val_data_feat, val_data_labels = load_data( validation_split, actions_dict, GT_folder, DATA_folder, datatype = split) #Get features and labels
def train(log_interval, model, device, train_loader, optimizer, epoch):
model.train()
losses = []
scores = []
for batch_idx, (in_feature, label, seq_length) in enumerate(train_loader):
in_feature = in_feature.to(device)
label = label.to(device)
seq_length = seq_length.to(device)
packed_input = u.rnn.pack_padded_sequence(in_feature, seq_length, batch_first=True, enforce_sorted=False).to(
device)