def get_dataset(args):
# process text
if args.val_data_folder == 'validation':
text_proc, raw_data = get_vocab_and_sentences(args.dataset_file,
args.max_sentence_len)
else:
raw_data = {}
text_proc, _ = get_vocab_and_sentences(
'/data6/users/xuran7/myresearch/dense_videocap/third_party/densecap/data/anet/anet_annotations_trainval.json',
args.max_sentence_len)
with open(args.dataset_file, 'r') as data_file:
data_all = json.load(data_file)
for i in range(len(data_all)):
raw_data[data_all[i][2:]] = dict(subset='testing', annotations={})
# Create the dataset and data loader instance
test_dataset = ANetTestDataset(args.feature_root,
args.slide_window_size,
text_proc,
raw_data,
args.val_data_folder,
learn_mask=args.learn_mask)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=anet_test_collate_fn)
return test_loader, text_proc
def get_dataset(args):
# process text
text_proc, raw_data = get_vocab_and_sentences(args.dataset_file,
args.max_sentence_len)
# Create the dataset and data loader instance
train_dataset = ANetDataset(args,
args.train_data_folder,
text_proc,
raw_data,
test=False)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers)
valid_dataset = ANetDataset(args,
args.val_data_folder,
text_proc,
raw_data,
test=False)
valid_loader = DataLoader(valid_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers)
return train_loader, valid_loader
def get_dataset(args):
# process text
text_proc, raw_data = get_vocab_and_sentences(args.dataset_file,
args.max_sentence_len)
# Create the dataset and data loader instance
train_dataset = ANetDataset(args,
args.train_data_folder,
text_proc,
raw_data,
test=False)
train_sampler = None
# batch size forced to be 1 here, and batch is implemented in training with periodical zero_grad
train_loader = DataLoader(train_dataset,
batch_size=1,
shuffle=True,
num_workers=args.num_workers)
if args.calc_pos_neg:
from model.loss_func import add_pem_cls_num, add_pem_reg_num, add_tem_num
bm_mask = get_mask(args.temporal_scale, args.max_duration)
num_pos_neg = [0 for i in range(7)]
for i in range(len(train_dataset)):
sentence, img_feat, label_confidence, label_start, label_end, _ = train_dataset[
i]
num_pos_neg[0:3] = add_pem_reg_num(label_confidence, bm_mask,
num_pos_neg[0:3])
num_pos_neg[3:5] = add_pem_cls_num(label_confidence, bm_mask,
num_pos_neg[3:5])
num_pos_neg[5:7] = add_tem_num(label_start, label_end,
num_pos_neg[5:7])
np.savetxt('results/num_pos_neg.txt', np.array(num_pos_neg))
else:
num_pos_neg = list(np.loadtxt('results/num_pos_neg.txt'))
valid_dataset = ANetDataset(args,
args.val_data_folder,
text_proc,
raw_data,
test=False)
valid_loader = DataLoader(valid_dataset,
batch_size=1,
shuffle=False,
num_workers=args.num_workers)
return train_loader, valid_loader, text_proc, train_sampler, num_pos_neg
def get_dataset(args):
# process text
text_proc, raw_data = get_vocab_and_sentences(args.dataset_file, args.max_sentence_len)
# Create the dataset and data loader instance
test_dataset = ANetTestDataset(args.feature_root,
args.slide_window_size,
text_proc, raw_data, args.val_data_folder,
learn_mask=args.learn_mask)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=False, num_workers=args.num_workers,
collate_fn=anet_test_collate_fn)
return test_loader, text_proc
def get_dataset(args):
# process text
text_proc, raw_data = get_vocab_and_sentences(args.dataset_file,
args.max_sentence_len)
# Create the dataset and data loader instance
train_dataset = ANetDataset(
args.feature_root,
args.train_data_folder,
args.slide_window_size,
args.dur_file,
args.kernel_list,
text_proc,
raw_data,
args.pos_thresh,
args.neg_thresh,
args.stride_factor,
args.dataset,
save_samplelist=args.save_train_samplelist,
load_samplelist=args.load_train_samplelist,
sample_listpath=args.train_samplelist_path,
)
# dist parallel, optional
args.distributed = args.world_size > 1
if args.distributed and args.cuda:
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
rank=args.rank,
world_size=args.world_size)
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = None
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
sampler=train_sampler,
num_workers=args.num_workers,
collate_fn=anet_collate_fn)
valid_dataset = ANetDataset(args.feature_root,
args.val_data_folder,
args.slide_window_size,
args.dur_file,
args.kernel_list,
text_proc,
raw_data,
args.pos_thresh,
args.neg_thresh,
args.stride_factor,
args.dataset,
save_samplelist=args.save_valid_samplelist,
load_samplelist=args.load_valid_samplelist,
sample_listpath=args.valid_samplelist_path)
valid_loader = DataLoader(valid_dataset,
batch_size=args.valid_batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=anet_collate_fn)
return train_loader, valid_loader, text_proc, train_sampler
def get_dataset(args):
# process text
text_proc, train_raw_data, val_raw_data = get_vocab_and_sentences(args.train_dataset_file, verbose=False)
print("len vocab:", len(text_proc.vocab))
# Create the dataset and data loader instance
en_train_dataset = ANetDataset(args.feature_root,
args.train_data_folder,
text_proc, train_raw_data,
language="en",
save_samplelist=args.save_train_samplelist,
load_samplelist=args.load_train_samplelist,
sample_listpath=args.train_samplelist_path,
verbose=False)
ch_train_dataset = ANetDataset(args.feature_root,
args.train_data_folder,
text_proc, train_raw_data,
language="ch",
save_samplelist=args.save_train_samplelist,
load_samplelist=args.load_train_samplelist,
sample_listpath=args.train_samplelist_path,
verbose=False)
print("size of English train dataset:", len(en_train_dataset))
print("size of Chinese train dataset:", len(ch_train_dataset))
# dist parallel, optional
args.distributed = args.world_size > 1
if args.distributed and torch.cuda.is_available():
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
world_size=args.world_size)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
else:
train_sampler = None
en_train_loader = DataLoader(en_train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None), sampler=train_sampler,
num_workers=args.num_workers,
collate_fn=anet_collate_fn)
ch_train_loader = DataLoader(ch_train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None), sampler=train_sampler,
num_workers=args.num_workers,
collate_fn=anet_collate_fn)
en_valid_dataset = ANetDataset(args.feature_root,
args.val_data_folder,
text_proc, val_raw_data,
language="en",
dset="validation",
save_samplelist=args.save_valid_samplelist,
load_samplelist=args.load_valid_samplelist,
sample_listpath=args.valid_samplelist_path,
verbose=False)
en_valid_loader = DataLoader(en_valid_dataset,
batch_size=args.valid_batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=anet_collate_fn)
ch_valid_dataset = ANetDataset(args.feature_root,
args.val_data_folder,
text_proc, val_raw_data,
language="ch",
dset="validation",
save_samplelist=args.save_valid_samplelist,
load_samplelist=args.load_valid_samplelist,
sample_listpath=args.valid_samplelist_path,
verbose=False)
ch_valid_loader = DataLoader(ch_valid_dataset,
batch_size=args.valid_batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=anet_collate_fn)
return ({"entrain":en_train_loader, "chtrain":ch_train_loader,
"envalid":en_valid_loader, "chvalid":ch_valid_loader}, text_proc, train_sampler)