# prepare directory structure
Path(args.out).mkdir(parents=True, exist_ok=True)
# prepare dataset
n_workers = 1
seq_len = 6
overlap = seq_len - 1
print('seq_len = {}, overlap = {}'.format(seq_len, overlap))
# loop over sequences
for seq in args.sequences:
# create ds to iterate
df = get_data_info(image_dir,
pose_dir,
folder_list=[seq],
seq_len_range=[seq_len, seq_len],
overlap=overlap,
sample_times=1,
shuffle=False,
sort=False)
df = df.loc[df.seq_len == seq_len] # drop last
dataset = PoseSequenceDataset(df)
dataloader = DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=n_workers)
# loop over sequence
answer = [
[0.0] * 6,
]
n_batch = len(dataloader)
else:
print('Create new data info')
if par.partition != None:
partition = par.partition
train_df, valid_df = get_partition_data_info(
partition,
par.train_video,
par.seq_len,
overlap=1,
sample_times=par.sample_times,
shuffle=True,
sort=True)
else:
print("going into get_data_info")
train_df = get_data_info(folder_list=par.train_video,
seq_len_range=par.seq_len,
overlap=1,
sample_times=par.sample_times)
valid_df = get_data_info(folder_list=par.valid_video,
seq_len_range=par.seq_len,
overlap=1,
sample_times=par.sample_times)
# save the data info
train_df.to_pickle(par.train_data_info_path)
valid_df.to_pickle(par.valid_data_info_path)
train_sampler = SortedRandomBatchSampler(train_df,
par.batch_size,
drop_last=True)
train_dataset = ImageSequenceDataset(train_df, par.resize_mode,
(par.img_w, par.img_h), par.img_means,
par.img_stds, par.minus_point_5)
print('Load model from: ', load_model_path)
# Data
n_workers = 6 #pmedina (original = 1)
seq_len = int((par.seq_len[0] + par.seq_len[1]) / 2)
overlap = seq_len - 1
print('seq_len = {}, overlap = {}'.format(seq_len, overlap))
batch_size = par.batch_size
fd = open('test_dump.txt', 'w')
fd.write('\n' + '=' * 50 + '\n')
for test_video in videos_to_test:
df = get_data_info(folder_list=[test_video],
seq_len_range=[seq_len, seq_len],
overlap=overlap,
sample_times=1,
shuffle=False,
sort=False)
df = df.loc[df.seq_len == seq_len] # drop last
dataset = ImageSequenceDataset(df, par.resize_mode,
(par.img_w, par.img_h), par.img_means,
par.img_stds, par.minus_point_5)
df.to_csv('test_df.csv')
dataloader = DataLoader(dataset,
batch_size=batch_size,
shuffle=False,
num_workers=n_workers)
gt_pose = np.load('{}{}.npy'.format(par.pose_dir,
test_video)) # (n_images, 6)
format(args.partition))
train_df, valid_df = get_partition_data_info(image_dir,
pose_dir,
args.partition,
args.train_sequences,
par.seq_len,
overlap=1,
sample_times=par.sample_times,
max_step=par.max_step)
else: # case: create training and validtion dataset from given list of sequences
print('make train data from sequences: {} (dataset: {})'.format(
args.train_sequences, args.dataset))
train_df = get_data_info(image_dir,
pose_dir,
folder_list=args.train_sequences,
seq_len_range=par.seq_len,
overlap=1,
sample_times=par.sample_times,
max_step=par.max_step)
print('make validation data from sequences: {} (dataset: {})'.format(
args.valid_sequences, args.dataset))
valid_df = get_data_info(image_dir,
pose_dir,
folder_list=args.valid_sequences,
seq_len_range=par.seq_len,
overlap=1,
sample_times=par.sample_times,
max_step=par.max_step)
print('Create Dataset Loaders')
train_sampler = SortedRandomBatchSampler(train_df,
