end = (n_batch + 1) * N_batch
cont_sentences = np.zeros([T, N_batch], dtype=np.float32)
input_sentences = np.zeros([T, N_batch], dtype=np.float32)
target_sentences = np.zeros([T, N_batch], dtype=np.float32)
bbox_coordinates = np.zeros([N_batch, 8], dtype=np.float32)
fc7_context = np.zeros([N_batch, 4096], dtype=np.float32)
fc7_local = np.zeros([N_batch, 4096], dtype=np.float32)
query = []
for n_pair in range(begin, end):
# Append 0 as dummy label
imcrop_path = resized_imcrop_dir + train_pairs[n_pair][0] + '.png 0'
imcrop_list.append(imcrop_path)
# Append 0 as dummy label
wholeim_path = image_dir + train_pairs[n_pair][3] + '.jpg 0'
wholeim_list.append(wholeim_path)
stream = retriever.sentence2vocab_indices(train_pairs[n_pair][1],
vocab_dict)
query.append(train_pairs[n_pair][1])
if len(stream) > T-1:
stream = stream[:T-1]
pad = T - 1 - len(stream)
cont_sentences[:, n_pair-begin] = [0] + [1] * len(stream) + [0] * pad
input_sentences[:, n_pair-begin] = [0] + stream + [-1] * pad
target_sentences[:, n_pair-begin] = stream + [0] + [-1] * pad
bbox_coordinates[n_pair-begin, :] = np.squeeze(train_pairs[n_pair][2])
fc7_context[n_pair-begin, :] = train_pairs[n_pair][4]
fc7_local[n_pair-begin, :] = train_pairs[n_pair][5]
h5_training_filename = './data/training/50_training_data/' + 'training_%d_to_%d' % (begin, end)
np.savez(h5_training_filename, raw_query=query, bbox_coordinates = bbox_coordinates, fc7_context = fc7_context, fc7_local = fc7_local)
# generate hdf5 files
if not os.path.isdir(save_hdf5_dir):
os.mkdir(save_hdf5_dir)
for n_batch in range(num_batch):
if (n_batch+1) % 10 == 0:
print('writing batch %d / %d' % (n_batch+1, num_batch))
begin = n_batch * N_batch
end = (n_batch + 1) * N_batch
cont_sentences = np.zeros([T, N_batch], dtype=np.float32)
input_sentences = np.zeros([T, N_batch], dtype=np.float32)
target_sentences = np.zeros([T, N_batch], dtype=np.float32)
for n_pair in range(begin, end):
# Append 0 as dummy label
image_path = image_dir + train_pairs[n_pair][0] + '.JPEG 0' # 0 as dummy label
image_list.append(image_path)
stream = retriever.sentence2vocab_indices(train_pairs[n_pair][1], vocab_dict)
if len(stream) > T-1:
stream = stream[:T-1]
pad = T - 1 - len(stream)
cont_sentences[:, n_pair-begin] = [0] + [1] * len(stream) + [0] * pad
input_sentences[:, n_pair-begin] = [0] + stream + [-1] * pad
target_sentences[:, n_pair-begin] = stream + [0] + [-1] * pad
h5_filename = save_hdf5_dir + '%d_to_%d.h5' % (begin, end)
retriever.write_batch_to_hdf5(h5_filename, cont_sentences, input_sentences, target_sentences)
hdf5_list.append(h5_filename)
util.io.save_str_list(image_list, save_image_list_file)
util.io.save_str_list(hdf5_list, save_hdf5_list_file)