def preprocess_dataset(split_name, data_split_name, batch_stream_length, aligned=False):
filenames = [(POOLFEAT_FILE_PATTERN.format(data_split_name),
SENTS_FILE_PATTERN.format(data_split_name))]
vocab_filename = VOCAB
output_basis_path = OUTPUT_BASIS_DIR_PATTERN % split_name
aligned = True
fsg = fc7SequenceGenerator(filenames, data_split_name, VIDEO_STREAMS,
max_frames=MAX_FRAMES, align=aligned, shuffle=True, pad=aligned,
truncate=aligned)
fsg.batch_stream_length = batch_stream_length
writer = HDF5SequenceWriter(fsg, output_dir=output_basis_path)
writer.write_to_exhaustion()
writer.write_filelists()
output_text_path = OUTPUT_TEXT_DIR_PATTERN % split_name
fsg_lines = fsg.lines
tsg = TextSequenceGenerator(fsg_lines, vocab_filename, BUFFER_SIZE,
max_words=MAX_WORDS, pad=aligned, truncate=aligned)
tsg.batch_stream_length = batch_stream_length
writer = HDF5SequenceWriter(tsg, output_dir=output_text_path)
writer.write_to_exhaustion()
writer.write_filelists()
# if not os.path.isfile(vocab_filename):
# fsg.dump_vocabulary(vocab_out_path)
out_path = OUT_CORPUS_PATH.format(data_split_name)
vid_id_order_outpath = '%s/yt_pool_%s_vidid_order_%d_%d.txt' % \
(out_path, data_split_name, BUFFER_SIZE, MAX_WORDS)
frame_sequence_outpath = '%s/yt_prepool_%s_sequence_%d_%d_recurrent.txt' % \
(out_path, data_split_name, BUFFER_SIZE, MAX_WORDS)
fsg.dump_video_file(vid_id_order_outpath, frame_sequence_outpath)
def preprocess_dataset(split_name, coco_split_name, batch_stream_length,
vocab=None, aligned=True):
with open(SPLITS_PATTERN % split_name, 'r') as split_file:
split_image_ids = [int(line) for line in split_file.readlines()]
output_dataset_name = split_name
if aligned:
output_dataset_name += '_aligned_%d' % MAX_WORDS
else:
output_dataset_name += '_unaligned'
output_path = OUTPUT_DIR_PATTERN % output_dataset_name
coco = COCO(COCO_ANNO_PATH % coco_split_name)
sg = CocoSequenceGenerator(coco, BUFFER_SIZE, split_ids=split_image_ids,
vocab=vocab, align=aligned, pad=aligned, truncate=aligned)
sg.batch_stream_length = batch_stream_length
writer = HDF5SequenceWriter(sg, output_dir=output_path)
writer.write_to_exhaustion()
writer.write_filelists()
if vocab is None:
vocab_out_path = '%s/vocabulary.txt' % OUTPUT_DIR
sg.dump_vocabulary(vocab_out_path)
image_out_path = '%s/image_list.txt' % output_path
image_dummy_labels_out_path = '%s/image_list.with_dummy_labels.txt' % output_path
sg.dump_image_file(image_out_path, image_dummy_labels_out_path)
num_outs = sg.num_outs
num_pads = sg.num_pads
num_truncates = sg.num_truncates
print 'Padded %d/%d sequences; truncated %d/%d sequences' % \
(num_pads, num_outs, num_truncates, num_outs)
return sg.vocabulary_inverted
def process_dataset(experiment_paths,
experiment_config,
pre_existing_file=None):
vocab = None
split_names = ['train', 'val']
if hasattr(experiment_config, 'debug') and experiment_config.debug:
split_names = ['train_debug']
for split_name in split_names:
if split_name == 'val' and vocab is None:
raise StandardError(
'Need a vocabulary constructed from train split')
#TODO: Does saving and reading the dataset from a pickle file save time?
if experiment_config.dataset == 'Google_RefExp':
dataset = GoogleRefExp(split_name, experiment_paths)
elif experiment_config.dataset == 'UNC_RefExp':
dataset = UNCRefExp(split_name, experiment_paths)
else:
raise Exception('Unknown dataset: %s' % experiment_config.dataset)
print "Processing dataset %s" % dataset.dataset_name
if experiment_config.exp_name == 'baseline':
sg = BaselineSequenceGenerator(experiment_paths,
experiment_config,
dataset,
vocab=vocab,
pre_existing_file=pre_existing_file)
elif experiment_config.exp_name.startswith('max_margin'):
sg = MaxMarginSequenceGenerator(
experiment_paths,
experiment_config,
dataset,
vocab=vocab,
pre_existing_file=pre_existing_file)
elif experiment_config.exp_name.startswith('mil_context'):
sg = MILContextSequenceGenerator(
experiment_paths,
experiment_config,
dataset,
vocab=vocab,
pre_existing_file=pre_existing_file)
else:
raise StandardError("Unknown experiment name %s" %
experiment_config.exp_name)
output_dir = "%s/buffer_%d/%s_%s" % (
experiment_paths.h5_data, experiment_config.train.batch_size,
dataset.dataset_name, experiment_config.train.tag)
writer = HDF5SequenceWriter(sg, output_dir=output_dir)
writer.write_to_exhaustion()
writer.write_filelists()
sg.dataset.image_features.close()
if vocab is None:
sg.dump_vocabulary(experiment_config.vocab_file)
vocab = sg.vocabulary_inverted
print 'Padded %d/%d sequences; truncated %d/%d sequences' % \
(sg.num_pads, sg.num_outs, sg.num_truncates, sg.num_outs)
def process_dataset(split_name,
coco_split_name,
batch_stream_length,
vocab=None,
aligned=True):
with open(SPLITS_PATTERN % split_name, 'r') as split_file:
split_image_ids = [
line.strip().replace('.jpg', '')
for line in split_file.readlines()
] # line.strip(): remove the '\n' in each line
output_dataset_name = split_name
if aligned:
output_dataset_name += '_aligned_%d' % MAX_WORDS
else:
output_dataset_name += '_unaligned'
output_path = OUTPUT_DIR_PATTERN % output_dataset_name
coco = COCO()
image_root = COCO_IMAGE_PATTERN % coco_split_name
#pdb.set_trace()
#print image_root
sg = CocoSequenceGenerator(coco,
split_name,
BUFFER_SIZE,
image_root,
split_ids=split_image_ids,
vocab=vocab,
align=aligned,
pad=aligned,
truncate=aligned)
sg.batch_stream_length = batch_stream_length #batch_stream_length 100000
writer = HDF5SequenceWriter(sg, output_dir=output_path)
writer.write_to_exhaustion() # call hdf5_sequence_generator
writer.write_filelists()
if vocab is None:
vocab_out_path = '%s/vocabulary.txt' % OUTPUT_DIR
sg.dump_vocabulary(vocab_out_path)
image_out_path = '%s/image_list.txt' % output_path
image_dummy_labels_out_path = '%s/image_list.with_dummy_labels.txt' % output_path
sg.dump_image_file(image_out_path, image_dummy_labels_out_path)
num_outs = sg.num_outs
num_pads = sg.num_pads
num_truncates = sg.num_truncates
print 'Padded %d/%d sequences; truncated %d/%d sequences' % \
(num_pads, num_outs, num_truncates, num_outs)
return sg.vocabulary_inverted
def process_dataset(split_name,
coco_split_name,
batch_stream_length,
vocab=None,
aligned=True,
vocab_tag=''):
with open(SPLITS_PATTERN % split_name, 'r') as split_file:
split_image_ids = [int(line) for line in split_file.readlines()]
output_dataset_name = split_name
if aligned:
output_dataset_name += '_aligned_%d' % MAX_WORDS
else:
output_dataset_name += '_unaligned'
output_path = OUTPUT_DIR_PATTERN % output_dataset_name
#coco (I think) just puts annotations together for train/test set. Want to use trainval set for my own images (I think... need to ask Jeff)
coco = COCO(
COCO_ANNO_PATH % split_name
) #THIS IS CHANGED!!! This means that you have to have a json file with the captions for your split!
image_root = COCO_IMAGE_PATTERN % coco_split_name
sg = CocoSequenceGenerator(coco,
BUFFER_SIZE,
image_root,
split_ids=split_image_ids,
vocab=vocab,
align=aligned,
pad=aligned,
truncate=aligned)
if vocab is None:
vocab_out_path = '%s/%svocabulary.txt' % (OUTPUT_DIR, vocab_tag)
sg.dump_vocabulary(vocab_out_path)
sg.batch_stream_length = batch_stream_length
writer = HDF5SequenceWriter(sg, output_dir=output_path)
writer.write_to_exhaustion()
writer.write_filelists()
image_out_path = '%s/image_list.txt' % output_path
image_dummy_labels_out_path = '%s/image_list.with_dummy_labels.txt' % output_path
sg.dump_image_file(image_out_path, image_dummy_labels_out_path)
num_outs = sg.num_outs
num_pads = sg.num_pads
num_truncates = sg.num_truncates
print 'Padded %d/%d sequences; truncated %d/%d sequences' % \
(num_pads, num_outs, num_truncates, num_outs)
return sg.vocabulary_inverted