vocab=vocab, max_history_length=args.max_history_length,
merge_source=args.merge_source)
logging.info('Loading validation data from ' + args.valid_set)
valid_data = dh.load(args.fea_type, args.valid_path, args.valid_set,
include_caption=args.include_caption, separate_caption=args.separate_caption,
vocab=vocab, max_history_length=args.max_history_length,
merge_source=args.merge_source)
if args.fea_type[0] == 'none':
feature_dims = 0
else:
feature_dims = dh.feature_shape(train_data)
logging.info("Detected feature dims: {}".format(feature_dims));
# report data summary
logging.info('#vocab = %d' % len(vocab))
# make batchset for training
train_indices, train_samples = dh.make_batch_indices(train_data, args.batch_size,
max_length=args.max_length, separate_caption=args.separate_caption)
logging.info('#train sample = %d' % train_samples)
logging.info('#train batch = %d' % len(train_indices))
# make batchset for validation
valid_indices, valid_samples = dh.make_batch_indices(valid_data, args.batch_size,
max_length=args.max_length, separate_caption=args.separate_caption)
logging.info('#validation sample = %d' % valid_samples)
logging.info('#validation batch = %d' % len(valid_indices))
# create_model
model = make_model(len(vocab), len(vocab),
N=args.nb_blocks, d_model=args.d_model, d_ff=args.d_ff,
h=args.att_h, dropout=args.dropout,
separate_his_embed=args.separate_his_embed,
separate_cap_embed=args.separate_cap_embed,
ft_sizes=feature_dims,
diff_encoder=args.diff_encoder,
if train_args.dictmap != '':
dictmap = json.load(open(train_args.dictmap, 'r'))
else:
dictmap = None
# report data summary
logging.info('#vocab = %d' % len(vocab))
# prepare test data
logging.info('Loading test data from ' + args.test_set)
test_data = dh.load(train_args.fea_type,
args.test_path,
args.test_set,
vocab=vocab,
dictmap=dictmap,
include_caption=train_args.include_caption)
test_indices, test_samples = dh.make_batch_indices(test_data, 1)
logging.info('#test sample = %d' % test_samples)
# generate sentences
logging.info('-----------------------generate--------------------------')
start_time = time.time()
result = generate_response(model,
test_data,
test_indices,
vocab,
maxlen=args.maxlen,
beam=args.beam,
penalty=args.penalty,
nbest=args.nbest)
logging.info('----------------')
logging.info('wall time = %f' % (time.time() - start_time))
if args.output:
dec_hsize=args.dec_hsize,
att_size=args.att_size,
sos=1,
eos=2,
ignore_label=3)
else:
print("Unknonw model type '{}' is specified.".format(args.type))
sys.exit(1)
# report data summary
print('#vocab =', len(vocab))
# make batchset for training
print('Making mini batches for training data')
train_indices = [None] * len(data)
train_indices[0], train_samples = dh.make_batch_indices(
data[0], args.train, batchsize, max_length=args.max_length)
for n in six.moves.range(1, len(data)):
train_indices[n], _ = dh.make_batch_indices(data[n],
args.train,
batchsize,
max_length=args.max_length,
reference=train_indices[0])
print('#train sample =', train_samples, ' #train batch =',
len(train_indices[0]))
if args.valid != '':
print('Making mini batches for validation data')
valid_indices = [None] * len(data)
valid_indices[0], valid_samples = dh.make_batch_indices(
data[0], args.valid, batchsize, max_length=args.max_length)
for n in six.moves.range(1, len(data)):
HLSTMDecoder(args.dec_layers,
len(vocab),
len(vocab),
args.embed_size,
args.hist_out_size + args.in_enc_hsize,
args.dec_hsize,
args.dec_psize,
independent=False,
dropout=dropout,
embed=embed_model))
initialize_model_weights(model, "he", "xavier")
# report data summary
logging.info('#vocab = %d' % len(vocab))
# make batchset for training
logging.info('Making mini batches for training data')
train_indices, train_samples = dh.make_batch_indices(
train_data, args.batch_size, max_length=args.max_length)
logging.info('#train sample = %d' % train_samples)
logging.info('#train batch = %d' % len(train_indices))
# make batchset for validation
logging.info('Making mini batches for validation data')
valid_indices, valid_samples = dh.make_batch_indices(
valid_data, args.batch_size, max_length=args.max_length)
logging.info('#validation sample = %d' % valid_samples)
logging.info('#validation batch = %d' % len(valid_indices))
# copy model to gpu
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
# save meta parameters
path = args.model + '.conf'
with open(path, 'wb') as f:
pickle.dump((vocab, args), f, -1)
model = nn.DataParallel(model)
model.to(device)
# prepare test data
print 'Making mini batches for test data from', args.test
data = []
for n, feafile in enumerate(args.feafile):
feature_data = dh.load(feafile, args.capfile, vocab=vocab)
data.extend(dh.check_feature_shape(feature_data))
test_indices = [None] * len(args.feafile)
for n, feafile in enumerate(args.feafile):
test_indices[n], test_samples = dh.make_batch_indices(data[n],
args.test,
1,
test=True)
print 'Feature[%d]: #test sample = %d #test batch = %d' % (
n, test_samples, len(test_indices[n]))
print '#vocab =', len(vocab)
# generate sentences
print '-----------------------generate--------------------------'
start_time = time.time()
result = generate_caption(model,
data,
test_indices,
vocab,
train_args.in_size,
stride=train_args.frame_stride,
maxlen=args.maxlen,
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
# report data summary
logging.info('#vocab = %d' % len(vocab))
# prepare test data
logging.info('Loading test data from ' + args.test_set)
test_data = dh.load(train_args.fea_type,
args.test_path,
args.test_set,
vocab=vocab,
include_caption=train_args.include_caption,
separate_caption=train_args.separate_caption,
max_history_length=train_args.max_history_length,
merge_source=train_args.merge_source,
undisclosed_only=args.undisclosed_only)
test_indices, test_samples = dh.make_batch_indices(
test_data, 1, separate_caption=train_args.separate_caption)
logging.info('#test sample = %d' % test_samples)
# generate sentences
logging.info('-----------------------generate--------------------------')
start_time = time.time()
labeled_test = None
if args.undisclosed_only and args.labeled_test is not None:
labeled_test = json.load(open(args.labeled_test, 'r'))
result = generate_response(model,
test_data,
test_indices,
vocab,
maxlen=args.maxlen,
beam=args.beam,
penalty=args.penalty,
nbest=args.nbest,