def score(batch_size=16):
valid_stream = parrot_stream('blizzard',
False, ('valid', ),
batch_size,
noise_level=0.,
labels_type='phonemes',
seq_size=100000)
itr = valid_stream.get_epoch_iterator()
costs = []
times = []
for val_X, val_mask, val_ctx, val_reset, val_noise_level in itr:
val_X = val_X.transpose((1, 0, 2))
val_mask = val_mask.T
val_ctx = val_ctx.T
start = time.time()
_loss = test_fn(val_X, val_ctx, val_mask)
costs.append(_loss)
times.append(time.time() - start)
print "\n\nValidation Completed!"
print "\tMean cost: ", np.mean(np.asarray(costs), axis=0) / EMB_DIM
print "\tMean time: ", np.mean(times)
return np.mean(np.asarray(costs), axis=0) / EMB_DIM
params_mode = 'last_'
else:
params_mode = 'best_'
args.samples_name = params_mode + args.samples_name
with open(
os.path.join(args.save_dir, "pkl",
params_mode + args.experiment_name + ".tar"),
'rb') as src:
parameters = load_parameters(src)
test_stream = parrot_stream(args.dataset,
saved_args.use_speaker, ('test', ),
args.num_samples,
args.num_steps,
sorting_mult=1,
labels_type=saved_args.labels_type,
raw_data=args.plot_raw)
data_tr = next(test_stream.get_epoch_iterator())
data_tr = {source: data for source, data in zip(test_stream.sources, data_tr)}
print "Loaded sources from test_stream: ", data_tr.keys()
features_tr = data_tr.get('features', None)
features_mask_tr = data_tr.get('features_mask', None)
speaker_tr = data_tr.get('speaker_index', None)
labels_tr = data_tr.get('labels', None)
labels_mask_tr = data_tr.get('labels_mask', None)
start_flag_tr = data_tr.get('start_flag', None)
raw_audio_tr = data_tr.get('raw_audio', None)
def sampler(save_dir, samples_name, do_post_filtering):
test_stream = datasets.parrot_stream(DATASET,
use_speaker=False,
which_sets=('test', ),
batch_size=BATCH_SIZE,
seq_size=10000)
test_iterator = test_stream.get_epoch_iterator()
latents_generated = numpy.random.normal(size=(NUM_REPEAT, LATENT_DIM))
latents_generated = lib.floatX(
numpy.tile(latents_generated, (BATCH_SIZE, 1)))
actual_so_far_raw, mask_raw, text_features_raw, reset = next(test_iterator)
text_features_raw_repeated = numpy.repeat(text_features_raw,
NUM_REPEAT,
axis=1)
samples_so_far = sample_fn(text_features_raw_repeated, latents_generated)
mask_so_far = mask_raw
actual_so_far = actual_so_far_raw.transpose((1, 0, 2))
mask_so_far_repeated = numpy.repeat(mask_so_far, NUM_REPEAT, axis=1)
norm_info_file = os.path.join(data_dir, DATASET,
'norm_info_mgc_lf0_vuv_bap_63_MVN.dat')
if not os.path.exists(os.path.join(save_dir, 'samples')):
os.makedirs(os.path.join(save_dir, 'samples'))
if not os.path.exists(os.path.join(save_dir, 'actual_samples')):
os.makedirs(os.path.join(save_dir, 'actual_samples'))
"""
TODO: Remove this commented section.
"""
for i, this_sample in enumerate(actual_so_far):
this_sample = this_sample[:int(mask_so_far.sum(axis=0)[i])]
generate_wav(this_sample,
os.path.join(save_dir, 'actual_samples'),
samples_name + '_' + str(i),
sptk_dir=SPTK_DIR,
world_dir=WORLD_DIR,
norm_info_file=norm_info_file,
do_post_filtering=do_post_filtering)
for i, this_sample in enumerate(samples_so_far):
this_sample = this_sample[:int(mask_so_far_repeated.sum(axis=0)[i])]
generate_wav(this_sample,
os.path.join(save_dir, 'samples'),
samples_name + '_' + str(i // NUM_REPEAT) + '_latent_' +
str(i % NUM_REPEAT),
sptk_dir=SPTK_DIR,
world_dir=WORLD_DIR,
norm_info_file=norm_info_file,
do_post_filtering=do_post_filtering)
generate_wav(this_sample,
os.path.join(save_dir, 'samples'),
samples_name + '_' + str(i // NUM_REPEAT) + '_latent_' +
str(i % NUM_REPEAT),
sptk_dir=SPTK_DIR,
world_dir=WORLD_DIR,
norm_info_file=norm_info_file,
do_post_filtering=do_post_filtering)
sampler(os.path.join(OUT_DIR, "samples", "initial_samples"), "sample", False)
train_stream = datasets.parrot_stream(DATASET,
use_speaker=False,
which_sets=('train', ),
batch_size=BATCH_SIZE,
seq_size=10000)
valid_stream = datasets.parrot_stream(DATASET,
use_speaker=False,
which_sets=('valid', ),
batch_size=BATCH_SIZE,
seq_size=10000)
total_iters = 0
total_time = 0.
train_costs = []
valid_costs = []
args = train_parse()
exp_name = args.experiment_name
save_dir = args.save_dir
print "Saving config ..."
with open(os.path.join(save_dir, 'config', exp_name + '.pkl'), 'w') as f:
cPickle.dump(args, f)
print "Finished saving."
w_init = initialization.IsotropicGaussian(0.01)
b_init = initialization.Constant(0.)
train_stream = parrot_stream(
args.dataset, args.use_speaker, ('train',), args.batch_size,
noise_level=args.feedback_noise_level, labels_type=args.labels_type,
seq_size=args.seq_size, raw_data=args.raw_output)
if args.feedback_noise_level is None:
val_noise_level = None
else:
val_noise_level = 0.
valid_stream = parrot_stream(
args.dataset, args.use_speaker, ('valid',), args.batch_size,
noise_level=val_noise_level, labels_type=args.labels_type,
seq_size=args.seq_size, raw_data=args.raw_output)
example_batch = next(train_stream.get_epoch_iterator())
for idx, source in enumerate(train_stream.sources):
on_unused_input='warn')
test_fn = theano.function([X, ctx, mask],
test_cost,
on_unused_input='warn')
lib.load_params(
'/data/lisa/exp/kumarrit/ppgn-speech/blizzard_tanh_recognizer_best.pkl'
)
print "Compiled Function!"
i = 0
train_stream = parrot_stream('blizzard',
False, ('train', ),
BATCH_SIZE,
noise_level=0.,
labels_type='phonemes',
seq_size=100000)
i = 0
best_score = 1000.
for i in xrange(i, NB_EPOCHS):
costs = []
iter = 0
times = []
itr = train_stream.get_epoch_iterator()
for train_X, train_mask, train_ctx, train_reset, train_noise_level in itr:
train_X = train_X.transpose((1, 0, 2))
train_mask = train_mask.T
train_ctx = train_ctx.T
iter += 1
(args.num_samples, max_length, saved_args.input_dim), dtype='float32')
for i, sample in enumerate(labels_tr):
padded_labels_tr[i, :len(sample)] = sample
features_mask_tr[i, :len(sample)] = 1.
labels_tr = padded_labels_tr
features_mask_tr = features_mask_tr.swapaxes(0, 1)
labels_tr = labels_tr.swapaxes(0, 1)
else:
test_stream = parrot_stream(args.new_dataset,
False, ('test', ),
args.num_samples,
10000,
sorting_mult=1,
labels_type=saved_args.labels_type,
quantize_features=saved_args.quantized_input)
data_tr = next(test_stream.get_epoch_iterator())
data_tr = {
source: data
for source, data in zip(test_stream.sources, data_tr)
}
print "Loaded sources from test_stream: ", data_tr.keys()
features_tr = data_tr.get('features', None)
features_mask_tr = data_tr.get('features_mask', None)
speaker_tr = data_tr.get('speaker_index', None)
labels_tr = data_tr.get('labels', None)
(args.num_samples, max_length, saved_args.input_dim), dtype='float32')
for i, sample in enumerate(labels_tr):
padded_labels_tr[i, :len(sample)] = sample
features_mask_tr[i, :len(sample)] = 1.
labels_tr = padded_labels_tr
features_mask_tr = features_mask_tr.swapaxes(0, 1)
labels_tr = labels_tr.swapaxes(0, 1)
else:
test_stream = parrot_stream(args.dataset,
saved_args.use_speaker, ('test', ),
args.num_samples,
args.num_steps,
sorting_mult=1,
labels_type=saved_args.labels_type,
quantize_features=saved_args.quantized_input,
raw_data=saved_args.raw_output)
data_tr = next(test_stream.get_epoch_iterator())
data_tr = {
source: data
for source, data in zip(test_stream.sources, data_tr)
}
print "Loaded sources from test_stream: ", data_tr.keys()
features_tr = data_tr.get('features', None)
features_mask_tr = data_tr.get('features_mask', None)
speaker_tr = data_tr.get('speaker_index', None)
labels_tr = data_tr.get('labels', None)