speaker,
start_flag,
args.num_samples,
raw_audio=raw_audio)
model = Model(cost)
model.set_parameter_values(parameters)
print "Successfully loaded the parameters."
if args.sample_one_step:
gen_x, gen_k, gen_w, gen_pi, gen_phi, gen_pi_att = \
parrot.sample_using_input(data_tr, args.num_samples)
else:
gen_x, gen_k, gen_w, gen_pi, gen_phi, gen_pi_att = parrot.sample_model(
labels_tr, labels_mask_tr, features_mask_tr, speaker_tr,
args.num_samples, args.num_steps)
print "Successfully sampled the parrot."
gen_x = gen_x.swapaxes(0, 1)
gen_phi = gen_phi.swapaxes(0, 1)
features_lengths = []
labels_lengths = []
for idx in range(args.num_samples):
# Heuristic for deciding when to end the sampling.
this_phi = gen_phi[idx]
this_labels_length = int(labels_mask_tr[idx].sum())
try:
latent_var_tr_old = numpy.tile(latent_var_tr_old, (args.num_samples, 1))
for i in range(10000):
lr_ = numpy.float32(0.001 / (1 + 3 * (i // 1000)))
cost_ = embed_learn_fn(features_tr, features_mask_tr, labels_tr,
labels_mask_tr, new_speakers_tr, 1., lr_)
costs.append(cost_)
if (i % 200) == 0:
# import ipdb; ipdb.set_trace()
# 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)
gen_x, gen_k, gen_w, gen_pi, gen_phi, gen_pi_att = parrot.sample_model(
labels_tr_old, labels_mask_tr_old, features_mask_tr_old,
new_speakers_tr_old, latent_var_tr_old,
LATENT_NUM * args.num_samples)
for j, this_sample in enumerate(gen_x.swapaxes(1, 0)):
this_sample = this_sample[:int(
features_mask_tr_old.sum(axis=0)[j])]
generate_wav(this_sample,
os.path.join(args.save_dir, 'samples',
'adaptation'),
"sample_{}_{}_iters_{}".format(
args.samples_name, j, i),
sptk_dir=args.sptk_dir,
world_dir=args.world_dir,
norm_info_file=norm_info_file,
do_post_filtering=args.do_post_filtering)
if len(costs) != 0: