left_context = int(sys.argv[5])
right_context = int(sys.argv[6])
lda_path = sys.argv[7]
output_path = sys.argv[8]
num_epochs = 400
batch_size = 256
learning_rate = 0.0015
model_type = 'SAT-LHUC'
utt_to_spk = load_utt_to_spk(utt2spk, lambda x: "".join(x.split("-")[1:-1]))
utt_to_pdfs = load_utt_to_pdfs(pdfs)
num_spks = max(utt_to_spk.values()) + 1
print "num_spks = %d" % num_spks
train_dataset = load_dataset(train_data, utt_to_spk, utt_to_pdfs, chunk_size=8, subsampling_factor=1, left_context=left_context, right_context=right_context, speaker_independent_prob=0.5)
train_dataset = train_dataset.batch(batch_size, drop_remainder=True)
train_dataset = train_dataset.prefetch(1024)
x, spk, y = train_dataset.make_one_shot_iterator().get_next()
val_dataset = load_dataset(val_data, utt_to_spk, utt_to_pdfs, chunk_size=8, subsampling_factor=1, left_context=left_context, right_context=right_context, speaker_independent_prob=1.0)
val_dataset = val_dataset.batch(batch_size, drop_remainder=True)
val_dataset = val_dataset.take(512).cache().repeat()
val_x, val_spk, val_y = val_dataset.make_one_shot_iterator().get_next()
if model_type == 'SAT-LHUC':
model = create_sat_model(600, lda_path, num_spks)
elif model_type == 'SAT-SPARSE-LHUC':
model = create_sparse_lhuc_sat_model(850, 0.1 / 850 / batch_size, lda_path, num_spks)
model.compile(
utt2spk = sys.argv[3]
pdfs = sys.argv[4]
left_context = int(sys.argv[5])
right_context = int(sys.argv[6])
output_path = sys.argv[7]
# 1 epoch is roughly 25270800*0.97 = 24512676 frames
# 1 epoch is roughly 61 iterations of 400000 frames each
num_iterations = 61*6
batch_size = 256
learning_rate = 0.0015
utt_to_spk = load_utt_to_spk(utt2spk)
utt_to_pdfs = load_utt_to_pdfs(pdfs)
train_dataset = load_dataset(train_data, utt_to_spk, utt_to_pdfs, chunk_size=1, subsampling_factor=1, left_context=left_context, right_context=right_context)
train_dataset = train_dataset.batch(batch_size, drop_remainder=True)
train_dataset = train_dataset.prefetch(512)
x, _, y = train_dataset.make_one_shot_iterator().get_next()
val_dataset = load_dataset(val_data, utt_to_spk, utt_to_pdfs, chunk_size=1, subsampling_factor=1, left_context=left_context, right_context=right_context)
val_dataset = val_dataset.batch(batch_size, drop_remainder=True)
val_dataset = val_dataset.take(256).cache().repeat()
val_x, _, val_y = val_dataset.make_one_shot_iterator().get_next()
model = create_model(800, 3976)
model.compile(
loss='sparse_categorical_crossentropy',
metrics=['accuracy'],
optimizer=Adam(lr=learning_rate, amsgrad=True, clipvalue=1.)
)