def test_trainable():
gpus = deepspeech.get_available_gpus(
) # Support both Multi and Single-GPU tests
base_configuration = DeepSpeech.get_configuration(
'tests/models/base/configuration.yaml')
base_configuration.model.pop('name')
base_model = deepspeech_custom(is_gpu=len(gpus) > 0,
**base_configuration.model)
fname = 'weights.hdf5'
base_model.save_weights(fname)
extended_configuration = DeepSpeech.get_configuration(
'tests/models/extended/configuration.yaml')
extended_configuration.model.pop('name')
extended_model = deepspeech_custom(is_gpu=len(gpus) > 0,
**extended_configuration.model)
weights_before_training = extended_model.get_weights()
assert all(not extended_model.get_layer(name).trainable
for name in ['base_1', 'base_2', 'base_3'])
assert all(
extended_model.get_layer(name).trainable
for name in ['extension_1', 'extension_2'])
assert all(not is_same(
base_model.get_layer(name).get_weights(),
extended_model.get_layer(name).get_weights())
for name in ['base_1', 'base_2', 'base_3'])
loss = DeepSpeech.get_loss()
optimizer = DeepSpeech.get_optimizer(**extended_configuration.optimizer)
parallel_model = DeepSpeech.distribute_model(extended_model, gpus)
DeepSpeech.compile_model(parallel_model, optimizer, loss)
extended_model.load_weights(fname, by_name=True)
assert all(
is_same(
base_model.get_layer(name).get_weights(),
extended_model.get_layer(name).get_weights())
for name in ['base_1', 'base_2', 'base_3'])
for i in range(10): # Dummy training (10 epochs / 10 batch_size)
X = np.random.rand(10, 100, 80)
y = np.random.randint(0, 35, size=[10, 20], dtype=np.int32)
parallel_model.train_on_batch(X, y)
assert all(
is_same(
base_model.get_layer(name).get_weights(),
extended_model.get_layer(name).get_weights())
for name in ['base_1', 'base_2', 'base_3'
]), "Freezed layers have to be unchangeable."
assert not is_same(
weights_before_training,
extended_model.get_weights()), "The base model updates weights."
assert is_close(extended_model.predict(X), parallel_model.predict(
X)), "The results are the same (compiled model)."
# assert is_same(extended_model.get_weights(), compiled_model.get_weights()) # Weights can not be compared (order changed)
os.remove(fname)
def config(config_path) -> Configuration:
return DeepSpeech.get_configuration(config_path)