def test_embeddings_everywhere(self):
hparams = ModelTrainer.create_hparams()
num_emb = 3
emb_dim = 12
in_dim = 42
out_dim = 12
hparams.add_hparam("f_get_emb_index", [lambda x: 0])
hparams.model_type = "RNNDYN-{}x{}_EMB_(-1)-3_RELU_128-2_BiLSTM_32-1_FC_12".format(
num_emb, emb_dim)
model = ModelFactory.create(hparams.model_type, (in_dim, ), out_dim,
hparams)
self.assertEqual(1, len(model.emb_groups))
self.assertEqual(torch.Size([num_emb, emb_dim]),
model.emb_groups[0].weight.shape)
self.assertEqual(torch.Size([128, in_dim - 1 + emb_dim]),
model[0].weight.shape)
self.assertEqual(torch.Size([128, 128 + emb_dim]),
model[1].weight.shape)
self.assertEqual(torch.Size([32 * 4, 128 + emb_dim]),
model[3].weight_ih_l0.shape)
self.assertEqual(torch.Size([32 * 4, 32 * 2 + emb_dim]),
model[4].weight_ih_l0_reverse.shape)
pass
def create_hparams(hparams_string=None, verbose=False):
hparams = ModelTrainer.create_hparams(hparams_string, verbose=False)
hparams.add_hparams(
thetas=None, # One initial theta value per filter.
k=2, # Order of the impulse response of the atoms.
min_atom_amp=
0.25, # Post-processing removes atoms with an absolute amplitude smaller than this.
complex_poles=True, # Comples poles possible.
phase_init=0.0, # Initial phase of the filters.
vuv_loss_weight=1.0, # Weight of the VUV RMSE.
L1_loss_weight=1.0, # Weight of the L1 loss on the spiking inputs.
weight_unvoiced=0.5, # Weight on unvoiced frames.
num_questions=None, # Dimension of the input questions.
dist_window_size=
51, # Size of distribution around spikes when training the AtomModel.
phrase_bias_init=
0.0, # Initial bias of neural filter, should be estimated mean of speaker's LF0.
atom_model_path=None, # Path to load a pre-trained atom model from.
hparams_atom=
None, # Hyper-parameter container used in the AtomModelTrainer
flat_model_path=
None, # Path to load a pre-trained atom neural filter model from (without phrase curve).
hparams_flat=
None, # Hyper-parameter container used in the AtomNeuralFilterModelTrainer.
)
if verbose:
logging.info(hparams.get_debug_string())
return hparams
def test_save_load_equality(self):
hparams = ModelTrainer.create_hparams()
hparams.out_dir = os.path.join(
self.out_dir,
"test_save_load_equality") # Add function name to path.
model_path = os.path.join(hparams.out_dir, "test_model.nn")
# Create a new model and save it.
dim_in, dim_out = 10, 4
total_epochs = 10
model_handler = ModelHandlerPyTorch()
model_handler.model = torch.nn.Sequential(
torch.nn.Linear(dim_in, dim_out))
model_handler.save_checkpoint(model_path, total_epochs)
# Create a new model handler and test load save.
hparams.model_type = None
model_handler = ModelHandlerPyTorch()
saved_total_epochs = model_handler.load_checkpoint(model_path, hparams)
self.assertEqual(total_epochs,
saved_total_epochs,
msg="Saved and loaded total epochs do not match")
model_copy_path = os.path.join(hparams.out_dir, "test_model_copy.nn")
model_handler.save_checkpoint(model_copy_path, total_epochs)
# self.assertTrue(filecmp.cmp(model_path, model_copy_path, False)) # This does not work.
self.assertTrue(equal_checkpoint(model_path, model_copy_path),
"Loaded and saved models are not the same.")
shutil.rmtree(hparams.out_dir)
def create_hparams(hparams_string=None, verbose=False):
hparams = ModelTrainer.create_hparams(hparams_string, verbose=False)
hparams.add_hparams(thetas=None,
k=None,
min_atom_amp=0.3,
num_questions=None)
if verbose:
logging.info(hparams.get_debug_string())
return hparams
def create_hparams(hparams_string=None, verbose=False):
hparams = ModelTrainer.create_hparams(hparams_string, verbose=False)
hparams.add_hparams( # exclude_begin_and_end_silence=False,
min_phoneme_length=50000,
phoneme_label_type="HTK full"
) # Specifies the format in which the .lab files are stored.
# Refer to PhonemeLabelGen.load_sample for a list of possible types.
if verbose:
logging.info(hparams.get_debug_string())
return hparams
def test_get_item(self):
hparams = ModelTrainer.create_hparams()
num_emb = 3
emb_dim = 12
in_dim = 42
out_dim = 12
hparams.add_hparam("f_get_emb_index", [lambda x: 0])
hparams.model_type = "RNNDYN-{}x{}_EMB_(0, 3, 5, 7)-5_RELU_128-3_BiLSTM_32-1_FC_12".format(
num_emb, emb_dim)
model = ModelFactory.create(hparams.model_type, (in_dim, ), out_dim,
hparams)
self.assertEqual(model.layer_groups[0][1], model[1])
self.assertEqual(model.layer_groups[1][0], model[3])
self.assertEqual(model.layer_groups[2][0], model[6])
def test_embeddings(self):
hparams = ModelTrainer.create_hparams()
num_emb = 3
emb_dim = 12
in_dim = 42 # Contains the embedding index.
out_dim = 12
hparams.variable_sequence_length_train = True
hparams.add_hparam("f_get_emb_index", [lambda x: 0])
hparams.model_type = "RNNDYN-{}x{}_EMB_(0, 3, 5, 7)-5_RELU_128-3_BiLSTM_32-1_FC_12".format(
num_emb, emb_dim)
# hparams.model_type = "RNNDYN-{}x{}_EMB_(-1)-5_RELU_128-2_BiLSTM_32-1_FC_12".format(num_emb, emb_dim)
model = ModelFactory.create(hparams.model_type, (in_dim, ), out_dim,
hparams)
self.assertEqual(1, len(model.emb_groups))
self.assertEqual(torch.Size([num_emb, emb_dim]),
model.emb_groups[0].weight.shape)
self.assertEqual(torch.Size([128, in_dim - 1 + emb_dim]),
model[0].weight.shape)
self.assertEqual(torch.Size([128, 128]), model[2].weight.shape)
self.assertEqual(torch.Size([128, 128 + emb_dim]),
model[3].weight.shape)
self.assertEqual(torch.Size([32 * 4, 128 + emb_dim]),
model[5].weight_ih_l0.shape)
self.assertEqual(torch.Size([32 * 4, 32 * 2 + emb_dim]),
model[7].weight_ih_l0_reverse.shape)
seq_length = torch.tensor((100, 75), dtype=torch.long)
batch_size = 2
test_input = torch.ones([seq_length[0], batch_size, in_dim])
model.init_hidden(batch_size)
output = model(test_input, None, seq_length, seq_length[0])
self.assertEqual(torch.Size([seq_length[0], batch_size, out_dim]),
output[0].shape)
seq_length = torch.tensor((100, ), dtype=torch.long)
batch_size = 1
test_input = torch.ones([seq_length[0], batch_size, in_dim])
model.init_hidden(batch_size)
output = model(test_input, None, seq_length, seq_length[0])
self.assertEqual(torch.Size([seq_length[0], batch_size, out_dim]),
output[0].shape)
def create_hparams(hparams_string=None, verbose=False):
"""Create model hyper parameter container. Parse non default from given string."""
hparams = ModelTrainer.create_hparams(hparams_string, verbose=False)
hparams.add_hparams(
num_questions=None,
question_file=None, # Used to add labels in plot.
num_coded_sps=60,
sp_type="mcep",
add_deltas=True,
synth_load_org_sp=False,
synth_load_org_lf0=False,
synth_load_org_vuv=False,
synth_load_org_bap=False)
if verbose:
logging.info(hparams.get_debug_string())
return hparams
def _get_hparams(self):
hparams = ModelTrainer.create_hparams()
# General parameters
hparams.add_hparam("num_questions", 409)
hparams.epochs = 0
hparams.test_set_perc = 0.05
hparams.val_set_perc = 0.05
hparams.optimiser_args["lr"] = 0.02
hparams.seed = None # Remove the default seed.
hparams.out_dir = os.path.join(
os.path.dirname(os.path.realpath(__file__)),
type(self).__name__)
hparams.num_coded_sps = 20
# Training parameters.
hparams.epochs = 0
hparams.model_name = "test_model.nn"
return hparams
def create_hparams(hparams_string=None, verbose=False):
"""Create model hyper-parameters. Parse non-default from given string."""
hparams = ModelTrainer.create_hparams(hparams_string, verbose=False)
hparams.synth_vocoder = "raw"
hparams.add_hparams(
batch_first=True,
frame_rate_output_Hz=16000,
mu=255,
bit_depth=16,
silence_threshold_quantized=
None, # Beginning and end of audio below the threshold are trimmed.
teacher_forcing_in_test=True,
ema_decay=0.9999,
# Model parameters.
input_type="mulaw-quantize",
hinge_regularizer=
True, # Only used in MoL prediction (input_type="raw").
log_scale_min=float(np.log(
1e-14)), # Only used for mixture of logistic distributions.
quantize_channels=256
) # 256 for input type mulaw-quantize, otherwise 65536
if hparams.input_type == "mulaw-quantize":
hparams.add_hparam("out_channels", hparams.quantize_channels)
else:
hparams.add_hparam("out_channels", 10 *
3) # num_mixtures * 3 (pi, mean, log_scale)
hparams.add_hparams(
layers=24, # 20
stacks=4, # 2
residual_channels=512,
gate_channels=512,
skip_out_channels=256,
dropout=1 - 0.95,
kernel_size=3,
weight_normalization=True,
use_cond=True, # Determines if conditioning is used.
cin_channels=63,
upsample_conditional_features=False,
upsample_scales=[5, 4, 2])
if hparams.upsample_conditional_features:
hparams.len_in_out_multiplier = reduce(mul,
hparams.upsample_scales, 1)
else:
hparams.len_in_out_multiplier = 1
hparams.add_hparams(freq_axis_kernel_size=3,
gin_channels=-1,
n_speakers=1,
use_speaker_embedding=False,
sp_type="mcep",
load_sp=True,
load_lf0=True,
load_vuv=True,
load_bap=True)
if verbose:
logging.info(hparams.get_debug_string())
return hparams