def fastspeech_load(path,
s3_path,
model,
name,
normalizer,
quantized=False,
**kwargs):
check_file(path[model], s3_path[model], quantized=quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
output_nodes = ['decoder_output', 'post_mel_outputs']
outputs = {n: g.get_tensor_by_name(f'import/{n}:0') for n in output_nodes}
return Fastspeech(
X=g.get_tensor_by_name('import/Placeholder:0'),
speed_ratios=g.get_tensor_by_name('import/speed_ratios:0'),
f0_ratios=g.get_tensor_by_name('import/f0_ratios:0'),
energy_ratios=g.get_tensor_by_name('import/energy_ratios:0'),
logits=outputs,
normalizer=normalizer,
sess=generate_session(graph=g, **kwargs),
model=model,
name=name,
)
def tacotron_load(path,
s3_path,
model,
name,
normalizer,
quantized=False,
**kwargs):
check_file(path[model], s3_path[model], quantized=quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
input_nodes = ['Placeholder', 'Placeholder_1']
output_nodes = [
'decoder_output', 'post_mel_outputs', 'alignment_histories'
]
eager_g, input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return Tacotron(
input_nodes=input_nodes,
output_nodes=output_nodes,
normalizer=normalizer,
sess=generate_session(graph=g, **kwargs),
eager_g=eager_g,
model=model,
name=name,
)
def load(model, module, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={'model': 'model.pb'},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
inputs = ['mel', 'ori_vector', 'target_vector', 'mel_lengths']
outputs = ['mel_before', 'mel_after']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
speaker_vector_model = '-'.join(model.split('-')[2:])
speaker_model = speaker_vector.deep_model(speaker_vector_model, **kwargs)
magnitudes = {
'vggvox-v2': lambda x: x * 30 - 3.5,
'speakernet': lambda x: x * 3,
}
return FastVC(
input_nodes=input_nodes,
output_nodes=output_nodes,
waveform_to_mel=universal_mel,
speaker_vector=speaker_model,
magnitude=magnitude[speaker_vector_model],
sess=generate_session(graph=g, **kwargs),
model=model,
name=module,
)
def fastspeech_load(
path, s3_path, model, name, normalizer, quantized = False, **kwargs
):
check_file(path[model], s3_path[model], quantized = quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
inputs = ['Placeholder', 'speed_ratios', 'f0_ratios', 'energy_ratios']
outputs = ['decoder_output', 'post_mel_outputs']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
stats = np.load(path[model]['stats'])
return Fastspeech(
input_nodes = input_nodes,
output_nodes = output_nodes,
normalizer = normalizer,
stats = stats,
sess = generate_session(graph = g, **kwargs),
model = model,
name = name,
)
def load_stft(path,
s3_path,
model,
name,
instruments,
quantized=False,
**kwargs):
check_file(path[model], s3_path[model], quantized=quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
inputs = ['Placeholder']
outputs = [f'logits_{i}' for i in range(len(instruments))]
eager_g, input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return UNETSTFT(
input_nodes=input_nodes,
output_nodes=output_nodes,
instruments=instruments,
sess=generate_session(graph=g, **kwargs),
eager_g=eager_g,
model=model,
name=name,
)
def ctc_load(path, s3_path, model, name, quantized = False, **kwargs):
check_file(path[model], s3_path[model], quantized = quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
with open(path[model]['vocab']) as fopen:
vocab = json.load(fopen) + ['{', '}', '[']
featurizer = STTFeaturizer(normalize_per_feature = True)
return STT(
X = g.get_tensor_by_name('import/Placeholder:0'),
X_len = g.get_tensor_by_name('import/Placeholder_1:0'),
logits = g.get_tensor_by_name('import/logits:0'),
seq_lens = g.get_tensor_by_name('import/seq_lens:0'),
featurizer = featurizer,
vocab = vocab,
sess = generate_session(graph = g, **kwargs),
model = model,
name = name,
)
def load(path, s3_path, model, name, extra, label, quantized=False, **kwargs):
check_file(path[model], s3_path[model], quantized=quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
vectorizer_mapping = {
'vggvox-v1':
featurization.vggvox_v1,
'vggvox-v2':
featurization.vggvox_v2,
'deep-speaker':
featurization.deep_speaker,
'speakernet':
featurization.SpeakerNetFeaturizer(**{
**speakernet_config,
**extra
}),
}
if name == 'speaker-vector':
if model == 'speakernet':
model_class = Speakernet
else:
model_class = Speaker2Vec
else:
if model == 'speakernet':
model_class = SpeakernetClassification
else:
model_class = Classification
if model == 'speakernet':
inputs = ['Placeholder', 'Placeholder_1']
outputs = ['logits']
else:
inputs = ['Placeholder']
outputs = ['logits']
eager_g, input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return model_class(
input_nodes=input_nodes,
output_nodes=output_nodes,
vectorizer=vectorizer_mapping[model],
sess=generate_session(graph=g, **kwargs),
eager_g=eager_g,
model=model,
extra=extra,
label=label,
name=name,
)
def transducer_load(path, s3_path, model, name, quantized = False, **kwargs):
check_file(path[model], s3_path[model], quantized = quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
vocab = subword_load(path[model]['vocab'].replace('.subwords', ''))
featurizer = STTFeaturizer(normalize_per_feature = True)
time_reduction_factor = {'small-conformer': 4, 'conformer': 4}
input_nodes = [
'X_placeholder',
'X_len_placeholder',
'encoded_placeholder',
'predicted_placeholder',
'states_placeholder',
]
output_nodes = [
'encoded',
'ytu',
'new_states',
'padded_features',
'padded_lens',
'initial_states',
]
inputs = {n: g.get_tensor_by_name(f'import/{n}:0') for n in input_nodes}
outputs = {n: g.get_tensor_by_name(f'import/{n}:0') for n in output_nodes}
return Transducer(
X_placeholder = inputs['X_placeholder'],
X_len_placeholder = inputs['X_len_placeholder'],
encoded_placeholder = inputs['encoded_placeholder'],
predicted_placeholder = inputs['predicted_placeholder'],
states_placeholder = inputs['states_placeholder'],
padded_features = outputs['padded_features'],
padded_lens = outputs['padded_lens'],
encoded = outputs['encoded'],
ytu = outputs['ytu'],
new_states = outputs['new_states'],
initial_states = outputs['initial_states'],
featurizer = featurizer,
vocab = vocab,
time_reduction_factor = time_reduction_factor[model],
sess = generate_session(graph = g, **kwargs),
model = model,
name = name,
)
def load(model, module, extra, label, quantized = False, **kwargs):
path = check_file(
file = model,
module = module,
keys = {'model': 'model.pb'},
quantized = quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
vectorizer_mapping = {
'vggvox-v1': featurization.vggvox_v1,
'vggvox-v2': featurization.vggvox_v2,
'deep-speaker': featurization.deep_speaker,
'speakernet': featurization.SpeakerNetFeaturizer(
**{**speakernet_config, **extra}
),
}
if module == 'speaker-vector':
if model == 'speakernet':
model_class = Speakernet
else:
model_class = Speaker2Vec
else:
if model == 'speakernet':
model_class = SpeakernetClassification
else:
model_class = Classification
if model == 'speakernet':
inputs = ['Placeholder', 'Placeholder_1']
outputs = ['logits']
else:
inputs = ['Placeholder']
outputs = ['logits']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return model_class(
input_nodes = input_nodes,
output_nodes = output_nodes,
vectorizer = vectorizer_mapping[model],
sess = generate_session(graph = g, **kwargs),
model = model,
extra = extra,
label = label,
name = module,
)
def transducer_load(model, module, quantized = False, **kwargs):
path = check_file(
file = model,
module = module,
keys = {'model': 'model.pb', 'vocab': TRANSDUCER_VOCAB},
quantized = quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
vocab = subword_load(path['vocab'].replace('.subwords', ''))
featurizer = STTFeaturizer(normalize_per_feature = True)
time_reduction_factor = {
'small-conformer': 4,
'conformer': 4,
'large-conformer': 4,
'alconformer': 4,
}
inputs = [
'X_placeholder',
'X_len_placeholder',
'encoded_placeholder',
'predicted_placeholder',
'states_placeholder',
]
outputs = [
'encoded',
'ytu',
'new_states',
'padded_features',
'padded_lens',
'initial_states',
'greedy_decoder',
'non_blank_transcript',
'non_blank_stime',
]
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return Transducer(
input_nodes = input_nodes,
output_nodes = output_nodes,
featurizer = featurizer,
vocab = vocab,
time_reduction_factor = time_reduction_factor.get(model, 4),
sess = generate_session(graph = g, **kwargs),
model = model,
name = module,
)
def load_1d(model, module, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={'model': 'model.pb'},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
inputs = ['Placeholder']
outputs = ['logits']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return UNET1D(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
model=model,
name=module,
)
def load(path, s3_path, model, name, quantized=False, **kwargs):
check_file(path[model], s3_path[model], quantized=quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
inputs = ['Placeholder']
outputs = ['logits']
eager_g, input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return UNET(
input_nodes=input_nodes,
output_nodes=output_nodes,
sess=generate_session(graph=g, **kwargs),
eager_g=eager_g,
model=model,
name=name,
)
def load_stft(model, module, instruments, quantized=False, **kwargs):
path = check_file(
file=model,
module=module,
keys={'model': 'model.pb'},
quantized=quantized,
**kwargs,
)
g = load_graph(path['model'], **kwargs)
inputs = ['Placeholder']
outputs = [f'logits_{i}' for i in range(len(instruments))]
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
return UNETSTFT(
input_nodes=input_nodes,
output_nodes=output_nodes,
instruments=instruments,
sess=generate_session(graph=g, **kwargs),
model=model,
name=module,
)
def load(path, s3_path, model, name, quantized=False, **kwargs):
check_file(path[model], s3_path[model], quantized=quantized, **kwargs)
if quantized:
model_path = 'quantized'
else:
model_path = 'model'
g = load_graph(path[model][model_path], **kwargs)
inputs = ['mel', 'ori_vector', 'target_vector', 'mel_lengths']
outputs = ['mel_before', 'mel_after']
input_nodes, output_nodes = nodes_session(g, inputs, outputs)
speaker_model = speaker_vector.deep_model('vggvox-v2', **kwargs)
return FastVC(
input_nodes=input_nodes,
output_nodes=output_nodes,
waveform_to_mel=universal_mel,
speaker_vector=speaker_model,
sess=generate_session(graph=g, **kwargs),
model=model,
name=name,
)