def load_and_setup_model(model_name, parser, checkpoint, device, logger, forward_is_infer=False, ema=True, jitable=False):
model_parser = models.parse_model_args(model_name, parser, add_help=False)
model_args, model_unk_args = model_parser.parse_known_args()
model_config = models.get_model_config(model_name, model_args)
model = models.get_model(model_name, model_config, device, logger, forward_is_infer=forward_is_infer, jitable=jitable)
model.eval()
if checkpoint is not None:
checkpoint_data = torch.load(checkpoint, map_location="cpu")
if 'state_dict' in checkpoint_data:
sd = checkpoint_data['state_dict']
if any(key.startswith('module.') for key in sd):
sd = {k.replace('module.', ''): v for k,v in sd.items()}
TEMP_NUM_SPEAKERS = 5
symbols_embedding_dim = 384
model.speaker_emb = nn.Embedding(TEMP_NUM_SPEAKERS, symbols_embedding_dim).to(device)
if "speaker_emb.weight" not in sd:
sd["speaker_emb.weight"] = torch.rand((TEMP_NUM_SPEAKERS, 384))
model.load_state_dict(sd, strict=False)
if 'model' in checkpoint_data:
model = checkpoint_data['model']
else:
model = checkpoint_data
print(f'Loaded {model_name}')
if model_name == "WaveGlow":
model = model.remove_weightnorm(model)
model.eval()
model.device = device
return model.to(device)
def load_and_setup_big_WN(model_name, parser, checkpoint, device, logger, forward_is_infer=False, ema=True, jitable=False):
model_parser = models.parse_model_args(model_name, parser, add_help=False)
model_args, model_unk_args = model_parser.parse_known_args()
model_config = models.get_model_config(model_name, model_args)
model = models.get_model(model_name, model_config, device, logger, forward_is_infer=forward_is_infer, jitable=jitable)
if checkpoint is not None:
checkpoint_data = torch.load(checkpoint, map_location="cpu")
status = ''
if 'state_dict' in checkpoint_data:
# torch.save(checkpoint_data['state_dict'], f'{"/".join(checkpoint.split("/")[:-1])}/minimal-{checkpoint.split("/")[-1].split("-")[-1]}')
sd = checkpoint_data['state_dict']
if any(key.startswith('module.') for key in sd):
sd = {k.replace('module.', ''): v for k,v in sd.items()}
status += ' ' + str(model.load_state_dict(sd, strict=True))
else:
model = checkpoint_data['model']
print(f'Loaded {model_name}{status}')
if model_name == "WaveGlow":
model = model.remove_weightnorm(model)
model.device = device
model.eval()
return model.to(device)
def example_imagenet_train():
import tensorflow as tf
import keras
from python import models
from python import approx_conv_v2 as aconv
# both are necessary to actually get consistent output
np.random.seed(123)
tf.random.set_random_seed(123)
model = models.get_model(models.VGG16, weights=None, input_shape=(224, 224, 3))
# swap out normal conv layer with our custom layer
model = models.replace_layer_classes(
model, {keras.layers.Conv2D: aconv.MyConv2D})
model.compile(loss='categorical_crossentropy', optimizer='adam',
metrics=['accuracy'])
model.summary()
train_generator, val_generator, test_generator = load_flow_example()
model.fit_generator(train_generator, steps_per_epoch=10, epochs=1,
validation_steps=1, validation_data=val_generator)
model.evaluate_generator(test_generator, steps=2)