batched = args.batched
samples = args.samples
target = args.target
overlap = args.overlap
file = args.file
gta = args.gta
print('\nInitialising Model...\n')
model = Model(rnn_dims=hp.voc_rnn_dims,
fc_dims=hp.voc_fc_dims,
bits=hp.bits,
pad=hp.voc_pad,
upsample_factors=hp.voc_upsample_factors,
feat_dims=hp.num_mels,
compute_dims=hp.voc_compute_dims,
res_out_dims=hp.voc_res_out_dims,
res_blocks=hp.voc_res_blocks,
hop_length=hp.hop_length,
sample_rate=hp.sample_rate).cuda()
paths = Paths(hp.data_path, hp.voc_model_id, hp.tts_model_id)
restore_path = args.weights if args.weights else paths.voc_latest_weights
model.restore(restore_path)
simple_table([('Generation Mode', 'Batched' if batched else 'Unbatched'),
('Target Samples', target if batched else 'N/A'),
('Overlap Samples', overlap if batched else 'N/A')])
msg = f'| Epoch: {e}/{epochs} ({i}/{total_iters}) | Loss: {avg_loss:#.4} | {speed:#.2} steps/s | Step: {k}k | '
stream(msg)
model.save(paths.latest_weights)
model.log(paths.log, msg)
print(' ')
print('\nInitialising Model...\n')
model = Model(rnn_dims=hp.rnn_dims,
fc_dims=hp.fc_dims,
bits=hp.bits,
pad=hp.pad,
upsample_factors=hp.upsample_factors,
feat_dims=hp.num_mels,
compute_dims=hp.compute_dims,
res_out_dims=hp.res_out_dims,
res_blocks=hp.res_blocks,
hop_length=hp.hop_length,
sample_rate=hp.sample_rate).cuda()
paths = Paths(hp.data_path, hp.model_id)
model.restore(paths.latest_weights)
optimiser = optim.Adam(model.parameters())
train_set, test_set = get_datasets(paths.data, batch_size)
train_loop(model, optimiser, train_set, test_set, lr)
batched = args.batched
target = args.target
overlap = args.overlap
input_text = args.input_text
weights_path = args.weights_path
print('\nInitialising WaveRNN Model...\n')
# Instantiate WaveRNN Model
voc_model = Model(rnn_dims=hp.voc_rnn_dims,
fc_dims=hp.voc_fc_dims,
bits=hp.bits,
pad=hp.voc_pad,
upsample_factors=hp.voc_upsample_factors,
feat_dims=hp.num_mels,
compute_dims=hp.voc_compute_dims,
res_out_dims=hp.voc_res_out_dims,
res_blocks=hp.voc_res_blocks,
hop_length=hp.hop_length,
sample_rate=hp.sample_rate).cuda()
voc_model.restore('quick_start/voc_weights/latest_weights.pyt')
print('\nInitialising Tacotron Model...\n')
# Instantiate Tacotron Model
tts_model = Tacotron(r=hp.tts_r,
embed_dims=hp.tts_embed_dims,
num_chars=len(symbols),
encoder_dims=hp.tts_encoder_dims,
args = parser.parse_args()
batch_size = args.batch_size
force_train = args.force_train
train_gta = args.gta
lr = args.lr
print('\nInitialising Model...\n')
# Instantiate WaveRNN Model
voc_model = Model(rnn_dims=hp.voc_rnn_dims,
fc_dims=hp.voc_fc_dims,
bits=hp.bits,
pad=hp.voc_pad,
upsample_factors=hp.voc_upsample_factors,
feat_dims=hp.num_mels,
compute_dims=hp.voc_compute_dims,
res_out_dims=hp.voc_res_out_dims,
res_blocks=hp.voc_res_blocks,
hop_length=hp.hop_length,
sample_rate=hp.sample_rate).cuda()
# Check to make sure the hop length is correctly factorised
assert np.cumprod(hp.voc_upsample_factors)[-1] == hp.hop_length
paths = Paths(hp.data_path, hp.voc_model_id, hp.tts_model_id)
voc_model.restore(paths.voc_latest_weights)
optimiser = optim.Adam(voc_model.parameters())