TTS/notebooks/ExtractTTSpectrogram.ipynb

This is a notebook to generate mel-spectrograms from a TTS model to be used for WaveRNN training.

In [ ]:

TTS_PATH = "/home/erogol/projects/"

In [ ]:

import os
import sys
sys.path.append(TTS_PATH)
import torch
import importlib
import numpy as np
from tqdm import tqdm as tqdm
from torch.utils.data import DataLoader
from TTS.models.tacotron2 import Tacotron2
from TTS.datasets.TTSDataset import MyDataset
from TTS.utils.audio import AudioProcessor
from TTS.utils.visual import plot_spectrogram
from TTS.utils.generic_utils import load_config
from TTS.datasets.preprocess import ljspeech
%matplotlib inline

import os
os.environ['CUDA_VISIBLE_DEVICES']='0'

In [ ]:

def set_filename(wav_path, out_path):
    wav_file = os.path.basename(wav_path)
    file_name = wav_file.split('.')[0]
    os.makedirs(os.path.join(out_path, "quant"), exist_ok=True)
    os.makedirs(os.path.join(out_path, "mel"), exist_ok=True)
    os.makedirs(os.path.join(out_path, "wav_gl"), exist_ok=True)
    wavq_path = os.path.join(out_path, "quant", file_name)
    mel_path = os.path.join(out_path, "mel", file_name)
    wav_path = os.path.join(out_path, "wav_gl", file_name)
    return file_name, wavq_path, mel_path, wav_path

In [ ]:

OUT_PATH = "/home/erogol/Data/LJSpeech-1.1/wavernn_4152/"
DATA_PATH = "/home/erogol/Data/LJSpeech-1.1/"
METADATA_FILE = "metadata_train.csv"
CONFIG_PATH = "/media/erogol/data_ssd/Data/models/ljspeech_models/4258_nancy/config.json"
MODEL_FILE = "/home/erogol/checkpoint_92000.pth.tar"
DRY_RUN = True   # if False, does not generate output files, only computes loss and visuals.
BATCH_SIZE = 16

use_cuda = torch.cuda.is_available()

C = load_config(CONFIG_PATH)
ap = AudioProcessor(bits=9, **C.audio)

In [ ]:

dataset = MyDataset(DATA_PATH, METADATA_FILE, C.r, C.text_cleaner, ap, ljspeech, use_phonemes=C.use_phonemes,  phoneme_cache_path=C.phoneme_cache_path)
loader = DataLoader(dataset, batch_size=BATCH_SIZE, num_workers=4, collate_fn=dataset.collate_fn, shuffle=False, drop_last=False)

In [ ]:

from utils.text.symbols import symbols, phonemes
from utils.generic_utils import sequence_mask
from layers.losses import L1LossMasked
# load the model
MyModel = importlib.import_module('TTS.models.'+C.model.lower())
MyModel = getattr(MyModel, C.model)
num_chars = len(phonemes) if C.use_phonemes else len(symbols)
model = MyModel(num_chars, C.r, attn_win=False)
checkpoint = torch.load(MODEL_FILE)
model.load_state_dict(checkpoint['model'])
print(checkpoint['step'])
model.eval()
if use_cuda:
    model = model.cuda()

Generate model outputs

In [ ]:

import pickle

file_idxs = []
losses = []
postnet_losses = []
criterion = L1LossMasked()
for data in tqdm(loader):
    # setup input data
    text_input = data[0]
    text_lengths = data[1]
    linear_input = data[2]
    mel_input = data[3]
    mel_lengths = data[4]
    stop_targets = data[5]
    item_idx = data[6]
    
    # dispatch data to GPU
    if use_cuda:
        text_input = text_input.cuda()
        text_lengths = text_lengths.cuda()
        mel_input = mel_input.cuda()
        mel_lengths = mel_lengths.cuda()
#         linear_input = linear_input.cuda()
        stop_targets = stop_targets.cuda()
    
    mask = sequence_mask(text_lengths)
    mel_outputs, mel_postnet_outputs, alignments, stop_tokens = model.forward(text_input, text_lengths, mel_input, mask)
    
    loss = criterion(mel_outputs, mel_input, mel_lengths)
    loss_postnet = criterion(mel_postnet_outputs, mel_input, mel_lengths)
    losses.append(loss.item())
    postnet_losses.append(loss_postnet.item())
    if not DRY_RUN:
        for idx in range(text_input.shape[0]):
            wav_file_path = item_idx[idx]
            wav = ap.load_wav(wav_file_path)
            file_name, wavq_path, mel_path, wav_path = set_filename(wav_file_path, OUT_PATH)
            file_idxs.append(file_name)

            # quantize and save wav
            wavq = ap.quantize(wav)
            np.save(wavq_path, wavq)

            # save TTS mel
            mel = mel_postnet_outputs[idx]
            mel = mel.data.cpu().numpy()
            mel_length = mel_lengths[idx]
            mel = mel[:mel_length, :].T
            np.save(mel_path, mel)

            # save GL voice
    #         wav_gen = ap.inv_mel_spectrogram(mel.T) # mel to wav
    #         wav_gen = ap.quantize(wav_gen)
    #         np.save(wav_path, wav_gen)

if not DRY_RUN:
    pickle.dump(file_idxs, open(OUT_PATH+"/dataset_ids.pkl", "wb"))      
    

print(np.mean(losses))
print(np.mean(postnet_losses))

Check model performance

In [ ]:

idx = 1
mel_example = mel_postnet_outputs[idx].data.cpu().numpy()
plot_spectrogram(mel_example[:mel_lengths[idx], :], ap);
print(mel_example[:mel_lengths[1], :].shape)

In [ ]:

wav = ap.load_wav(item_idx[idx])
melt = ap.melspectrogram(wav)
print(melt.shape)
plot_spectrogram(melt.T, ap);

In [ ]:

from matplotlib import pylab as plt
mel_diff = mel_outputs[idx] - mel_postnet_outputs[idx]
plt.figure(figsize=(16, 10))
plt.imshow(abs(mel_diff.detach().cpu().numpy()[:mel_lengths[idx],:]).T,aspect="auto", origin="lower");
plt.colorbar()
plt.tight_layout()

In [ ]:

from matplotlib import pylab as plt
# mel = mel_poutputs[idx].detach().cpu().numpy()
mel = mel_postnet_outputs[idx].detach().cpu().numpy()
mel_diff2 = melt.T - mel[:melt.shape[1]]
plt.figure(figsize=(16, 10))
plt.imshow(abs(mel_diff2).T,aspect="auto", origin="lower");
plt.colorbar()
plt.tight_layout()