inference_e2e.py

from __future__ import absolute_import, division, print_function, unicode_literals

import glob
import os
import numpy as np
import argparse
import json
import torch
from scipy.io.wavfile import write
from env import AttrDict
from meldataset import MAX_WAV_VALUE
from models import Generator

h = None
device = None


def load_checkpoint(filepath, device):
    assert os.path.isfile(filepath)
    print("Loading '{}'".format(filepath))
    checkpoint_dict = torch.load(filepath, map_location=device)
    print("Complete.")
    return checkpoint_dict


def scan_checkpoint(cp_dir, prefix):
    pattern = os.path.join(cp_dir, prefix + '*')
    cp_list = glob.glob(pattern)
    if len(cp_list) == 0:
        return ''
    return sorted(cp_list)[-1]


def inference(a):
    generator = Generator(h).to(device)

    state_dict_g = load_checkpoint(a.checkpoint_file, device)
    generator.load_state_dict(state_dict_g['generator'])

    filelist = os.listdir(a.input_mels_dir)

    os.makedirs(a.output_dir, exist_ok=True)

    generator.eval()
    generator.remove_weight_norm()
    with torch.no_grad():
        for i, filname in enumerate(filelist):
            if not (filname.endswith(".pkl") or filname.endswith(".npy")):
                continue
            x = np.load(os.path.join(a.input_mels_dir, filname), allow_pickle=True)
            if type(x) == dict:
                x = x["mel"]

            if type(x) is torch.Tensor:
                x = x.detach().numpy()

            if len(x.shape) != 3:
                x = np.expand_dims(x,0)

            if x.shape[1] != h.num_mels:
                x = np.swapaxes(x,1,2)
                assert x.shape[1] == h.num_mels

            x = x - x.max()
            assert len(x.shape) == 3
            assert type(x) is np.ndarray

            print(f"mel: min: {x.min():.2f}, mean: {x.mean():.2f}, median: {np.median(x):.2f}, max: {x.max():.2f}")

            x = torch.FloatTensor(x).to(device)
            y_g_hat = generator(x)
            audio = y_g_hat.squeeze()
            audio = audio * MAX_WAV_VALUE
            if MAX_WAV_VALUE > 32768:
                bitdepth = 'int32'
            else:
                bitdepth = 'int16'
            audio = audio.cpu().numpy().astype(bitdepth)

            output_file = os.path.join(a.output_dir, os.path.splitext(filname)[0] + a.file_suffix + '.wav')
            write(output_file, h.sampling_rate, audio)
            print(output_file)


def main():
    print('Initializing Inference Process..')

    parser = argparse.ArgumentParser()
    parser.add_argument('--input_mels_dir', default='test_mel_files')
    parser.add_argument('--output_dir', default='generated_files_from_mel')
    parser.add_argument('--checkpoint_file', required=True)
    parser.add_argument('--file_suffix', default="_generated_e2e")
    a = parser.parse_args()

    config_file = os.path.join(os.path.split(a.checkpoint_file)[0], 'config.json')
    with open(config_file) as f:
        data = f.read()

    global h
    json_config = json.loads(data)
    h = AttrDict(json_config)

    torch.manual_seed(h.seed)
    global device
    if torch.cuda.is_available():
        torch.cuda.manual_seed(h.seed)
        device = torch.device('cuda')
    else:
        device = torch.device('cpu')

    inference(a)


if __name__ == '__main__':
    main()