kxngluca/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

data and prep improvements

Browse Source
This commit is contained in:
James Betker 2022-03-12 15:10:11 -07:00
parent 1e87b934db
commit 896accb71f
3 changed files with 44 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
codes/data/__init__.py
View File

@ -9,6 +9,7 @@ from utils.util import opt_get
def create_dataloader(dataset, dataset_opt, opt=None, sampler=None, collate_fn=None, shuffle=True):
phase = dataset_opt['phase']
pin_memory = opt_get(dataset_opt, ['pin_memory'], True)
if phase == 'train':
if opt_get(opt, ['dist'], False):
world_size = torch.distributed.get_world_size()
@ -20,11 +21,11 @@ def create_dataloader(dataset, dataset_opt, opt=None, sampler=None, collate_fn=N
batch_size = dataset_opt['batch_size']
return torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=shuffle,
num_workers=num_workers, sampler=sampler, drop_last=True,
pin_memory=True, collate_fn=collate_fn)
pin_memory=pin_memory, collate_fn=collate_fn)
else:
batch_size = dataset_opt['batch_size'] or 1
return torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=0,
pin_memory=True, collate_fn=collate_fn)
pin_memory=pin_memory, collate_fn=collate_fn)
def create_dataset(dataset_opt, return_collate=False):

20
codes/scripts/audio/preparation/phase_2_sample_and_filter.py
View File

@ -4,6 +4,7 @@ import os
import shutil
import sys
from multiprocessing.pool import ThreadPool
from random import shuffle
import torch
import torch.nn as nn
@ -53,18 +54,19 @@ class AudioFolderDataset(torch.utils.data.Dataset):
def process_folder(folder, output_path, base_path, progress_file, max_files):
classifier = load_model_from_config(args.classifier_model_opt, model_name='classifier', also_load_savepoint=True).cuda().eval()
dataset = AudioFolderDataset(folder, sampling_rate=22050, pad_to=600000)
if len(dataset) == 0:
return
dataloader = DataLoader(dataset, batch_size=32, shuffle=True, num_workers=2, pin_memory=True)
spec_injector = MelSpectrogramInjector({'in': 'clip', 'out': 'mel'}, {})
with torch.no_grad():
total_count = 0
for batch in tqdm(dataloader):
try:
max_len = max(batch['samples'])
clips = batch['clip'][:, :max_len].cuda()
paths = batch['path']
mels = spec_injector({'clip': clips})['mel']
padding = ceil_multiple(mels.shape[-1], 16)
mels = F.pad(mels, (0, padding))
def get_spec_mags(clip):
stft = torch.stft(clip, n_fft=22000, hop_length=1024, return_complex=True)
@ -91,19 +93,22 @@ def process_folder(folder, output_path, base_path, progress_file, max_files):
break
if total_count >= max_files:
break
except:
print("Exception encountered. Will ignore and continue. Exception info follows.")
print(sys.exc_info())
with open(progress_file, 'a', encoding='utf-8') as pf:
pf.write(output_path + "\n")
pf.write(folder + "\n")
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-path', type=str, help='Path to search for split files (should be the direct output of phase 1)',
default='Y:\\split\\big_podcast')
parser.add_argument('-progress_file', type=str, help='Place to store all folders that have already been processed', default='Y:\\split\\big_podcast\\already_processed.txt')
parser.add_argument('-progress_file', type=str, help='Place to store all folders that have already been processed', default='Y:\\filtered\\big_podcast\\already_processed.txt')
parser.add_argument('-output_path', type=str, help='Path where sampled&filtered files are sent', default='Y:\\filtered\\big_podcast')
parser.add_argument('-num_threads', type=int, help='Number of concurrent workers processing files.', default=1)
parser.add_argument('-max_samples_per_folder', type=int, help='Maximum number of clips that can be extracted from each folder.', default=2000)
parser.add_argument('-num_threads', type=int, help='Number of concurrent workers processing files.', default=6)
parser.add_argument('-max_samples_per_folder', type=int, help='Maximum number of clips that can be extracted from each folder.', default=1000)
parser.add_argument('-classifier_model_opt', type=str, help='Train/test options file that configures the model used to classify the audio clips.',
default='../options/test_noisy_audio_clips_classifier.yml')
args = parser.parse_args()
@ -114,12 +119,13 @@ if __name__ == '__main__':
fullpath = os.path.join(args.path, cast_dir)
if os.path.isdir(fullpath):
all_split_files.append(fullpath)
shuffle(all_split_files)
all_split_files = set(all_split_files)
# Load the already processed files, if present, and get the set difference.
if os.path.exists(args.progress_file):
with open(args.progress_file, 'r', encoding='utf-8') as pf:
processed = set(pf.readlines())
processed = set([l.strip() for l in pf.readlines()])
orig_len = len(all_split_files)
all_split_files = all_split_files - processed
print(f'All folders: {orig_len}, processed files: {len(processed)}; {len(all_split_files)/orig_len}% of files remain to be processed.')

4
codes/scripts/hugging_face_hub_upload.py
View File

@ -18,5 +18,5 @@ if __name__ == '__main__':
'''
# Build tokenizer vocab
mapping = tacotron_symbol_mapping()
print(json.dumps(mapping))
#mapping = tacotron_symbol_mapping()
#print(json.dumps(mapping))