Upload run.py with huggingface_hub

64850d8 verified 20 days ago

5.44 kB

import os, sys, subprocess, shutil

WORK = '/kaggle/working'
os.chdir(WORK)

# 1) system deps
subprocess.run(['apt-get', 'update', '-qq'], check=False, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
subprocess.run(['apt-get', 'install', '-y', '-qq', 'espeak-ng'], check=False, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)

# 2) python deps
subprocess.run([sys.executable, '-m', 'pip', 'install', '-q',
                'torch', 'torchaudio', 'soundfile', 'huggingface_hub', 'misaki',
                'transformers', 'numpy', 'tqdm', 'loguru', 'attrs', 'phonemizer', 'num2words'], check=False)

from huggingface_hub import hf_hub_download, snapshot_download, upload_file

# 3) download code files from HF dataset
for fn in ['train_student.py', 'dump_keme.py', 'config_keme_student.json', '_env.py']:
    p = os.path.join(WORK, fn)
    if not os.path.exists(p):
        hf_hub_download('0best/kemetone-data', fn, repo_type='dataset', local_dir=WORK)
sys.path.insert(0, WORK)

# 4) clone StyleTTS2
st = os.path.join(WORK, 'StyleTTS2')
if not os.path.exists(st):
    subprocess.run(['git', 'clone', '--depth', '1',
                    'https://github.com/yl4579/StyleTTS2.git', st], check=True)
sys.path.insert(0, st)
sys.path.insert(0, WORK)

# 5) download KemeTone (model + g2p package)
model_dir = os.path.join(WORK, 'kemetone_model')
if not os.path.exists(model_dir):
    snapshot_download('Rabe3/kemetone', local_dir=model_dir, ignore_patterns=['*.wav','*.png','*.jpg'])
if model_dir not in sys.path:
    sys.path.insert(0, model_dir)
# patch g2p for misaki return arity
g2p_path = os.path.join(model_dir, 'kemetone', 'g2p.py')
if os.path.exists(g2p_path):
    src = open(g2p_path).read()
    src = src.replace("        raw, _ = self._g2p(text)\n",
                      "        _r = self._g2p(text)\n        raw = _r[0] if isinstance(_r, (tuple, list)) else _r\n")
    open(g2p_path, 'w').write(src)

# 6) download patched kokoro from Nabra-Saudi-7M
patched = os.path.join(WORK, 'kokoro_patched')
if not os.path.exists(patched):
    snapshot_download('oddadmix/Nabra-Saudi-7M', local_dir=os.path.join(WORK,'_nabra'),
                      allow_patterns=['kokoro_patched/**'])
    src = os.path.join(WORK, '_nabra', 'kokoro_patched')
    if os.path.exists(src):
        shutil.move(src, patched)
sys.path.insert(0, patched)

# 7) create styletts2_gan shim
shim = os.path.join(WORK, 'styletts2_gan')
os.makedirs(shim, exist_ok=True)
open(os.path.join(shim, '__init__.py'), 'w').write("import sys\nsys.path.insert(0,'/kaggle/working/StyleTTS2')\n")
open(os.path.join(shim, 'discriminators.py'), 'w').write(
    "import sys\nsys.path.insert(0,'/kaggle/working/StyleTTS2')\n"
    "from Modules.discriminators import MultiPeriodDiscriminator, MultiResSpecDiscriminator, WavLMDiscriminator\n")
open(os.path.join(shim, 'losses_styletts2.py'), 'w').write(
    "import sys\nsys.path.insert(0,'/kaggle/working/StyleTTS2')\n"
    "from losses import GeneratorLoss, DiscriminatorLoss, MultiResolutionSTFTLoss, WavLMLoss\n")
if WORK not in sys.path:
    sys.path.insert(0, WORK)

# 8) dump teacher (if not already done)
dump_dir = os.path.join(WORK, 'dump')
index_path = os.path.join(dump_dir, 'index.jsonl')
data_file = os.path.join(WORK, 'data.txt')
if not os.path.exists(data_file):
    data_file = hf_hub_download('0best/kemetone-data', 'kemetone_TTS_CLEAN_DEDUPED_PART1.txt',
                                repo_type='dataset', local_dir=WORK)
if not os.path.exists(index_path):
    os.environ['TEXT_FILE'] = data_file
    os.environ['OUT_DIR'] = dump_dir
    os.environ['MODEL_DIR'] = model_dir
    subprocess.run([sys.executable, 'dump_keme.py'], check=True, cwd=WORK)

# 9) train student in chunks of 1000 steps, resuming each time
out_dir = os.path.join(WORK, 'runs', 'egypt7m')
os.makedirs(out_dir, exist_ok=True)

CHUNK = 1000
TOTAL_CHUNKS = 10  # 10 * 1000 = 10000 steps on part 1

for chunk in range(1, TOTAL_CHUNKS + 1):
    target_steps = chunk * CHUNK
    print(f"\n=== Chunk {chunk}/{TOTAL_CHUNKS}: training to step {target_steps} ===", flush=True)

    cmd = [sys.executable, 'train_student.py',
           '--index', index_path,
           '--config', 'config_keme_student.json',
           '--out', out_dir,
           '--steps', str(target_steps),
           '--batch', '8',
           '--lr', '2e-4',
           '--lr-decay',
           '--mel-weight', '5',
           '--sil-weight', '100',
           '--slm-weight', '1',
           '--slm-adv-weight', '0.2',
           '--ckpt-every', '1000',
           '--resume']
    subprocess.run(cmd, check=True, cwd=WORK)

    # Upload checkpoints to HF
    for ckpt_file in [os.path.join(out_dir, 'last.pt'), os.path.join(out_dir, 'student.pth')]:
        if os.path.exists(ckpt_file):
            try:
                token = os.environ.get('HF_TOKEN')
                upload_file(
                    path_or_fileobj=ckpt_file,
                    path_in_repo=f"runs/egypt7m/chunk{chunk}/" + os.path.basename(ckpt_file),
                    repo_id="0best/kemetone-distill-ckpts",
                    repo_type="model",
                    token=token,
                )
                print(f"[*] uploaded {os.path.basename(ckpt_file)} to HF (chunk {chunk})", flush=True)
            except Exception as e:
                print(f"[*] upload failed for {ckpt_file}: {e}", flush=True)

print('TRAINING DONE - all 10 chunks completed (10,000 steps on part 1)')

	import os, sys, subprocess, shutil

	WORK = '/kaggle/working'
	os.chdir(WORK)

	# 1) system deps
	subprocess.run(['apt-get', 'update', '-qq'], check=False, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
	subprocess.run(['apt-get', 'install', '-y', '-qq', 'espeak-ng'], check=False, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)

	# 2) python deps
	subprocess.run([sys.executable, '-m', 'pip', 'install', '-q',
	'torch', 'torchaudio', 'soundfile', 'huggingface_hub', 'misaki',
	'transformers', 'numpy', 'tqdm', 'loguru', 'attrs', 'phonemizer', 'num2words'], check=False)

	from huggingface_hub import hf_hub_download, snapshot_download, upload_file

	# 3) download code files from HF dataset
	for fn in ['train_student.py', 'dump_keme.py', 'config_keme_student.json', '_env.py']:
	p = os.path.join(WORK, fn)
	if not os.path.exists(p):
	hf_hub_download('0best/kemetone-data', fn, repo_type='dataset', local_dir=WORK)
	sys.path.insert(0, WORK)

	# 4) clone StyleTTS2
	st = os.path.join(WORK, 'StyleTTS2')
	if not os.path.exists(st):
	subprocess.run(['git', 'clone', '--depth', '1',
	'https://github.com/yl4579/StyleTTS2.git', st], check=True)
	sys.path.insert(0, st)
	sys.path.insert(0, WORK)

	# 5) download KemeTone (model + g2p package)
	model_dir = os.path.join(WORK, 'kemetone_model')
	if not os.path.exists(model_dir):
	snapshot_download('Rabe3/kemetone', local_dir=model_dir, ignore_patterns=['.wav','.png','*.jpg'])
	if model_dir not in sys.path:
	sys.path.insert(0, model_dir)
	# patch g2p for misaki return arity
	g2p_path = os.path.join(model_dir, 'kemetone', 'g2p.py')
	if os.path.exists(g2p_path):
	src = open(g2p_path).read()
	src = src.replace(" raw, _ = self._g2p(text)\n",
	" _r = self._g2p(text)\n raw = _r[0] if isinstance(_r, (tuple, list)) else _r\n")
	open(g2p_path, 'w').write(src)

	# 6) download patched kokoro from Nabra-Saudi-7M
	patched = os.path.join(WORK, 'kokoro_patched')
	if not os.path.exists(patched):
	snapshot_download('oddadmix/Nabra-Saudi-7M', local_dir=os.path.join(WORK,'_nabra'),
	allow_patterns=['kokoro_patched/**'])
	src = os.path.join(WORK, '_nabra', 'kokoro_patched')
	if os.path.exists(src):
	shutil.move(src, patched)
	sys.path.insert(0, patched)

	# 7) create styletts2_gan shim
	shim = os.path.join(WORK, 'styletts2_gan')
	os.makedirs(shim, exist_ok=True)
	open(os.path.join(shim, '__init__.py'), 'w').write("import sys\nsys.path.insert(0,'/kaggle/working/StyleTTS2')\n")
	open(os.path.join(shim, 'discriminators.py'), 'w').write(
	"import sys\nsys.path.insert(0,'/kaggle/working/StyleTTS2')\n"
	"from Modules.discriminators import MultiPeriodDiscriminator, MultiResSpecDiscriminator, WavLMDiscriminator\n")
	open(os.path.join(shim, 'losses_styletts2.py'), 'w').write(
	"import sys\nsys.path.insert(0,'/kaggle/working/StyleTTS2')\n"
	"from losses import GeneratorLoss, DiscriminatorLoss, MultiResolutionSTFTLoss, WavLMLoss\n")
	if WORK not in sys.path:
	sys.path.insert(0, WORK)

	# 8) dump teacher (if not already done)
	dump_dir = os.path.join(WORK, 'dump')
	index_path = os.path.join(dump_dir, 'index.jsonl')
	data_file = os.path.join(WORK, 'data.txt')
	if not os.path.exists(data_file):
	data_file = hf_hub_download('0best/kemetone-data', 'kemetone_TTS_CLEAN_DEDUPED_PART1.txt',
	repo_type='dataset', local_dir=WORK)
	if not os.path.exists(index_path):
	os.environ['TEXT_FILE'] = data_file
	os.environ['OUT_DIR'] = dump_dir
	os.environ['MODEL_DIR'] = model_dir
	subprocess.run([sys.executable, 'dump_keme.py'], check=True, cwd=WORK)

	# 9) train student in chunks of 1000 steps, resuming each time
	out_dir = os.path.join(WORK, 'runs', 'egypt7m')
	os.makedirs(out_dir, exist_ok=True)

	CHUNK = 1000
	TOTAL_CHUNKS = 10 # 10 * 1000 = 10000 steps on part 1

	for chunk in range(1, TOTAL_CHUNKS + 1):
	target_steps = chunk * CHUNK
	print(f"\n=== Chunk {chunk}/{TOTAL_CHUNKS}: training to step {target_steps} ===", flush=True)

	cmd = [sys.executable, 'train_student.py',
	'--index', index_path,
	'--config', 'config_keme_student.json',
	'--out', out_dir,
	'--steps', str(target_steps),
	'--batch', '8',
	'--lr', '2e-4',
	'--lr-decay',
	'--mel-weight', '5',
	'--sil-weight', '100',
	'--slm-weight', '1',
	'--slm-adv-weight', '0.2',
	'--ckpt-every', '1000',
	'--resume']
	subprocess.run(cmd, check=True, cwd=WORK)

	# Upload checkpoints to HF
	for ckpt_file in [os.path.join(out_dir, 'last.pt'), os.path.join(out_dir, 'student.pth')]:
	if os.path.exists(ckpt_file):
	try:
	token = os.environ.get('HF_TOKEN')
	upload_file(
	path_or_fileobj=ckpt_file,
	path_in_repo=f"runs/egypt7m/chunk{chunk}/" + os.path.basename(ckpt_file),
	repo_id="0best/kemetone-distill-ckpts",
	repo_type="model",
	token=token,
	)
	print(f"[*] uploaded {os.path.basename(ckpt_file)} to HF (chunk {chunk})", flush=True)
	except Exception as e:
	print(f"[*] upload failed for {ckpt_file}: {e}", flush=True)

	print('TRAINING DONE - all 10 chunks completed (10,000 steps on part 1)')