diff --git a/custom_nodes/rvc_nodes.py b/custom_nodes/rvc_nodes.py index f76005b..78bbe9a 100644 --- a/custom_nodes/rvc_nodes.py +++ b/custom_nodes/rvc_nodes.py @@ -19,10 +19,10 @@ from ..config import config from ..lib.model_utils import load_hubert from .utils import MultipleTypeProxy, increment_filename_no_overwrite, model_downloader -from .settings import PITCH_EXTRACTION_OPTIONS, SR_MAP +from .settings import PITCH_EXTRACTION_OPTIONS from .settings.downloader import PRETRAINED_MODELS, RVC_DOWNLOAD_LINK, RVC_INDEX, RVC_MODELS, download_file, extract_zip_without_structure -from ..lib.audio import SUPPORTED_AUDIO, audio_to_bytes, load_input_audio, save_input_audio, get_audio +from ..lib.audio import SUPPORTED_AUDIO, audio_to_bytes, load_input_audio, save_input_audio, get_audio, SR_MAP from ..vc_infer_pipeline import get_vc, vc_single import folder_paths @@ -38,7 +38,7 @@ weights_path = os.path.join(BASE_MODELS_DIR, "uvr5") dataset_path = os.path.join(input_path,"datasets") device = get_optimal_torch_device() CATEGORY = "🌺RVC-Studio/rvc" -LOG_DIR = os.path.join(BASE_DIR,"logs") +MUTE_DIR = os.path.join(BASE_DIR,"dataset") class LoadPitchExtractionParams: @classmethod @@ -259,31 +259,31 @@ class RVCProcessDatasetNode: if "crepe" in f0_method: cached_params.append(crepe_hop_length) cache_name = get_hash(*cached_params) - model_log_dir = os.path.join(output_path,"logs",cache_name) - os.makedirs(model_log_dir,exist_ok=True) + dataset_dir = os.path.join(output_path,"dataset",cache_name) + os.makedirs(dataset_dir,exist_ok=True) - filelist_path = os.path.join(model_log_dir, "filelist.txt") + filelist_path = os.path.join(dataset_dir, "filelist.txt") if not os.path.isfile(filelist_path): - dataset_dir = os.path.join(input_path,"datasets",dataset.split(".")[0]) + input_dir = os.path.join(dataset_path,dataset.split(".")[0]) if dataset.endswith("zip"): - files = extract_zip_without_structure(os.path.join(dataset_path,dataset),dataset_dir) + files = extract_zip_without_structure(os.path.join(dataset_path,dataset),input_dir) assert len(files), "Failed to extract zip file..." - assert preprocess_trainset(dataset_dir,SR_MAP[sr],n_threads,model_log_dir,audio_processor,period,overlap,max_volume), "Failed to preprocess audio..." + assert preprocess_trainset(input_dir,SR_MAP[sr],n_threads,dataset_dir,audio_processor,period,overlap,max_volume), "Failed to preprocess audio..." - assert extract_features_trainset(hubert_model(), model_log_dir,n_p=n_threads,f0method=f0_method,device=device,if_f0=bool(f0_method),version="v2",crepe_hop_length=crepe_hop_length), "Failed to extract features..." + assert extract_features_trainset(hubert_model(), dataset_dir,n_p=n_threads,f0method=f0_method,device=device,if_f0=bool(f0_method),version="v2",crepe_hop_length=crepe_hop_length), "Failed to extract features..." - gt_wavs_dir = os.path.join(model_log_dir,"0_gt_wavs") - feature_dir = os.path.join(model_log_dir,"3_feature768") + gt_wavs_dir = os.path.join(dataset_dir,"0_gt_wavs") + feature_dir = os.path.join(dataset_dir,"3_feature768") os.makedirs(gt_wavs_dir, exist_ok=True) os.makedirs(feature_dir, exist_ok=True) # add training data if f0_method: - f0_dir = os.path.join(model_log_dir,"2a_f0") - f0nsf_dir = os.path.join(model_log_dir,"2b-f0nsf") + f0_dir = os.path.join(dataset_dir,"2a_f0") + f0nsf_dir = os.path.join(dataset_dir,"2b-f0nsf") names = ( set([os.path.splitext(name)[0] for name in os.listdir(feature_dir)]) & set([os.path.splitext(name)[0] for name in os.listdir(f0_dir)]) @@ -328,16 +328,16 @@ class RVCProcessDatasetNode: for _ in range(num_mute): if f0_method: data = "|".join([ - os.path.join(LOG_DIR,"mute","0_gt_wavs",f"mute{sr}.wav"), - os.path.join(LOG_DIR,"mute",f"3_feature{fea_dim}","mute.npy"), - os.path.join(LOG_DIR,"mute","2a_f0","mute.wav.npy"), - os.path.join(LOG_DIR,"mute","2b-f0nsf","mute.wav.npy"), + os.path.join(MUTE_DIR,"mute","0_gt_wavs",f"mute{sr}.wav"), + os.path.join(MUTE_DIR,"mute",f"3_feature{fea_dim}","mute.npy"), + os.path.join(MUTE_DIR,"mute","2a_f0","mute.wav.npy"), + os.path.join(MUTE_DIR,"mute","2b-f0nsf","mute.wav.npy"), str(0) ]) else: data = "|".join([ - os.path.join(LOG_DIR,"mute","0_gt_wavs",f"mute{sr}.wav"), - os.path.join(LOG_DIR,"mute",f"3_feature{fea_dim}","mute.npy"), + os.path.join(MUTE_DIR,"mute","0_gt_wavs",f"mute{sr}.wav"), + os.path.join(MUTE_DIR,"mute",f"3_feature{fea_dim}","mute.npy"), str(0) ]) opt.append(data) @@ -349,7 +349,7 @@ class RVCProcessDatasetNode: print("write filelist done") return (dict( sample_rate=sr, - model_dir=model_log_dir, + dataset_dir=dataset_dir, name=model_name, training_files=filelist_path, if_f0=bool(f0_method), @@ -377,7 +377,7 @@ class RVCTrainModelNode: }, "optional": dict( gpu=(DEVICES, {"default": DEVICES[0]}), - batch_size=("INT",dict(default=4,min=0,max=64,step=4)), + batch_size=("INT",dict(default=4,min=1,max=64,step=1)), total_epoch=("INT",dict(default=100,min=10,max=1000,step=10)), save_every_epoch=("INT",dict(default=0,min=0,max=100)), pretrained_G=(PRETRAINED_G,{"default": PRETRAINED_G[0]}), @@ -388,7 +388,10 @@ class RVCTrainModelNode: train_index=("BOOLEAN",{"default": True}), retrain=("BOOLEAN",{"default": False}), save_best_model=("BOOLEAN",{"default": True}), - log_every_epoch=("FLOAT",dict(default=1.,min=0.,max=2.,step=.5)) + log_every_epoch=("FLOAT",dict(default=1.,min=0.,max=2.,step=.5)), + gradient_lambda=("FLOAT",dict(default=0.,min=0.,max=100.,step=.1)), + timbre_lambda=("FLOAT",dict(default=0.,min=0.,max=100.,step=.1)), + num_workers=("INT",dict(default=1,min=1,max=16)) ) } @@ -415,11 +418,16 @@ class RVCTrainModelNode: train_index=True, retrain=False, save_best_model=True, - log_every_epoch=1.): + log_every_epoch=1., + gradient_lambda=0., + timbre_lambda=0., + num_workers=1): sample_rate = rvc_dataset_pipe["sample_rate"] name = rvc_dataset_pipe["name"] - model_dir = rvc_dataset_pipe["model_dir"] + dataset_dir = rvc_dataset_pipe["dataset_dir"] + cache_name = get_hash(batch_size,pretrained_G,pretrained_D,gradient_lambda,timbre_lambda) + model_dir = os.path.join(output_path,"logs",cache_name) if_f0 = rvc_dataset_pipe["if_f0"] config_path = os.path.join(BASE_DIR,"configs",f"{sample_rate}{'' if sample_rate=='40k' else '_v2'}.json") @@ -427,7 +435,8 @@ class RVCTrainModelNode: config = json.load(f) hparams = HParams(**config) - hparams.model_dir = hparams.experiment_dir = model_dir + hparams.experiment_dir = dataset_dir + hparams.model_dir = model_dir hparams.save_every_epoch = save_every_epoch hparams.name = name hparams.total_epoch = total_epoch @@ -444,8 +453,11 @@ class RVCTrainModelNode: hparams.data.training_files = rvc_dataset_pipe["training_files"] hparams.save_best_model = save_best_model hparams.log_every_epoch = log_every_epoch + hparams.train.gradient_lambda = gradient_lambda + hparams.train.num_workers = num_workers + hparams.train.timbre_lambda = timbre_lambda - file_index = self.train_index(model_dir, sample_rate, name) if train_index else None + file_index = self.train_index(dataset_dir, sample_rate, name) if train_index else None model_path = os.path.join(BASE_MODELS_DIR,"RVC",f"{name}_{sample_rate}.pth") if os.path.isfile(model_path) and retrain: model_path = increment_filename_no_overwrite(model_path) hparams.model_path = model_path @@ -455,9 +467,9 @@ class RVCTrainModelNode: return (lambda: get_vc(model_path, file_index), name, rvc_dataset_pipe["hubert_model"], rvc_dataset_pipe["pitch_extraction_params"]) - def train_index(self, model_log_dir, sr, name): + def train_index(self, dataset_dir, sr, name): - key = get_hash(model_log_dir, sr, name) + key = get_hash(dataset_dir, sr, name) index_file = os.path.join(BASE_MODELS_DIR,"RVC",".index",f"{name}_v2_{sr}_{key}.index") try: @@ -466,8 +478,7 @@ class RVCTrainModelNode: from sklearn.cluster import MiniBatchKMeans import faiss - feature_dir = os.sep.join([model_log_dir, "3_feature768"]) - os.makedirs(feature_dir, exist_ok=True) + feature_dir = os.path.join(dataset_dir, "3_feature768") npys = [] listdir_res = list(os.listdir(feature_dir)) diff --git a/custom_nodes/settings/__init__.py b/custom_nodes/settings/__init__.py index 1493e8b..a620f6e 100644 --- a/custom_nodes/settings/__init__.py +++ b/custom_nodes/settings/__init__.py @@ -11,5 +11,4 @@ PITCH_EXTRACTION_OPTIONS = ["crepe","mangio-crepe","rmvpe","rmvpe+"] MERGE_OPTIONS=["median","mean","min","max"] TTS_MODELS = ["edge","speecht5"] N_THREADS_OPTIONS=[1,2,4,8,12,16] -SR_MAP = {"32k": 32000,"40k": 40000, "48k": 48000} SUPPORTED_LANGUAGES = ['en', 'fr', 'es', "ja", "zh"] \ No newline at end of file diff --git a/logs/.gitignore b/dataset/.gitignore similarity index 100% rename from logs/.gitignore rename to dataset/.gitignore diff --git a/logs/mute/.gitignore b/dataset/mute/.gitignore similarity index 100% rename from logs/mute/.gitignore rename to dataset/mute/.gitignore diff --git a/logs/mute/0_gt_wavs/mute32k.wav b/dataset/mute/0_gt_wavs/mute32k.wav similarity index 100% rename from logs/mute/0_gt_wavs/mute32k.wav rename to dataset/mute/0_gt_wavs/mute32k.wav diff --git a/logs/mute/0_gt_wavs/mute40k.wav b/dataset/mute/0_gt_wavs/mute40k.wav similarity index 100% rename from logs/mute/0_gt_wavs/mute40k.wav rename to dataset/mute/0_gt_wavs/mute40k.wav diff --git a/logs/mute/0_gt_wavs/mute48k.wav b/dataset/mute/0_gt_wavs/mute48k.wav similarity index 100% rename from logs/mute/0_gt_wavs/mute48k.wav rename to dataset/mute/0_gt_wavs/mute48k.wav diff --git a/logs/mute/1_16k_wavs/mute.wav b/dataset/mute/1_16k_wavs/mute.wav similarity index 100% rename from logs/mute/1_16k_wavs/mute.wav rename to dataset/mute/1_16k_wavs/mute.wav diff --git a/logs/mute/2a_f0/mute.wav.npy b/dataset/mute/2a_f0/mute.wav.npy similarity index 100% rename from logs/mute/2a_f0/mute.wav.npy rename to dataset/mute/2a_f0/mute.wav.npy diff --git a/logs/mute/2b-f0nsf/mute.wav.npy b/dataset/mute/2b-f0nsf/mute.wav.npy similarity index 100% rename from logs/mute/2b-f0nsf/mute.wav.npy rename to dataset/mute/2b-f0nsf/mute.wav.npy diff --git a/logs/mute/3_feature256/mute.npy b/dataset/mute/3_feature256/mute.npy similarity index 100% rename from logs/mute/3_feature256/mute.npy rename to dataset/mute/3_feature256/mute.npy diff --git a/logs/mute/3_feature768/mute.npy b/dataset/mute/3_feature768/mute.npy similarity index 100% rename from logs/mute/3_feature768/mute.npy rename to dataset/mute/3_feature768/mute.npy diff --git a/lib/audio.py b/lib/audio.py index f237ed3..ba18337 100644 --- a/lib/audio.py +++ b/lib/audio.py @@ -27,6 +27,7 @@ AUTOTUNE_NOTES = np.array([ 2093.00, 2217.46, 2349.32, 2489.02, 2637.02, 2793.83, 2959.96, 3135.96, 3322.44, 3520.00, 3729.31, 3951.07 ]) +SR_MAP = {"32k": 32000,"40k": 40000, "48k": 48000} class AudioProcessor: def __init__(self, normalize=True, threshold_silence=True, dynamic_threshold=True, sample_size=16000, multiplier=2.0, fill_method="median", kernel_size=5, silence_threshold_db=-50, normalize_threshold_db=-1): diff --git a/lib/train/utils.py b/lib/train/utils.py index 0f45bc7..d320c9f 100644 --- a/lib/train/utils.py +++ b/lib/train/utils.py @@ -382,6 +382,10 @@ def get_hparams(init=True): hparams.if_cache_data_in_gpu = args.if_cache_data_in_gpu hparams.data.training_files = os.path.join(experiment_dir,"filelist.txt") hparams.save_best_model = True + hparams.train.gradient_lambda = 0. + hparams.train.num_workers = 4 + hparams.train.timbre_alpha = 0. + hparams.model_path = None return hparams @@ -476,6 +480,10 @@ class HParams: def __repr__(self): return self.__dict__.__repr__() + def get(self, key, default=None): + try: return self[key] + except: return default + def sync_log_interval(self, dataset_length): log_interval = getattr(self,"log_every_epoch",1) - self.train.log_interval = int(self.train.batch_size*dataset_length*log_interval) \ No newline at end of file + self.train.log_interval = int(dataset_length*log_interval) \ No newline at end of file diff --git a/training_cli.py b/training_cli.py index f64a901..39affeb 100644 --- a/training_cli.py +++ b/training_cli.py @@ -3,14 +3,11 @@ import json import os import shutil import traceback - +from .lib.audio import SR_MAP from .lib.train import utils import datetime - from random import shuffle, randint - import torch - torch.backends.cudnn.deterministic = False torch.backends.cudnn.benchmark = False from torch.nn import functional as F @@ -30,7 +27,6 @@ from .lib.train.data_utils import ( TextAudioCollate, DistributedBucketSampler, ) - from .lib.train.losses import generator_loss, discriminator_loss, feature_loss, kl_loss from .lib.train.mel_processing import mel_spectrogram_torch, spec_to_mel_torch @@ -122,7 +118,7 @@ def train_model(hps: "utils.HParams"): def run(rank, n_gpus, hps, device): print(f"{__name__=}") - global global_step, least_loss, loss_file + global global_step, least_loss, loss_file, best_model_name global_step = 0 loss_file = os.path.join(hps.model_dir,"losses.json") @@ -130,7 +126,10 @@ def run(rank, n_gpus, hps, device): with open(loss_file,"r") as f: data: dict = json.load(f) least_loss = data.get("least_loss",40) - else: least_loss = 40 + best_model_name = data.get("best_model_name","") + else: + least_loss = 40 + best_model_name = "" if hps.version == "v1": from .lib.infer_pack.models import ( @@ -304,16 +303,16 @@ def run(rank, n_gpus, hps, device): def train_and_evaluate( - rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers, cache + rank, epoch, hps, nets, optims, _, scaler, loaders, logger, writers, cache ): net_g, net_d = nets optim_g, optim_d = optims - train_loader, eval_loader = loaders + train_loader, _ = loaders if writers is not None: - writer, writer_eval = writers + writer, _ = writers train_loader.batch_sampler.set_epoch(epoch) - global global_step, least_loss, loss_file + global global_step, least_loss, loss_file, best_model_name net_g.train() net_d.train() @@ -479,13 +478,41 @@ def train_and_evaluate( ) # slice # Discriminator - y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach()) + gen_wave = y_hat.detach() + y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, gen_wave) + + gradient_penalty = torch.Tensor([0.]).to(wave.device) + if hps.train.get("gradient_lambda",0.)>0: + # Compute the gradient penalty + # Randomly interpolate between real and generated data + size = [1]*wave.ndim + size[0] = wave.size(0) + alpha = torch.rand(*size, device=wave.device) + interpolated = alpha * wave + (1 - alpha) * gen_wave + interpolated.requires_grad_(True) + # Get the discriminator output for the interpolated data + _, disc_interpolated_output, _, _ = net_d(wave, interpolated) + # Compute gradients of discriminator output w.r.t. interpolated data + for output in disc_interpolated_output: + gradient = torch.autograd.grad( + outputs=output, + inputs=interpolated, + grad_outputs=torch.ones(output.size(), device=wave.device), + create_graph=True, + only_inputs=True + )[0] + gradient = gradient.view(gradient.size(0), -1) + gradient_penalty += ((gradient.norm(2,dim=-1) - 1) ** 2).mean() + gradient_penalty = gradient_penalty.squeeze() + with autocast(enabled=False): loss_disc, losses_disc_r, losses_disc_g = discriminator_loss( y_d_hat_r, y_d_hat_g ) + if hps.train.get("gradient_lambda",0.)>0: gradient_penalty*=hps.train.gradient_lambda + optim_d.zero_grad() - scaler.scale(loss_disc).backward() + scaler.scale(loss_disc+gradient_penalty).backward() scaler.unscale_(optim_d) grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None) scaler.step(optim_d) @@ -507,38 +534,35 @@ def train_and_evaluate( scaler.update() if rank == 0: - if loss_gen_allmin(hps.total_epoch*.2,20): #start saving after 20% or 20 epochs - if hasattr(net_g, "module"): - ckpt = net_g.module.state_dict() - else: - ckpt = net_g.state_dict() - best_model_name = f"{hps.name}_e{epoch}_s{global_step}_loss{loss_gen_all:2.2f}" - status = save_checkpoint(ckpt,best_model_name,epoch,hps) + if epoch>=max(hps.total_epoch*.2,20): #start saving best model after 20% or 20 epochs + if loss_gen_all+loss_disc0 and global_step % hps.train.log_interval == 0: lr = optim_g.param_groups[0]["lr"] - logger.info( - "Train Epoch: {} [{:.0f}%]".format(epoch, 100.0 * batch_idx / len(train_loader)) - ) + logger.info(f"Train Epoch: {epoch} [{100.0 * batch_idx / len(train_loader):.0f}%]") # Amor For Tensorboard display if loss_mel > 75: loss_mel = 75 if loss_kl > 9: loss_kl = 9 - logger.info(f"epoch={epoch} steps={global_step} lr={lr} total_loss={loss_gen_all}") + logger.info(f"{epoch=} {global_step=} {lr=:.2E} {loss_disc=:.3f} {gradient_penalty.item()=:.2E}") + logger.info(f"{loss_gen_all=:.3f}: {loss_gen=:.3f}, {loss_fm=:.3f}, {loss_mel=:.3f}, {loss_kl=:.3f}") - logger.info( - f"loss_disc={loss_disc:.3f}, loss_gen={loss_gen:.3f}, loss_fm={loss_fm:.3f},loss_mel={loss_mel:.3f}, loss_kl={loss_kl:.3f}" - ) - scalar_dict = { "total/loss/all": loss_gen_all, "total/loss/g": loss_gen, @@ -547,35 +571,31 @@ def train_and_evaluate( "total/loss/mel": loss_mel, "total/loss/kl": loss_kl, "learning_rate": lr, - "grad_norm_d": grad_norm_d, - "grad_norm_g": grad_norm_g, + "gradient/grad_norm_d": grad_norm_d, + "gradient/grad_norm_g": grad_norm_g, + "gradient/gradient_penalty": gradient_penalty, + **{f"loss/g/{i}": v for i, v in enumerate(losses_gen)}, + **{f"loss/d_r/{i}": v for i, v in enumerate(losses_disc_r)}, + **{f"loss/d_g/{i}": v for i, v in enumerate(losses_disc_g)} } - scalar_dict.update( - {"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)} - ) - scalar_dict.update( - {"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)} - ) - scalar_dict.update( - {"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)} - ) image_dict = { - "slice/mel_org": utils.plot_spectrogram_to_numpy( - y_mel[0].data.cpu().numpy() - ), - "slice/mel_gen": utils.plot_spectrogram_to_numpy( - y_hat_mel[0].data.cpu().numpy() - ), - "all/mel": utils.plot_spectrogram_to_numpy( - mel[0].data.cpu().numpy() - ), + "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()), + "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()), + "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()), + } + + audio_dict = { + "slice/wave_org": wave.flatten(), + "slice/wave_gen": y_hat.flatten() } utils.summarize( writer=writer, global_step=global_step, images=image_dict, scalars=scalar_dict, + audios=audio_dict, + audio_sampling_rate=SR_MAP[hps.sample_rate] ) global_step += 1 # /Run steps @@ -588,14 +608,14 @@ def train_and_evaluate( optim_g, hps.train.learning_rate, epoch, - os.path.join(hps.model_dir, f"G_23333{hps.total_epoch}.pth"), + os.path.join(hps.model_dir, "G_23333.pth"), ) utils.save_checkpoint( net_d, optim_d, hps.train.learning_rate, epoch, - os.path.join(hps.model_dir, f"D_23333{hps.total_epoch}.pth"), + os.path.join(hps.model_dir, "D_23333.pth"), ) else: utils.save_checkpoint( @@ -603,14 +623,14 @@ def train_and_evaluate( optim_g, hps.train.learning_rate, epoch, - os.path.join(hps.model_dir, f"G_{global_step}.pth"), + os.path.join(hps.model_dir, f"G_{epoch}.pth"), ) utils.save_checkpoint( net_d, optim_d, hps.train.learning_rate, epoch, - os.path.join(hps.model_dir, f"D_{global_step}.pth"), + os.path.join(hps.model_dir, f"D_{epoch}.pth"), ) if rank == 0 and hps.save_every_weights: if hasattr(net_g, "module"): @@ -621,7 +641,7 @@ def train_and_evaluate( logger.info(f"saving ckpt {hps.name}_e{epoch}: {status}") if rank == 0: - logger.info("====> Epoch: {} {}".format(epoch, epoch_recorder.record())) + logger.info(f"====> Epoch {epoch}: {loss_gen_all=:.3f} {loss_disc=:.3f} {epoch_recorder.record()}") if epoch >= hps.total_epoch and rank == 0: logger.info("Training is done. The program is closed.") @@ -632,12 +652,15 @@ def train_and_evaluate( if hps.save_best_model: with open(loss_file,"r") as f: data = json.load(f) - best_model = data["best_model"] - if os.path.isfile(best_model): - shutil.copy(best_model,os.path.join(os.path.dirname(hps.model_path),f"{hps.name}_{hps.sample_rate}-best.pth")) + best_model_name = data.get("best_model_name","") + best_model_path = os.path.join(hps.model_dir,f"{best_model_name}.pth") + if os.path.isfile(best_model_path): + shutil.copy(best_model_path,os.path.join( + os.path.dirname(hps.model_path), + f"{os.path.basename(hps.model_path).split('.')[0]}-best.pth")) status = save_checkpoint(ckpt,hps.name,epoch,hps,model_path=hps.model_path) - logger.info(f"saving final ckpt: {status}") + logger.info(f"saving final ckpt {hps.model_path}: {status}") sleep(1) os._exit(0)