add tensorboard and add export sample for mel and audio

src/f5_tts/model/trainer.py

@@ -3,7 +3,11 @@ from __future__ import annotations
 import os
 import gc
 from tqdm import tqdm
-import wandb
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except ImportError:
+    print("TensorBoard is not installed")
 
 import torch
 from torch.optim import AdamW
@@ -19,9 +23,26 @@ from f5_tts.model import CFM
 from f5_tts.model.utils import exists, default
 from f5_tts.model.dataset import DynamicBatchSampler, collate_fn
 
-
+import numpy as np
+import matplotlib.pyplot as plt
 # trainer
 
+# audio imports
+import torchaudio
+import soundfile as sf
+from vocos import Vocos
+import warnings
+
+warnings.filterwarnings("ignore", category=FutureWarning)
+
+# -----------------------------------------
+target_sample_rate = 24000
+hop_length = 256
+nfe_step = 16
+cfg_strength = 2.0
+sway_sampling_coef = -1.0
+# -----------------------------------------
+
 
 class Trainer:
     def __init__(
@@ -39,6 +60,8 @@ class Trainer:
         max_grad_norm=1.0,
         noise_scheduler: str | None = None,
         duration_predictor: torch.nn.Module | None = None,
+        logger: str = "wandb",  # Add logger parameter wandb,tensorboard , none
+        log_dir: str = "logs",  # Add log directory parameter
         wandb_project="test_e2-tts",
         wandb_run_name="test_run",
         wandb_resume_id: str = None,
@@ -46,24 +69,24 @@ class Trainer:
         accelerate_kwargs: dict = dict(),
         ema_kwargs: dict = dict(),
         bnb_optimizer: bool = False,
+        export_samples=False,
     ):
         ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
 
-        logger = "wandb" if wandb.api.api_key else None
-        print(f"Using logger: {logger}")
-
-        self.accelerator = Accelerator(
-            log_with=logger,
-            kwargs_handlers=[ddp_kwargs],
-            gradient_accumulation_steps=grad_accumulation_steps,
-            **accelerate_kwargs,
-        )
+        self.logger = logger
+        if self.logger == "wandb":
+            self.accelerator = Accelerator(
+                log_with="wandb",
+                kwargs_handlers=[ddp_kwargs],
+                gradient_accumulation_steps=grad_accumulation_steps,
+                **accelerate_kwargs,
+            )
 
-        if logger == "wandb":
             if exists(wandb_resume_id):
                 init_kwargs = {"wandb": {"resume": "allow", "name": wandb_run_name, "id": wandb_resume_id}}
             else:
                 init_kwargs = {"wandb": {"resume": "allow", "name": wandb_run_name}}
+
             self.accelerator.init_trackers(
                 project_name=wandb_project,
                 init_kwargs=init_kwargs,
@@ -80,12 +103,37 @@ class Trainer:
                     "noise_scheduler": noise_scheduler,
                 },
             )
+        elif self.logger == "tensorboard":
+            self.accelerator = Accelerator(
+                kwargs_handlers=[ddp_kwargs],
+                gradient_accumulation_steps=grad_accumulation_steps,
+                **accelerate_kwargs,
+            )
+            if self.is_main:
+                path_log_dir = os.path.join(log_dir, wandb_project)
+                os.makedirs(path_log_dir, exist_ok=True)
+                existing_folders = [folder for folder in os.listdir(path_log_dir) if folder.startswith("exp")]
+                next_number = len(existing_folders) + 2
+                folder_name = f"exp{next_number}"
+                folder_path = os.path.join(path_log_dir, folder_name)
+                os.makedirs(folder_path, exist_ok=True)
+
+                self.writer = SummaryWriter(log_dir=folder_path)
+
+        # export audio and mel
+        self.export_samples = export_samples
+        if self.export_samples:
+            self.path_ckpts_project = checkpoint_path
+
+            self.vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
+            self.vocos.to("cpu")
+            self.file_path_samples = os.path.join(self.path_ckpts_project, "samples")
+            os.makedirs(self.file_path_samples, exist_ok=True)
 
         self.model = model
 
         if self.is_main:
             self.ema_model = EMA(model, include_online_model=False, **ema_kwargs)
-
             self.ema_model.to(self.accelerator.device)
 
         self.epochs = epochs
@@ -175,6 +223,82 @@ class Trainer:
         gc.collect()
         return step
 
+    def log(self, metrics, step):
+        """Unified logging method for both WandB and TensorBoard"""
+        if self.logger == "none":
+            return
+        if self.logger == "wandb":
+            self.accelerator.log(metrics, step=step)
+        elif self.is_main:
+            for key, value in metrics.items():
+                self.writer.add_scalar(key, value, step)
+
+    def export_add_log(self, global_step, mel_org, text_inputs):
+        try:
+            generated_wave_org = self.vocos.decode(mel_org.unsqueeze(0).cpu())
+            generated_wave_org = generated_wave_org.squeeze().cpu().numpy()
+            file_wav_org = os.path.join(self.file_path_samples, f"step_{global_step}_org.wav")
+            sf.write(file_wav_org, generated_wave_org, target_sample_rate)
+
+            audio, sr = torchaudio.load(file_wav_org)
+            audio = audio.to("cuda")
+
+            ref_audio_len = audio.shape[-1] // hop_length
+            text = [text_inputs[0] + [" . "] + text_inputs[0]]
+            duration = int((audio.shape[1] / 256) * 2.0)
+
+            with torch.inference_mode():
+                generated_gen, _ = self.model.sample(
+                    cond=audio,
+                    text=text,
+                    duration=duration,
+                    steps=nfe_step,
+                    cfg_strength=cfg_strength,
+                    sway_sampling_coef=sway_sampling_coef,
+                )
+
+            generated_gen = generated_gen.to(torch.float32)
+            generated_gen = generated_gen[:, ref_audio_len:, :]
+            generated_mel_spec_gen = generated_gen.permute(0, 2, 1)
+            generated_wave_gen = self.vocos.decode(generated_mel_spec_gen.cpu())
+            generated_wave_gen = generated_wave_gen.squeeze().cpu().numpy()
+            file_wav_gen = os.path.join(self.file_path_samples, f"step_{global_step}_gen.wav")
+            sf.write(file_wav_gen, generated_wave_gen, target_sample_rate)
+
+            if self.logger == "tensorboard":
+                self.writer.add_audio("Audio/original", generated_wave_org, global_step, sample_rate=target_sample_rate)
+
+                self.writer.add_audio("Audio/generate", generated_wave_gen, global_step, sample_rate=target_sample_rate)
+
+            mel_org = mel_org
+            mel_min, mel_max = mel_org.min(), mel_org.max()
+            mel_norm = (mel_org - mel_min) / (mel_max - mel_min + 1e-8)
+            mel_colored = plt.get_cmap("viridis")(mel_norm.detach().cpu().numpy())[:, :, :3]
+            mel_colored = np.transpose(mel_colored, (2, 0, 1))
+
+            if self.logger == "tensorboard":
+                self.writer.add_image("Mel/oginal", mel_colored, global_step, dataformats="CHW")
+
+            mel_colored_hwc = np.transpose(mel_colored, (1, 2, 0))
+            file_gen_org = os.path.join(self.file_path_samples, f"step_{global_step}_org.png")
+            plt.imsave(file_gen_org, mel_colored_hwc)
+
+            mel_gen = generated_mel_spec_gen[0]
+            mel_min, mel_max = mel_gen.min(), mel_gen.max()
+            mel_norm = (mel_gen - mel_min) / (mel_max - mel_min + 1e-8)
+            mel_colored = plt.get_cmap("viridis")(mel_norm.detach().cpu().numpy())[:, :, :3]
+            mel_colored = np.transpose(mel_colored, (2, 0, 1))
+
+            if self.logger == "tensorboard":
+                self.writer.add_image("Mel/generate", mel_colored, global_step, dataformats="CHW")
+
+            mel_colored_hwc = np.transpose(mel_colored, (1, 2, 0))
+            file_gen_gen = os.path.join(self.file_path_samples, f"step_{global_step}_gen.png")
+            plt.imsave(file_gen_gen, mel_colored_hwc)
+
+        except Exception as e:
+            print("An error occurred:", e)
+
     def train(self, train_dataset: Dataset, num_workers=16, resumable_with_seed: int = None):
         if exists(resumable_with_seed):
             generator = torch.Generator()
@@ -270,6 +394,15 @@ class Trainer:
                     loss, cond, pred = self.model(
                         mel_spec, text=text_inputs, lens=mel_lengths, noise_scheduler=self.noise_scheduler
                     )
+
+                    # save 4 audio per save step
+                    if (
+                        self.accelerator.is_local_main_process
+                        and self.export_samples
+                        and global_step % (int(self.save_per_updates * 0.25) * self.grad_accumulation_steps) == 0
+                    ):
+                        self.export_add_log(global_step, batch["mel"][0], text_inputs)
+
                     self.accelerator.backward(loss)
 
                     if self.max_grad_norm > 0 and self.accelerator.sync_gradients:
@@ -285,7 +418,7 @@ class Trainer:
                 global_step += 1
 
                 if self.accelerator.is_local_main_process:
-                    self.accelerator.log({"loss": loss.item(), "lr": self.scheduler.get_last_lr()[0]}, step=global_step)
+                    self.log({"loss": loss.item(), "lr": self.scheduler.get_last_lr()[0]}, step=global_step)
 
                 progress_bar.set_postfix(step=str(global_step), loss=loss.item())
 

src/f5_tts/train/finetune_cli.py

@@ -56,6 +56,14 @@ def parse_args():
         help="Path to custom tokenizer vocab file (only used if tokenizer = 'custom')",
     )
 
+    parser.add_argument(
+        "--export_samples",
+        type=bool,
+        default=False,
+        help="Export 4 audio and spect samples for the checkpoint audio, per step.",
+    )
+    parser.add_argument("--logger", type=str, default="wandb", choices=["none", "wandb", "tensorboard"], help="logger")
+
     return parser.parse_args()
 
 
@@ -64,6 +72,7 @@ def parse_args():
 
 def main():
     args = parse_args()
+
     checkpoint_path = str(files("f5_tts").joinpath(f"../../ckpts/{args.dataset_name}"))
 
     # Model parameters based on experiment name
@@ -136,6 +145,8 @@ def main():
         wandb_run_name=args.exp_name,
         wandb_resume_id=wandb_resume_id,
         last_per_steps=args.last_per_steps,
+        logger=args.logger,
+        export_samples=args.export_samples,
     )
 
     train_dataset = load_dataset(args.dataset_name, tokenizer, mel_spec_kwargs=mel_spec_kwargs)

src/f5_tts/train/finetune_gradio.py

@@ -447,6 +447,8 @@ def start_training(
 
     cmd += f" --tokenizer {tokenizer_type} "
 
+    cmd += " --export_samples True --logger wandb "
+
     print(cmd)
 
     save_settings(
@@ -1223,6 +1225,27 @@ def get_checkpoints_project(project_name, is_gradio=True):
     return files_checkpoints, selelect_checkpoint
 
 
+def get_audio_project(project_name, is_gradio=True):
+    if project_name is None:
+        return [], ""
+    project_name = project_name.replace("_pinyin", "").replace("_char", "")
+
+    if os.path.isdir(path_project_ckpts):
+        files_audios = glob(os.path.join(path_project_ckpts, project_name, "samples", "*.wav"))
+        files_audios = sorted(files_audios, key=lambda x: int(os.path.basename(x).split("_")[1].split(".")[0]))
+
+        files_audios = [item.replace("_gen.wav", "") for item in files_audios if item.endswith("_gen.wav")]
+    else:
+        files_audios = []
+
+    selelect_checkpoint = None if not files_audios else files_audios[0]
+
+    if is_gradio:
+        return gr.update(choices=files_audios, value=selelect_checkpoint)
+
+    return files_audios, selelect_checkpoint
+
+
 def get_gpu_stats():
     gpu_stats = ""
 
@@ -1290,6 +1313,21 @@ def get_combined_stats():
     return combined_stats
 
 
+def get_audio_select(file_sample):
+    select_audio_org = file_sample
+    select_audio_gen = file_sample
+    select_image_org = file_sample
+    select_image_gen = file_sample
+
+    if file_sample is not None:
+        select_audio_org += "_org.wav"
+        select_audio_gen += "_gen.wav"
+        select_image_org += "_org.png"
+        select_image_gen += "_gen.png"
+
+    return select_audio_org, select_audio_gen, select_image_org, select_image_gen
+
+
 with gr.Blocks() as app:
     gr.Markdown(
         """
@@ -1511,6 +1549,47 @@ If you encounter a memory error, try reducing the batch size per GPU to a smalle
 
             ch_stream = gr.Checkbox(label="stream output experiment.", value=True)
             txt_info_train = gr.Text(label="info", value="")
+
+            list_audios, select_audio = get_audio_project(projects_selelect, False)
+
+            select_audio_org = select_audio
+            select_audio_gen = select_audio
+            select_image_org = select_audio
+            select_image_gen = select_audio
+
+            if select_audio is not None:
+                select_audio_org += "_org.wav"
+                select_audio_gen += "_gen.wav"
+                select_image_org += "_org.png"
+                select_image_gen += "_gen.png"
+
+            with gr.Row():
+                ch_list_audio = gr.Dropdown(
+                    choices=list_audios,
+                    value=select_audio,
+                    label="audios",
+                    allow_custom_value=True,
+                    scale=6,
+                    interactive=True,
+                )
+                bt_stream_audio = gr.Button("refresh", scale=1)
+                bt_stream_audio.click(fn=get_audio_project, inputs=[cm_project], outputs=[ch_list_audio])
+                cm_project.change(fn=get_audio_project, inputs=[cm_project], outputs=[ch_list_audio])
+
+            with gr.Row():
+                audio_org_stream = gr.Audio(label="original", type="filepath", value=select_audio_org)
+                mel_org_stream = gr.Image(label="original", type="filepath", value=select_image_org)
+
+            with gr.Row():
+                audio_gen_stream = gr.Audio(label="generate", type="filepath", value=select_audio_gen)
+                mel_gen_stream = gr.Image(label="generate", type="filepath", value=select_image_gen)
+
+            ch_list_audio.change(
+                fn=get_audio_select,
+                inputs=[ch_list_audio],
+                outputs=[audio_org_stream, audio_gen_stream, mel_org_stream, mel_gen_stream],
+            )
+
             start_button.click(
                 fn=start_training,
                 inputs=[