better hyperparams

add progress bar to process
keep the unused variable to process the data!!

dataset/process.py

@@ -3,6 +3,7 @@
 from typing import List
 
 from pydub import AudioSegment
+from tqdm import tqdm
 
 import dataset.exceptions
 from dataset.aeneas_wrapper import AeneasWrapper
@@ -160,6 +161,7 @@ def process(self, remove: bool = False) -> None:
         """
 
         nbm_files = len(os.listdir(self.audio_path))
+        progress_bar = tqdm(total=nbm_files)
         for i, audio_f in enumerate(os.listdir(self.audio_path)):
             if not audio_f.endswith(".ogg") and not audio_f.endswith(".mp4"):
                 continue
@@ -180,10 +182,13 @@ def process(self, remove: bool = False) -> None:
             self._export_audio(audio_segments, audio_f.split(".")[0])
             self._export_lyric(lyric_segments, audio_f.split(".")[0])
 
-            print(
-                f"Processed {i}/ {nbm_files} - {round(i/nbm_files*100, 2)}%", end="\r"
-            )
+            # print(
+            #     f"Processed {i}/ {nbm_files} - {round(i/nbm_files*100, 2)}%", end="\r"
+            # )
 
             if remove:
                 os.remove(lyric_path)
                 os.remove(audio_path)
+
+            progress_bar.update(1)
+        progress_bar.close()

train.py

@@ -1,9 +1,9 @@
-from datasets import concatenate_datasets, Dataset
+import argparse
+
+from datasets import Dataset
 
 from training import utils
 from training.train import Trainer
-import argparse
-import glob
 
 parser = argparse.ArgumentParser(
     description="Process the dataset and train the model",
@@ -51,4 +51,4 @@
 print(dataset)
 
 trainer.train()
-trainer.save_model(args.model_path)
+trainer.model.save_pretrained(args.model_path)

train2.py

@@ -1,19 +1,18 @@
 import librosa
-import numpy as np
-from datasets import Audio, DatasetDict, load_from_disk
+from datasets import load_from_disk
 
 from training import utils
 from training.train import Trainer
 
-DS_PATH = "dataset/"
+DS_PATH = "dataset/export"
 
-dataset = utils.gather_dataset(DS_PATH)
 trainer = Trainer()
 
 is_prepared = False
 
 
 if not is_prepared:
+    dataset = utils.gather_dataset(DS_PATH)
     target_sr = trainer.processor.feature_extractor.sampling_rate
 
     def prepare_dataset(batch):

training/train.py

@@ -1,6 +1,7 @@
 """
 This module contains the Trainer class which is responsible for training whisper on predicting lyrics.
 """
+from functools import partial
 
 import evaluate
 from transformers import (
@@ -79,28 +80,30 @@ def train(self, dataset):
         :return:
         """
 
+        self.model.generate = partial(
+            self.model.generate, task="transcribe", use_cache=True
+        )
         training_args = Seq2SeqTrainingArguments(
             output_dir=self._ouput_dir,  # change to a repo name of your choice
             per_device_train_batch_size=8,
             gradient_accumulation_steps=1,  # increase by 2x for every 2x decrease in batch size
             learning_rate=1e-5,
-            warmup_steps=500,
-            max_steps=4000,
-            gradient_checkpointing=True,
+            warmup_steps=50,
+            max_steps=8000,
+            gradient_checkpointing=False,
             fp16=True,
             eval_strategy="steps",
             per_device_eval_batch_size=8,
             predict_with_generate=True,
             generation_max_length=225,
-            save_steps=80,
-            eval_steps=80,
+            save_steps=1000,
+            eval_steps=1000,
             logging_steps=25,
             report_to=["tensorboard"],
-            load_best_model_at_end=True,
+            # load_best_model_at_end=True,
             metric_for_best_model="wer",
             greater_is_better=False,
             push_to_hub=False,
-            gradient_checkpointing_kwargs={"use_reentrant": False},
         )
 
         trainer = Seq2SeqTrainer(

training/utils.py

@@ -18,6 +18,7 @@ def gather_dataset(path: str) -> Dataset:
     """
 
     def gen():
+        i = 2
         audios = glob.glob(path + "/audio/*")
         lyrics = glob.glob(path + "/lyrics/*.txt")
         for audio, lyric in zip(audios, lyrics):