针对1000条数据优化超参

examples/asr/asr_eou/speech_to_text_rnnt_eou_train_number.py

@@ -254,6 +254,13 @@ def main(cfg):
     trainer = pl.Trainer(**resolve_trainer_cfg(cfg.trainer))
     exp_manager(trainer, cfg.get("exp_manager", None))
 
+    class PrintLossCallback(pl.Callback):
+        def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
+            if outputs is not None and "loss" in outputs:
+                print(f"[Step {trainer.global_step}] Loss: {outputs['loss'].item():.4f}")
+
+    trainer.callbacks.append(PrintLossCallback())
+
     if cfg.model.get("adapter", None) is not None:
         update_model_config_to_support_adapter(cfg.model)
 
@@ -362,21 +369,36 @@ if __name__ == '__main__':
         'exp_manager.resume_ignore_no_checkpoint=true',
         'exp_manager.exp_dir=results/',
         'exp_manager.checkpoint_callback_params.save_top_k=1',
-        '++trainer.check_val_every_n_epoch=1',
+        '++trainer.check_val_every_n_epoch=200',
         '++model.encoder.conv_norm_type=layer_norm',
         f'model.tokenizer.dir={tokenizer_dir}',
         f'model.train_ds.manifest_filepath={manifest_path}',
-        '~model.train_ds.augmentor.noise',
+        '~model.train_ds.augmentor.noise', # 恢复禁用 noise，因为它需要额外的背景噪声数据集
         f'model.validation_ds.manifest_filepath={manifest_path}',
         f'model.test_ds.manifest_filepath={manifest_path}',
-        'trainer.max_epochs=1',
+        
+        # --- 针对 1000 条小数据集的超参数配置 ---
+        'model.optim.lr=0.001',                      # 学习率 (默认通常为 1e-3 或 2e-3)
+        'model.optim.sched.warmup_steps=200',        # 1000条数据步数少，必须调小预热步数 (NeMo默认可能过万)
+        'model.optim.sched.min_lr=1e-6',             # 最小学习率限制
+        'trainer.max_epochs=500',                     # 数据量小，1个 epoch 远远不够，建议增加至 50-100 轮
+        
+        # --- 正则化配置 (防止模型死记硬背) ---
+        'model.optim.weight_decay=0.01',             # 增加权重衰减 (默认通常为 1e-3)
+        'model.encoder.dropout=0.15',                 # 增加 Dropout (默认 0.1)
+        'model.encoder.dropout_pre_encoder=0.2',     # 增加 Pre-encoder Dropout (默认 0.1)
+        'model.encoder.dropout_att=0.2',             # 增加 Attention Dropout (默认 0.1)
+        'model.spec_augment.freq_masks=2',           # 确保开启 SpecAugment (频域掩码)
+        'model.spec_augment.time_masks=10',          # 确保开启 SpecAugment (时域掩码)
+        # ----------------------------------------
+        
         'trainer.devices=1',
         'trainer.accelerator=gpu',
         'trainer.strategy=auto',
-        'trainer.log_every_n_steps=1',
-        'model.train_ds.batch_size=1', # Reduced batch size for transducer/EOU as it consumes more memory
-        'model.validation_ds.batch_size=1',
-        'model.test_ds.batch_size=1',
+        'trainer.log_every_n_steps=25',
+        'model.train_ds.batch_size=16', # Reduced batch size for transducer/EOU as it consumes more memory
+        'model.validation_ds.batch_size=8',
+        'model.test_ds.batch_size=8',
         'model.train_ds.num_workers=0',
         'model.validation_ds.num_workers=0',
         'model.test_ds.num_workers=0',