Upload finetune_speech.py

examples/finetune_speech.py

@@ -25,36 +25,6 @@ import soundfile as sf
 from datasets import Audio
 import random
 
-class MultipleTokenBatchStoppingCriteria(StoppingCriteria):
-    """Stopping criteria capable of receiving multiple stop-tokens and handling batched inputs."""
-
-    def __init__(self, stop_tokens: torch.LongTensor, batch_size: int = 1) -> None:
-        """Initialize the multiple token batch stopping criteria.
-
-        Args:
-            stop_tokens: Stop-tokens.
-            batch_size: Batch size.
-
-        """
-
-        self.stop_tokens = stop_tokens
-        self.max_stop_tokens = stop_tokens.shape[-1]
-        self.stop_tokens_idx = torch.zeros(batch_size, dtype=torch.long, device=stop_tokens.device)
-
-    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
-        # Only gather the maximum number of inputs compatible with stop tokens
-        # and checks whether generated inputs are equal to `stop_tokens`
-        generated_inputs = torch.eq(input_ids[:, -self.max_stop_tokens :].unsqueeze(1), self.stop_tokens)
-        equal_generated_inputs = torch.all(generated_inputs, dim=2)
-
-        # Mark the position where a stop token has been produced for each input in the batch,
-        # but only if the corresponding entry is not already set
-        sequence_idx = torch.any(equal_generated_inputs, dim=1)
-        sequence_set_mask = self.stop_tokens_idx == 0
-        self.stop_tokens_idx[sequence_idx & sequence_set_mask] = input_ids.shape[-1]
-
-        return torch.all(self.stop_tokens_idx)
-
 class BaseAudioDataset(Dataset):
     def __init__(self, processor, split, sampling_rate=16000, debug=False):
         self.processor = processor
@@ -581,65 +551,6 @@ def create_model(model_name_or_path, revision="main", use_flash_attention = Fals
 
     return model
 
-@torch.no_grad()
-def evaluate(model, processor, eval_dataset, save_path=None, disable_tqdm=False, eval_batch_size=1):
-    model.eval()
-    all_generated_texts = []
-    all_labels = []
-
-    eval_dataloader = torch.utils.data.DataLoader(
-        eval_dataset,
-        batch_size=eval_batch_size,
-        collate_fn=covost_collate_fn,
-        shuffle=False,
-        drop_last=False,
-        num_workers=8,
-        prefetch_factor=2,
-        pin_memory=True,
-    )
-    stop_tokens = [processor.tokenizer.eos_token]
-    stop_tokens_ids = processor.tokenizer(stop_tokens, add_special_tokens=False, padding="longest", return_tensors="pt")["input_ids"]
-    stop_tokens_ids = stop_tokens_ids.to('cuda')
-
-    for inputs in tqdm(
-        eval_dataloader, disable= disable_tqdm, desc='running eval'
-    ):
-        stopping_criteria=StoppingCriteriaList([MultipleTokenBatchStoppingCriteria(stop_tokens_ids, batch_size=inputs.input_ids.size(0))])
-        inputs = inputs.to('cuda').to(model.dtype)
-        generated_ids = model.generate(
-            **inputs, eos_token_id=processor.tokenizer.eos_token_id, max_new_tokens=64,
-            stopping_criteria=stopping_criteria,
-        )
-
-        stop_tokens_idx = stopping_criteria[0].stop_tokens_idx.reshape(inputs.input_ids.size(0), -1)[:, 0]
-
-        stop_tokens_idx = torch.where(
-            stop_tokens_idx > 0,
-            stop_tokens_idx - stop_tokens_ids.shape[-1],
-            generated_ids.shape[-1],
-        )
-        generated_text = [
-            processor.decode(_pred_ids[inputs["input_ids"].shape[1] : _stop_tokens_idx], skip_special_tokens=True, clean_up_tokenization_spaces=False)
-            for _pred_ids, _stop_tokens_idx in zip(generated_ids, stop_tokens_idx)
-        ]
-        all_generated_texts.extend(generated_text)
-        labels = [processor.decode(_label_ids[_label_ids != 0]).removesuffix(ANSWER_SUFFIX) for _label_ids in inputs["labels"]]
-        all_labels.extend(labels)
-
-    assert len(all_generated_texts) == len(all_labels)
-    bleu = sacrebleu.corpus_bleu(all_generated_texts, [all_labels])
-    print(bleu)
-    if save_path:
-        with open(save_path, 'w') as f:
-            save_dict = {
-                'all_generated_texts': all_generated_texts,
-                'all_labels': all_labels,
-                'score': bleu.score,
-            }
-            json.dump(save_dict, f)
-
-    return bleu.score
-
 os.environ["TOKENIZERS_PARALLELISM"] = "false" 
 
 INSTRUCTION = {