fix bug

游雁

2024-06-10 82530ddf974a706df5a6a1e258d80c8dbc3f1d72

 funasr/models/llm_asr/model.py

@@ -19,6 +19,7 @@
from funasr.utils.datadir_writer import DatadirWriter
from funasr.register import tables
from funasr.train_utils.device_funcs import to_device
import traceback


@tables.register("model_classes", "LLMASR")
@@ -489,6 +490,7 @@
            fbank_fake_len = fbank_fake_lens[batch_idx].item()
            fbank_beg_idx = fbank_beg[batch_idx, 0].item()
            min_len = min(fbank_fake_len, inputs_embeds.shape[1] - fbank_beg_idx)

            try:
                inputs_embeds[batch_idx, fbank_beg_idx : fbank_beg_idx + min_len, :] = encoder_out[
                    batch_idx, :min_len, :
@@ -496,10 +498,10 @@
            except Exception as e:
                logging.error(f"{str(e)}, {traceback.format_exc()}")
                logging.info(
                    f"batch_idx: {batch_idx}, inputs_embeds: {inputs_embeds.shape}, fbank_beg_idx: {fbank_beg_idx}, min_len: {min_len}, fbank_fake_len: {fbank_fake_len}"
                    f"batch_idx: {batch_idx}, inputs_embeds: {inputs_embeds.shape}, fbank_beg_idx: {fbank_beg_idx}, min_len: {min_len}, fbank_fake_len: {fbank_fake_len}, encoder_out: {encoder_out.shape}, encoder_out_lens: {encoder_out_lens[batch_idx].item()}"
                )
                fbank_fake_len = encoder_out_lens[batch_idx].item()
                min_len = min(fbank_fake_len, inputs_embeds.shape[1] - fbank_beg_idx)
                min_len = min(fbank_fake_len, min_len)
                inputs_embeds[batch_idx, fbank_beg_idx : fbank_beg_idx + min_len, :] = encoder_out[
                    batch_idx, :min_len, :
                ]
@@ -554,7 +556,7 @@

        return contents

    def data_load_speech(self, contents: dict, tokenizer, frontend, **kwargs):
    def data_load_speech(self, contents: dict, tokenizer, frontend, meta_data={}, **kwargs):

        system = contents["system"]
        user = contents["user"]
@@ -592,7 +594,10 @@
                    )
                    if sub_str.startswith("!"):
                        try:
                            time1 = time.perf_counter()
                            data_src = load_audio_text_image_video(sub_str[1:], fs=frontend.fs)
                            time2 = time.perf_counter()
                            meta_data["load_data"] = f"{time2 - time1:0.3f}"
                        except Exception as e:
                            logging.error(f"Loading wav failed! {str(e)}, {traceback.format_exc()}")

@@ -602,6 +607,15 @@
                            frontend=frontend,
                            is_final=True,
                        )  # speech: [b, T, d]

                        time3 = time.perf_counter()
                        meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
                        meta_data["batch_data_time"] = (
                            speech_lengths.sum().item()
                            * frontend.frame_shift
                            * frontend.lfr_n
                            / 1000
                        )

                        if kwargs.get("permute", True):
                            speech = speech.permute(0, 2, 1)
@@ -664,7 +678,7 @@
            raise NotImplementedError("batch decoding is not implemented")

        contents = self.data_template(data_in[0])
        output = self.data_load_speech(contents, tokenizer, frontend, **kwargs)
        output = self.data_load_speech(contents, tokenizer, frontend, meta_data=meta_data, **kwargs)
        batch = to_device(output, kwargs["device"])

        # audio encoder
@@ -692,19 +706,20 @@
                batch_idx, :min_len, :
            ]

        label = contents["assistant"][0]
        if not kwargs.get("tearchforing", False):

            generated_ids = self.llm.generate(
                inputs_embeds=inputs_embeds, max_new_tokens=kwargs.get("max_length", 512)
            )
            generated_ids = [
                output_ids[len(input_id) :]
                for input_id, output_ids in zip(input_ids, generated_ids)
            ]
            # generated_ids = [
            #     output_ids[len(input_id) :]
            #     for input_id, output_ids in zip(input_ids, generated_ids)
            # ]
            response = tokenizer.batch_decode(
                generated_ids, skip_special_tokens=kwargs.get("skip_special_tokens", True)
            )[0]
            label = contents["assistant"][0]

            loss = None
        else:

@@ -715,13 +730,13 @@
                inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids
            )

            preds = torch.argmax(model_outputs.logits, -1)[:, source_ids.shape[1]]
            preds = torch.argmax(model_outputs.logits, -1)[:, source_ids.shape[1] :]
            response = tokenizer.batch_decode(
                preds,
                add_special_tokens=False,
                skip_special_tokens=kwargs.get("skip_special_tokens", True),
            )[0]
            loss = model_outputs.loss
            loss = model_outputs.loss.item()

        ibest_writer = None
        if kwargs.get("output_dir") is not None:
@@ -730,7 +745,8 @@
            ibest_writer = self.writer[f"{0 + 1}best_recog"]

        results = []
        result_i = {"key": key[0], "text": response, "label": label}
        response_clean = re.sub("[^\w\s\u3000\u4e00-\u9fff]+", "", response)
        result_i = {"key": key[0], "text": response, "text_tn": response_clean, "label": label}
        if loss is not None:
            result_i["loss"] = loss
        results.append(result_i)
@@ -738,5 +754,6 @@
        if ibest_writer is not None:
            ibest_writer["text"][key[0]] = response
            ibest_writer["label"][key[0]] = label
            ibest_writer["text_tn"][key[0]] = response_clean

        return results, meta_data