auto frontend

游雁

2024-06-07 4a99b828a834d1d3870abbe3ee477518470f3dd9

auto frontend

 funasr/datasets/audio_datasets/samplers.py

@@ -364,6 +364,8 @@
        self.sort_size = sort_size * num_replicas
        self.max_token_length = kwargs.get("max_token_length", 2048)
        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
        self.start_step = kwargs.get("start_step", 2048)

        super().__init__(
            dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle, drop_last=drop_last
        )
@@ -415,13 +417,17 @@
            rank_batches[i % self.num_replicas].append(batch)

        # Assign all batches for the current rank directly
        final_batches = rank_batches[self.rank]
        final_batches = rank_batches[self.rank][self.start_step :]
        self.batch_num = len(final_batches)

        logging.info(
            f"rank: {self.rank}, dataloader start from step: {self.start_step}, batch_num: {self.batch_num}"
        )
        return iter(final_batches)

    def __len__(self):

        return 1
        # Calculate the number of batches per epoch for the current rank
        return self.batch_num

    def set_epoch(self, epoch):
        self.epoch = epoch

 funasr/datasets/openai_datasets/datasets.py

@@ -51,7 +51,7 @@
        self.batch_size = kwargs.get("batch_size")
        self.batch_type = kwargs.get("batch_type")
        self.prompt_ids_len = 0
        self.retry = kwargs.get("retry", 5)
        self.retry = kwargs.get("retry", 10)

        self.permute = False
        from funasr.frontends.whisper_frontend import WhisperFrontend
@@ -60,6 +60,8 @@
            self.permute = True

        self.pattern = re.compile(r"(<\|startofspeech\|>.*?<\|endofspeech\|>)")
        # self.kwargs = kwargs
        self.max_token_length = kwargs.get("max_token_length", 1024)

    def get_source_len(self, index):
        item = self.index_ds[index]
@@ -77,7 +79,9 @@
        # pdb.set_trace()

        output = None

        for idx in range(self.retry):
            badcase_flag = False
            if idx == 0:
                index_cur = index
            else:
@@ -112,9 +116,14 @@
                            "<|endofspeech|>", ""
                        )
                        if sub_str.startswith("!"):

                            data_src = load_audio_text_image_video(sub_str[1:], fs=self.fs)

                            try:
                                data_src = load_audio_text_image_video(sub_str[1:], fs=self.fs)
                            except Exception as e:
                                logging.error(
                                    f"Loading wav failed! {str(e)}, {traceback.format_exc()}"
                                )
                                badcase_flag = True
                                continue
                            speech, speech_lengths = extract_fbank(
                                data_src,
                                data_type=self.data_type,
@@ -134,6 +143,8 @@
                            source_ids += sub_token
                            fbank_mask_i += [1] * len(sub_token)

                if badcase_flag:
                    continue
                source_mask = [-100] * len(source_ids)
                target_out = f"{target_out}<|im_end|>"
                target_ids = self.tokenizer.encode(target_out)
@@ -142,6 +153,10 @@
                fbank_mask += fbank_mask_i
                fbank_beg.append(fbank_beg_i)

            if len(input_ids) > self.max_token_length:
                badcase_flag = True
            if badcase_flag:
                continue
            input_ids = torch.tensor(input_ids, dtype=torch.int64)
            attention_mask = torch.tensor([len(input_ids)], dtype=torch.int32)
            labels = torch.tensor(labels, dtype=torch.int64)
@@ -186,9 +201,9 @@
                    data_list, batch_first=True, padding_value=pad_value
                )

        if self.batch_type != "example":
            for i in range(10):
                outputs = self._filter_badcase(outputs, i=i)
        # if self.batch_type != "example":
        #     for i in range(10):
        #         outputs = self._filter_badcase(outputs, i=i)

        return outputs


 funasr/datasets/openai_datasets/index_ds.py

@@ -48,7 +48,10 @@
        for file_json in file_list:
            with open(file_json.strip(), encoding="utf-8") as fin:
                for line in fin:
                    data = json.loads(line.strip())["messages"]
                    data_dict = json.loads(line.strip())
                    data = data_dict["messages"]
                    speech_length = data_dict.get("speech_length", -1) // 8
                    text_length = data_dict.get("text_length", 0)

                    system, user, assistant = [], [], []
                    for i, item in enumerate(data):
@@ -63,7 +66,12 @@

                    system = system * len(user)

                    contents_i = {"system": system, "user": user, "assistant": assistant}
                    contents_i = {
                        "system": system,
                        "user": user,
                        "assistant": assistant,
                        "source_len": speech_length + text_length,
                    }
                    contents.append(contents_i)

        self.contents = contents
@@ -80,11 +88,14 @@
        return data

    def get_source_len(self, data_dict):
        return len(data_dict["system"]) + len(data_dict["user"])
        source_len = data_dict.get("source_len", -1)
        if source_len < 0:
            source_len = len(data_dict["system"]) + len(data_dict["user"])
        return source_len

    def get_target_len(self, data_dict):

        return len(data_dict["assistant"])
        return 0


if __name__ == "__main__":