decoding

游雁

2024-06-14 08114ae27d85949106aeab03b3fa5d764d100b33

 funasr/models/llm_asr/model.py

@@ -6,7 +6,7 @@
import torch.nn as nn
import torch.nn.functional as F
from torch.cuda.amp import autocast

import re
from funasr.models.scama.utils import sequence_mask
from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
from funasr.models.ctc.ctc import CTC
@@ -19,6 +19,9 @@
from funasr.utils.datadir_writer import DatadirWriter
from funasr.register import tables
from funasr.train_utils.device_funcs import to_device
import traceback

dtype_map = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}


@tables.register("model_classes", "LLMASR")
@@ -395,7 +398,9 @@
            # frontend = model.kwargs.get("frontend")
            audio_encoder_output_size = model.model.encoder_output_size

            audio_encoder = model.model.model.encoder
            audio_encoder = (
                model.model.model.encoder if hasattr(model.model, "model") else model.model.encoder
            )

            # self.frontend = frontend

@@ -406,38 +411,60 @@
            audio_encoder = encoder_class(input_size=input_size, **audio_encoder_conf)
            audio_encoder_output_size = audio_encoder.output_size()
        freeze = audio_encoder_conf.get("freeze", True)
        freeze_layer_num = int(audio_encoder_conf.get("freeze_layer_num", -1))
        # if freeze_layer_num > 0:
        #     freeze_layer_num = range(freeze_layer_num)

        if freeze:
            for name, param in audio_encoder.named_parameters():
                param.requires_grad = False
                if freeze_layer_num > 0:
                    idx = re.search(r"\.\d+\.", name)
                    if idx is not None:
                        beg, end = idx.regs[0]
                        layer_id = int(name[beg + 1 : end - 1])
                        if layer_id < freeze_layer_num:
                            param.requires_grad = False
                    elif "ln_post." not in name:
                        param.requires_grad = False
                else:
                    param.requires_grad = False

            audio_encoder.eval()

        self.audio_encoder = audio_encoder

        # llm
        hub = llm_conf.get("hub", "hf")
        self.llm = None
        if hub == "hf":
            from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig

            init_param_path = llm_conf.get("init_param_path", "vicuna-7b-v1.5")
        from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig

            model = AutoModelForCausalLM.from_pretrained(
                init_param_path,
                load_in_8bit=None,
                device_map=None,
                use_cache=None,
            )
            freeze = llm_conf.get("freeze", True)
            if freeze:
                for name, param in model.named_parameters():
                    param.requires_grad = False
                model.eval()
            self.llm = model
        init_param_path = llm_conf.get("init_param_path", "vicuna-7b-v1.5")

        model = AutoModelForCausalLM.from_pretrained(
            init_param_path,
            load_in_8bit=None,
            device_map=None,
            use_cache=None,
        )
        freeze = llm_conf.get("freeze", True)
        if freeze:
            for name, param in model.named_parameters():
                param.requires_grad = False
            model.eval()
        self.llm_dtype = llm_conf.get("llm_dtype", "fp32")
        self.llm = model.to(dtype_map[self.llm_dtype])
        llm_dim = model.get_input_embeddings().weight.shape[-1]

        # adaptor
        adaptor_class = tables.adaptor_classes.get(audio_adaptor)
        audio_adaptor_conf["encoder_dim"] = audio_encoder_output_size
        audio_adaptor_conf["llm_dim"] = llm_dim
        audio_adaptor = adaptor_class(**audio_adaptor_conf)
        init_param_path = audio_adaptor_conf.get("init_param_path", None)
        if init_param_path is not None:
            src_state = torch.load(init_param_path, map_location="cpu")
            flag = audio_adaptor.load_state_dict(src_state, strict=False)
            logging.info(f"Loading audio_adaptor ckpt: {init_param_path}, status: {flag}")

        self.audio_adaptor = audio_adaptor

@@ -471,11 +498,12 @@

        batch_size, frames, _ = speech.shape

        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)
        with torch.cuda.amp.autocast(enabled=False):
            # audio encoder
            encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

        # audio_adaptor
        encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)
            # audio_adaptor
            encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)

        input_ids[input_ids < 0] = 0
        inputs_embeds = self.llm.model.get_input_embeddings()(input_ids)
@@ -489,6 +517,7 @@
            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, :
@@ -496,20 +525,25 @@
            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, :
                ]

        labels_ids[labels_ids == -1] = -100

        model_outputs = self.llm(
            inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids
        )
        loss = model_outputs.loss
        with torch.cuda.amp.autocast(
            enabled=True if self.llm_dtype != "fp32" else False, dtype=dtype_map[self.llm_dtype]
        ):
            labels_ids[labels_ids == -1] = -100
            attention_mask[attention_mask < 0] = 0
            model_outputs = self.llm(
                inputs_embeds=inputs_embeds.to(dtype_map[self.llm_dtype]),
                attention_mask=attention_mask,
                labels=labels_ids,
            )
            loss = model_outputs.loss

        stats = {}
        with torch.no_grad():
@@ -532,7 +566,13 @@
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight

    def data_template(self, data_in):
    def encode(self, speech, speech_lengths):
        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)

        return encoder_out, encoder_out_lens

    def data_template(self, data):
        system, user, assistant = [], [], []
        for i, item in enumerate(data):
            role = item["role"]
@@ -554,27 +594,37 @@

        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"]
        assistant = contents["assistant"]
        pattern = re.compile(r"(<\|startofspeech\|>.*?<\|endofspeech\|>)")
        input_ids, labels, fbank, fbank_lens, fbank_mask, fbank_beg = [], [], [], [], [], []
        input_ids, labels, source_ids, target_ids, fbank, fbank_lens, fbank_mask, fbank_beg = (
            [],
            [],
            [],
            [],
            [],
            [],
            [],
            [],
        )

        for i, (system_prompt, user_prompt, target_out) in enumerate(zip(system, user, assistant)):

            source_input = f"<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{user_prompt}<|im_end|>\n<|im_start|>assistant\n"

            splits = pattern.split(source_input)
            source_ids = []
            source_ids_i = []
            fbank_mask_i = []
            fbank_beg_i = []
            fbank_lens_i = []
            # target_ids_i = []
            for k, sub_str in enumerate(splits):
                if not sub_str.startswith("<|startofspeech|>"):
                    sub_token = tokenizer.encode(sub_str)
                    source_ids += sub_token
                    source_ids_i += sub_token
                    fbank_mask_i += [0] * len(sub_token)
                else:
                    sub_str = sub_str.replace("<|startofspeech|>", "").replace(
@@ -582,7 +632,10 @@
                    )
                    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()}")

@@ -593,6 +646,15 @@
                            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)

@@ -600,14 +662,14 @@
                        olens = 1 + (olens - 3 + 2 * 1) // 2
                        sub_token_len = (olens - 1) // 2 + 1
                        sub_token = [0] * sub_token_len
                        fbank_beg_i = [len(source_ids)]
                        source_ids += sub_token
                        fbank_beg_i = [len(source_ids_i)]
                        source_ids_i += sub_token
                        fbank_mask_i += [1] * len(sub_token)

            source_mask = [-100] * len(source_ids)
            source_mask = [-100] * len(source_ids_i)
            target_out = f"{target_out}<|im_end|>"
            target_ids = tokenizer.encode(target_out)
            input_ids += source_ids + target_ids
            input_ids += source_ids_i + target_ids
            labels += source_mask + target_ids
            fbank_mask += fbank_mask_i
            fbank_beg.append(fbank_beg_i)
@@ -615,7 +677,7 @@
        input_ids = torch.tensor(input_ids, dtype=torch.int64)  # [: self.max_token_length]
        attention_mask = torch.tensor([1] * len(input_ids), dtype=torch.int32)
        labels = torch.tensor(labels, dtype=torch.int64)  # [: self.max_token_length]
        source_ids = torch.tensor(source_ids, dtype=torch.int64)
        source_ids = torch.tensor(source_ids_i, dtype=torch.int64)
        target_ids = torch.tensor(target_ids, dtype=torch.int64)

        fbank = speech[0, :, :]
@@ -653,21 +715,27 @@
        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")

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

        # audio encoder
        speech = batch["speech"]
        speech_lengths = batch["speech_lengths"][:, 0]
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)
        # fp16
        if kwargs.get("fp16", False):
            speech = speech.to(torch.float16)
        elif kwargs.get("bf16", False):
            speech = speech.to(torch.bfloat16)
        # audio encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

        # audio_adaptor
        encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)

        input_ids = batch["input_ids"]
        source_ids = batch["source_ids"]
        if kwargs.get("tearchforing", False):
        if not kwargs.get("tearchforing", False):
            input_ids = source_ids
        input_ids[input_ids < 0] = 0
        inputs_embeds = self.llm.model.get_input_embeddings()(input_ids)
@@ -682,19 +750,49 @@
                batch_idx, :min_len, :
            ]

        if not kwargs.get("tearchforing", False):
        llm_dtype = kwargs.get("llm_dtype", "fp32")
        if llm_dtype == "fp32":
            llm_dtype = "fp16" if kwargs.get("fp16", False) else llm_dtype
            llm_dtype = "bf16" if kwargs.get("bf16", False) else llm_dtype

            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)
            ]
            response = tokenizer.batch_decode(
                generated_ids, skip_special_tokens=kwargs.get("skip_special_tokens", True)
            )[0]
        with torch.cuda.amp.autocast(
            enabled=True if llm_dtype != "fp32" else False, dtype=dtype_map[llm_dtype]
        ):
            label = contents["assistant"][0]
            self.llm = self.llm.to(dtype_map[llm_dtype])
            inputs_embeds = inputs_embeds.to(dtype_map[llm_dtype])

            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)
                # ]
                response = tokenizer.batch_decode(
                    generated_ids, skip_special_tokens=kwargs.get("skip_special_tokens", True)
                )[0]

                loss = None
            else:

                labels_ids = batch["labels_ids"]
                labels_ids[labels_ids == -1] = -100
                attention_mask = batch.get("attention_mask", None)
                # attention_mask = attention_mask.to(dtype_map[llm_dtype])
                model_outputs = self.llm(
                    inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids
                )

                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.item()

        ibest_writer = None
        if kwargs.get("output_dir") is not None:
@@ -703,11 +801,33 @@
            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)

        if ibest_writer is not None:
            ibest_writer["text"][key[0]] = text
            ibest_writer["text"][key[0]] = response
            ibest_writer["label"][key[0]] = label
            ibest_writer["text_tn"][key[0]] = response_clean

        return results, meta_data


@tables.register("model_classes", "LLMASR3")
class LLMASR3(LLMASR2):
    """ """

    def __init__(
        self,
        *args,
        **kwargs,
    ):

        super().__init__(*args, **kwargs)

    def encode(self, speech, speech_lengths):
        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech, speech_lengths)
        return encoder_out, encoder_out_lens