关于LLM-ASR-NAR的训练疑问

[NiniAndy]

❓ Questions and Help

###问题1 请问有llm-asr-nar的conf吗，我感觉我自己配置的文件不是很对：

model: LLMASRNAR
model_conf:
    lsm_weight: 0.1     # label smoothing option
    length_normalized_loss: true

encoder: SANMEncoder
encoder_conf:
    hub: funasr
    init_param_path: "/ssd/zhuang/code/LLM/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/"
    freeze: true

llm: Qwen2.5-1.5B-Instruct
llm_conf:
  hub: hf
  freeze: true
  init_param_path: "/ssd/zhuang/code/LLM/qwen2.5-1.5B-Instruct/"

adaptor: Linear
adaptor_conf:
  downsample_rate: 1
  llm_dim: 1536
  encoder_dim: 512

frontend: WavFrontend
frontend_conf:
    fs: 16000
    window: hamming
    n_mels: 80
    frame_length: 25
    frame_shift: 10
    lfr_m: 7
    lfr_n: 6
    cmvn_file: "/ssd/zhuang/code/LLM/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/am.mvn"


specaug: SpecAugLFR
specaug_conf:
    apply_time_warp: false
    time_warp_window: 5
    time_warp_mode: bicubic
    apply_freq_mask: true
    freq_mask_width_range:
    - 0
    - 30
    lfr_rate: 6
    num_freq_mask: 1
    apply_time_mask: true
    time_mask_width_range:
    - 0
    - 12
    num_time_mask: 1

train_conf:
  accum_grad: 1
  grad_clip: 5
  max_epoch: 150
  keep_nbest_models: 10
  log_interval: 150

optim: adamw
optim_conf:
   lr: 0.0001
   weight_decay: 0.000001
scheduler: warmuplr
scheduler_conf:
   warmup_steps: 1500

dataset: AudioLLMQwenAudioDataset
dataset_conf:
    index_ds: IndexDSJsonl
    batch_sampler: CustomDistributedBatchSampler
    batch_type: example # example or length
    batch_size: 16 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
    max_token_length: 3000 
    shuffle: True
    num_workers: 4
    preprocessor_text: TextPreprocessRemovePunctuation  
    audio_adaptor_downsample_rate: ${adaptor_conf.downsample_rate}
    audio_encoder_downsample_rate: 1

tokenizer: HuggingfaceTokenizer
tokenizer_conf:
  unk_symbol: <unk>
  init_param_path: "/ssd/zhuang/code/LLM/qwen2.5-1.5B-Instruct/"

###问题2

请问在funasr/model/llm_asr_nar.py中192-199中：

if audio_mask is not None:
     # audio_mask: [b, 1*enc_len+0*(prompt_len+label_len), 1]
    batch_size, token_num, dims = inputs_embeds.shape  # [b, 0*enc_len+prompt+label, dim]
    _, l, _ = encoder_out.shape # [b, enc_len,dim]
    encoder_outs_pad = F.pad(encoder_out, (0, 0, token_num - l - 1, 1, 0, 0), value=0.0)  
    # [b, 0*(prompt_len+label_len-1)+enc_len+0,dim]
    inputs_embeds = encoder_outs_pad * audio_mask[:, :, None] + inputs_embeds * (1.0 - audio_mask[:, :, None] )
    inputs_embeds = F.pad(inputs_embeds[:, 1:, :], (0, 0, 0, 1, 0, 0), value=0.0)

在这里inputs_embeds这里encoder_outs_pad在上一步已经被0填充了一段位移为什么还用原来没位移前的audio_mask去乘呢？ 用这个代码训练很快就loss=0了，感觉不是很对呢。