tadaconv-main训练报错及解决方案

tadaconv训练报错：RuntimeError: Error(s) in loading state_dict for BaseVideoModel: size mismatch for head.out.weight: copying a param with shape torch.Size([400, 768]) from checkpoint, the shape in current model is torch.Size([5, 768]). size mismatch for head.out.bias: copying a param with shape torch.Size([400]) from checkpoint, the shape in current model is torch.Size([5]).*

1.首先看配置文件

_BASE_RUN: ../../pool/run/training/from_scratch_large.yaml
_BASE_MODEL: ../../pool/backbone/tadaconvnextv2_tiny.yaml

PRETRAIN:
  ENABLE: false
  GENERATOR:
TRAIN:
  ENABLE: true
  DATASET: ucf101
  BATCH_SIZE: 32 #total batch size: 128x4=512
  FINE_TUNE: true
  LR_REDUCE: true
  INIT: in1k
  CHECKPOINT_FILE_PATH: "/TAdaConv-main/tadaconvnextv2_tiny_in1k_k400_79.6.pyth"
TEST:
  ENABLE: true
  DATASET: ucf101
  BATCH_SIZE: 32
DATA:
  DATA_ROOT_DIR: /mmaction/data/ucf5/videos
  ANNO_DIR: mma/mmaction/data/ucf5/videos
  SAMPLING_RATE: 5
  NUM_INPUT_FRAMES: 16
  TRAIN_JITTER_SCALES: [0.08, 1.0]
  TRAIN_CROP_SIZE: 224
  TEST_SCALE: 256
  TEST_CROP_SIZE: 256
VIDEO:
  BACKBONE:
    DROP_PATH: 0.2
  HEAD:
    NUM_CLASSES: 5
    DROPOUT_RATE: 0.5

OUTPUT_DIR: /yxli/mma/TAdaConv-main/out/ucf5_pretrain

OPTIMIZER:
  BASE_LR: 5e-4
  ADJUST_LR: false
  LR_POLICY: cosine
  MAX_EPOCH: 300
  MOMENTUM: 0.9
  WEIGHT_DECAY: 0.02
  WARMUP_EPOCHS: 8
  WARMUP_START_LR: 1e-6
  OPTIM_METHOD: adamw
  DAMPENING: 0.0
  NESTEROV: true
  HEAD_LRMULT: 10
  NEW_PARAMS: ["dwconv_rf", "norm_avgpool"]
  NEW_PARAMS_MULT: 10
AUGMENTATION: 
  COLOR_AUG: true
  GRAYSCALE: 0.2
  COLOR_P: 0.0
  CONSISTENT: true
  SHUFFLE: true
  GRAY_FIRST: false
  IS_SPLIT: false
  USE_GPU: false
  SSV2_FLIP: true
  RATIO: [0.75, 1.333]
  MIXUP: 
    ENABLE: false
  CUTMIX: 
    ENABLE: false
  RANDOM_ERASING:
    ENABLE: false
  LABEL_SMOOTHING: 0.0
  AUTOAUGMENT:
    ENABLE: true
    BEFORE_CROP: true
    TYPE: rand-m9-n4-mstd0.5-inc1
NUM_GPUS: 2

DATA_LOADER:
  NUM_WORKERS: 0
  PIN_MEMORY: true
  SHARD_ID: [0,1]

2.因为作者使用了自己的数据集且数据集只有5个分类，但是加载的预训练模型tadaconvnextv2_tiny_in1k_k400_79.6.pyth是由Kinetics 400训练得到的，所以我们输出checkpoint[“model_state”][“head.out.weight”] ，checkpoint[“model_state”][“head.out.bias”]可以看到形状与构建的模型不符。
所以在/TAdaConv-main/tadaconv/utils/checkpoint.py下找到加载预训练模型代码：

mismatch = ms.load_state_dict(checkpoint["model_state"], strict=False)

在这行代码前加入：

del checkpoint["model_state"]["head.out.weight"] 
del checkpoint["model_state"]["head.out.bias"]

运行，解决问题。