在FACED数据集上不加载预训练权重训练模型，结果异常

[trefoil0219]

您好！感谢CbraMod项目中脑电领域做出的突出贡献。我对数据集FACED进行不加载预训练权重的训练，得到的结果与您的论文中报告的结果差异过大。我现在怀疑我的参数设置有误。
我运行finetune_main.py文件进行微调，以下是我的代码运行日志：
faced.txt
参数设置如下(日志太长不看版)：

Namespace(seed=3407, cuda=0, epochs=50, batch_size=64, lr=0.0001, weight_decay=0.05, optimizer='AdamW', clip_value=1, dropout=0.1, classifier='all_patch_reps', downstream_dataset='FACED', datasets_dir='/data/faced/processed_for_finetuning', num_of_classes=9, model_dir='./finetuned_model/', num_workers=16, label_smoothing=0.1, multi_lr=True, frozen=False, use_pretrained_weights=False, foundation_dir='pretrained_weights/pretrained_weights.pth')
The downstream dataset is FACED
6720 1680 1932
10332
Model(
  (backbone): CBraMod(
    (patch_embedding): PatchEmbedding(
      (positional_encoding): Sequential(
        (0): Conv2d(200, 200, kernel_size=(19, 7), stride=(1, 1), padding=(9, 3), groups=200)
      )
      (proj_in): Sequential(
        (0): Conv2d(1, 25, kernel_size=(1, 49), stride=(1, 25), padding=(0, 24))
        (1): GroupNorm(5, 25, eps=1e-05, affine=True)
        (2): GELU(approximate='none')
        (3): Conv2d(25, 25, kernel_size=(1, 3), stride=(1, 1), padding=(0, 1))
        (4): GroupNorm(5, 25, eps=1e-05, affine=True)
        (5): GELU(approximate='none')
        (6): Conv2d(25, 25, kernel_size=(1, 3), stride=(1, 1), padding=(0, 1))
        (7): GroupNorm(5, 25, eps=1e-05, affine=True)
        (8): GELU(approximate='none')
      )
      (spectral_proj): Sequential(
        (0): Linear(in_features=101, out_features=200, bias=True)
        (1): Dropout(p=0.1, inplace=False)
      )
    )
    (encoder): TransformerEncoder(
      (layers): ModuleList(
        (0-11): 12 x TransformerEncoderLayer(
          (self_attn_s): MultiheadAttention(
            (out_proj): NonDynamicallyQuantizableLinear(in_features=100, out_features=100, bias=True)
          )
          (self_attn_t): MultiheadAttention(
            (out_proj): NonDynamicallyQuantizableLinear(in_features=100, out_features=100, bias=True)
          )
          (linear1): Linear(in_features=200, out_features=800, bias=True)
          (dropout): Dropout(p=0.1, inplace=False)
          (linear2): Linear(in_features=800, out_features=200, bias=True)
          (norm1): LayerNorm((200,), eps=1e-05, elementwise_affine=True)
          (norm2): LayerNorm((200,), eps=1e-05, elementwise_affine=True)
          (dropout1): Dropout(p=0.1, inplace=False)
          (dropout2): Dropout(p=0.1, inplace=False)
        )
      )
    )
    (proj_out): Identity()
  )
  (classifier): Sequential(
    (0): Rearrange('b c s d -> b (c s d)')
    (1): Linear(in_features=64000, out_features=2000, bias=True)
    (2): ELU(alpha=1.0)
    (3): Dropout(p=0.1, inplace=False)
    (4): Linear(in_features=2000, out_features=200, bias=True)
    (5): ELU(alpha=1.0)
    (6): Dropout(p=0.1, inplace=False)
    (7): Linear(in_features=200, out_features=9, bias=True)
  )
)
结果：
Epoch 50 : Training Loss: 2.19850, acc: 0.11111, kappa: 0.00000, f1: 0.03571, LR: 0.00000, Time elapsed 0.47 mins
[[  0   0   0   0 180   0   0   0   0]
 [  0   0   0   0 180   0   0   0   0]
 [  0   0   0   0 180   0   0   0   0]
 [  0   0   0   0 180   0   0   0   0]
 [  0   0   0   0 240   0   0   0   0]
 [  0   0   0   0 180   0   0   0   0]
 [  0   0   0   0 180   0   0   0   0]
 [  0   0   0   0 180   0   0   0   0]
 [  0   0   0   0 180   0   0   0   0]]

这与论文中的结果差距过大，混淆矩阵也很异常，您能否帮我查看是参数设置或者其他地方出现了问题或者错误？另外您在SEED-V数据集上的不加载预训练权重的训练参数设置是怎么样的，非常感谢！

[ChenHu-ML]

@trefoil0219 我在很多数据集都出现过这个情况，有时候加载预训练权重也会出现。请问你目前是否有解决方案？