预训练模型参数不匹配

(34条消息) strict=False 但还是size mismatch for []: copying a param with shape [] from checkpoint,the shape in cur_Adenialzz的博客-CSDN博客

model = ViT(num_classes=10)
ckpt = torch.load('vit-in1k.pth', map_location='cpu')
msg = model.load_state_dict(ckpt, strict=False)
print(msg)

# 剔除掉，不加载即可
ckpt.pop('head.weight')
ckpt.pop('head.bias')