摘要
加载预训练权重，或者别人训练的权重的时候，更改了类别的个数，导致尺寸不匹配。这种情况非常常见，那么如何解决呢？
详细问题
问题如下：
出现了尺寸不匹配的情况。打印权重的，详细结果：
odict_keys([‘downsample_layers.0.0.weight’,
‘downsample_layers.0.0.bias’, ‘downsample_layers.0.1.weight’,
‘downsample_layers.0.1.bias’, ‘downsample_layers.1.0.weight’,
‘downsample_layers.1.0.bias’, ‘downsample_layers.1.1.weight’,
‘downsample_layers.1.1.bias’, ‘downsample_layers.2.0.weight’,
‘downsample_layers.2.0.bias’, ‘downsample_layers.2.1.weight’,
‘downsample_layers.2.1.bias’, ‘downsample_layers.3.0.weight’,
‘downsample_layers.3.0.bias’, ‘downsample_layers.3.1.weight’,
‘downsample_layers.3.1.bias’, ‘stages.0.0.dkl.cv1.conv.weight’,
‘stages.0.0.dkl.cv1.bn.weight’, ‘stages.0.0.dkl.cv1.bn.bias’,
‘stages.0.0.dkl.cv1.bn.running_mean’,
‘stages.0.0.dkl.cv1.bn.running_var’,
‘stages.0.0.dkl.cv1.bn.num_batches_tracked’,
‘stages.0.0.dkl.cv2.conv.weight’, ‘stages.0.0.dkl.cv2.bn.weight’,
‘stages.0.0.dkl.cv2.bn.bias’, ‘stages.0.0.dkl.cv2.bn.running_mean’,
‘stages.0.0.dkl.cv2.bn.running_var’,
‘stages.0.0.dkl.cv2.bn.num_batches_tracked’,
‘stages.0.0.dkl.conv1.weight’, ‘stages.0.0.ffn.fc1.weight’,
‘stages.0.0.ffn.fc1.bias’, ‘stages.0.0.ffn.dwconv.weight’,
‘stages.0.0.ffn.dwconv.bias’, ‘stages.0.0.ffn.fc2.weight’,
‘stages.0.0.ffn.fc2.bias’, ‘stages.0.0.ffn.decompose.weight’,
‘stages.0.0.ffn.decompose.bias’, ‘stages.0.0.ffn.sigma.scale’,
‘stages.0.1.dkl.cv1.conv.weight’, ‘stages.0.1.dkl.cv1.bn.weight’,
‘stages.0.1.dkl.cv1.bn.bias’, ‘stages.0.1.dkl.cv1.bn.running_mean’,
‘stages.0.1.dkl.cv1.bn.running_var’,
‘stages.0.1.dkl.cv1.bn.num_batches_tracked’,
‘stages.0.1.dkl.cv2.conv.weight’, ‘stages.0.1.dkl.cv2.bn.weight’,
‘stages.0.1.dkl.cv2.bn.bias’, ‘stages.0.1.dkl.cv2.bn.running_mean’,
‘stages.0.1.dkl.cv2.bn.running_var’,
‘stages.0.1.dkl.cv2.bn.num_batches_tracked’,
‘stages.0.1.dkl.conv1.weight’, ‘stages.0.1.ffn.fc1.weight’,
‘stages.0.1.ffn.fc1.bias’, ‘stages.0.1.ffn.dwconv.weight’,
‘stages.0.1.ffn.dwconv.bias’, ‘stages.0.1.ffn.fc2.weight’,
‘stages.0.1.ffn.fc2.bias’, ‘stages.0.1.ffn.decompose.weight’,
‘stages.0.1.ffn.decompose.bias’, ‘stages.0.1.ffn.sigma.scale’,
‘stages.0.2.dkl.cv1.conv.weight’, ‘stages.0.2.dkl.cv1.bn.weight’,
‘stages.0.2.dkl.cv1.bn.bias’, ‘stages.0.2.dkl.cv1.bn.running_mean’,
‘stages.0.2.dkl.cv1.bn.running_var’,
‘stages.0.2.dkl.cv1.bn.num_batches_tracked’,
‘stages.0.2.dkl.cv2.conv.weight’, ‘stages.0.2.dkl.cv2.bn.weight’,
‘stages.0.2.dkl.cv2.bn.bias’, ‘stages.0.2.dkl.cv2.bn.running_mean’,
‘stages.0.2.dkl.cv2.bn.running_var’,
‘stages.0.2.dkl.cv2.bn.num_batches_tracked’,
‘stages.0.2.dkl.conv1.weight’, ‘stages.0.2.ffn.fc1.weight’,
‘stages.0.2.ffn.fc1.bias’, ‘stages.0.2.ffn.dwconv.weight’,
‘stages.0.2.ffn.dwconv.bias’, ‘stages.0.2.ffn.fc2.weight’,
‘stages.0.2.ffn.fc2.bias’, ‘stages.0.2.ffn.decompose.weight’,
‘stages.0.2.ffn.decompose.bias’, ‘stages.0.2.ffn.sigma.scale’,
‘stages.1.0.dkl.cv1.conv.weight’, ‘stages.1.0.dkl.cv1.bn.weight’,
‘stages.1.0.dkl.cv1.bn.bias’, ‘stages.1.0.dkl.cv1.bn.running_mean’,
‘stages.1.0.dkl.cv1.bn.running_var’,
‘stages.1.0.dkl.cv1.bn.num_batches_tracked’,
‘stages.1.0.dkl.cv2.conv.weight’, ‘stages.1.0.dkl.cv2.bn.weight’,
‘stages.1.0.dkl.cv2.bn.bias’, ‘stages.1.0.dkl.cv2.bn.running_mean’,
‘stages.1.0.dkl.cv2.bn.running_var’,
‘stages.1.0.dkl.cv2.bn.num_batches_tracked’,
‘stages.1.0.dkl.conv1.weight’, ‘stages.1.0.ffn.fc1.weight’,
‘stages.1.0.ffn.fc1.bias’, ‘stages.1.0.ffn.dwconv.weight’,
‘stages.1.0.ffn.dwconv.bias’, ‘stages.1.0.ffn.fc2.weight’,
‘stages.1.0.ffn.fc2.bias’, ‘stages.1.0.ffn.decompose.weight’,
‘stages.1.0.ffn.decompose.bias’, ‘stages.1.0.ffn.sigma.scale’,
‘stages.1.1.dkl.cv1.conv.weight’, ‘stages.1.1.dkl.cv1.bn.weight’,
‘stages.1.1.dkl.cv1.bn.bias’, ‘stages.1.1.dkl.cv1.bn.running_mean’,
‘stages.1.1.dkl.cv1.bn.running_var’,
‘stages.1.1.dkl.cv1.bn.num_batches_tracked’,
‘stages.1.1.dkl.cv2.conv.weight’, ‘stages.1.1.dkl.cv2.bn.weight’,
‘stages.1.1.dkl.cv2.bn.bias’, ‘stages.1.1.dkl.cv2.bn.running_mean’,
‘stages.1.1.dkl.cv2.bn.running_var’,
‘stages.1.1.dkl.cv2.bn.num_batches_tracked’,
‘stages.1.1.dkl.conv1.weight’, ‘stages.1.1.ffn.fc1.weight’,
‘stages.1.1.ffn.fc1.bias’, ‘stages.1.1.ffn.dwconv.weight’,
‘stages.1.1.ffn.dwconv.bias’, ‘stages.1.1.ffn.fc2.weight’,
‘stages.1.1.ffn.fc2.bias’, ‘stages.1.1.ffn.decompose.weight’,
‘stages.1.1.ffn.decompose.bias’, ‘stages.1.1.ffn.sigma.scale’,
‘stages.1.2.dkl.cv1.conv.weight’, ‘stages.1.2.dkl.cv1.bn.weight’,
‘stages.1.2.dkl.cv1.bn.bias’, ‘stages.1.2.dkl.cv1.bn.running_mean’,
‘stages.1.2.dkl.cv1.bn.running_var’,
‘stages.1.2.dkl.cv1.bn.num_batches_tracked’,
‘stages.1.2.dkl.cv2.conv.weight’, ‘stages.1.2.dkl.cv2.bn.weight’,
‘stages.1.2.dkl.cv2.bn.bias’, ‘stages.1.2.dkl.cv2.bn.running_mean’,
‘stages.1.2.dkl.cv2.bn.running_var’,
‘stages.1.2.dkl.cv2.bn.num_batches_tracked’,
‘stages.1.2.dkl.conv1.weight’, ‘stages.1.2.ffn.fc1.weight’,
‘stages.1.2.ffn.fc1.bias’, ‘stages.1.2.ffn.dwconv.weight’,
‘stages.1.2.ffn.dwconv.bias’, ‘stages.1.2.ffn.fc2.weight’,
‘stages.1.2.ffn.fc2.bias’, ‘stages.1.2.ffn.decompose.weight’,
‘stages.1.2.ffn.decompose.bias’, ‘stages.1.2.ffn.sigma.scale’,
‘stages.2.0.dkl.cv1.conv.weight’, ‘stages.2.0.dkl.cv1.bn.weight’,
‘stages.2.0.dkl.cv1.bn.bias’, ‘stages.2.0.dkl.cv1.bn.running_mean’,
‘stages.2.0.dkl.cv1.bn.running_var’,
‘stages.2.0.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.0.dkl.cv2.conv.weight’, ‘stages.2.0.dkl.cv2.bn.weight’,
‘stages.2.0.dkl.cv2.bn.bias’, ‘stages.2.0.dkl.cv2.bn.running_mean’,
‘stages.2.0.dkl.cv2.bn.running_var’,
‘stages.2.0.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.0.dkl.conv1.weight’, ‘stages.2.0.ffn.fc1.weight’,
‘stages.2.0.ffn.fc1.bias’, ‘stages.2.0.ffn.dwconv.weight’,
‘stages.2.0.ffn.dwconv.bias’, ‘stages.2.0.ffn.fc2.weight’,
‘stages.2.0.ffn.fc2.bias’, ‘stages.2.0.ffn.decompose.weight’,
‘stages.2.0.ffn.decompose.bias’, ‘stages.2.0.ffn.sigma.scale’,
‘stages.2.1.dkl.cv1.conv.weight’, ‘stages.2.1.dkl.cv1.bn.weight’,
‘stages.2.1.dkl.cv1.bn.bias’, ‘stages.2.1.dkl.cv1.bn.running_mean’,
‘stages.2.1.dkl.cv1.bn.running_var’,
‘stages.2.1.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.1.dkl.cv2.conv.weight’, ‘stages.2.1.dkl.cv2.bn.weight’,
‘stages.2.1.dkl.cv2.bn.bias’, ‘stages.2.1.dkl.cv2.bn.running_mean’,
‘stages.2.1.dkl.cv2.bn.running_var’,
‘stages.2.1.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.1.dkl.conv1.weight’, ‘stages.2.1.ffn.fc1.weight’,
‘stages.2.1.ffn.fc1.bias’, ‘stages.2.1.ffn.dwconv.weight’,
‘stages.2.1.ffn.dwconv.bias’, ‘stages.2.1.ffn.fc2.weight’,
‘stages.2.1.ffn.fc2.bias’, ‘stages.2.1.ffn.decompose.weight’,
‘stages.2.1.ffn.decompose.bias’, ‘stages.2.1.ffn.sigma.scale’,
‘stages.2.2.dkl.cv1.conv.weight’, ‘stages.2.2.dkl.cv1.bn.weight’,
‘stages.2.2.dkl.cv1.bn.bias’, ‘stages.2.2.dkl.cv1.bn.running_mean’,
‘stages.2.2.dkl.cv1.bn.running_var’,
‘stages.2.2.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.2.dkl.cv2.conv.weight’, ‘stages.2.2.dkl.cv2.bn.weight’,
‘stages.2.2.dkl.cv2.bn.bias’, ‘stages.2.2.dkl.cv2.bn.running_mean’,
‘stages.2.2.dkl.cv2.bn.running_var’,
‘stages.2.2.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.2.dkl.conv1.weight’, ‘stages.2.2.ffn.fc1.weight’,
‘stages.2.2.ffn.fc1.bias’, ‘stages.2.2.ffn.dwconv.weight’,
‘stages.2.2.ffn.dwconv.bias’, ‘stages.2.2.ffn.fc2.weight’,
‘stages.2.2.ffn.fc2.bias’, ‘stages.2.2.ffn.decompose.weight’,
‘stages.2.2.ffn.decompose.bias’, ‘stages.2.2.ffn.sigma.scale’,
‘stages.2.3.dkl.cv1.conv.weight’, ‘stages.2.3.dkl.cv1.bn.weight’,
‘stages.2.3.dkl.cv1.bn.bias’, ‘stages.2.3.dkl.cv1.bn.running_mean’,
‘stages.2.3.dkl.cv1.bn.running_var’,
‘stages.2.3.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.3.dkl.cv2.conv.weight’, ‘stages.2.3.dkl.cv2.bn.weight’,
‘stages.2.3.dkl.cv2.bn.bias’, ‘stages.2.3.dkl.cv2.bn.running_mean’,
‘stages.2.3.dkl.cv2.bn.running_var’,
‘stages.2.3.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.3.dkl.conv1.weight’, ‘stages.2.3.ffn.fc1.weight’,
‘stages.2.3.ffn.fc1.bias’, ‘stages.2.3.ffn.dwconv.weight’,
‘stages.2.3.ffn.dwconv.bias’, ‘stages.2.3.ffn.fc2.weight’,
‘stages.2.3.ffn.fc2.bias’, ‘stages.2.3.ffn.decompose.weight’,
‘stages.2.3.ffn.decompose.bias’, ‘stages.2.3.ffn.sigma.scale’,
‘stages.2.4.dkl.cv1.conv.weight’, ‘stages.2.4.dkl.cv1.bn.weight’,
‘stages.2.4.dkl.cv1.bn.bias’, ‘stages.2.4.dkl.cv1.bn.running_mean’,
‘stages.2.4.dkl.cv1.bn.running_var’,
‘stages.2.4.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.4.dkl.cv2.conv.weight’, ‘stages.2.4.dkl.cv2.bn.weight’,
‘stages.2.4.dkl.cv2.bn.bias’, ‘stages.2.4.dkl.cv2.bn.running_mean’,
‘stages.2.4.dkl.cv2.bn.running_var’,
‘stages.2.4.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.4.dkl.conv1.weight’, ‘stages.2.4.ffn.fc1.weight’,
‘stages.2.4.ffn.fc1.bias’, ‘stages.2.4.ffn.dwconv.weight’,
‘stages.2.4.ffn.dwconv.bias’, ‘stages.2.4.ffn.fc2.weight’,
‘stages.2.4.ffn.fc2.bias’, ‘stages.2.4.ffn.decompose.weight’,
‘stages.2.4.ffn.decompose.bias’, ‘stages.2.4.ffn.sigma.scale’,
‘stages.2.5.dkl.cv1.conv.weight’, ‘stages.2.5.dkl.cv1.bn.weight’,
‘stages.2.5.dkl.cv1.bn.bias’, ‘stages.2.5.dkl.cv1.bn.running_mean’,
‘stages.2.5.dkl.cv1.bn.running_var’,
‘stages.2.5.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.5.dkl.cv2.conv.weight’, ‘stages.2.5.dkl.cv2.bn.weight’,
‘stages.2.5.dkl.cv2.bn.bias’, ‘stages.2.5.dkl.cv2.bn.running_mean’,
‘stages.2.5.dkl.cv2.bn.running_var’,
‘stages.2.5.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.5.dkl.conv1.weight’, ‘stages.2.5.ffn.fc1.weight’,
‘stages.2.5.ffn.fc1.bias’, ‘stages.2.5.ffn.dwconv.weight’,
‘stages.2.5.ffn.dwconv.bias’, ‘stages.2.5.ffn.fc2.weight’,
‘stages.2.5.ffn.fc2.bias’, ‘stages.2.5.ffn.decompose.weight’,
‘stages.2.5.ffn.decompose.bias’, ‘stages.2.5.ffn.sigma.scale’,
‘stages.2.6.dkl.cv1.conv.weight’, ‘stages.2.6.dkl.cv1.bn.weight’,
‘stages.2.6.dkl.cv1.bn.bias’, ‘stages.2.6.dkl.cv1.bn.running_mean’,
‘stages.2.6.dkl.cv1.bn.running_var’,
‘stages.2.6.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.6.dkl.cv2.conv.weight’, ‘stages.2.6.dkl.cv2.bn.weight’,
‘stages.2.6.dkl.cv2.bn.bias’, ‘stages.2.6.dkl.cv2.bn.running_mean’,
‘stages.2.6.dkl.cv2.bn.running_var’,
‘stages.2.6.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.6.dkl.conv1.weight’, ‘stages.2.6.ffn.fc1.weight’,
‘stages.2.6.ffn.fc1.bias’, ‘stages.2.6.ffn.dwconv.weight’,
‘stages.2.6.ffn.dwconv.bias’, ‘stages.2.6.ffn.fc2.weight’,
‘stages.2.6.ffn.fc2.bias’, ‘stages.2.6.ffn.decompose.weight’,
‘stages.2.6.ffn.decompose.bias’, ‘stages.2.6.ffn.sigma.scale’,
‘stages.2.7.dkl.cv1.conv.weight’, ‘stages.2.7.dkl.cv1.bn.weight’,
‘stages.2.7.dkl.cv1.bn.bias’, ‘stages.2.7.dkl.cv1.bn.running_mean’,
‘stages.2.7.dkl.cv1.bn.running_var’,
‘stages.2.7.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.7.dkl.cv2.conv.weight’, ‘stages.2.7.dkl.cv2.bn.weight’,
‘stages.2.7.dkl.cv2.bn.bias’, ‘stages.2.7.dkl.cv2.bn.running_mean’,
‘stages.2.7.dkl.cv2.bn.running_var’,
‘stages.2.7.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.7.dkl.conv1.weight’, ‘stages.2.7.ffn.fc1.weight’,
‘stages.2.7.ffn.fc1.bias’, ‘stages.2.7.ffn.dwconv.weight’,
‘stages.2.7.ffn.dwconv.bias’, ‘stages.2.7.ffn.fc2.weight’,
‘stages.2.7.ffn.fc2.bias’, ‘stages.2.7.ffn.decompose.weight’,
‘stages.2.7.ffn.decompose.bias’, ‘stages.2.7.ffn.sigma.scale’,
‘stages.2.8.dkl.cv1.conv.weight’, ‘stages.2.8.dkl.cv1.bn.weight’,
‘stages.2.8.dkl.cv1.bn.bias’, ‘stages.2.8.dkl.cv1.bn.running_mean’,
‘stages.2.8.dkl.cv1.bn.running_var’,
‘stages.2.8.dkl.cv1.bn.num_batches_tracked’,
‘stages.2.8.dkl.cv2.conv.weight’, ‘stages.2.8.dkl.cv2.bn.weight’,
‘stages.2.8.dkl.cv2.bn.bias’, ‘stages.2.8.dkl.cv2.bn.running_mean’,
‘stages.2.8.dkl.cv2.bn.running_var’,
‘stages.2.8.dkl.cv2.bn.num_batches_tracked’,
‘stages.2.8.dkl.conv1.weight’, ‘stages.2.8.ffn.fc1.weight’,
‘stages.2.8.ffn.fc1.bias’, ‘stages.2.8.ffn.dwconv.weight’,
‘stages.2.8.ffn.dwconv.bias’, ‘stages.2.8.ffn.fc2.weight’,
‘stages.2.8.ffn.fc2.bias’, ‘stages.2.8.ffn.decompose.weight’,
‘stages.2.8.ffn.decompose.bias’, ‘stages.2.8.ffn.sigma.scale’,
‘stages.3.0.dkl.cv1.conv.weight’, ‘stages.3.0.dkl.cv1.bn.weight’,
‘stages.3.0.dkl.cv1.bn.bias’, ‘stages.3.0.dkl.cv1.bn.running_mean’,
‘stages.3.0.dkl.cv1.bn.running_var’,
‘stages.3.0.dkl.cv1.bn.num_batches_tracked’,
‘stages.3.0.dkl.cv2.conv.weight’, ‘stages.3.0.dkl.cv2.bn.weight’,
‘stages.3.0.dkl.cv2.bn.bias’, ‘stages.3.0.dkl.cv2.bn.running_mean’,
‘stages.3.0.dkl.cv2.bn.running_var’,
‘stages.3.0.dkl.cv2.bn.num_batches_tracked’,
‘stages.3.0.dkl.conv1.weight’, ‘stages.3.0.ffn.fc1.weight’,
‘stages.3.0.ffn.fc1.bias’, ‘stages.3.0.ffn.dwconv.weight’,
‘stages.3.0.ffn.dwconv.bias’, ‘stages.3.0.ffn.fc2.weight’,
‘stages.3.0.ffn.fc2.bias’, ‘stages.3.0.ffn.decompose.weight’,
‘stages.3.0.ffn.decompose.bias’, ‘stages.3.0.ffn.sigma.scale’,
‘stages.3.1.dkl.cv1.conv.weight’, ‘stages.3.1.dkl.cv1.bn.weight’,
‘stages.3.1.dkl.cv1.bn.bias’, ‘stages.3.1.dkl.cv1.bn.running_mean’,
‘stages.3.1.dkl.cv1.bn.running_var’,
‘stages.3.1.dkl.cv1.bn.num_batches_tracked’,
‘stages.3.1.dkl.cv2.conv.weight’, ‘stages.3.1.dkl.cv2.bn.weight’,
‘stages.3.1.dkl.cv2.bn.bias’, ‘stages.3.1.dkl.cv2.bn.running_mean’,
‘stages.3.1.dkl.cv2.bn.running_var’,
‘stages.3.1.dkl.cv2.bn.num_batches_tracked’,
‘stages.3.1.dkl.conv1.weight’, ‘stages.3.1.ffn.fc1.weight’,
‘stages.3.1.ffn.fc1.bias’, ‘stages.3.1.ffn.dwconv.weight’,
‘stages.3.1.ffn.dwconv.bias’, ‘stages.3.1.ffn.fc2.weight’,
‘stages.3.1.ffn.fc2.bias’, ‘stages.3.1.ffn.decompose.weight’,
‘stages.3.1.ffn.decompose.bias’, ‘stages.3.1.ffn.sigma.scale’,
‘stages.3.2.dkl.cv1.conv.weight’, ‘stages.3.2.dkl.cv1.bn.weight’,
‘stages.3.2.dkl.cv1.bn.bias’, ‘stages.3.2.dkl.cv1.bn.running_mean’,
‘stages.3.2.dkl.cv1.bn.running_var’,
‘stages.3.2.dkl.cv1.bn.num_batches_tracked’,
‘stages.3.2.dkl.cv2.conv.weight’, ‘stages.3.2.dkl.cv2.bn.weight’,
‘stages.3.2.dkl.cv2.bn.bias’, ‘stages.3.2.dkl.cv2.bn.running_mean’,
‘stages.3.2.dkl.cv2.bn.running_var’,
‘stages.3.2.dkl.cv2.bn.num_batches_tracked’,
‘stages.3.2.dkl.conv1.weight’, ‘stages.3.2.ffn.fc1.weight’,
‘stages.3.2.ffn.fc1.bias’, ‘stages.3.2.ffn.dwconv.weight’,
‘stages.3.2.ffn.dwconv.bias’, ‘stages.3.2.ffn.fc2.weight’,
‘stages.3.2.ffn.fc2.bias’, ‘stages.3.2.ffn.decompose.weight’,
‘stages.3.2.ffn.decompose.bias’, ‘stages.3.2.ffn.sigma.scale’,
‘norm.weight’, ‘norm.bias’, ‘head.weight’, ‘head.bias’])
出现不匹配的key是最后两个'head.weight', 'head.bias'，将这两个权重改了才能用。
修改方法：
model = torch.load(resume)
print(model['state_dict'].keys())
model['state_dict']["head.weight"].resize_(classes, 768)
model['state_dict']["head.bias"].resize_(classes)
model_ft.load_state_dict(model['state_dict'],strict=True)
然后，就可以了！
