yaoyao-liu/meta-transfer-learning

how to fix feature encoder weights during SS process

mengruwg opened this issue · 2 comments

mengruwg commented

during pythorch version
new_weight = self.weight.mul(new_mtl_weight)(line 95 in conv2d_mtl.py)
self.weight = Parameter(torch.Tensor(out_channels, in_channels // groups, *kernel_size))(line 42 in conv2d_mtl.py)
How to load pretrained and fix feature encoder weights during SS process?

yaoyao-liu commented

How to freeze the convolution weights?

If you set mtl=True in the following line:

meta-transfer-learning/pytorch/models/resnet_mtl.py

Line 165 in fe189c9

def __init__(self, layers=[4, 4, 4], mtl=True):

It means that you're using _ConvNdMtl function, so

meta-transfer-learning/pytorch/models/conv2d_mtl.py

Line 45 in fe189c9

self.weight.requires_grad=False

i.e., the convolution weights are frozen.

How to load the pre-trained model?

You may directly load normal checkpoints to MTL models like this:

meta-transfer-learning/pytorch/trainer/meta.py

Line 69 in fe189c9

pretrained_dict = torch.load(self.args.init_weights)['params']

If you have any further questions, feel free to ask.

mengruwg commented