How to transfer weight/bias parameters from one (architecture A) to another (architecture A')?

1

There is a network that has completed training with VGG16. It was saved using torch.save.
e.g.

        state = {
            'net': net.module if use_cuda else net,
            'acc': acc,
        }
        if not os.path.isdir('checkpoint'):
            os.mkdir('checkpoint')
        torch.save(state, './checkpoint/ckpt_20190312_cifar10_vgg16.t7')
        best_acc = acc

I would like to implement VGG16_v2 with the same structure as VGG16 and using other convolution operations.

At this time, I want to import the weight / bias values of the pretrained VGG16.

Once I have implemented the following:

checkpoint = torch.load('./checkpoint/ckpt_20190312_cifar10_vgg16.t7') # Load your own neural network models
net = checkpoint['net']

model = VGG16_v2.VGG()

model.load_state_dict(net, strict=False)

When I implemented this, I got the following error.

Traceback (most recent call last):
File “main_20190325_cifar10_vgg16_alpha_training_v1.py”, line 75, in
model.load_state_dict(net, strict=False)
File “/home/mhha/.conda/envs/pytorchmh/lib/python3.6/site-packages/torch/nn/modules/module.py”, line 695, in load_state_dict
state_dict = state_dict.copy()
File “/home/mhha/.conda/envs/pytorchmh/lib/python3.6/site-packages/torch/nn/modules/module.py”, line 532, in getattr
type(self).name, name))
AttributeError: ‘VGG’ object has no attribute ‘copy’

How can I solve this problem?

2

This seems like it’s related to the way you saved the model, rather than the changes in the architecture.

Instead of doing model.load_state_dict(net), try model.load_state_dict(net.state_dict)! The state_dict of the model is the weights and biases of the network, whereas the model itself is the entire object, with the architecture, weights and other elements included.

3

Unfortunately, it doesn’t work…

Is there any method that is not using torch.load??

4

Haha, using state_dict(), it works!

1 Like
5

Sorry about that, I forgot the parentheses :sweat_smile: glad it works!

6

Your solution helps a lot!!

Thanks!