RuntimeError: Only Tensors created explicitly by the user (graph leaves) support the deepcopy protocol at the moment

[wengooooo]

I tried the sample code available in README, but i get a error

byol-pytorch  0.5.4
torch              1.7.1
kornia             0.4.1

Code:

import torch
from byol_pytorch import BYOL
from torchvision import models

resnet = models.resnet50(pretrained=True)

learner = BYOL(
    resnet,
    image_size = 256,
    hidden_layer = 'avgpool'
)

opt = torch.optim.Adam(learner.parameters(), lr=3e-4)

def sample_unlabelled_images():
    return torch.randn(20, 3, 256, 256)

for _ in range(100):
    images = sample_unlabelled_images()
    loss = learner(images)
    opt.zero_grad()
    loss.backward()
    opt.step()
    learner.update_moving_average() # update moving average of target encoder

# save your improved network
torch.save(resnet.state_dict(), './improved-net.pt')

Error

Traceback (most recent call last):
  File "demo.py", line 12, in <module>
    hidden_layer = 'avgpool'
  File "/cloud/byol-pytorch/byol_pytorch/byol_pytorch.py", line 224, in __init__
    self.forward(torch.randn(2, 3, image_size, image_size, device=device))
  File "/cloud/byol-pytorch/byol_pytorch/byol_pytorch.py", line 256, in forward
    target_encoder = self._get_target_encoder() if self.use_momentum else self.online_encoder
  File "/cloud/byol-pytorch/byol_pytorch/byol_pytorch.py", line 28, in wrapper
    instance = fn(self, *args, **kwargs)
  File "/cloud/byol-pytorch/byol_pytorch/byol_pytorch.py", line 229, in _get_target_encoder
    target_encoder = copy.deepcopy(self.online_encoder)
  File "/environment/python/versions/miniconda3-4.7.12/lib/python3.7/copy.py", line 180, in deepcopy
    y = _reconstruct(x, memo, *rv)
  File "/environment/python/versions/miniconda3-4.7.12/lib/python3.7/copy.py", line 280, in _reconstruct
    state = deepcopy(state, memo)
  File "/environment/python/versions/miniconda3-4.7.12/lib/python3.7/copy.py", line 150, in deepcopy
    y = copier(x, memo)
  File "/environment/python/versions/miniconda3-4.7.12/lib/python3.7/copy.py", line 240, in _deepcopy_dict
    y[deepcopy(key, memo)] = deepcopy(value, memo)
  File "/environment/python/versions/miniconda3-4.7.12/lib/python3.7/copy.py", line 150, in deepcopy
    y = copier(x, memo)
  File "/environment/python/versions/miniconda3-4.7.12/lib/python3.7/copy.py", line 240, in _deepcopy_dict
    y[deepcopy(key, memo)] = deepcopy(value, memo)
  File "/environment/python/versions/miniconda3-4.7.12/lib/python3.7/copy.py", line 161, in deepcopy
    y = copier(memo)
  File "/environment/python/versions/miniconda3-4.7.12/lib/python3.7/site-packages/torch/tensor.py", line 47, in __deepcopy__
    raise RuntimeError("Only Tensors created explicitly by the user "
RuntimeError: Only Tensors created explicitly by the user (graph leaves) support the deepcopy protocol at the moment

[dmitrav]

Did you resolve it? What helped?

[eufrizz]

In my Jupyter notebook, this happens to me if I run the BYOL constructor twice, reusing the network passed to it.
The solution is to recreate the resnet as well.
i.e. keep these lines together

resnet = models.resnet50(pretrained=True)

learner = BYOL(
    resnet,
    image_size = 256,
    hidden_layer = 'avgpool'
)