AttributeError: cannot assign parameters before Module.__init__() call

Describe the bug

---

AttributeError Traceback (most recent call last) <ipython-input-39-048b789dea33> in <module> ----> 1 crop = K.RandomResizedCrop((224, 224), scale=p([0.08, 1.0]))

~/anaconda3/envs/dino/lib/python3.8/site-packages/kornia/augmentation/_2d/geometric/resized_crop.py in init(self, size, scale, ratio, resample, same_on_batch, align_corners, p, keepdim, cropping_mode, return_transform) 81 p=1.0, return_transform=return_transform, same_on_batch=same_on_batch, p_batch=p, keepdim=keepdim 82 ) —> 83 self._param_generator = cast(rg.ResizedCropGenerator, rg.ResizedCropGenerator(size, scale, ratio)) 84 self.flags = dict( 85 size=size, resample=Resample.get(resample), align_corners=align_corners, cropping_mode=cropping_mode

~/anaconda3/envs/dino/lib/python3.8/site-packages/kornia/augmentation/random_generator/base.py in call(cls, *args, **kwargs) 9 10 def call(cls, *args, **kwargs): —> 11 obj = type.call(cls, *args, **kwargs) 12 obj.post_init() 13 return obj

~/anaconda3/envs/dino/lib/python3.8/site-packages/kornia/augmentation/random_generator/_2d/crop.py in init(self, output_size, scale, ratio) 160 ratio: Union[torch.Tensor, Tuple[float, float]], 161 ) -> None: –> 162 self.scale = scale 163 self.ratio = ratio 164 self.output_size = output_size

~/anaconda3/envs/dino/lib/python3.8/site-packages/torch/nn/modules/module.py in setattr(self, name, value) 960 if isinstance(value, Parameter): 961 if params is None: –> 962 raise AttributeError( 963 “cannot assign parameters before Module.init() call”) 964 remove_from(self.dict, self._buffers, self._modules, self._non_persistent_buffers_set)

AttributeError: cannot assign parameters before Module.init() call

Reproduction steps

(this is code slightly adapted from the paper "Differentiable Data Augmentation with Kornia": https://arxiv.org/pdf/2011.09832.pdf)

import kornia.augmentation as K
import torch
import torch.nn as nn
t = lambda x: torch.tensor(x)
p = lambda x: nn.Parameter(t(x))
import torchvision
import matplotlib.pyplot as plt

from PIL import Image
img = Image.open('/path/to/image.jpg')
img = torchvision.transforms.functional.pil_to_tensor(img).float()
img = (img - img.min()) / (img.max() - img.min())
images = img.clone().detach().requires_grad_(True)

crop = K.RandomResizedCrop((224, 224), scale=p([0.08, 1.0]))

Expected behavior

If scale is a static tensor, there is no problem

# p([0.08, 1.0])) is changed to t([0.08, 1.0])) crop = K.RandomResizedCrop((224, 224), scale=t([0.08, 1.0]))

I would expect K.RandomResizedCrop to work when parameters are fed as nn.Parameters and not only with static tensors

Environment

PyTorch version: 1.8.1
Is debug build: False
CUDA used to build PyTorch: 10.2
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.3 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: 10.0.0-4ubuntu1 
CMake version: Could not collect
Libc version: glibc-2.31

Python version: 3.8.2 (default, Mar 26 2020, 15:53:00)  [GCC 7.3.0] (64-bit runtime)
Python platform: Linux-5.13.0-37-generic-x86_64-with-glibc2.10
Is CUDA available: True
CUDA runtime version: Could not collect
GPU models and configuration: GPU 0: NVIDIA GeForce GTX 1650
Nvidia driver version: 470.103.01
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

Versions of relevant libraries:
[pip3] numpy==1.20.2
[pip3] torch==1.8.1
[pip3] torchaudio==0.8.0a0+e4e171a
[pip3] torchvision==0.9.1
[pip3] torchviz==0.0.2
[conda] blas                      1.0                         mkl  
[conda] cudatoolkit               10.2.89              hfd86e86_1  
[conda] ffmpeg                    4.3                  hf484d3e_0    pytorch
[conda] mkl                       2021.2.0           h06a4308_296  
[conda] mkl-service               2.3.0            py38h27cfd23_1  
[conda] mkl_fft                   1.3.0            py38h42c9631_2  
[conda] mkl_random                1.2.1            py38ha9443f7_2  
[conda] numpy                     1.20.2           py38h2d18471_0  
[conda] numpy-base                1.20.2           py38hfae3a4d_0  
[conda] pytorch                   1.8.1           py3.8_cuda10.2_cudnn7.6.5_0    pytorch
[conda] torchaudio                0.8.1                      py38    pytorch
[conda] torchvision               0.9.1                py38_cu102    pytorch
[conda] torchviz                  0.0.2                    pypi_0    pypi

Thank you in advance!

[shijianjian]

@tileb1 Thanks for reporting. I will fix it this weekend.

The documented methods shall work with earlier version like 0.6.0, before the refactor of the new random generator.

[tileb1]

@shijianjian then, I get:

tensor([0., 0.])
tensor([0., 0.])

So now at least I have gradient, but they are 0 for some reason. If you check the loss I use

loss = nn.MSELoss()(out, images[:, :, :224, :224])

(so I’m enforcing the output of crop to be the upper left crop of size 224x224) I don’t see why the gradients are 0. In fact, if I change the loss to simpler things, I still get gradients of 0 e.g.

loss = nn.MSELoss()(out, torch.tensor(0.0))