Results are inconsistent with expectations

[dongxuanlb]

Platform: xavier. jetpack4.4.1 cuda:10.2 pytorch:1.6

the correlation_packege in this repo is for pytorch0.4 which is not suite for pytorch 1.6, so I used flownet2's (https://github.com/NVIDIA/flownet2-pytorch/tree/master/networks/correlation_package)

then everything is ok.

the output is :

/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py:3384: UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details. warnings.warn("Default grid_sample and affine_grid behavior has changed "

but the result is not same in the repository.

[ltkong218]

You can try the Pytorch Correlation module, which supports newer versions of PyTorch, such as 1.2 and 1.6, and refer to the issue Installing Correlation package. Also, please note the implementation of warping layer.

[dongxuanlb]

Hi, I had installed the Pytorch Correlation module, but I noticed that the "Pytorch Correlation module " construct is not equal with origin.

SpatialCorrelationSampler(1, 9, 1, 0, 1) <-> Correlation(pad_size=4, kernel_size=1, max_displacement=4, stride1=1, stride2=1, corr_multiply=1)

[ltkong218]

import torch
from spatial_correlation_sampler import SpatialCorrelationSampler

input1 = torch.randn(2, 32, 48, 64).cuda()
input2 = torch.randn(2, 32, 48, 64).cuda()

# define a correlation module
correlation_sampler = SpatialCorrelationSampler(1, 9, 1, 0, 1)

output = correlation_sampler(input1, input2)

# reshape output to be a 3D cost volume
b, c, h, w = input1.shape
output = output.view(b, -1, h, w) / c

print(output.shape)

I have checked that the Correlation module in my repository (PyTorch=0.4.1) and SpatialCorrelationSampler module (PyTorch=1.2) can output the same results. Remember to reshape the direct output of SpatialCorrelationSampler module and divide it by channel numbers. Also, you should check whether the input tensors are the same.