Vision: AdaptiveAvgPool2d in resnet not working with small images eg.64x64

[lsqshr]

When I try to forward a tensor of size Bx3x64x64 through resnet.ResNet(...),
at x = self.avgpool(x)
it complains

RuntimeError: Given input size: (512x2x2). Calculated output size: (512x-4x-4). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47

which does not make sense.

I found that the errors were gone if the AdaptivePool2d is intialized as
self.avgpool = nn.AdaptiveAvgPool2d(1) instead of self.avgpool = nn.AdaptiveAvgPool2d((1, 1))

It might be a bug in nn.AdaptiveAvgPool2d

[fmassa]

This is very weird, because the error message that you have is from AvgPool2d. Maybe you have an older version of torchvision that doesn't have the nn.AdaptiveAvgPool2d improvement?

[githubssd]

I think you should update the torchvision version and see if the issue is solved.

[lsqshr]

Thnaks. Will try on a clean conda environment shortly and confirm it when
I'm back from a short trip.

On Sun., 23 Dec. 2018, 12:19 am Dhirendra Kumar <[email protected]
wrote:

I think you should update the torchvision version and see if the issue is
solved.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/pytorch/vision/issues/696#issuecomment-449615225, or mute
the thread
https://github.com/notifications/unsubscribe-auth/ABfshdfg04ZSaKjflUwsVrMRM8Bniy7zks5u7xJRgaJpZM4ZbLpz
.

[lsqshr]

Just confirmed the bug on a clean conda environment with Python 3.6.8, Ubuntu18.04, cuda9.1

outputs of pip freeze | grep torch

torch==1.0.0
torchvision==0.2.1

To reproduce,

from torchvision.models.resnet import resnet18, resnet50, resnet101
import torch

for sz in (64, 96, 128, 256):
    for func in (resnet18, resnet50, resnet101):
        try:
            print('== Testing 3x{}x{} input for {}'.format(sz, sz, func))
            model = func()
            model.cuda()
            x = torch.FloatTensor(1, 3, sz, sz).cuda()
            model(x)
            print('!! Passed !!')
        except RuntimeError as e:
            print(e)

Output log on my side:

== Testing 3x64x64 input for <function resnet18 at 0x7fcac6297730>
Given input size: (512x2x2). Calculated output size: (512x-4x-4). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x64x64 input for <function resnet50 at 0x7fcac6297bf8>
Given input size: (2048x2x2). Calculated output size: (2048x-4x-4). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x64x64 input for <function resnet101 at 0x7fcac6297c80>
Given input size: (2048x2x2). Calculated output size: (2048x-4x-4). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x96x96 input for <function resnet18 at 0x7fcac6297730>
Given input size: (512x3x3). Calculated output size: (512x-3x-3). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x96x96 input for <function resnet50 at 0x7fcac6297bf8>
Given input size: (2048x3x3). Calculated output size: (2048x-3x-3). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x96x96 input for <function resnet101 at 0x7fcac6297c80>
Given input size: (2048x3x3). Calculated output size: (2048x-3x-3). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x128x128 input for <function resnet18 at 0x7fcac6297730>
Given input size: (512x4x4). Calculated output size: (512x-2x-2). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x128x128 input for <function resnet50 at 0x7fcac6297bf8>
Given input size: (2048x4x4). Calculated output size: (2048x-2x-2). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x128x128 input for <function resnet101 at 0x7fcac6297c80>
Given input size: (2048x4x4). Calculated output size: (2048x-2x-2). Output size is too small at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:47
== Testing 3x256x256 input for <function resnet18 at 0x7fcac6297730>
size mismatch, m1: [1 x 2048], m2: [512 x 1000] at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THC/generic/THCTensorMathBlas.cu:266
== Testing 3x256x256 input for <function resnet50 at 0x7fcac6297bf8>
size mismatch, m1: [1 x 8192], m2: [2048 x 1000] at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THC/generic/THCTensorMathBlas.cu:266
== Testing 3x256x256 input for <function resnet101 at 0x7fcac6297c80>
size mismatch, m1: [1 x 8192], m2: [2048 x 1000] at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THC/generic/THCTensorMathBlas.cu:266

It is should be caused by a bug in SpatialAveragePooling.cu. But the simple fix in https://github.com/pytorch/vision/pull/697 should work it around for torchvision.

[timonbimon]

i think torchvision==0.2.1 does not have the AdaptiveAvgPool2d layer yet...

[fmassa]

@timonbimon is right, the latest torchvision version available doesn't have AdaptiveAvgPool2d.

We will be cutting a new release soon, which will have this fix.

[wangjinjie722]

Still cannot solve this problem in the latest or 0.2.1 version torchvision.

[fmassa]

@wangjinjie722 if you want to use a very small image size, you need to modify the network architecture, as we don't provide models for cifar-sized images, only for imagenet-sized

[Quetzalcohuatl]

This is annoying and I experience the same problem when using 64x64 image size. Replacing with AdaptiveAvgPool2d fixed it.

[fmassa]

@Quetzalcohuatl we already use AdaptiveAvgPool2d in resnet, so this shouldn't be a problem with torchvision 0.4 or 0.5

[Quetzalcohuatl]

@fmassa I encountered this problem when dealing with se-resnext in the package pretrainedmodels. So I made the post on the wrong Python package. My bad, I was not distinguishing between pretrainedmodels and torchvision

[kimdn]

Just reporting, 64x64 images do not work for AvgPool2d as well.

self.Avergepool = nn.Sequential( nn.BatchNorm2d(512), nn.ReLU(inplace=True), nn.AvgPool2d(kernel_size=6, stride=1) )

```x = self.Avergepool(x)

File "/people/kimd999/bin/Miniconda3-latest-Linux-x86_64/envs/pytorch/lib/python3.8/site-packages/torch/nn/modules/module.py", line 550, in __call__
result = self.forward(input, *kwargs)

File "/people/kimd999/bin/Miniconda3-latest-Linux-x86_64/envs/pytorch/lib/python3.8/site-packages/torch/nn/modules/container.py", line 100, in forward
input = module(input)

File "/people/kimd999/bin/Miniconda3-latest-Linux-x86_64/envs/pytorch/lib/python3.8/site-packages/torch/nn/modules/module.py", line 550, in __call__
result = self.forward(input, *kwargs)

File "/people/kimd999/bin/Miniconda3-latest-Linux-x86_64/envs/pytorch/lib/python3.8/site-packages/torch/nn/modules/pooling.py", line 553, in forward
return F.avg_pool2d(input, self.kernel_size, self.stride,
RuntimeError: Given input size: (512x4x4). Calculated output size: (512x-1x-1). Output size is too small```