How to train Inceptionv3 model with high-resolution images (ex. 512x512, 2048x2048)?

[hongyiyu]

I'd like to train the Inceptionv3 model with high-resolution images like 512x512.
As I know, the Inceptionv3 is designed to work with variable size of images.
However, if I train it with images which is not 299x299 size, it return with errors like:
size mismatch, m1: [8 x 19200], m2: [768 x 1000]

It's looks like have some problem with this line.

if self.training and self.aux_logits:
aux = self.AuxLogits(x)`

vision/torchvision/models/inception.py

Line 108 in 62cbf0b

if self.training and self.aux_logits:

Anybody can help me?

Thanks a lot in advance

[fmassa]

Hi,

The issue is that we have a fixed avg_pool2d in https://github.com/pytorch/vision/blob/master/torchvision/models/inception.py#L117 instead of a global avg pooling.

The solution for now is to copy-paste the implementation and modify that line to use a F.adaptive_avg_pool2d(x, (1, 1)).

[hongyiyu]

@fmassa
Thank you for your reply.
It seems like also need to modify the class InceptionAux
https://github.com/pytorch/vision/blob/master/torchvision/models/inception.py#L297
or
https://github.com/pytorch/vision/blob/master/torchvision/models/inception.py#L308
I want to change the kernel_size of conv1 adatively to output 1x1x768 size. However, due to I am new to Pytorch, I tried several ways, all give me errors.
Do you have any idea?

[fmassa]

No, it looks like you'd need to change https://github.com/pytorch/vision/blob/master/torchvision/models/inception.py#L304 to be F.adaptive_avg_pool2d(x, (5, 5)), or something like that.

[hongyiyu]

@fmassa

I saw someone change the auxlogits part to the following line to make the inceptionv3 available for variable image size.
https://github.com/ncoudray/DeepPATH/blob/master/DeepPATH_code/01_training/xClasses/inception/slim/inception_model.py#L257
https://github.com/ncoudray/DeepPATH/blob/master/DeepPATH_code/01_training/xClasses/inception/slim/inception_model.py#L321

I think there are three available choices, first the L117 should change to F.adaptive_avg_pool2d(x, (1, 1)), then:

1. Change the L304 to F.adaptive_avg_pool2d(x, (5, 5)) as you said,

2. Add a line F.adaptive_avg_pool2d(x, (1, 1)) below L308 x = self.conv1(x)

3. Change the kernel_size of conv1 in auxlogits (L297) adatively to output 1x1x768 size

which one do you think is most reasonable?

Thanks.

[fmassa]

I think I'd follow option 1.

[ekagra-ranjan]

I guess this can be closed due to #744.