Resnet fc size does not match about ddsm-visual-primitives HOT 7 CLOSED

Where is the test.py file you are running? I'm looking at line 164 of generate_unitsegments.py and it's not the same.

from ddsm-visual-primitives.

The test.py is just generate_unitsegments.py with a different name. I added some print statements try to debug myself so the number might not be correct. The error occurs everytime I run model(input). I have also checked input shape, which is (32, 3,227,227), which looks correct.

from ddsm-visual-primitives.

Hmm, could you attach your test.py file, and also the .../virtual_envs/ML/lib/python3.5/site-packages/torchvision/models/resnet.py file?

from ddsm-visual-primitives.

Yes, of course.
test.py.txt
resnet.py.txt

from ddsm-visual-primitives.

I changed the image size to 224,224 and fixed the issue. I also added a resize line in DDSM transforms.

from ddsm-visual-primitives.

Okay, sounds good. Please note though that resizing the image to 224x224 from 227x227 will produce different outputs and lower AUC numbers than reported in the paper, since the resizing blurs the image.

If you are curious why you were getting the original error, it is because the torchvision resnet was changed in newer versions. I took a look at your resnet.py and it is the newer version. In particular, nn.AvgPool2d(7) was changed to nn.AvgPool2d(7, stride=1) in this commit.

If you install torchvision 0.1.9, the error should go away and the code should work as-is. This pitfall was described in the README.

from ddsm-visual-primitives.

Yes, of course.
test.py.txt
resnet.py.txt

can you share full code?

from ddsm-visual-primitives.