Implementation of VGG16 on Pytorch giving size mismatch error

Question

Snippet of my code implementation on PyTorch is:

model = models.vgg16(pretrained = False)
classifier = nn.Sequential(
        nn.Linear(25088, 128),
        nn.ReLU(True),
        nn.Dropout(),
        nn.Linear(128, 128),
        nn.ReLU(True),
        nn.Dropout(),
        nn.Linear(128, 20)
)
model.classifier = classifier

I'm feeding images of input size (60x60x3) and batch_size = 30.

When I run the code from Linux (Ubuntu) Terminal (with PyTorch Version: 1.0.0, Torchvision Version: 0.2.1) it gives me, the following error message:

RuntimeError: size mismatch, m1: [30 x 512], m2: [25088 x 128]

While, when I run it from Spyder (Anaconda) on Windows (with PyTorch Version: 1.0.1, Torchvision Version: 0.2.2), it runs perfectly.

Am I missing something or is this because of some version mismatch in Pytorch and Torchvision? Both, I'm running on Python 3.6. Please suggest.

[UPDATE: Mistakenly interchanged the version numbers for the error-case and error-free case. Thanks @Manoj Mohan for pointing it out]

Manoj Mohan · Accepted Answer · 2019-03-13 19:11:33Z

1

It's probably the other way around. Things run perfectly on torchvision 0.2.2 and fails on torch vision 0.2.1.

This change of using AdaptiveAvgPool2d that went into 0.2.2 is why you don't see the error. https://github.com/pytorch/vision/commit/83b2dfb2ebcd1b0694d46e3006ca96183c303706

>>> import torch
>>> model = models.vgg16(pretrained = False)
>>> x = torch.randn(1,3,60,60) # random image
>>> feat = model.features(x)
>>> flat_feat = feat.view(feat.size(0), -1) # flatten
>>> flat_feat.shape
torch.Size([1, 512])
>>> model.classifier(flat_feat)

RuntimeError: size mismatch, m1: [1 x 512], m2: [25088 x 4096] at /pytorch/aten/src/TH/generic/THTensorMath.cpp:940

You see the error of size mismatch. After, adaptive average pooling, things work fine.

>>> import torch.nn.functional as F
>>> avg = F.adaptive_avg_pool2d(feat, (7,7))
>>> avg = avg.view(avg.size(0), -1)
>>> output = model.classifier(avg)
>>> output.shape
torch.Size([1, 1000])

edited Mar 13, 2019 at 19:11

answered Mar 13, 2019 at 19:00

Manoj Mohan

6,0741 gold badge24 silver badges26 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

Prithwish Jana Over a year ago

Yes...sorry, I just checked...I had mistakenly interchanged the version numbers while writing... Thanks for your answer :)

Collectives™ on Stack Overflow

Implementation of VGG16 on Pytorch giving size mismatch error

1 Answer 1

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related