Can't pass a single image through a model?

fastai fastai dev
1

I’m running a model on the PETS dataset, following the tutorial at A walk with fastai2 - Vision - Study Group and Online Lectures Megathread. I want to be able to pass a single image through a resnet-based model that I’ve defined as -

class MyPetsNetwork(nn.Module):

    def __init__(self, arch= resnet18):
        super().__init__()
        self.cnn = create_body(arch)
        self.head = create_head(num_features_model(self.cnn) * 2, 37) # for 37 categories of PETS

    def forward(self, image):     
        x = self.cnn(image)
        x = self.head(x)
        return 2 * (x.sigmoid_() - 0.5)

I’ve created a model from this as model = MyPetsNetwork(resnet18).

And my single image is got by -

img = PILImage.create(fnames[0])
img = image2tensor(img)
img = img[None, :, :, :] # to create the batch dimension
img = img / 255.0

But passing it to the model as model(img) gives the error -

> /usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
>     720             result = self._slow_forward(*input, **kwargs)
>     721         else:
> --> 722             result = self.forward(*input, **kwargs)
>     723         for hook in itertools.chain(
>     724                 _global_forward_hooks.values(),
> 
> <ipython-input-35-1dc3616dbfb7> in forward(self, image)
>       7     def forward(self, image):
>       8         x = self.cnn(image)
> ----> 9         x = self.head(x)
>      10         return 2 * (x.sigmoid_() - 0.5)
>      11 
> 
> /usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
>     720             result = self._slow_forward(*input, **kwargs)
>     721         else:
> --> 722             result = self.forward(*input, **kwargs)
>     723         for hook in itertools.chain(
>     724                 _global_forward_hooks.values(),
> 
> /usr/local/lib/python3.6/dist-packages/torch/nn/modules/container.py in forward(self, input)
>     115     def forward(self, input):
>     116         for module in self:
> --> 117             input = module(input)
>     118         return input
>     119 
> 
> /usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
>     720             result = self._slow_forward(*input, **kwargs)
>     721         else:
> --> 722             result = self.forward(*input, **kwargs)
>     723         for hook in itertools.chain(
>     724                 _global_forward_hooks.values(),
> 
> /usr/local/lib/python3.6/dist-packages/torch/nn/modules/batchnorm.py in forward(self, input)
>     134             self.running_mean if not self.training or self.track_running_stats else None,
>     135             self.running_var if not self.training or self.track_running_stats else None,
> --> 136             self.weight, self.bias, bn_training, exponential_average_factor, self.eps)
>     137 
>     138 
> 
> /usr/local/lib/python3.6/dist-packages/torch/nn/functional.py in batch_norm(input, running_mean, running_var, weight, bias, training, momentum, eps)
>    2010                 bias=bias, training=training, momentum=momentum, eps=eps)
>    2011     if training:
> -> 2012         _verify_batch_size(input.size())
>    2013 
>    2014     return torch.batch_norm(
> 
> /usr/local/lib/python3.6/dist-packages/torch/nn/functional.py in _verify_batch_size(size)
>    1993         size_prods *= size[i + 2]
>    1994     if size_prods == 1:
> -> 1995         raise ValueError('Expected more than 1 value per channel when training, got input size {}'.format(size))
>    1996 
>    1997 
> 
> ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 1024])

The error occurs because the model is running a batch norm since it’s expecting a proper batch (size>1) for input. How do I set it to not do a batch normalization for inference?

2

You need to do learn.model.eval() before performing the inference. (or in this case model.eval())

3

So running model.eval() before inference does give me the same value each time now. But, if I create the model again and do the inference (with model.eval() done before the inference), it gives a different result. (I feel like I’m missing a basic concept in all this, but I can’t figure that out)

4

Yes. All models weights are random at the start.

1 Like
5

That makes a lot of sense, thanks!