RuntimeError: Output size is too small for torch v1.3.1

1

Hi! I have a strange runtime problem when using torch v1.3.1 to run this code

import torch
import hiddenlayer


class LeNet5(torch.nn.Module):
    
    def __init__(self):
        
        super(LeNet5, self).__init__()
        
        self.convnet = torch.nn.Sequential(
            # Conv Block 1
            torch.nn.Conv2d(
                in_channels=1,
                out_channels=6,
                kernel_size=(5, 5),
                stride=1,
                bias=True
            ),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(
                kernel_size=(2, 2),
                stride=2
            ),
            # Conv Block 2
            torch.nn.Conv2d(
                in_channels=6,
                out_channels=16,
                kernel_size=(5, 5),
                stride=1,
                bias=True
            ),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(
                kernel_size=(2, 2),
                stride=2
            ),
            # Conv Block 3
            torch.nn.Conv2d(
                in_channels=16,
                out_channels=120,
                kernel_size=(5, 5),
                stride=1,
                bias=True
            ),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(
                kernel_size=(2, 2),
                stride=2
            )
        )
        
        self.fcn = torch.nn.Sequential(
            # Fully Connected Layer 1
            torch.nn.Linear(
                in_features=120,
                out_features=84,
                bias=True
            ),
            torch.nn.ReLU(),
            # Classifier Layer 2
            torch.nn.Linear(
                in_features=84,
                out_features=10,
                bias=True
            ),
            torch.nn.Softmax(dim=1)
        )

    def forward(self, batch):
        ret = self.convnet(batch)
        ret = ret.view(batch.size(0), -1)
        ret = self.fcn(ret)
        return ret

hiddenlayer.build_graph(LeNet5(), torch.zeros(1, 1, 32, 32)) # 1 image, greyscaled (1-channeled), 32 over 32

The error itself:


---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-4-844794a5d10c> in <module>()
      2 
      3 
----> 4 hiddenlayer.build_graph(LeNet5(), torch.zeros(1, 1, 32, 32)) # 1 image, shape-1, because greyscaled, 32 over 32

15 frames
/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py in _max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode, return_indices)
    486         stride = torch.jit.annotate(List[int], [])
    487     return torch.max_pool2d(
--> 488         input, kernel_size, stride, padding, dilation, ceil_mode)
    489 
    490 max_pool2d = boolean_dispatch(

RuntimeError: Given input size: (120x1x1). Calculated output size: (120x0x0). Output size is too small

However, when I run the same code with pytorch 1.1.0 - there is no error, the graph is generated with no problem.
Please, advise what can be the issue with v1.3.1 and how to solve it?

2

I think you should try to isolate this error. I personally don’t know how to reproduce this easily because of the import hiddenlayer dependency.

I mean: write an almost-dependency-free script that runs correctly on torch 1.1.0 and throws an error on torch 1.3.1. The script should only use very common libraries, like torch or numpy.

3

Indeed, you’re right. But the Error itself rise only while using hiddenlayer library with torch 1.3.1, and doesn’t arise while using hiddenlayer library with torch 1.1.0. So, probably in torch 1.3.1 there is different graph construction procedure or something else is changed, that affects how hiddenlayer library responds. That is what I want to understand and fix.