How to apply 4D maxpool in pytorch?

PyTorch Live
1

I want to convert a 4d maxpool from TensorFlow to PyTorch, but I could not find the equivalent function:

tf.nn.max_pool(first_layer
ksize=[1, 1, 5, 1],
strides=[1, 1, 5, 1],
padding=‘SAME’)
to
torch.nn.MaxPool4d(kernel_size=(1, 1, 5, 1), stride=(1, 1, 5, 1))
the first layer is a 4d tensor
But, it seems that PyTorch doesn’t have maxpool4d.
How can I do this conversion?

2

I’m not sure if this means your input tensor has 4 dimensions, but if so you could use nn.MaxPool2d assuming the input tensor dimensions are defined as [batch_size, channels, height, width] and specify the kernel_size as well as the stride for the spatial dimensions only (the first two are set to 1 so don’t have an effect).

1 Like
3

Thanks, it works: torch.nn.MaxPool2d(kernel_size=( 5, 1), stride=( 5, 1))

4

But, after applying a conv2d on the output of maxpool2d, a tensor with reduced dimension(32, 5, 3000, 1)
nn.Conv2d(5, 5, kernel_size=(5, 4, 1, 4), stride=(1, 5, 1, 1), bias=True)
it returns the following error:
RuntimeError: expected padding to be a single integer value or a list of 4 values to match the convolution dimensions, but got padding=[0, 0]

5

2d layers are using the kernel_sizeand stride on the spatial dimensions.
Since you are again defining a 4D kernel_size, could you explain your use case a bit more and give more information about the input dimensions to these layers?

6

I think, I should use kernel_size = (5,4) instead of (5,4,14).

7

It didn’t work. this is the information about input dimensions and layers:
s = 150,000
my input is [32,5,s, 1] and I need to apply the following maxpool and conv functions:

  1. maxpool, kernel size = (50,1), stride (50,1) → output = (32, 5, 3000, 1) so: torch.nn.MaxPool2d(kernel_size=( 50,1), stride=( 50,1))
  2. conv2d, kernel size = (5,4), stride(5,1) → output = (32, 1, 3000, 4): torch.nn.Conv2d(5, 1, kernel_size=(1,4),stride=(1, 1), bias=True)
  3. conv2d, kernel size = (1,4), stide (1,1) ->output = (32, 1, 3000, 4): torch.nn.Conv2d(1,1, kernel_size=(1,4), stride=(1,1))
  4. maxpool2d, kernel size (1,6), stride (1,4) → output = (32, 1, 750, 4): torch.nn.MaxPool2d((1, 6), stride=(1, 4))
  5. reshape to (32, 3000)
  6. fully connected (32, 512)
    After applying the first conv2d at the second line it returns the following error:
    RuntimeError: Calculated padded input size per channel: (3000 x 1). Kernel size: (5 x 4). Kernel size can’t be greater than actual input size
8

This layer won’t work as the kernel_size is specified as (5, 4) while the input seems to have a shape of [batch_size, channels, 3000, 1] which is too small in the width.