[HELP] convert TensorFlow convolutional layer into PyTorch

1

Hi have this layer written in python with the TensorFlow library:

Conv2D(64, 
       (3, 6), 
       strides=(1, 1), 
       input_shape= (10, 6, 1), 
       activation='relu'
)

and I’m trying to convert it into PyTorch, for now what i achieved is this:

self.conv2d = nn.Sequential(
                nn.Conv2d(1, 1, (3,6), (1, 1)),
                nn.ReLU()
            )

The problem is that in the TF implementation i have as output dimension [None, 8, 1, 64] and in PyTorch [-1, 1, 8, 1] so i think that the last 1 should be 64. This is even validated by the number of parameters that i have in TF and PyTorch respectively: 1216 and 19.
(19*64 = 1216)

Am I missing something?
Thanks.

2

Set output channels to 64.

self.conv2d = nn.Sequential(
                nn.Conv2d(1, 64, (3,6), (1, 1)),
                nn.ReLU()
            )

In pytorch channel dimension is second not last

3

You can check the tutorial for PyTorch.

4

ok, I’m trying with that, now i have some problems with the input, I’ll check to solve this and I’ll update you. Thanks

5

can you link it to me please?

6

https://pytorch.org/tutorials/beginner/pytorch_with_examples.html

Scroll here, lots of material

7

and thanks for the clarification, this is helping me a lot on understanding conditionality on pytorch.