Creating a Simple 1D CNN in PyTorch with Multiple Channels

Question

The dimensionality of the PyTorch inputs are not what the model expects, and I am not sure why.

To my understanding...

in_channels is first the number of 1D inputs we would like to pass to the model, and is the previous out_channel for all subsequent layers.

out_channels is the desired number of kernels (filters).

kernel_size is the number of parameters per filter.

Therefore, we would expect, as data passed to forward, a dataset with 7 1D channels (i.e. a 2D input).

However, the following code throws an error that is not consistent with what I expect, where this code:

import numpy
import torch

X = numpy.random.uniform(-10, 10, 70).reshape(-1, 7)
# Y = np.random.randint(0, 9, 10).reshape(-1, 1)

class Simple1DCNN(torch.nn.Module):
    def __init__(self):
        super(Simple1DCNN, self).__init__()
        self.layer1 = torch.nn.Conv1d(in_channels=7, out_channels=20, kernel_size=5, stride=2)
        self.act1 = torch.nn.ReLU()
        self.layer2 = torch.nn.Conv1d(in_channels=20, out_channels=10, kernel_size=1)
    def forward(self, x):
        x = self.layer1(x)
        x = self.act1(x)
        x = self.layer2(x)

        log_probs = torch.nn.functional.log_softmax(x, dim=1)

        return log_probs

model = Simple1DCNN()
print(model(torch.tensor(X)).size)

Throws the following error:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-5-eca5856a2314> in <module>()
     21 
     22 model = Simple1DCNN()
---> 23 print(model(torch.tensor(X)).size)

~/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    487             result = self._slow_forward(*input, **kwargs)
    488         else:
--> 489             result = self.forward(*input, **kwargs)
    490         for hook in self._forward_hooks.values():
    491             hook_result = hook(self, input, result)

<ipython-input-5-eca5856a2314> in forward(self, x)
     12         self.layer2 = torch.nn.Conv1d(in_channels=20, out_channels=10, kernel_size=1)
     13     def forward(self, x):
---> 14         x = self.layer1(x)
     15         x = self.act1(x)
     16         x = self.layer2(x)

~/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    487             result = self._slow_forward(*input, **kwargs)
    488         else:
--> 489             result = self.forward(*input, **kwargs)
    490         for hook in self._forward_hooks.values():
    491             hook_result = hook(self, input, result)

~/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/torch/nn/modules/conv.py in forward(self, input)
    185     def forward(self, input):
    186         return F.conv1d(input, self.weight, self.bias, self.stride,
--> 187                         self.padding, self.dilation, self.groups)
    188 
    189 

RuntimeError: Expected 3-dimensional input for 3-dimensional weight [20, 7, 5], but got 2-dimensional input of size [10, 7] instead

Edit: See below for solution, motivated by Shai.

import numpy
import torch

X = numpy.random.uniform(-10, 10, 70).reshape(1, 7, -1)
# Y = np.random.randint(0, 9, 10).reshape(1, 1, -1)

class Simple1DCNN(torch.nn.Module):
    def __init__(self):
        super(Simple1DCNN, self).__init__()
        self.layer1 = torch.nn.Conv1d(in_channels=7, out_channels=20, kernel_size=5, stride=2)
        self.act1 = torch.nn.ReLU()
        self.layer2 = torch.nn.Conv1d(in_channels=20, out_channels=10, kernel_size=1)
    def forward(self, x):
        x = self.layer1(x)
        x = self.act1(x)
        x = self.layer2(x)

        log_probs = torch.nn.functional.log_softmax(x, dim=1)

        return log_probs

model = Simple1DCNN().double()
print(model(torch.tensor(X)).shape)

Answer 1

You are forgetting the "minibatch dimension", each "1D" sample has indeed two dimensions: the number of channels (7 in your example) and length (10 in your case). However, pytorch expects as input not a single sample, but rather a minibatch of B samples stacked together along the "minibatch dimension".
So a "1D" CNN in pytorch expects a 3D tensor as input: BxCxT. If you only have one signal, you can add a singleton dimension:

 out = model(torch.tensor(X)[None, ...])