多个Conv1D图层：由于'conv1d_2 / convolution / Conv2D从1中减去8而导致的负尺寸大小

Question

I'm still quite novice with regard to convolutional networks. I'm trying to implement multiple Conv1D layers in Keras. Unfortunately, after the very first layer, any subsequent layers throw the following error:

tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative dimension size caused by subtracting 8 from 1 for 'conv1d_2/convolution/Conv2D' (op: 'Conv2D') with input shapes: [?,1,1,32], [1,8,32,32].

I had thought it may have something to do with the size reduction due to strides, but it still does not work after setting strides=1 for both Conv1D lines. Here is my code. If the for loop runs, then the error is thrown.

#State branch
x = Conv1D(layerSize,8,strides=1)(inputState)
x = Activation("relu")(x)

for l in range(conv1Layer-1):
    x = Conv1D(layerSize,8,strides=1)(x)
    x = Activation("relu")(x)

x = MaxPooling1D(pool_size=1)(x)
x = Flatten()(x)
x = Model(inputs=inputState, outputs=x)

Any help or advice would be greatly appreciated. Thank you!

Answer 1

If you do not want the length to change after the convolution consider specifying padding='same' in the constructor of Conv1d .

For more info see the docs .

Answer 2

The kernel_size must be changed to 1 after the first layer.

EDIT: Or the padding must be set to the same! Thanks.