1

Currently, I am tuning my model by testing the Kernel size.

I have the following code :

x = embedding_layer(input_4)                
x = Conv1D(FILTERS, KERNEL, activation='relu')(x)
x = Dropout(DROPOUT)(x)            
x = Conv1D(FILTERS, KERNEL, activation='relu')(x)       
x = Dropout(DROPOUT)(x)
x = Conv1D(FILTERS, KERNEL, activation='relu')(x)       
x = Dropout(DROPOUT)(x)
x = Conv1D(FILTERS, KERNEL, activation='relu')(x) 
x = Dropout(DROPOUT)(x)            
x = Conv1D(FILTERS, KERNEL, activation='relu')(x) 
x = Dropout(DROPOUT)(x)            
x = MaxPooling1D(3)(x)

x = Conv1D(FILTERS, KERNEL, activation='relu')(x)
x = Dropout(DROPOUT)(x)            
x = Conv1D(FILTERS, KERNEL, activation='relu')(x)       
x = Dropout(DROPOUT)(x)
x = Conv1D(FILTERS, KERNEL, activation='relu')(x)       
x = Dropout(DROPOUT)(x)
x = Conv1D(FILTERS, KERNEL, activation='relu')(x) 
x = Dropout(DROPOUT)(x)            
x = Conv1D(FILTERS, KERNEL, activation='relu')(x) 
x = Dropout(DROPOUT)(x)            
x = MaxPooling1D(3)(x)

When the Kernel is 2 or 3, the network runs fine, but from 4 onwards it runs into an error about the dimensionality. I suspect that it has to do with the stride length. However, the Keras website (https://keras.io/layers/convolutional/) does not say what the default stride length is.

My question now is: what is default stride length in Keras' Conv1D? And what would be a good stride length for a kernel size of 4 and for a kernel size of 5?

BattleTested_закалённый в бою
  • 4,037
  • 5
  • 26
  • 47
Emil
  • 1,531
  • 3
  • 22
  • 47

2 Answers2

4

From Conv1D, the default stride length is 1. Unless you have a concrete justification for another length, a stride length of 1 is usually appropriate.

The error you get is probably because the output dimension of a 1D convolutional layer is:

output_dim = 1 + (input_dim - kernel_size)/stride

And after stacking several 1D convolutional layers, you might be reaching a layer in which the input dimensionality is smaller than the kernel size. This happens because the default value for the argument padding is 'valid', which means that the input is not padded.

If instead you want to preserve the input dimensionality at each convolutional layer, setting padding='same' results in padding the input such that the output has the same length as the original input.

rvinas
  • 11,824
  • 36
  • 58
1

The default stride is really (1,1), meaning that the filter will one pixel right for each horizontal movement of the filter and one pixel down for each vertical movement of the filter.