一维卷积神经网络

Question

I need to test CNN on EEG data, and I have heard that 1D-CNN is useful for real-time application. I have 5 test subjects with data from 3 sessions each. Each file contain signal from 56 electrodes/channels (56, 260).

I am struggling to find how to set up the CNN and how to input data should be transformed.

model = Sequential()
model.add(Conv1D(filters=64, kernel_size=3, activation='relu', input_shape=input_shape))

How do I choose the input shape for 1DCNN from (15, 56, 260)?

Answer 1

So, for a Keras convolution, you should keep it this way: (examples, time_steps, features) .

Where:

• examples (usually called samples, but not to be confused with your sensor samples) - they are the number of examples you have for your model: 15
• time_steps: the continuous sequential measure of your sensors: 260
• features (also channels): the number of independent/parallel channels: 56

Your data then should be shaped as (15,260,56) .

If you already have it organized as (15,56,260) , you need to permute or transpose it, not reshape. You can try numpy.swapaxes() .

Once your data is organized corretly as (15, 260, 56) , you can create your network with input_shape=(260,56) , or even input_shape=(None, 56) in case you want variable length sequences.

---

This is also the same shape you need if you want to try recurrent networks, or even mix recurrent with conv1d.