CNN example

[JNaranjo-Alcazar]

Hi community, excellent work. I am curious if this activation layer can be applied in a CNN network. Is there any example similar to this https://www.nengo.ai/keras-spiking/examples/spiking-fashion-mnist.html but using CNNs?

The input to the network would be of (T,28,28,1) dimensionality if we follow the previous example. How would the network have to be modified?

Best regards

[drasmuss]

Yes, the spiking activation layers can be used with any other layer types. In that example you linked you could replace the Dense layers with Conv2D layers (and remove the Flatten/Reshape layers, as they are no longer needed). Or you could just take any Keras CNN example, and use SpikingActivation layers instead of the standard activations.

[JNaranjo-Alcazar]

Okey, thanks for the quick answer. My question is the following, I have implemented a CNN with Conv2D layers inside TimeDistributed followed by Pooling layers inside TimeDistributed. As you can see, my image is 64 x 50 and it is repeated 10 times as explained in the MNIST manual. How do I do the last pooling before the classification layer? I have solved this with a Flatten layer but I think it is not the right thing to do. Any suggestions?

Find the architecture below:

model = tf.keras.Sequential(
        [

          TimeDistributed(Conv2D(32, 3, padding='same', input_shape=(10, 64, 50, 1))),
          keras_spiking.SpikingActivation("relu", spiking_aware_training=False),
          TimeDistributed(MaxPool2D(pool_size=2)),
          
          TimeDistributed(Conv2D(64, 3, padding='same')),
          keras_spiking.SpikingActivation("relu", spiking_aware_training=False),
          TimeDistributed(MaxPool2D(pool_size=2)),
          
          TimeDistributed(Conv2D(128, 3, padding='same')),
          keras_spiking.SpikingActivation("relu", spiking_aware_training=False),
          TimeDistributed(MaxPool2D(pool_size=2)),
          
          TimeDistributed(Conv2D(256, 3, padding='same')),
          keras_spiking.SpikingActivation("relu", spiking_aware_training=False),
          TimeDistributed(MaxPool2D(pool_size=2)),
          
          Flatten(),
          Dense(10)
        ]

[JNaranjo-Alcazar]

Hi, I manage to train the CNN I explained in the previous comment. However, when launching check_output I had to change this line:

if has_global_average_pooling:
        # check test accuracy using average output over all timesteps
        predictions = np.argmax(output.mean(axis=1), axis=-1)
    else:
        # check test accuracy using output from only the last timestep
        # predictions = np.argmax(output[:, -1], axis=-1) # ORIGINAL
        predictions = np.argmax(output, axis=-1) # MY LINE

output had a size of (N_SAMPLES, 10) so I had to do this modification to get the label information.

I do not understand why I had to make this change. I the shape of output correct?

Does the modification make sense? Thanks for the support