Questions about FCN

[tengerye]

May I ask two questions of your FCN code please?

1. Traditional sequential data is of three dimensions: (batch_size, sequence_length, sequence_dimension). Why the data is four dimension in your code?

2. In your paper, you stated that "the features are fed into a global average pooling layer instead of a fully connected layer." But you implemented the fully connected layer
   out = keras.layers.Dense(nb_classes, activation='softmax')(full). Why is that?

Thank you in advance.

[cauchyturing]

1. As the code is using conv2d, so just pad 1 more dimension with 1d kernel. It is the same with (batch_size, sequence_length, sequence_dimension) with conv1d.

2. This is a softmax layer to output the label, not a fc layer.

[tengerye]

Hi, just one more question. In the global average pooling, the number of padding will affect the forward result, e.g. for the sequence [1, 2, 3], if the length of the longest sequence is 9, then the average is 0.67, but if the length is 6, then the average is 0.5.

Do you think it is fine?

[ResonWang]

Hi，about the code "x_train_mean = x_train.mean() x_train_std = x_train.std() " , I have tested that the output is a scalar; It shouldn't be a vector with the same length as the time sequence?

[ResonWang]

One more question, the code "conv1 = keras.layers.Conv2D(128, 8, 1, padding='same')(x)" is the order of "filters=128, kernel_size=(8,1)" yeah ？

[InputBlackBoxOutput]

Hi，about the code "x_train_mean = x_train.mean() x_train_std = x_train.std() " , I have tested that the output is a scalar; It shouldn't be a vector with the same length as the time sequence?

mean() and std() methods in numpy calculate the mean and std respectively of all values in the array and returns a scalar by default

One more question, the code "conv1 = keras.layers.Conv2D(128, 8, 1, padding='same')(x)" is the order of "filters=128, kernel_size=(8,1)" yeah ？

No, it means kernel size=8 and stride=1