Henry Li

Hi, of course, thanks for using our code! I couldn't tell from your question, but do you have an initial embedding dimensionality of 500, or a SpectralNet embedding dimensionality of...

Which tensorflow version are both of you using? I have tested it to work up to tf1.13

Late response, but the h5 pretrained models are only for the autoencoder. To level the playing field, we use the same autoencoder as several of the papers that we compare...

Hi, thanks for using SpecrtalNet, which version of TensorFlow are you using?

This is old, but have you made sure to use keras==2.3 as well? We realized that this was an important choice.

Another tip is reducing the number of clusters, if possible. One requirement of SpectralNet is that the orthonormalization layer is of rank equal to the number of clusters you set....

I see. Yeah, this could be the reason why you had problems, especially if the classes are not balanced, unfortunately.

Hi, Thanks for looking at our code. Please note that the weights are tuned by QR decomposition, which means that it is not a 'trainable weight' in the traditional sense,...

Note that the orthogonalization layer orthogonalizes the input, and _nothing_ else. Unlike fully connected, recurrent, convolutional, etc., layers, which are functions that depend both on the input AND a set...

Hi, you can actually create your own n x n affinity matrix as a function of the n inputs to the affinity matrix per batch. See https://github.com/KlugerLab/SpectralNet/blob/43b0fca784491f234489b860fc35832697ad20c2/src/core/networks.py#L108 The easiest reference...