TDeVries

I only implemented the macro architecture search for CNN (no micro search), but I tried to copy the official TensorFlow implementation as best as I could. I ran some tests...

Yes, the first number in each row are the operation types, which are defined here: https://github.com/TDeVries/enas_pytorch/blob/a3e54b290787572a0fab920d9f3b3149e09dbc82/models/shared_cnn.py#L73 The following numbers indicate whether there is a skip connection between the current layer...

> RT, is this a complete implementation? Do you get the reported performance on CIFAR and Imagenet? I'm actually not sure if I implemented the cosine learning rate properly, so...

> If I understand correctly, you mean that for these two nodes (index start from 0) in the table : > [3 0 0] #line 3 (node 2) > [3...

Okay, I think it looks correct. Sorry to say I am no longer working on this project.

Each ENAS layer contains the weights for several different branches, and these weights are reused each time we sample a new architecture. When a random path through the architecture is...

Yup, this is definitely a possibility!

I don't think the experiment had anything fancy. It was a pretty straightforward implementation of an MLP + confidence branch trained on a 2D dataset. The dataset was sampled from...

This is quite strange, I haven't seen this behaviour before. Is it possible that self.embed(y) is receiving values greater than the number of classes in the dataset? That seems to...

It should be fairly easy to apply to custom datasets. You need to create a dataset object that contains your data and then pass it to the select_instances function. See...