Alexander Rakowski

The `Dense` layers `Dec_VAE_VDraw_Mean` and `Dec_VAE_VDraw_Var` should just be Identity layers taking the first and last 128 outputs of `Dec_VAE_VD_Dense`, as per section 3.3 of the paper: _"Starting from the...

Shouldn't the `sample_weight_test` parameter be uncommented in dropout networks (eg. in https://github.com/yaringal/DropoutUncertaintyCaffeModels/blob/master/cifar10_uncertainty/lenet_all_dropout_deploy.prototxt)?

Just a small issue: `message DropoutParameter` in `src/caffe/proto/caffe.proto` is not in line 471 anymore (line 691 in current version of Caffe).

What is the correct way of running the code (and in the best case, reproducing the paper results)? The arguments listed in README result in ` AssertionError: Invalid experiment defaults`

I am confused with how many reduction (pooling?) layers a micro model should have. Figure 4 in the paper shows 1 reduction layer per 1 convolutional block, whereas in the...

Shapes3D dataset is not included in the data download script

5

Also the `.npz` file listed in https://github.com/google-research/disentanglement_lib/blob/a070670e6589ed45c0b0bd95ea8913e7067eb3ae/disentanglement_lib/data/ground_truth/shapes3d.py#L28 is not to be found anywhere on the web.

Is the encoder architecture ( `conv_encoder` ) from [here](https://github.com/google-research/disentanglement_lib/blob/master/disentanglement_lib/methods/shared/architectures.py) the one used in the experiments in the paper? It seems to be a discrepancy in terms of the size of...

Normalizing the images doesn't work

2

Normalizing the dataset to values in range [0,1], done by this line: `faces = (faces/255.)` doesn't seem to work properly, at least not with the visualizations. Is it meant to...