Gavia Gray

Looking at the paper's code Tim sent me, have found the following differences: - Their architecture uses 4 dense layers in total, with 1024 hidden units in each (apart from...

Have implemented these changes in the code, and the results match a lot better. Unfortunately, there are still some problems; mostly that the variance is increasing after training for 100...

Weight Initialisation --------------------------- Both sample from uniform distributions and then scale. The pytorch implementation [uses the default](https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/conv.py#L40-L47). The scaling is `1./sqrt(c_in*k*k)`. The tensorflow version uses `xavier_initializer_conv2d` by default, which [is...

Batchnorm config ---------------- In tensorflow, we use [batch_norm with default params](https://www.tensorflow.org/api_docs/python/tf/contrib/layers/batch_norm). Going to assume the placeholder `is_train` is working; assuming this isn't a problem with running averages being calculated on...

Training set augmentation ------------------------------------- Tensorflow code in this repo does the following for `distrorted_inputs`: * Randomly crop `32x32` out of `32x32` image. So, it's not actually cropping anything (my fault)....

Doing all of the above, performance plateaus at 92.5% accuracy on the test set.

Seems likely a problem with regularisation, this tensorflow model is able to push the cross entropy loss much closer to zero in the final iterations than the pytorch model was...

Oh no my mistake, you've recently changed name of the directory containing the code from `torch_trust_ncg` to `Code` and that breaks the `transfer_model` script in the smplx repo. I'll rollback...

Looks good! I tried running it and it hits this error: ``` > python IncrementalPCA.py Using device: cpu Traceback (most recent call last): File "/Users/gngdb/gngdb/pytorch-pca/IncrementalPCA.py", line 128, in ipca =...

Made a PR to your branch here: https://github.com/YRYoung/pytorch-pca/pull/1 I think this is the right way to do open source collaboration?