chrischute / glow Goto Github PK

Hello,
I can run the code without conda, but it takes me about 6 hours to train an epoch with 2 1080Ti GPUs. I can't figure out what happened. Is this the influence of conda?

Hi,

I am taking the code from the repo as-is and running it. The only difference is I am running it a different PyTorch version (1.6.0).

For some test_runs (for e.g. 10, 20 epochs), the BPD score is abnormally high,

50000/50000 [16:59<00:00, 49.03it/s, bpd=9.49e+11, lr=0.0003, nll=2.02e+15]

For 2/3 runs even after a 50 epochs I am getting such scores.
Is there something wrong that I am doing during the training or the pytorch version is affecting the results so much?

The version of pytorch in the environment.yml gives a cublas error:

RuntimeError: cublas runtime error : the GPU program failed to execute at /opt/conda/conda-bld/pytorch_1544176307774/work/aten/src/THC/THCBlas.cu:511

Do you have any idea as to why it occurs?

In my implementation my BPD and NLL seem to be going down but the quality of the images produced seems to be indistinguishable from plain white noise.

Epoch 0: bpd=4.84,nll=3.44e+3
Epoch 1: bpd=3.58, nll=2.54e+3
Epoch 20: bpd=2.88, nll=2.04e+3
Epoch 29: bpd=2.83,nll=2.01e+3
This is for the training set , validation remains fairly the same bpd±0.02,nll±0.02
So, I don't think it is overfitting (N=512,L=3,K=32)
The dataset is FashionMNIST .

Do I let it train for longer or is it maybe something else?
Would appreciate your input

I could not understand how do you compute the loss function? Especially the computation of prior_ll part.
It would be great if you could explain what you are doing there

The original glow code uses gradient checkpointing, a very efficient way of reducing peak memory consumption. The following single line adds gradient checkpointing in away that memory consumption from 11 GB to 2 GB. It allowed me to increased batch size from 64 to 256 with no issue. I think 512 is possible, maybe even 1024 if we use float16 for some of the layers.

def forward(self, x, ldj, reverse=False):
        x_change, x_id = x.chunk(2, dim=1)

        #st = self.nn(x_id) # change this line to the one below. 
        st = torch.utils.checkpoint.checkpoint(self.nn, x_id)
        s, t = st[:, 0::2, ...], st[:, 1::2, ...]
        s = self.scale * torch.tanh(s)

Hi!

I wonder, why there is no addition of log(p(z_i)) (i=1..L-1) to the sum of log-det jacobians after split operation in lines 107-111 of glow.py file. Also I have a question about line 110 of the same file: why does one concatenate back already split tensors? Should we let only one part of the split flow further, don't we? Or do I misunderstand something?

Thanks!

Should not line num 38 in coupling .py be

x_change = (x_change * s.exp()) + t

instead of

x_change = (x_change + t) * s.exp()

Hey, thank you very much for this nice repository! I was wondering, what would be the most convenient way to condition Glow model on image classes?

It seems in the paper, the description of the NN in the coupling layer is three convolutions with relu activations. Here, you also include an actnorm or batchnorm in between the convolutions. The other pytorch implementations + the tensorflow implementations don't include this. Should the normalization be there?

Very nice implementation, the code structure is very nice.

Is it possible you are willing to share pre trained weights?