flows_ood
flows_ood copied to clipboard
PolinaKirichenko
Hi, Thanks for your work and code. Is there any code snippet that reproduces section 8 (figure 5b)? 1. Currently, if I use EfficientNet embeddings on RealNVPTabular, the log_determinants are...
Hi, thank you for providing all the code and the hyperparameter specifications for the experiments. Unfortunately, I'm having issues reproducing your results for the Glow architecture with MNIST and FashionMNIST....
In case this may be helpful for someone: The iLogits class uses a constant factor, which is set in the constructor, that is applied to the input image. Thus, the...
Hi, thanks for making the code available. I'd like to produce results from fig 1 a) in the paper. Can i do this with the code provided (are the pretrained...
Hi Polina, Did you report any BPD (bit per dimension) or log-likelihood performance of the model trained by the scripts? E.g., that of Glow trained on CIFAR-10. Thanks.
Hi Polina, It seems only naive `invertible conv1x1` is implemented for which [the original Glow model](https://github.com/openai/glow/blob/master/model.py#L438) (implemented by TensorFlow) also provides another efficient LU decomposition. The reference: https://github.com/PolinaKirichenko/flows_ood/blob/e9755db3454bbf9e8d46086446ddb9caa8870173/flow_ssl/glow/glow_utils.py#L6 Thanks.
Hi Polina, Thanks for your excellent code. May I know whether SyncBN is used in the mode of multiple GPUs? See the reference: https://github.com/PolinaKirichenko/flows_ood/blob/e9755db3454bbf9e8d46086446ddb9caa8870173/flow_ssl/invertible/normalizations.py#L13 Thx & Regards, Glory
Hi, I'm running baseline realNVP training with CIFAR10 with following command: ``` python3 train_unsup.py --dataset=CIFAR10 --data_path=DATA_PATH --logdir=LOG_DIR \ --ckptdir=CKPTS_DIR --num_epochs=101 --lr=1e-4 --batch_size=32 --num_blocks=8 \ --weight_decay=5e-5 --num_scales=3 ``` I'm printing out...
Hello, I have not yet seen code where GLOW is combined with checkerboard masking. It is always done on RealNVP. I am trying to do this however I am getting...
