Yi-Lun Wu
Yi-Lun Wu
Is this your custom dataset?
Sure, I just update the download link for precalculated statistics of CIFAR10. I suggest that you can clone again since the config file is also updated for new filename.
Please see the latest readme to download the checkpoint.
It seems correct. This is my loss curve.  All the hyperparameters are copied from the official implementation.
I didn't record the FID and IS during training since it costs too much time to generate images for evaluation. However I can give the example at step 10k, 50k...
Can you provide a minimum executable example to let me reproduce the error?
No, that is abnormal. To train CIFAR-10, an 11G VRAM like the 2080 Ti is sufficient. However, if you use a larger model, the VRAM requirements may increase.
Apologies for the delayed response. To the best of my recollection, you do not need to update the `GaussianDiffusionTrainer` and `GaussianDiffusionSampler` when training with images of different sizes. These components...
The height and width of the image do not need to be equal, and this can be accomplished by updating the UNet model. Another thing to consider is the downsampling...
This implementation follows the official SN-GAN implementation. For details, please see Table 4 and Table 5 in [paper](https://arxiv.org/pdf/1802.05957.pdf).