u1ug
u1ug
What dataset and training code do you use?
Here is a sample code for training ``` import torch from dalle2_pytorch.tokenizer import SimpleTokenizer from torch.utils.data import DataLoader from dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, OpenAIClipAdapter, Unet, Decoder, \ DecoderTrainer from...
https://data.caltech.edu/records/65de6-vp158
You need to pass an embedding from prior, not a CLIP.
> Decoder loss is resulting "nan" values, when training with flikr8k dataset. Had a similar issue. Fixed with setting learned_variance to False in the decoder
You just pass random data to the model. ``` text = torch.randint(0, 49408, (4, 256)).cuda() images = torch.randn(4, 3, 256, 256).cuda() ``` Also you are not using optimizer so the...
> @u1ug are there ready-made solutions? Yes, check them on Huggingface