Jan-Hendrik Bastek issues

Repositories
Issues
Comments

Results 2 issues of


                                            Jan-Hendrik Bastek

Potential bug in cross-attention layer for class-conditioning

While trying to work through the details of the code, I noticed a potential bug in the forward call of the **CrossAttention** class (in ldm/modules/attention.py, code provided below). Namely, since...

Wrong order of dataset scale and shift in model training

In `train_ddpm/runners/diffusion_tub.py` the loss is computed as `loss = loss_registry[config.model.type](model, x, t, e, b, x_offset.item(), x_scale.item())`. This calls `def conditional_noise_estimation_loss(model, x0: torch.Tensor, t: torch.LongTensor, e: torch.Tensor, b: torch.Tensor, x_scale, x_offset,...