杨奇(yann qi)

是相符合的，没有不符。一开始被你也带跑了，跟着又回顾了下代码。重点看下述代码的注释： ``` for compactor_param, mask in compactor_mask_dict.items(): compactor_param.grad.data = mask * compactor_param.grad.data # equ 14， 对应公式第一项，Compactor参数损失梯度* Mask lasso_grad = compactor_param.data * ((compactor_param.data ** 2).sum(dim=(1, 2, 3), keepdim=True) ** (-0.5))equ 14,...

> Hi @XiaoxxWang, our PPC loss takes the form of InfoNCE, which is in essence a cross-entropy loss. In our context, you can understand PPC as, for each pixel, we...

Another feasible way, you can change 'reshape_heads_to_batch_dim' to 'head_to_batch_dim' According to https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py