error: AttributeError: 'float' object has no attribute 'backward'

[KeranLi]

Hi, I'm confused about the error: AttributeError: 'float' object has no attribute 'backward'
It occurs during the image-to-image translation. The details about the error information are:

────────── Traceback (most recent call last) ────────────────────────────────╮
│ /pix2pix-zero/src/edit_synthetic.py:51 in │
│ │
│ 48 │ pipe = EditingPipeline.from_pretrained(args.model_path, torch_dtype=torch_dtype).to( │
│ 49 │ pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config) │
│ 50 │ │
│ ❱ 51 │ rec_pil, edit_pil = pipe(args.prompt_str, │
│ 52 │ │ num_inference_steps=args.num_ddim_steps, │
│ 53 │ │ x_in=x, │
│ 54 │ │ edit_dir=construct_direction(args.task_name), │
│ │
│ /pix2pix-zero/src/utils/edit_pipeline.py:150 in call │
│ │
│ 147 │ │ │ │ │ │ ref = d_ref_t2attn[t.item()][name].detach().to(device) │
│ 148 │ │ │ │ │ │ cal = ((curr-ref)**2).sum((1,2)).mean(0) │
│ 149 │ │ │ │ │ │ loss += torch.tensor(cal, requires_grad=True) │
│ ❱ 150 │ │ │ │ loss.backward(retain_graph=False) │
│ 151 │ │ │ │ opt.step() │
│ 152 │ │ │ │ │
│ 153 │ │ │ │ # recompute the noise │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
AttributeError: 'float' object has no attribute 'backward'

I think the problem is caused by the line 149 in edit_pipline.py so I transfer it into a tensor, but it doesn't work. If possible, I wanna know what's wrong here. Thanks a lot!

[Pudding-0503]

I added this line of code loss = torch.tensor(0.0, requires_grad=True) after line 142 in the edit_pipline.py file, and it will work. Maybe you can try it?

[KeranLi]

Thanks, it's really helpful