Other ways to training the model
66ling66 opened this issue · comments
Can you provide the full command you used for running the experiments and the error message? Thanks.
Thanks you,my gpu memory is not encough,and I just change my training strategy.
Line 167 in dadf0e3
if I don't want to add this when training the unet model,what should I do?Is there a very simple way to do this?
Sorry for the late reply. I am a little bit confused, you mean you want to train the model without any grounding inputs? Our model training needs to have instance/part-level location and captioning inputs, otherwise, it should be equivalent to directly fine-tuning the Stable Diffusion model.
Yes I want to train the model without any grounding inputs,I just train a lora model ,when sampling image,add the grounding inputs,it seems work a litle bit.
Oh, I see. Maybe you can provide zero tensors as a placeholder for bbox, masks and instance caption embeddings. The easies way might be calling 'self.grounding_tokenizer_input.get_null_input()'. You should manually set 'self.set=True', and provide the 'self.device', 'self.dtype', 'self.max_box', etc.
You can find more details on this function at 'grounding_input/text_grounding_tokinzer_input.py'