ultimatech-cn

It is the problem of accelerate. You can try use python command directly, not using accelerate. Just change the code of app.py ` 'accelerate', 'launch', f'{project_dir}/facechain/train_text_to_image_lora.py',` to ` 'python', f'{project_dir}/facechain/train_text_to_image_lora.py',`...

btw, you should install mmcv-full by pip install mmcv-full on windows.

It is seems the preprocess doesn't produce the right output. You may need to add all the output of console or just try to train the lora by direct python...

I encounter the same error ``` File "E:\training\python_embeded\Lib\site-packages\llama_cpp\llama_chat_format.py", line 2651, in __call__ raise ValueError(f"Prompt exceeds n_ctx: {llama.n_tokens + embed.contents.n_image_pos} > {llama.n_ctx()}") ValueError: Prompt exceeds n_ctx: 2892 > 2048 ``` I...

In my comfyUI, the base sample workflow set like following the error of ``` File "..\custom_nodes\deforum-comfy-nodes\deforum_nodes\nodes\deforum_iteration_nodes.py", line 253, in get args.height, args.width = latent["samples"].shape[2] * 8, latent["samples"].shape[3] * 8 ~~~~~~^^^^^^^^^^^...

可以参考https://github.com/modelscope/facechain/issues/527 需要加一段代码

I resovle the error by adding "loss.requires_grad = True" after line 1033 in train_text_to_image_lora.py ``` train_loss += avg_loss.item() / args.gradient_accumulation_steps loss.requires_grad = True ```

You need to assure your PC can access "huggingface.com" directly or by proxy which you set is "hf-mirror.com". In china mainland it is normal issue. 你确定你的电脑能正常访问huggingface.com或者通过你设置的代理“hf-mirror.com能直接访问才行，不是程序本身的问题。

+1

Sure， so far there are three ways as I know 1. update your diffusers to latest 2. idm的问题，是需要把节点目录下的idm文件夹拷贝到site-package下 3. changing the name of idm( not tried) I used the second...