train custom dataset for inpainting task

[fushen1228]

how to train my custom dataset, where picture size is 128 * 128 or 64*64?

[GuHuangAI]

Please refer to the config file, and you need to modify line 3, 39 and 69.

[fushen1228]

thank you very much for your reply. and don't I need to change others, like in_channels or out_channels?

[GuHuangAI]

There may be some channel mismatch problems. If it occurs, please give me feedback.

[fushen1228]

thanks. Please tell me the specific steps of the training inpainting task. Do I need to train the autoencoder? when i executive command: accelerate launch train_cond_dpm.py --cfg ./configs/inpainting/celebahq_ddm_const4_condunet5_6202.yaml, which exists errors. TypeError: init() missing 1 required positional argument: 'auto_encoder'

[GuHuangAI]

@fushen1228 Hello, you can follow the training of super resolution.
Firstly, you need to train the autoencoder as the first stage model. Then, you should use the train_cond_ldm.py for the second stage training.

[qiongshuai]

@GuHuangAI What does the dataset look like in the inpainting task? Does celeba_256 only include ground truth and damaged images? Is there a mask image included in celeba_0q_256?

[GuHuangAI]

@GuHuangAI What does the dataset look like in the inpainting task? Does celeba_256 only include ground truth and damaged images? Is there a mask image included in celeba_0q_256?

We generate random masks for the images, please see data.py.

[qiongshuai]

@GuHuangAI Hi，I made this error while training the model, what might have caused it?

[GuHuangAI]

@qiongshuai
Hello, the training log shows no problem with the iter range from 0-63412.
I can not find the real reason for the error from your picture. Could you provide more detailed error information?

[qiongshuai]

@GuHuangAI Thank you very much for your prompt reply, I found the cause of the problem.My storage space is full

[qiongshuai]

@GuHuangAI Hello, is the more training times in the second stage the better? I found that when it （train_num_steps） is set to 200000, the training effect is not very good. Must it be set to 400000 or greater?

[GuHuangAI]

@GuHuangAI Hello, is the more training times in the second stage the better? I found that when it （train_num_steps） is set to 200000, the training effect is not very good. Must it be set to 400000 or greater?

Generally speaking, the performance may be related to many factors. Since I have no background in your project, I suggest that you can adjust the learning rate, model size, data size, training steps and so on.

[qiongshuai]

@GuHuangAI Hello，after I adjusted the image resolution of the training dataset (521x512->256x256), the following error appeared, may I ask how to solve it? But when the image resolution is 521x512, the program works fine. This is very confusing to me, thank you for your help.