search

thuanz123 / realfill

Unofficial implementation of RealFill
MIT License
359 stars 27 forks source link

Why is Use "accelerate launch train_realfill.py" super slower than directly use "python " #2

Closed zj19921221 closed 1 year ago

zj19921221 commented 1 year ago

Have you ever tested how long it takes to train an image? Also, why is the max_train_steps set to 2000 when it was previously set to 400?

thuanz123 commented 1 year ago

Why is Use "accelerate launch train_realfill.py" super slower than directly use "python "

I'm not sure, this is problem of accelerate, you can ask them

Have you ever tested how long it takes to train an image?

About 1 hour on a 32GB V100

Also, why is the max_train_steps set to 2000 when it was previously set to 400?

It is just what the author use in the paper, you can change to whatever you like

zj19921221 commented 1 year ago

When will the inference code be available? I'm really looking forward to it~

thanks～

zj19921221 commented 1 year ago

Why is Use "accelerate launch train_realfill.py" super slower than directly use "python "

I'm not sure, this is problem of accelerate, you can ask them

Have you ever tested how long it takes to train an image?

About 1 hour on a 32GB V100

Also, why is the max_train_steps set to 2000 when it was previously set to 400?

It is just what the author use in the paper, you can change to whatever you like

thanks～

thuanz123 commented 1 year ago

When will the inference code be available? I'm really looking forward to it~

thanks～

inference code is super simple actually, I just add it in

zj19921221 commented 1 year ago

When will the inference code be available? I'm really looking forward to it~ thanks～

inference code is super simple actually, I just add it in

May I ask, in the inference script, does "model_path" refer to the address of lora after training, or the address of the original stable-diffusion inpainting? If it is the address of lora, an error will be reported: 「OSError: Error no file named config.json found in directory /home/nick.zj/pro/realfill-main/flowerwoman-model.」

thuanz123 commented 1 year ago

Oh sorry, I didn't commit some fix. After update the new code and run the training again, you should be able to use the inference code

zj19921221 commented 1 year ago

Oh sorry, I didn't commit some fix. After update the new code and run the training again, you should be able to use the inference code

Okay, I'll re-pull the latest training code and retrain it. Which path should I use for the model_path in the inference script? Is it the one after training?

zj19921221 commented 1 year ago

I make it~ thanks~