search

TencentARC / T2I-Adapter

T2I-Adapter
3.45k stars 206 forks source link

When should training be stopped? #133

Open YY-RR-ZZ opened 1 month ago

YY-RR-ZZ commented 1 month ago

Hello,Thank you very much for your open-source work. I have a question regarding the visualization of the loss generated while running train_sketch.py. Below is the visualization of the validation loss. The training reached 290,000 steps and epoch 7, but the loss does not show signs of convergence. I am unsure when the training should be considered complete. Could you please provide some guidance on this? loss

Yonghao-Yu commented 4 weeks ago

Where did you download the training data from? data/LAION_6plus data/WebDataset

cuhkfz commented 3 weeks ago

@YY-RR-ZZ Hello, I encountered the same problem. Have you finished the training successfully?

cuhkfz commented 3 weeks ago

@YY-RR-ZZ Hello, I encountered the same problem. Have you finished the training successfully?

Yonghao-Yu commented 3 weeks ago

@YY-RR-ZZ Hello, I encountered the same problem. Have you finished the training successfully?

hi，how you get the train data？ data/LAION_6plus data/WebDataset

cuhkfz commented 3 weeks ago

@Yonghao-Yu Hello yonghao, I use COCO dataset, the original T2I Adapter with sd v1-4 as backbone.

Yonghao-Yu commented 3 weeks ago

@Yonghao-Yu Hello yonghao, I use COCO dataset, the original T2I Adapter with sd v1-4 as backbone.

Yes, the paper mentions using the COCO17 dataset and using pidinet to extract edges. But I don’t know the details, such as the data format, etc. Did you make the dataset yourself? Is there a tutorial I can follow? Or where can I download the dataset so that I can use it directly? Thank you very much!

cuhkfz commented 3 weeks ago

@Yonghao-Yu Hello yonghao, I use COCO dataset, the original T2I Adapter with sd v1-4 as backbone.

Yes, the paper mentions using the COCO17 dataset and using pidinet to extract edges. But I don’t know the details, such as the data format, etc. Did you make the dataset yourself? Is there a tutorial I can follow? Or where can I download the dataset so that I can use it directly? Thank you very much!

Just download the required coco dataset, and run train_sketch.py. The only pre-process I did was create train_color/val_color.

Yonghao-Yu commented 3 weeks ago

@Yonghao-Yu Hello yonghao, I use COCO dataset, the original T2I Adapter with sd v1-4 as backbone.

Yes, the paper mentions using the COCO17 dataset and using pidinet to extract edges. But I don’t know the details, such as the data format, etc. Did you make the dataset yourself? Is there a tutorial I can follow? Or where can I download the dataset so that I can use it directly? Thank you very much!

Just download the required coco dataset, and run train_sketch.py. The only pre-process I did was creating train_color/val_color. But the training does not converge.

I see, I'll try again, thanks again.

Windrain7 commented 3 weeks ago

Just download the required coco dataset, and run train_sketch.py. The only pre-process I did was create train_color/val_color.

Hello, could you tell me how to set the parameters of train_sketch.py?