lllyasviel
How do I test my trained model
Hello! Your work is excellent! I trained my own data set with your method, but I don't know how to test the effect of my trained checkpoint, can I only check the results saved during training in the folder /image_log? Looking forward to your reply.
Same here. What is the difference between these images in the image_log: reconstruction_gs-011070_e-000135_b-000000 samples_cfg_scale_9.00_gs-011070_e-000135_b-000000
Samples are predicted images and the reconstruction is the ground truth
hi @mahdip72 @kako523 @engrmusawarali ,
Did you guys fired out what exactly is" reconstruction" and "samples"?
Image logs @engrmusawarali, are you 100% sure that "reconstruction" is ground truth (dataset input data).? Although both are very similar, I do see a difference between my input and "reconstruction".... My problem is that reconstruction images look fantastic, and samples look way too terrible and dont match anywhere close the input images...
Saved model When training if finished, I dont find the model saved to the models folder....
For the Image logs part, the recontruction images I think is basde on the output of VQ-VAE. Thats why it looks slightly different from the original image. In other words, the original image goes to the encoder of VQ-VAE to create the latent of the image. Then, it feeds to the decoder of VQ-VAE to reconstruct the original one. Since it goes through a compression and decompression process, it changes slightly. As VQ-VAE is frozen during training and LDMs uses latent space of images to do forward and backward diffusion steps, it is rational to consider reconstructed images as your targets, not the original targets.
thanks for the detailed explanation @mahdip72 !
did you manage to get your model properly trained?
@YunChangLi I have the idea but I just want to confirm.
My model is still on training. After the training completion, I'll try to execute that.
@YunChangLi I have the idea but I just want to confirm.
My model is still on training. After the training completion, I'll try to execute that.
Do you have any update? I have trained my model, now I want to use the train model to test
@YunChangLi I have the idea but I just want to confirm. My model is still on training. After the training completion, I'll try to execute that.
Do you have any update? I have trained my model, now I want to use the train model to test
I have showed the correct output with my testing code. The issue that occurred earlier was related to color space thresholds. After correction, my results turned out perfectly. Thanks for your reply
Training and testing configuration should be same for training and testing. Deep models do depend upon training circumstances. If you test with different (or sometimes slightly) circumstances, then the results can be different.
On Thu, 16 Nov 2023, 2:32 pm YunChangLi, @.***> wrote:
@YunChangLi https://github.com/YunChangLi I have the idea but I just want to confirm. My model is still on training. After the training completion, I'll try to execute that.
Do you have any update? I have trained my model, now I want to use the train model to test
I have showed the correct output with my testing code. The issue that occurred earlier was related to color space thresholds. After correction, my results turned out perfectly. Thanks for your reply
— Reply to this email directly, view it on GitHub https://github.com/lllyasviel/ControlNet/issues/304#issuecomment-1814083932, or unsubscribe https://github.com/notifications/unsubscribe-auth/AY2GYVWGO6TVLUCJDEIYJR3YEXMTTAVCNFSM6AAAAAAWF7HODWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMJUGA4DGOJTGI . You are receiving this because you commented.Message ID: @.***>
@YunChangLi I have the idea but I just want to confirm. My model is still on training. After the training completion, I'll try to execute that.
Do you have any update? I have trained my model, now I want to use the train model to test
I have showed the correct output with my testing code. The issue that occurred earlier was related to color space thresholds. After correction, my results turned out perfectly. Thanks for your reply
How you test your trained model? i could not see any script for inference... i want to evaluate my trained model on the validation set.
looking for your reply please... @YunChangLi @mahdip72 @Khaihuyennguyen @HassanBinHaroon
