MixNMatch
MixNMatch copied to clipboard
WisconsinAIVision
Result is very bad on custom dataset
Hii @Yuheng-Li @utkarshojha @kkanshul @Johnson-yue
I am trying to generate full body human using this model but when we train this model on custom dataset result was bad after training completion, so can you suggest me how can we improve results on custom datasets. I am share some details which you will be able to understand easily.
Model configuration : SUPER_CATEGORIES = 1 FINE_GRAINED_CATEGORIES = 1 FIRST_MAX_EPOCH = 600 SECOND_MAX_EPOCH = 400
Here it's our model configuration as you can see above, now I am sharing two picture first one is ref image and another one is result of our model.
Hi Ayush, First, what is the second image here? I don't think there are final results. Could you show all fake images? Also, You should set SUPER_CATEGORIES and FINE_GRAINED_CATEGORIES much higher. SUPER_CATEGORIES means the potential shape and FINE_GRAINED_CATEGORIES represents different kinds of texture among foreground objects
@Yuheng-Li second image is the final epoch result of second stage training and also I'm sharing last epoch result of both stages training, please have a look. https://drive.google.com/drive/folders/1Qr96uKdJbUqpDSIEBNT7K-ZfokkE1FhQ?usp=sharing
@Yuheng-Li second image is the final epoch result of second stage training and also I'm sharing last epoch result of both stages training, please have a look. https://drive.google.com/drive/folders/1Qr96uKdJbUqpDSIEBNT7K-ZfokkE1FhQ?usp=sharing
Have you found the reason for the poor results？
You should not set super categories or fine-grained categories to be 1. The disentanglement will not learn.
i am trying to train on a similar dataset . i set SUPER_CATEGORIES = 30 and FINE_GRAINED_CATEGORIES = 200 unfortunately after the first step (1200 epochs) i get bad results ( noise) only after train the second stage ( 600 epochs) i start to get nice generated images
the problem is when i run eval at code ( only first stage) i will get noise and if i choose feature mode the generated image ignore the background source image and texture source image and generate image base on shape source image only ( i donwt know why it's happen , but i the background and texture generated image are black) so my question it's ok i didn't get any good results from stage one? there is any more parmas i should consider to change ? thanks
Hello, @ayush-angelium and @orydatadudes @Yuheng-Li I am also trying to use the same methodology for my custom dataset, I wanted to know how did you prepare your dataset like in CUB dataset. Basically i have used Roboflow for the annotations but still while training my shapes are not separated well from background. For this i am sharing some images. Your assistance in this matter would greatly support my efforts.
