WangXin93 / fashion_compatibility_mcn

Outfit Compatibility Prediction and Diagnosis with Multi-Layered Comparison Network
MIT License
82 stars 31 forks source link

How did you split train, test, val in paper? #13

Open shivamtundele opened 1 year ago

shivamtundele commented 1 year ago

You mention in paper that you used "Graph segmentation is used to split train, validation and test dataset."

What exactly is that? I am trying to re-create your model training pipeline on my custom data. But the loss is not decreasing and is everywhere. I have tried everything from diagnosing image tensors, changing parameters, adjusting regularization.

I am thinking my problem might be because I randomly split the outfits in train, test and val set.

Please help me on how Graph Segmentation can be done? I am dealing in Python FYI.

TIA