Anson Yu

So all the nets should be trained alone, but input data for training higher resolution nets are dependent on lower resolution ones. Calibration nets are not dependent on each other,...

Hi N G K Sai, I think using a size slightly larger than the model's input size would be best. For example, I chose 15 pixels for 12-net, since when...

Hi, Please refer to issue #25 . Thanks!

I think you should be able to get the prob similarly by `prob = self.model.predict([img1, img2])` I recall using GTX 970, it took me around 8 hours to train.

I'm not sure why your results are quite different.. Regarding how to save the model, you can check this [link](https://keras.io/getting-started/faq/#how-can-i-save-a-keras-model). Hope it helps, thanks!

Hi, Here's the baiduyun weights link: [https://pan.baidu.com/s/1jNa5hKC_e9OgxW234oiR-g](https://pan.baidu.com/s/1jNa5hKC_e9OgxW234oiR-g) The dataset was obtained from here: [http://scikit-learn.org/stable/datasets/olivetti_faces.html](http://scikit-learn.org/stable/datasets/olivetti_faces.html) Thanks, Anson

Hi, Can you try if this works? [https://stackoverflow.com/questions/45078392/failedpreconditionerror-attempting-to-use-uninitialized-value-with-keras](https://stackoverflow.com/questions/45078392/failedpreconditionerror-attempting-to-use-uninitialized-value-with-keras) Maybe try changing the **_W_init** and **_b_init** methods to return `keras.initializers.RandomNormal(mean=0.0, stddev=1e-2, seed=None)` instead of K.variable Thanks! Anson