search

akarshzingade / image-similarity-deep-ranking

369 stars 103 forks source link

No convergence #19

Closed christophesmet closed 6 years ago

christophesmet commented 6 years ago

My model isn't converging for some reason. I'm checking out what could be the cause, testing with a lower amount of classes ( 50 ). I noticed the batch_y might be incorrectly generated.

batch_y = np.zeros((len(batch_x), self.num_class), dtype=K.floatx())
        for i, label in enumerate(self.classes):
                batch_y[i, label] = 1.

Since self.classes is always zero, it will always generate an array where only the first one is 1. But I don't think the cause is this, probably being self.classes being incorrect, as its all zero's This wouldn't have allowed you to generate the results in the blog. Do you have a hint where this might be ? Thanks

christophesmet commented 6 years ago

After some alterations, its starting too work. Accuracy is still poor, but I think I'll have to create something to generate a more effective dataset.