Code gives better results than that mentioned in the paper about hdml HOT 8 CLOSED

@trungpham2606 I got the above results in 20K iterations, which is part of the default configuration. No, I haven't tried with unknown data.

from hdml.

@sampathchanda Hello, Iam also trying this repository. Can you tell me how many epochs you needed to achieve that results :-? Btw, have you tried to test this model with unknown data to verify whether or not this model can distinguish 2 images (at the same class or not) ?

from hdml.

@sampathchanda Can you please send me your pre-trained weights ?, Iam working in distinguish images on my own dataset.

from hdml.

@sampathchanda Very sorry for the late reply. We did achieve better results than those in the paper since we only reported average numbers. So your results are not surprising :) Btw, the best result we got is about 0.81 of recall@1. Also, we changed the initialization method from tf.truncated_normal_initializer(stddev=0.1) to tf.glorot_uniform_initializer() since the latter leads to more stable performance.
For CUB dataset, there should be no problem. I will include hyperparameters for CUB soon, hopefully before sunset.

from hdml.

@wzzheng When can you public your pre-trained weights :3 Iam so excited looking forward to it

from hdml.

@trungpham2606 Hi, if you could tell me your email address, I will send a pretrained model to you.

from hdml.

@sampathchanda Hi, I've added the recommended hyperparameters for CUB-200-2011, I hope they will solve your problem.

from hdml.

@wzzheng here's my gmail: [email protected] Thank you so much! Iam also using your model to solve images classification problem by image similarity. But the results I got so far are not good enough, maybe due to the lack of training iterations.

from hdml.