Filipp Ozinov

BTW, what accuracy do you have with your dataset? Could you share your model (or how you gather it)?

> Should I train it only with people photos or should I provide with all kind of different samples as negative ones? Not only people - just arbitrary images. The...

> Would it be smart to double image size? Would it be enough to change this constant? > IMG_SIZE = 128 > to > IMG_SIZE = 256 It's not that...

Not sure. Try to increase to 3K and check if quality is better or worse than with 1.5k.

> It really takes lot of ram. I could only run 40k sample on 32gb ram. Seems like it currently not optimized for large datests. Currently it is loading the...

`W_fc1 = tf.get_variable("W_fc1", shape=[8 * 8 * 24, 1024], initializer=xavier())`

> By the way in new TensorFlow libarry you have to call init_ops.zeros_initializer Haven't yet ported to a new version. Will do it soon.

Sory, I can't understand what's the problem. 93% is rother good accuracy. All dataset is spliited into train (80%) and test (20%) one, accuracy is calculated over a test set....

Ok, I'll add some later.

Hi, only official tensorflow [tutorial](https://www.tensorflow.org/tutorials/) and little experiments with network architecture.