reported error results about vision_networks

[ikhlestov]

Yes, maybe answer should be a little bit different - keep prob will not be set to 1.0 tf.cond will just not apply dropout in case of test mode. It will return unchanged _input

[ikhlestov]

No, error mean prediction accuracy. Percent of how many classes were labeled incorrect. For example if we have 100 images and error 7% - it means that 93 images were classified correct, and 7 images receive another class.

[balodhi]

how about the keep_prob in your code? it is set to 0.8.
doesnt it needs to change to 1 while testing

[ikhlestov]

Yes, it should be set to 1.0, so on lines https://github.com/ikhlestov/vision_networks/blob/master/models/dense_net.py#L298-L302 it will be disable, if network in test mode.

[balodhi]

yes i can see that but i dont find any line which sets keep_prob=1
self.keep_prob is always 0.8