Bad accuracy on MNIST dataset

d4nst / RotNet

https://d4nst.github.io/2017/01/12/image-orientation/

MIT License

539 stars 189 forks source link

Bad accuracy on MNIST dataset #12

Open d4nst opened 6 years ago

d4nst commented 6 years ago

Commit 2a4fe4be1ad7f2a6900408641d128e73469edbd8 made the accuracy on the MNIST dataset decrease from 6-7 degrees to 26-27 degrees.

The version before that commit can be used as a workaround for now: https://github.com/d4nst/RotNet/tree/1d934ba4e810806d264cbbbb94d165515887988a

suzhenghang commented 5 years ago

@d4nst Hi, I got the accuracy 29-30 degrees after 50 epoch. It is right?

d4nst commented 5 years ago

Sounds about the same that I get with the latest version of the code. As mentioned in the first comment, I used to get an error of 6-7 degrees before commit 2a4fe4be1ad7f2a6900408641d128e73469edbd8. I suspect that the old rotate method introduced some artefacts that made it easier for the network to figure out the angle. I'm not planning to investigate the issue further but if you find the reason please let me know :)

AlexZot commented 5 years ago

@d4nst Hi. It's off topic, but can You explain why you are using Sigmoid activation in output Dense layer in regression task? I've always thought that regression uses no activations on output.

d4nst commented 5 years ago

Yes, you are right. It is not recommended to use sigmoid as an output when training with L1/L2 loss. I believe that I tried to train the model without the sigmoid and the result was very similar though.

luisfelipe18 commented 4 years ago

My tensorflow stops without erros at epoch 6.