'Striving for Simplicity: The All Convolutional Net' The original paper can be found here.
We have also tried explaning the important concept of the paper on our Medium Blog.
In this implementation we are using the Cifar10 dataset. Either you can import the dataset from keras.datasets
Or
You can Download the dataset from here.
If you want to run the model without using the pretrained weights then: Run python allconv.py
You can download the weights from here: weights.994-0.56.hdf5 And use the weights to retrain your model.
The above model easily achieves more than 90% accuracy after the first 350 iterations. If you want to increase the accuracy then you can try much more heavy data augmentation at the cost of computation time.
MIT License
In the original paper learning rate of 'γ' and scheduler S = "e1 ,e2 , e3" were used in which γ is multiplied by a fixed multiplier of 0.1 after e1. e2 and e3 epochs respectively. (where e1 = 200, e2 = 250, e3 = 300) But in our implmentation we have went with a learning rate of 0.1, decay of 1e-6 and momentum of 0.9. This is done to make the model converge to a desirable accuracy in the first 100 epoch (Benificial for those who have a constrain on the computation power, feel free to play around with the learning rate and scheduler)
In the original paper very extensive data augmentation were used such as placing the cifar10 images of size 32 × 32 into larger 126 × 126 pixel images and can hence be heavily scaled, rotated and color augmented. In our implementation we have only done very mild data augmentation. We hope that the accuracy will increase if you increase the data augmentation.
