https://arxiv.org/abs/2009.06385

A method for training the size of convolutional kernels to provide varying size kernels in a single layer. The method utilizes a differentiable, and therefore backpropagation-trainable Gaussian envelope which can grow or shrink in a base grid.

First of all some of the experiments were done initially in Tensorflow 1.4. (actually we started with theano a long time ago) Later many of them was translated to Tensorflow 2+ Unfortunately moving to Tensorflow 2+, we had to give up separate learning rates for the sigma parameter.

There is a clipcallback which ensures sigma is positive.

Therefore there are two folders for tf1.4 and tf2.0 Inside there are scripts for four different experiments.

1. testfilters.py: Autoencoder experiments (tf2.0 up to date)
2. Conv2DAdaptive_k2.py: The main model adaptive convolution implementation and simple convolution network experiments (tf2.0 up to date, at least for mnist, cifar10)
3. resnet_acnn_test_trials_tf2.py: Resnet model experiments (tf2.0 working but paper experiments used tf1.4)
4. test_ox_pets_acnn_tf_data.py: U-net segmentation experiment (tf2.0 only)

If you are tf2 you can simply run the mnist example with calling:

python Conv2DAdaptive_k2.py

which will run the Input-CNN-CNN-MAxpool-Dense-Output layer MNIST example classification example with the following parameters.

kwargs = {'dset':'mnist', 'arch':'simple', 'repeats':1, 'test_layer':'conv', 'epochs':20, 'batch':128, 'exname':'noname', 'adaptive_kernel_size':7, 'nfilters':32, 'data_augmentation':False, 'lr_multiplier':1.0}

This will reach ~99.55 in just 20 epochs.

Some problems require separate learning rates for weights and sigma. In that case you need a custom training loop which uses a different optimizers for weights and sigma