An optimization algorithm for learning Piecewise Linear Convolutional Neural Networks. This is the implementation of the paper Trusting SVM for Piecewise Linear CNNs by Leonard Berrada, Andrew Zisserman and M. Pawan Kumar.
This code has been developped python 2.7 and requires the packages given in requirements.txt.
Contains the data set folders. The use of the following data sets is implemented:
- MNIST
- CIFAR-10
- CIFAR-100
- ImageNet
This contains the python code. To run the tests, run sh scripts/test.sh from the src working directory.
This contains the logs from the experiments as well as the saved models. All experiment logs are backed up in log/all.
First make sure your working directory is src, and that the standard data sets are in data.
Then to run experiments with gpu [x] with results output to directory log/my_dir:
sh scripts/mnist.sh gpu[x] my_dir
sh scripts/cifar10.sh gpu[x] my_dir
sh scripts/cifar100.sh gpu[x] my_dir
The following code assumes that:
- the images have been extracted and resized to a minimal size of 256, and are in the path specified in
datasets/imagenet.py. - the pre-trained VGG-16 model has been downloaded from Lasagne model zoo and placed in
my_dir, the directory for the experiment.
Launch data pipeline for training set:
python server.py --train
And in a different shell:
python lwsvm.py --dataset imagenet --in_name vgg16 --device gpu[x] --xp_dir my_dir
When training is complete, launch data pipeline for validation set:
python server.py --val
And launch evaluation in a different shell:
python extract.py --imagenet --weights_file imagenet_lwsvm_15
The architectures can easily be changed through the addLayer-type methods of NeuralNet. Note that the code assumes a restricted search on the conditional gradient of convolutional layers (Observation 1 in section 4 of the paper), since they are usually followed by an important number of non-linearities (max-pooling).