Achille, A., & Soatto, S. (2016). Information Dropout: learning optimal representations through noise. arXiv:1611.01353
Keras implementation of the Information Dropout (arXiv:1611.01353) paper.
$ luajit download_mnist.lua
$ luajit make_cluttered_dataset.lua
$ python t7_to_hdf5.py
$ python train.py /path/to/hdf/data/file
usage: train.py [-h] [--img-size IMG_SIZE IMG_SIZE] [--beta BETA]
[--kernel-size KERNEL_SIZE KERNEL_SIZE]
[--nb-filters NB_FILTERS NB_FILTERS NB_FILTERS NB_FILTERS]
[--model-path MODEL_PATH] [--weights-path WEIGHTS_PATH]
[--lr LR] [--nb-epoch NB_EPOCH] [--batch-size BATCH_SIZE]
[--name NAME] [--rootdir ROOTDIR]
datapath
Training All-CNN-96 network.
positional arguments:
datapath /path/to/hdf/data/file
optional arguments:
-h, --help show this help message and exit
--img-size IMG_SIZE IMG_SIZE
pixel size of input image (default: (96, 96))
--beta BETA hyper parameter for KL penalty (default: 1.0)
--kernel-size KERNEL_SIZE KERNEL_SIZE
kernel size of Conv2D layer (default: (3, 3))
--nb-filters NB_FILTERS NB_FILTERS NB_FILTERS NB_FILTERS
number of filters for Conv2D blocks (default: [32, 64,
96, 192])
--model-path MODEL_PATH
json file defining model architecture (default: None)
--weights-path WEIGHTS_PATH
hdf5 file containing network parameters (default:
None)
--lr LR initial learning rate (default: 0.01)
--nb-epoch NB_EPOCH number of epochs (default: 80)
--batch-size BATCH_SIZE
batch size (default: 128)
--name NAME experiment name (default: None)
--rootdir ROOTDIR root directory for saving results (default: ./)