Home

Setup

The new gpu back end of theano is not supported by this version.

$ pip install -r requirements.txt

Configuration scripts depend on the utility scripts dataset_utils.py modeltemplate.py for their function, it is suggested to have them in the same directory as the configuration scripts or the one up (i.e cd ..).

Run (Clustering)

Clustering configurations depend on *.npy files as input (see more on data access), for all experiment configurations we assume that the config file for the running experiment exist (see more on data access). In order to assume availability some configurations (e.g Stacked autoencoders) have each component running individually.

Denoising autoencoder (Shallow AE)

$ python autoencoder.py -i <config_path>.ini -t train/test | tee <experiment_name>.txt

Stacked denoising autoencoders (Deep AE)

One liner

$ for i in <range from zero to num of autoencoders>;do python autoencoder.py -i <config_path>$i.ini -t train/test | tee autoenc_$i.txt;done && python stackedautoencoders.py -i <config_path>.ini -n <number of autoencoders> -t train/test -p <prefix_of_invidual_autoenc_config_files> | tee sda.txt

Script (./script.sh <num_of_autoencoders>)

#! /bin/bash
let END=$1-1
range=$(seq 0 $END)
for i in $range;do
    python autoencoder.py -i <config_path>$i.ini -t train/test | tee autoenc_$i.txt
done
python stackedautoencoders.py -i <config_path>.ini -n $1 -t train/test -p <prefix_of_invidual_autoenc_config_files> | tee sda.txt

Convolutional autoencoder (Conv AE)

$ python convolutional_ae.py -i <config_path>.ini -t train/test | tee <experiment_name>.txt

Run (Classification)

Clustering configurations depend on *.npz files as input (see more on data access), for all experiment configurations we assume that the config file for the running experiment exist (see more on data access).

Simple supervised classification inspired by inception model (Shallow)

$ python supervised.py -i <config_path>.ini -t train/test | tee <experiment_name>.txt

Deep supervised classification inspired by inception model (Deep or Vase)

$ python supervised_vase.py -i <config_path>.ini -t train/test | tee <experiment_name>.txt

Deep supervised classification inspired by inception model with multiple channels (Deep MC or Vase MC)

$ python supervised_vase_mult.py -i <config_path>.ini -t train/test | tee <experiment_name>.txt

Example

Example of stacked denoising autoencoders using the MNIST dataset.

Run

$ cd ncsr-atmo-learn/examples/sda/
$ ./start.sh 4

Monitoring train loss

$ ./plot_error_lines.py <tee experiment file> <label defined in the tee file>