iRBM

Infinite Restricted Boltzmann Machine

Paper on arxiv and at ICML2015 - Deep Learning Workshop.

Dependencies:

• python == 2.7
• numpy >= 1.7
• scipy >= 0.11
• theano >= 0.6
• texttable

Usage

Experiments are saved in : ./experiments/{experiment_name}/.

Datasets will be downloaded and saved in : ./datasets/{dataset_name}/.

Train

See python train_model.py --help

Binarized MNIST

Training a model (infinite RBM) on binarized MNIST.

python train_model.py --name "best_irbm_mnist" --max-epoch 100 --batch-size 64 --ADAGRAD 0.03 irbm binarized_mnist --beta 1.01 --PCD --cdk 10

CalTech101 Silhouettes

Training a model (infinite RBM) on CalTech101 Shilhouettes.

python train_model.py --name "best_irbm_caltech101" --max-epoch 1000 --batch-size 64 --ADAGRAD 0.03 irbm caltech_silhouettes28 --beta 1.01 --PCD --cdk 10

Evaluate

See python eval_model.py --help

Binarized MNIST

Evaluating a model trained on binarized MNIST (assuming the one above).

python eval_model.py experiments/best_irbm_mnist/

CalTech101 Silhouettes

Evaluating a model trained on CalTech101 Silhouettes (assuming the one above).

python eval_model.py experiments/best_irbm_caltech101/

Sample

See python sample_model.py --help

Binarized MNIST

Generating 16 binarized MNIST digits images sampled from a trained model (assuming the one above).

python -u sample_model.py experiments/best_irbm_mnist/ --nb-samples 16 --view

CalTech101 Silhouettes

Generating 16 silhouette images sampled from a trained model (assuming the one above).

python -u sample_model.py experiments/best_irbm_caltech101/ --nb-samples 16 --view

Visualize filters

See python show_filters.py --help

Binarized MNIST

Visualizing filters of a model trained on binarized MNIST (assuming the one above).

python show_filters.py experiments/best_irbm_mnist/

CalTech101 Silhouettes

Visualizing filters of a model trained on CalTech101 Silhouettes (assuming the one above).

python show_filters.py experiments/best_irbm_caltech101/

Datasets

The datasets are automatically downloaded and processed. Available datasets are:

• binarized MNIST
• CalTech101 Silhouettes (28x28 pixels)

Troubleshooting

• I got a weird cannot convert int to float error. TypeError: Cannot convert Type TensorType(float32, matrix) (of Variable Subtensor{int64:int64:}.0) into Type TensorType(float64, matrix)

Have you configured theano? Here is my .theanorc config (use cpu if you do not have a CUDA capable gpu):

[global]
device = gpu
floatX = float32
exception_verbosity=high

[nvcc]
fastmath = True

• I got an IO error about status.json. IOError: [Errno 2] No such file or directory: './experiments/.../status.json'

There is no status.json file, so it is impossible to resume the experiment. Use --force to restart the experiment form scratch.