marccote / irbm

Train an Infinite Restricted Boltzmann Machine

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.

irbm's Issues

question about evaluating

i evaluating a model trained on binarized MNIST,enter python eval_model.py experiments/best_irbm_mnist/
Then these appear

Loading dataset... (data: 50,000; 10,000; 10,000) 0.49 sec.
Loading model... (irbm with 852 hidden units) 0.40 sec.
Estimating model's partition function with AIS(5000) and 100,000 temperatures.... AIS run: 0/5000 (using batch size of 5000)

After that，There is no movement.always like this.

[ISSUE] Compatibility of non-MNIST dataset

I'm not sure how to rectify this without breaking the code.
DONE modify

question about the process to use your project

Thank you for your listen! I don't know how to run you project correctly. Can you give me some instructions
about how to run the project correctly. For example, when I try to run train_model.py, an error comes at with the warning that
usage: train_model.py [-h] (--nb-epochs N | --max-epoch N) [--batch-size M]
[--dataset-percent X]
(--ConstantLearningRate LR | --ADAGRAD LR [EPS=1e-6])
[--L1Regularization LAMBDA | --L2Regularization LAMBDA]
[--name NAME] [--seed SEED] [--keep K] [-f]
...
train_model.py: error: too few arguments
Can you give some instructions. Thank you very much.

object of class 'tasks' doesn't have an attribute 'GrowiRBM'.

tasks.GrowiRBM is called in train_model.py but does not exist in training.tasks