Control: tags -1 + pending The package is now in NEW and goes into experimental first.
Sneak peek on Mentors: http://mentors.debian.net/debian/pool/main/p/python-brainstorm/python-brainstorm_0.5-1.dsc If you want to test drive it, to create a classifier for the MNIST-dataset (the typical "hello world" dataset for deep learning applications) goes like this: https://github.com/pinae/MNIST-Brainstorm `load_dataset.py` gets the training data set and converts it into hdf5, then `train.py` trains the network. After 75-100 epochs the network (`mnist_pi_best.hdf5`, delete first to create a new one) should be sufficient (this takes a couple of minutes on a up-to-date multi core CPU, the dataset contains 60000 images), it could be stopped with CTRL-C then. The classifier then recognizes random images of hand written numbers (hardcoded is `test_4.jpg`): $ python3 ./classify.py 4 [ 0.00120161 0.01541831 0.00469815 0.00289947 0.86062551 0.00200154 0.02392112 0.01064807 0.07652133 0.00206483] If could read German, this example is explained in current C't 06/2016, p. 142 sq. If the datasets are getting bigger, GPU based processing is preferred. Brainstorm could use the CUDA API through Pycuda, but also needs scikit-cuda, which isn't available yet. There is lot going on with deep learning (multiple layered neural network based machine learning) on Debian currently (Caffe is being worked on, Tensorflow is ITP, Theano is in NEW), it's a very fascinating subject. DS -- 4096R/DF5182C8 http://www.danielstender.com/blog/
