Our default YAML file should actually produce an OK model. The current one runs, but that's about it (some of the monitoring numbers are concerning). Looking at the models in Pylearn2's examples from papers (pylearn2/scripts/papers/...) should be able to improve it. Also, using some knowledge about how to design convolutional neural nets. Some good resources could be Ilya's recommendations in this blog post and Yoshua Bengio's paper on practical recommendations.
This issue should've been more specific. YAML model file from Matt is improved. Will need issues for more specific ideas for models in future, closing.
Our default YAML file should actually produce an OK model. The current one runs, but that's about it (some of the monitoring numbers are concerning). Looking at the models in Pylearn2's examples from papers (
pylearn2/scripts/papers/...
) should be able to improve it. Also, using some knowledge about how to design convolutional neural nets. Some good resources could be Ilya's recommendations in this blog post and Yoshua Bengio's paper on practical recommendations.