This is part of a project for CS231N at Stanford University, written by Ankit Kumar, Mathematics major, Class of 2015
This is a repository that allows you to convert pretrained caffe models into models in Lasagne, a thin wrapper around Theano. You can also convert a caffe model's architecture to an equivalent one in Lasagne. You do not need caffe installed to use this module.
Currently, the following caffe layers are supported:
* Convolution * LRN * Pooling * Inner Product * Relu * Softmax
You can also load in a mean file using conversion's convert_mean_file function. Future work is to put that in the conversion of net architectures itself, which might have a data layer with transformparameter. That can be automated.
Right now, you have to put the cost layer on yourself, as well as do the backprop code. This is a future step for me, however, Lasagne is very easy to use and you can learn how to add your own stuff very easily. I want to keep this as configurable as possible because that's the benefit of theano.
Theano (http://deeplearning.net/software/theano/) needs to be bleeding-edge:
pip install --upgrade --no-deps git+git://github.com/Theano/Theano.git
Lasagne (https://github.com/benanne/Lasagne), Google protobuf. Pylearn2 for cuda convnet wrappers (see below). If you have Caffe installed, you already have google protobuf, otherwise see here: https://code.google.com/p/protobuf/
##USING CUDA CONVNET WRAPPERS:
Install pylearn2 as develop:
git clone git://github.com/lisa-lab/pylearn2.git cd pylearn2 python setup.py develop
The cuda-convnet wrappers in pylearn2 are much faster than the GPU implementations of convolutions in Theano. Lasagne has cuda-convnet layers, and I have created a caffe version of these layers. However, they require you to go into pylearn2 and change some of the files. I don't know what the best way to package that change in this repo is, so until someone tells me a better way I'll just describe what to do:
In this file: https://github.com/lisa-lab/pylearn2/blob/master/pylearn2/sandbox/cuda_convnet/base_acts.py, change:
class BaseActs(GpuOp): """ Shared code for wrapping various convnet operations. """ def __init__(self, pad=0, partial_sum=None, stride=1):
class BaseActs(GpuOp): """ Shared code for wrapping various convnet operations. """ def __init__(self, pad=0, partial_sum=None, stride=1, numGroups=1):
Then, in that init function, change:
self.dense_connectivity = True
if numGroups == 1: self.dense_connectivity = True else: self.dense_connectivity = False
and add a line:
self.numGroups = numGroups
then, in this file: https://github.com/lisa-lab/pylearn2/blob/master/pylearn2/sandbox/cuda_convnet/filter_acts.py, change FilterActs c_code function from:
if self.dense_connectivity: basic_setup += """ #define numGroups 1 """
if self.dense_connectivity: basic_setup += """ #define numGroups 1 """ else: basic_setup += """ #define numGroups %s """ % self.numGroups
You should be able to not have that if statement at all, but I kept it in. You will have to do the same change (the #define numGroups) change for ImageActs and WeightActs as well if you want to backprop through the convolutional layers; if you just want to run the forward function, that isn't needed.
You can test the repo by python tests.py. All the tests should pass, but terminal is out of GPUs, so I haven't been able to run the tests.py script on GPU. However, I have used this repo with GPUs, and it worked, the only question is if it still works after I moved files around. If it doesn't let me know.
The file conversion.py has a function
convert that takes a prototxt file and a caffemodel file, and returns a lasagne base model. I have plans to superclass base models with other models, for the purposes of training, but I haven't yet best figured out how to do this. This repo is still in active development while I use it for my project. If you have ideas for additions, let me know.
>>> from conversion import convert >>> from models import * >>> lmodel = convert('/path/to/deploy.prototxt', '/path/to/pretrained_model.caffemodel') >>> dump(lmodel, 'filename')