Stefan Lattner, Maarten Grachten, Carlos Eduardo Cancino Chacon
This is a python package that provides tools for learning representations from data, implemented in the context of the Lrn2Cre8.eu project. Its goal is to simplify the creation and training of Neural Network architectures using theano, as well as reducing the effort in the associated data management. It includes the nn_bricks package, providing classes which can be combined in a building block like manner making it easy to create custom Neural Network (NN) architectures. Pre-defined layers can be created, stacked and trained using a configuration file.
Among others, lrn2 includes [Convolutional] Restricted Boltzmann Machines (RBMs), [Convolutional] Neural Networks and Auto-Encoders, [Gated] RNNs, LSTMs, corresponding training routines and cost functions, as well as common regularizers. Overview of available bricks.
The structure of lrn2 is designed to facilitate the creation of an end-to-end work flow. This work flow involves the following four phases:
- Reading and preparing data from files
- Defining and creating models using a configuration file
- Training models to learning representations from the data
- Using the trained models in specific use case scenarios
Files are organized into several subdirectories (under lrn2):
- nn_bricks: This package provides brick-like classes (i.e. predefined NN, RNN and RBM layers, several unit types, cost functions, regularizers, training algorithms, plotting functions, serialization methods, ...) which can be freely combined into mix-in classes to define different NN layers or stacks.
- data: Data holding and data import. Interfaces for different domains and formats.
- util: Several utility modules specific for this framework
- application: Code which can be used for special applications (like classification or visualization)
- Documentation
- Tutorials
- Project website
For convolution, NVIDIA cuDNN is strongly recommended.
To install the lrn2 framework (Linux/Mac), run the following command in your terminal
(optional) cd /usr/local/lib/python2.7/dist-packages/
sudo pip install -e git+https://github.com/OFAI/lrn2.git#egg=lrn2
or manually
git clone https://github.com/OFAI/lrn2.git
cd lrn2
git submodule init
git submodule update
sudo python [setup.py | setup_mac.py] install
or download the source and use it in your IDE (feel free to fork). Running the above commands may still help to install the required dependencies.
Test the lrn2 framework:
- In the (extracted) tar archive, browse to an example folder examples/[example_name]. Run one of the examples by entering
python run_demo.py [any_run_keyword] config_model.ini [options], or with the--helpflag for help on the specific example. For the mnist_ examples, no further parameters are needed. The other two examples need the path to a .csv file with (monophonic) melodies (e.g. the Essen Folksong Collection) as additional parameter. You can generate such a .csv file by usingpython create_csv_filelist.pyinlrn2/util. - If some import errors are shown, ensure that all dependencies are included in the environment variable $PYTHONPATH.
Run the following command in your terminal: sudo pip uninstall lrn2
The example below shows the main workflow of the lrn2 framework (omitting data preparation for now).
lrn2.nn_bricks.make.make_net takes a configuration file and creates NN layers accordingly. The layers are then packed into a stack with cross-entropy cost (custom stacks and layers can be defined using some bricks of the nn_bricks collection). Binding data to variables (input and target are some default variables) is done using an ordered dictionary. lrn2.nn_bricks.train.train_cached is one of the two training methods of the framework, which trains a whole stack or a single layer. The other, lrn2.nn_bricks.train.train_layer_wise could be used for e.g. greedy, layer-wise pretraining of an RBM stack or for layer-wise pre-training of an Auto-Encoder stack.
# Create NN layers (according to a configuration)
layers = make_net(config, input_shape = (28, 28))
# Pack layers in stack with cross entropy cost
stack = FFNNCrossEntropy(layers, name = 'mnist_classifier')
# Define input data (mnist_data and mnist_labels are numpy arrays)
data = OrderedDict((('input', mnist_data), ('target', mnist_labels)))
# Train the whole stack with backpropagation in discriminative task
train_cached(stack, data, config, run_keyword = 'mnist_classifier', re_train = True)For an extended introduction, see the tutorial first steps.
A config file (using the ConfigObj library) determines the nets creation and training parameters. In the case below, we create a three-layered Convolutional Neural Network with Rectified Linear Units and Batch Normalization, a Max-Pooling layer between the first and the second layer, and a dense layer with Softmax units on top (with 10 neurons for 10 MNIST classes). The category [STACK] defines some training parameters for the full stack. Training parameters defined in the layer categories are only related to layer-wise training (except regularizers like the weight regularization defined via wl2, which remain part of the resulting cost function).
[OUTPUT]
output_path = nn_output # Resulting files will go nn_output/[run_keyword]
[INPUT]
ngram_size = 1 # 1 for images (e.g. MNIST)
[TOPOLOGY]
[[Layer1]]
type_net = NNConvReLUBN # class name of the layer in lrn.nn_bricks.make
# (Convolutional NN with ReLU units and Batch Normalization)
convolutional = True # convolutional flag
filter_shape = 5, 5 # shape of conv filter
n_hidden = 32 # hidden maps
wl2 = 1e-3 # wl2 weight regularization
[[LayerPooling1]]
type_net = MaxPooling # class name of the layer in lrn.nn_bricks.make
# (Max-pooling layer)
convolutional = True # convolutional flag
downsample = 2, 2 # downsample by 1/2 in each dimension
no_param_layer = True # layer has no params
[[Layer2]]
type_net = NNConvReLUBN # class name of the layer in lrn.nn_bricks.make
# (Convolutional NN with ReLU units and Batch Normalization)
convolutional = True # convolutional flag
filter_shape = 5, 5 # shape of conv filter
n_hidden = 32 # hidden maps
wl2 = 1e-3 # wl2 weight regularization
[[LayerToDense]]
type_net = ConvToDense # class name of the layer in lrn.nn_bricks.make
# (Convolution to dense transition layer)
convolutional = True # convolutional flag
no_param_layer = True # layer has no params
[[LayerTop]]
type_net = NNSoftmaxBN # class name of the layer in lrn.nn_bricks.make
# Dense NN layer with softmax units and batch normalization
n_hidden = 10 # 10 hidden units
wl2 = 0.001 # wl2 weight regularization
[STACK]
img_interval = 25 # Plot params, histograms, cost curves every 25 epochs
dump_interval = 50 # Backup model parameters every 50 epochs
learning_rate = 5e-3 # Learning rate
momentum = 0.85 # Momentum during training
batch_size = 100 # Mini-batch size
epochs = 151 # Number of epochs to train
reduce_lr = False # Don`t reduce learning rate (to zero over epochs)
The file can be imported using the method lrn2.util.config.get_config(config_file_path, specification_path), which returns a dictionary reflecting the config file content and structure. The specification_path (default is lrn2/util/config_spec.ini) points to a specification file, which defines the parameter defaults, variable types, and the config files structure (see default config_spec.ini). You can always add a new parameter (e.g. to be used with a custom layer) and use it immediately in a brick, as long as it receives **kwargs (kwargs['parameter_name']), however, if its variable type is not defined in util/config_spec.ini, it has to be cast to the desired type. In order to avoid casting, copy the config_spec.ini in your project folder, edit it accordingly and pass it in specification_path of get_config.
Specific layer parameters are forwarded to the respective layer constructors as **kwargs and should be passed on to most mix-in class constructors (those who take parameters at all). Parameters defined in the config file but not used by any mix-in class (e.g. wl2 is defined in config file, but the corresponding layer does not derive from WeightRegular) are ignored.
For further information on the configuration file, see Tutorial 3, Tutorial 1, or see the related Documentation.
lrn2 provides two classes for data management, the lrn2.data.Corpus class and the lrn2.data.LiveCorpus class. The Corpus class should be used for data which fits in the working memory, while the LiveCorpus is made for managing big data, which is stored and loaded batch by batch from a possibly fast (i.e. SSD) hard disk. However, both classes provide a very similar interface and their use do not essentially differ from each other. Note that data handling is fully independent from creation and training of the models, the final data fed into the models (corpus.ngram_data) is a simple numpy array.
Example
corpus = Corpus(file_loader = load_mnist_files, viewpoints = [MnistVP(shape = (28, 28))])
corpus.load_files(datafiles) # calls load_mnist_files(datafiles) for raw data and passes it to
# mnist_vp.raw_to_repr(raw_data) to obtain input representation
corpus.set_to_ngram(ngram_size = 1) # Prepares the actual data, here each row is a linearized image
corpus.shuffle_instances() # Shuffle rows
data = corpus.ngram_data # 2d numpy array: [instance_count, linearized_image_size]The formalization consists in the separation of the input data processing into two components:
- data loading function a function for loading data from files (format specific, e.g. MIDI)
- viewpoints a set of viewpoint objects that produce a vectorized representation of some aspect of the data returned by the data loading function, in the form of one or more instances (domain specific, e.g. Music)
When corpus.load_files(datafiles) is called, datafiles is directly passed on to the file loader (here: load_mnist_files), which reads the files and returns the raw data. The viewpoint (here: MnistVP) receives this raw data and turns it into a representation, suited for an input to a Neural Network. For more information on data handling, see the related Tutorial, refer to the Project website, the Documentation, or inspect some examples in the /examples folder.
The idea of nn_bricks is to write once and reuse NN building blocks, constituting typical components of NNs. In the following, find an overview of some pre-defined bricks in lrn2:
-
layers:
NN(Standard Neural Network layer),NN_BN(NN with batch normalization),NNAuto(Autoencoder layer),CNN(Convolutional NN layer),CNN_BN(Convolutional NN with batch normalization),DCNN(de-convolutional NN layer),RBM(Restricted Boltzmann Machine),CRBM(Convolutional Restricted Boltzmann Machine),CondCRBM(Conditional Convolutional Restricted Boltzmann Machine),RNN(Recurrent Neural Network),RNN_Gated(Gated RNN),LSTM(Long-short-term Memory),ToDense(Convolutional to dense layer),ConvShaping,MaxPooler, ... -
units:
UnitsNNSigmoid,UnitsNNTanh,UnitsNNLinear,UnitsNNLinearNonNegUnitsNNSoftmax,UnitsNNReLU,UnitsRBMSigmoid,UnitsRBMReLU,UnitsRBMGauss,UnitsNoisy(adds noise before activation function),UnitsDropOut(Adds dropout to any unit type), ... -
cost:
CostCrossEntropy,CostCategoricCrossEntropy,CostKL(Kullback-Leibler cost),CostMaxLikelihood,CostSquaredError,CostReconErr,CostCD(Contrastive Divergence cost),CostPCD(Persistent Contrastive Divergence cost),CostPCDFW(PCD cost with fast weights),CostRNNPred(Cross Entropy cost for prediction RNN, no need to define target data (as target = roll(input, -1)), ... -
regularizer:
MaxNormRegular(Max-norm regularization),SparsityLee(Lee et. al. sparsity regularizer),SparsityLeeConv(for convolutional layers),WeightRegular(w1/w2 weight regularization),SparsityGoh(Goh et. al. sparsity regularization),SparsityGOHConv(for convolutional layers),NonNegative(prevents negative weights),ActivationCrop(shuffles weights of neurons which are too active), ...
Those bricks are then used to get mixed in a class, constituting a layer or a stack (as both derive from FFBase, they share some properties, like input / output, cost, notifier,..), for example
class NNSigmoidBN(Notifier, UnitsNNSigmoid, UnitsDropOut, NN_BN, CostCrossEntropy, WeightRegular,
MaxNormRegular, SparsityLee, SerializeLayer, Monitor, Plotter):
"""
NN Layer with sigmoid activation and batch normalization.
"""
def __init__(self, **kwargs):
kwargs['variables'] = create_in_trg(kwargs['name'])
Notifier.__init__(self) # Notifier should be always added
UnitsNNSigmoid.__init__(self) # Sigmoid units
UnitsDropOut.__init__(self, **kwargs) # ..with Dropout
NN_BN.__init__(self, **kwargs) # Standard NN layer with batch normalization
CostCrossEntropy.__init__(self, **kwargs) # Cross Entropy cost
WeightRegular.__init__(self, **kwargs) # Weight regularization
MaxNormRegular.__init__(self, **kwargs) # Max-norm regularization
SparsityLee.__init__(self, **kwargs) # Lee sparsity regularization
SerializeLayer.__init__(self) # Adds ability to serialize layer
Monitor.__init__(self) # Adds ability to monitor the training
self.notify(Notifier.MAKE_FINISHED) # Those three notifications are recommended after...
self.notify(Notifier.COMPILE_FUNCTION) # ...initializing all mix-in classes.
self.notify(Notifier.REGISTER_PLOTTING)The order of calling the respective constructors should always be the following: the notifier, units, the actual layer logic, the cost function, regularizers, the serializer and the monitor. Finally, notifications should be sent, as in the example above. The constructor is called by the make_net method, which passes all configuration parameters of the corresponding section of the config file in kwargs, as well as other related information.
Several such predefined layers are available and can be used by defining them in a config file: type_net = [class name]. Custom layers can be defined, too (in that case use type_net = custom and the additional property custom_type = [class name]). For an example using custom layers, see example/custom_layer.
Layers created by make_net are not connected. In order to connect them, initialize a stack like
# Create NN layers (according to a configuration)
layers = make_net(config, input_shape = (28, 28))
# Pack layers in stack with cross entropy cost
stack = FFNNCrossEntropy(layers, name = 'mnist_classifier')A list and description of pre-defined stacks in lrn2.nn_bricks.stacks can be found here.
Stacks are build from mix-in classes, like single layers. In fact, they derive (indirectly) from FFBase, as they have things in common with single layers (e.g. symbolic input, symbolic target, cost function, variables, data binding and plot functions). Most importantly, stacks derive from NNStack, which is initialized with a list of layers. An NNStack creates a theano graph by connecting outputs with inputs, builds a global cost function by collecting regularizers of the layers, collects registered notifiers, and so on. Note that the current NNStack assumes layers with one input and one output each. For more complicated architectures, you can either encapsulate the complexity in custom (single) layers, or extend the NNStack class according to your needs.
class FFNNCrossEntropy(Notifier, NNStack, CostCrossEntropy, SerializeStack, Monitor, Plotter):
""" *Stack:* Feed forward neural network with cross entropy cost """
def __init__(self, layers, name, **kwargs):
Notifier.__init__(self)
NNStack.__init__(self, layers, layers[0].input, name)
CostCrossEntropy.__init__(self, **kwargs)
SerializeStack.__init__(self)
Monitor.__init__(self)
Plotter.__init__(self)
self.notify(Notifier.COMPILE_FUNCTIONS)
self.notify(Notifier.REGISTER_PLOTTING)Example of a feed-forward NN stack with cross-entropy cost as a combination of some mix-in classes.
The project Lrn2Cre8 acknowledges the financial support of the Future and Emerging Technologies (FET) programme within the Seventh Framework Programme for Research of the European Commission, under FET grant number 610859.