DeepLearning

Code to build MLP models for outdoor head orientation tracking

The following is the set of directories with a description of what is in each.

dl

The directory with the python files to train and test an MLP.

The file genmlp.py is based on the mlp.py that is part of the Theano documentation. It is more general purpose in that one can configure a network of arbitrary depth and number of nodes per layer. It also implements a sliding window for training that enables one to train data sets of arbitrary size on limited GPU memory.

The file logistic_sgd.py comes with a nice reporting function that builds a matrix with classification results on the test set where we show the number of correctly classified frames and the distribution of the incorrectly classified frames across all classes.

The file pickler.py has a number of helper methods that can be used to build files with the data that conform to the Theano input format. The file takes as input files in the MNIST IDX format. It can be used to chunk data sets into multiple sets of files, one for training, one for validation, and the last for test.

utils

The directory with C++ code that can be used to generate datasets in the MNIST IDX format from labeled data. The labels correspond to a partition of the space in front of a driver in a car, with the following values,

1. Driver window
2. Left of center
3. Straight ahead
4. Right of center
5. Passenger window

Given a video of the driver, an annotation file for that video has the following format,

<?xml version="1.0"?>
<annotations dir="/media/CESAR-EXT02/VCode/CESAR_May-Fri-11-11-00-50-2012" center="350,200">
  <frame>
    <frameNumber>1</frameNumber>
    <face>0,0</face>
    <zone>9</zone>
    <status>1</status>
    <intersection>4</intersection>
  </frame>
  <frame>
    <frameNumber>2</frameNumber>
    <face>0,0</face>
    <zone>9</zone>
    <status>1</status>
    <intersection>4</intersection>
  </frame>
  ...
  ...
</annotations>

where, the directory is expected to contain frames from the video with the following filenames 'frame_%d.png'%frameNumber. Each video frame is a 640x480 image file with the zone indicating the class, the status indicating the car status, and the intersection indicating the type of intersection. For the purposes of building the data sets, we only use the zone information at this point. The center is expected to be the rough center of the location of the face in each frame.

The pre-processing that is done on the images is as follows,

1. A Region of Interest (ROI) of configurable size (Globals.cpp) is picked around the image center.
2. A histogram equalization followed by edge detection is performed.
3. A DC suppression using a sigmoid is then done.
4. A gaussian window function is applied around the center.
5. The image is scaled and a vector generated from the image matrix in row-major.

Build

Do a make mode=opt in utils/src to build optimized executables. The dependencies is OpenCV. This builds everything and places them in an install directory under DeepLearning.

Data generation

To generate data sets, use the following commands,

xmlToIDX 
-o <outputFileNameSuffix> 
-r <training_fraction> 
-v <validation_fraction> 
-status <statusFilter> 
-inter <intersectionFilter> 
[-d <trainingDirectory>]+ 
[-b binaryThreshold] 
[-usebins] 
[-h for usage]

if the outputFileNameSuffix is ubyte, then run the following command to generate pickled numpy arrays from the IDX data sets

python pickler.py data-train-ubyte label-train-ubyte data-valid-ubyte label-valid-ubyte data-test-ubyte label-test-ubyte gaze_data.pkl

which will generate sets of training, validation, and test files with the prefix gaze_data.pkl. The number of files generated in each set will depend on the chunking size used in pickler.py. The data is broken up into chunks and files are generated one per chunk; as an example the set of test files will be gaze_data.pkl_test_%d.gz, with the integer argument in range(numberOfChunks). The first command builds the IDX format data sets. The second converts them into a numpy array of tuples, with each tuple being an array of data points and an array of labels. We have one tuple for the training data, one for validation, and one for test.

The options to xmlToIDX are as follows,

• -o is the suffix to use for all generated files
• -r is the training fraction in the interval [0, 1)
• -v is the validation fraction in the interval (0, 1)
• -usebins is used to bin the data based on their labels. We generate as many data points as argmin_{l \in labels} |D_l|, where D_l is the set of data points with label l; in other workds we pick as many data points as the cardinality of the smallest set of data points across all labels. This is to prevent our network from being biased to class label 3, which is straight ahead. A large fraction of the frames have the driver facing straight ahead which causes an enormous bias during training without binning.
• -d a directory of images for training. An annotation file called annotations.xml is expected to be present in each such directory.
• -b is used to specify a binary threshold that is used to generate image pixel data as binary values with all pixel values above the threshold considered as 1, with the rest being 0.
• -status is used to pick only those frames that have a car status annotation that matches what follows this flag
• -intersection is used to pick only those frames that have the intersection annotation that matches what follows this flag

The second command builds the tuples of numpy arrays as required by the Theano based trainer. This one takes as input the training, validation, and test data and label files with the prefix to use for the generated file names.

Training and classification

Training and classification can be done using genmlp.py. The following command will train a network and generate a report with the validation error rate, test error rate, and the distribution of the numbers of frames across all classes together with the expected number of frames per class.

THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python genmlp.py 
(-d datasetDir | -f datasetFileName) 
[-p prefix] 
[-b batchSize]
[-nt nTrainFiles] 
[-nv nValidFiles] 
[-ns nTestFiles]
[-i inputLength] 
[-o numClasses] 
[-gen modelFileName]
[-l [nLayer1Size, nLayer2Size, ...]] 
[-classify] 
[-useparams paramsFileName]
[-h help]

The options are,

• -d the directory that contains the data sets
• -f a single file that contains the complete pickled data. This is useful when the data sets are small enough to be pickled into one file
• -p the file name prefix for the files names that hold the data sets
• -nt the number of training files
• -nv the number of validation files
• -ns the number of test files
• -l the configuration of the hidden layers in the network, with as many hidden layers as the number of comma separated elements with the size of each hidden layer being the elements
• -o the number of labels
• -i the input dimension of the data
• -gen to generate the trained model for use outside Theano. This is as a text file. We also generate a pickled file called params.gz in the training data set directory that contains the numpy weights and biases of all hidden layers and the final logistic layer.

For questions please send mail to: vishwa.raman@west.cmu.edu

Thanks for looking.