How to use my Deep Learning library:

Documentation:

The Documentation can be found in the docs folder.

It contains:

• the doxygen-config File
• a pdf Documentation, explainin the Top-level concepts
• a ready compiled doxygen documentation

Python Bindings:

In order to use the python Bindings one most build the target 'libdl'. In the next step one can then import libdl into a python script as follows:

import sys
pathBuildDir = 'path/to/your/build/directory/'

sys.path.append(pathBuildDir+'library/bindings')


import libdl as dl

Problems that may Occur

in the Top Level CMakeLists.txt -march=native is set as CXX Flag, this causes problems for some processors resulting in SIGSEVs

Example Applications:

Project: LEGO Bricks Classification:

The Project aimed, to classify 16 different types of LEGO Bricks. For each Lego Brick Type there exist around 400 images showing the brick from a different angle e.g.:

Label: 2357-Brick corner 1x2x2: Label: 3022-Plate 2x2 Label: 3040 Roof Tile 1x2x45deg Label: 6632 Technic Lever 3M

There are in total three examples that run a network on this data:

• the presentation jupyter notebook in pythonApps/ :
  • mostly visualization and explanation
• pythonApps/example_lego.py :
  • playground in python with comments in the file
  • a pretrained Network is set
• example_LEGO.cpp :
  • similiar to the MNIST examples, comments on the parameters are in the beginning of main()
  • no pretrained network is set

Note: For the python examples please first apply the above provided instructions.

XOR Problem

The XOR problem is simulated in the build File Example_XOR, which is build from the Code in Example_XOR.cpp
When executing, it uses the following Mini-batch where one data point is represented by one row:

      Eigen::MatrixXf mX1(4, 2);
            mX1 << 1, 0,
                         0, 1,
                         0, 0,
                         1, 1;

with the Corresponding Classes stored in the Vector C

      Eigen::MatrixXf C(4, 1);
            C << 1,
                       1,
                       0,
                       0;

and randomly initialized Weights and Biases.

As a output it gives the prediction and loss after 10 epochs with learning Rate 10.

MNIST

The MNIST problem is simulated in the build File Example_MNIST, which is build from the Code in Example_MNIST.cpp
The settings are predefined so that pre-learned weights are read from the Storage and are then used to predict a subset of the MNIST testset.

In order to train the Model, the settings at the beginning of the main method (in Example_MNIST.cpp) have to be changed. See inline Comments.