jensen1026 / burlap_caffe

burlap_caffe

Caffe Dependencies

To make sure you have all the dependencies for Caffe, we recommend cloning the Caffe Github repo and compiling it (with CUDA) by following their installation instructions.

Additionally, make sure CUDA is added to the LD_LIBRARY_PATH. If CUDA was installed with apt-get (on a 64-bit machine), you can simply add this to your .bashrc

export LD_LIBRARY_PATH=/usr/local/lib:/usr/local/cuda/lib64:$LD_LIBRARY_PATH

And then running

source ~/.bashrc

JavaCPP Caffe with cuDNN

We are using Bytedeco's Java bindings for Caffe, the dependencies of which are all handled through maven, and should download and run smoothly. But, this repository compiles Caffe without cuDNN, which makes the GPU training much slower. To compile with cuDNN, first clone the JavaCPP presets repo:

git clone https://github.com/bytedeco/javacpp-presets

By default, this library does not compile caffe with cuDNN, so we have to change the cppbuild.sh script. Replace the javacpp-presets/caffe/cppbuild.sh with the file provided in the instalation_files directory. Now run these commands from the javacpp-presets directory to compile JavaCPP caffe and install it to maven:

./cppbuild.sh install caffe
mvn clean install --projects caffe

Now this library should be linked with maven.

Example Code

We provide two sets of example code within the project.

The first, AtariDQN, is the same architecture as DeepMind's DQN with the same hyper-parameters as specified in their [nature paper] (http://www.nature.com/nature/journal/v518/n7540/full/nature14236.html).

To run this code, you will need to ALE. To install ALE, follow the instructions on our burlap_ale repository. If you want to replicate Deep-Mind's results, be sure to use our fork of ALE, and set TERMINATE_ON_END_LIFE to true in AtariDQN.

NOTE: You may need to increase the Java max memory size to run this example. To do so, add -Xmx8192m as a Java VM argument (this sets the max memory to 8GB).

The second, GridWorldDQN, is a simple DQN implementation for a built-in BURLAP domain.