Scribd
Upload
198 views

MariFlow Manual

The document provides instructions for installing, training, and running MariFlow, a recurrent neural network for playing Super Mario Kart. It discusses requirements, placing files, capturing training data by running a Bizhawk script, modifying neural network inputs, additional capture features, training a neural network using configuration files, and validating the network periodically during training.

Uploaded by

cprobbiano
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
198 views

MariFlow Manual

The document provides instructions for installing, training, and running MariFlow, a recurrent neural network for playing Super Mario Kart. It discusses requirements, placing files, capturing training data by running a Bizhawk script, modifying neural network inputs, additional capture features, training a neural network using configuration files, and validating the network periodically during training.

Uploaded by

cprobbiano
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

MariFlow Manual

SethBling

Introduction
This manual explains, roughly, how to install, train and run MariFlow, a recurrent neural network
for playing Super Mario Kart. If you want to try anything but the sample, I recommend learning
about neural networks, theres a lot to know.

My video about MariFlow: https://www.youtube.com/watch?v=Ipi40cb_RsI

Introduction

Installation
Requirements
Placing Files

Capturing Training Data


Modifying the Neural Network Inputs
Additional Capture Features
L and R
+Screen
Capturing Mixed Data

Training a Neural Network


Config Files
Validation

Running a Neural Network


The Neural Network Server
The Emulator Client
Capturing Mixed Data
Installation

Requirements
The following software is required to run MariFlow. Ive included the version numbers that I have
installed on my computer. Some of the requirements are stricter than others. See h
ere for more
information on installing TensorFlow.
Python 3.5.0 (64bit)
CUDA Development Tools 8.0.61 (64bit)
CuNN 6 (64bit)
For some reason downloading this requires creating an account.
With the release of Tensorflow 1.4, you may need CuNN 7 instead.
Visual C++ 2015 Redistributable
tensorflow-gpu/tensorflow 1.3 (can be installed with p ip etc.)
pygame (can be installed with pip etc.)
Bizhawk 1.12.2
Super Mario Kart ROM.

Placing Files
Download this zip file: https://www.dropbox.com/s/m59prsodl2t8bec/MariFlow.zip?dl=1

Place the contents of the zip file anywhere youd like. Within your Bizhawk directory, theres a
folder called Lua. Copy the contents of the Lua folder in the zip file into the Lua folder in the
Bizhawk installation. When you start training models, make sure they arent in a Dropbox folder,
or the process will fail.

Capturing Training Data


To capture training data, youll need to run a Bizhawk script called
Capture.lua from the Bizhawk Lua Console (accessed via the
Tools menu). This will create a small window that you can use to
control the script. Load up a course in whatever cc youd like (or
even time trial) and pause the game. Click Start on the
Capture.lua window. Now you may unpause the game. Run for as
many courses as youd like. You can load save states, or reset the
emulator as much as youd like. At the end of each course the game
may freeze momentarily as it writes the training data to file. The
data will be saved in a .txt file in the <Bizhawk>\Lua\Capture
folder. Click Stop on the Capture.lua window to stop recording.
Modifying the Neural Network Inputs
Certain aspects of the neural network inputs can be modified within the source code of
Capture.lua. You can change the screen size, tilt shift properties, and add extra network
nodes. Anything thats part of the cfg dictionary in C apture.lua is modifiable. If you modify
any of those constants, they will be recorded in a header of any captured data. This header will
allow the data format to remain consistent for any networks that train or run using those
constants, even if you further change the constants in the Lua file.

Additional Capture Features

L and R
L and R are not captured as part of the training data. Instead, pressing L or R will mark those
frames as non-training frames. That is, they will still be seen during the training process, but no
weight will be given to the controller inputs on those frames to improve the neural network. You
can use L or R to intentionally drive the kart into a bad situation in order to create training data
that shows the network how to recover from that bad situation.

+Screen
This will toggle a display of the neural network inputs as theyll be captured to file.

Capturing Mixed Data


This is a mode where the script hands control between the player and a neural network. This is
covered below, and can only be used after youve trained a neural network.

Training a Neural Network


Training a Recurrent Neural Network can be done in a lot of different ways. You have to choose
the architecture of the network, and various facets of how the training occurs. Ive included
some defaults that worked well for me, but to understand how changing the settings will affect
things, youll need to have some understanding of neural networks in general.
Config Files
All the settings for defining and training MariFlow are contained in a *.cfg file. d
efaults.cfg
contains default values for every parameter, and s ample.cfg has a sample configuration. You
can create your own configuration files, too. Once youve created a configuration, open a
command window in the MariFlow folder and run:

python Train.py <config file>

Here are the parameters supported by the config file:

Section Parameter Description

Data Filename List of training file names, one per line (indent additional
lines)

Data SequenceLength Number of time steps to train at once for


backpropagation through time.

Data BatchSize Number of sequences to process per batch.

Data RecurButtons Whether to include the previous time steps button


presses in the input set.

RNN Layer<number> The size of each layer of the RNN. You can have as
many layers as you like, start numbering at 1.

Checkpoint Dir The directory to save checkpoints.

Train DropoutKeep The probability that each nodes output will be kept
during training.

Train MaxGradNorm The maximum gradient component value.

Train VariationalRecurrent A variation of dropout designed for RNNs.

Train NumPasses The number of passes through the training set (this
keeps sequences from always containing the same
training data).

Train ValidationPeriod The number of seconds between validation tests.

Train LossFunction Either mean squared error or cross entropy.


Validation
It may take several minutes for the training data to be loaded and the training process to begin.
The neural network will periodically be evaluated on a special batch thats set aside for
validation (it has the same batch size and sequence length as any other batch). When a new
best validation error is found, it will save the neural network to a checkpoint file in the checkpoint
directory. If you press ctrl+C during training, it will manually save a checkpoint file and continue
training. Press ctrl+Break to exit training without saving. You may resume training from the
latest checkpoint file by simply running the same python command again.

Running a Neural Network

The Neural Network Server


Once youve trained a neural network and its written a checkpoint file, you can run the neural
network live by connecting it to Bizhawk. On the command line, from the MariFlow folder, run
this command:

python RunLive.py <config file> [checkpoint number]

If you dont specify a checkpoint number, it will run the last checkpoint file saved. Otherwise itll
use the checkpoint saved at the batch number specified (this is the number at the end of the
checkpoint file).

It may take several minutes to load the neural network, but once you do itll tell you what
hostname and port its listening on for the emulator. Take note of the hostname, youll need it to
connect with the emulator.

The Emulator Client


To connect the emulator, run the RemotePlay.lua script in the
Bizhawk Lua Console. This will bring up a small window. Write
the hostname you noted from the server into the textbox and click
NN Start. Once youre in a level, it should start playing. You can
connect and disconnect from the server by clicking the button
again.
Network Display
Once the neural network is running, a window will pop up and
display the neural networks inputs, states of the hidden layers, and
predicted outputs. This window will freeze if you pause the
emulator. Its background is green so that it can be overlaid on a
livestream using a chroma key to make the background transparent.

Capturing Mixed Data


The Capture.lua script can be used to interactively create
training data. In this mode, the script hands control back and forth
between the player and the neural network. Once youve trained a
neural network and run the server, you can operate in this mode to
create training data that will cover a wider array of situations a self-driving kart might face. In the
Capture.lua window, you can write the hostname to connect to the NN server. There are
configurable minimum and maximum numbers of frames for handoff times. Remember that the
game runs at 60fps.

You might also like