Customise Local Training

This page lists out the files that you can modify your local training config.

Depending on which repository you are using, the file location may be different. This page use Matt's repository as example

Hyperparameters

rl_coach/rl_deepracer_coach_robomaker.py

...
estimator = RLEstimator(
    ...
    hyperparameters={
        ...
        "batch_size": 64,
        "beta_entropy": 0.01,
        "discount_factor": 0.999,
        "e_greedy_value": 0.05,
        "epsilon_steps": 10000,
        "exploration_type": "categorical",
        "loss_type": "mean squared error",
        "lr": 0.0003,
        "num_episodes_between_training": 20,
        "num_epochs": 10,
        "stack_size": 1,
        "term_cond_avg_score": 100000.0,
        "term_cond_max_episodes": 100000,
        ...
    },
    ...
)
...

Possible Values

Description of each hyperparameter

Items	Values
batch_size	64
stack_size	1
lr	0.0003
exploration_type	categorical
e_greedy_value	0.05
epsilon_steps	10000
beta_entropy	0.01
discount_factor	0.999
loss_type	Mean squared error or huber
num_episodes_between_training	20
term_cond_max_episodes	100000
term_cond_avg_score	100000

Track

data/minio/bucket/custom_files/training_params.yaml

---
WORLD_NAME: "LGSWide"
...

config.env

---
WORLD_NAME=LGSWide
...

Make sure the values in these 2 files are the same

Where can I find the track name

All tracks available for training can be found in deepracer-simapp repository

Simply put the filename (without .world) as value of WORLD_NAME

Action space

data/minio/bucket/custom_files/model_metadata.json

{
    "action_space": [
        {
            "steering_angle": -30,
            "speed": 0.4,
            "index": 0
        },
        {
            "steering_angle": -30,
            "speed": 0.8,
            "index": 1
        },
        ...
    ],
    ...
}

Make sure index is starting from 0 and in the increment of 1

Reward function

• Save your reward function into data/minio/bucket/custom_files/reward.py
• Learn more about reward function here: AWS DeepRacer Reward Function Reference

Vehicle mod specifications

For more detail, please read: Understand Sensors Enabling Racing Types Supported by AWS DeepRacer

data/minio/bucket/custom_files/model_metadata.json

{
    "action_space": [
        ...
    ],
    "sensor": [
        "LIDAR",
        "STEREO_CAMERAS"
    ],
    "neural_network": "DEEP_CONVOLUTIONAL_NETWORK_SHALLOW",
    ...
}

sensor

Option	Value
Single Camera	FRONT_FACING_CAMERA
Stereo camera	STEREO_CAMERAS
LIDAR sensor	LIDAR

Examples

1. Single camera with a LIDAR sensor

"sensor": [
    "LIDAR",
    "FRONT_FACING_CAMERA"
],

2. Single camera without LIDAR sensor

"sensor": [
    "FRONT_FACING_CAMERA"
],

3. Stereo camera with a LIDAR sensor

"sensor": [
    "LIDAR",
    "STEREO_CAMERAS"
],

neural_network

Option	Value
3 Layer CNN	DEEP_CONVOLUTIONAL_NETWORK_SHALLOW
5 Layer CNN	DEEP_CONVOLUTIONAL_NETWORK

Race type

For more detail, please read: Understand Sensors Enabling Racing Types Supported by AWS DeepRacer

data/minio/bucket/custom_files/training_params.yaml

(All values in this file should be enclosed by double quotes regardless of the data type)

---
...
RACE_TYPE: "TIME_TRIAL"
...

Setting race type

RACE_TYPE

Option	Value
Time trial	TIME_TRIAL
Object avoidance	OBJECT_AVOIDANCE
Head-to-head racing	HEAD_TO_HEAD_RACING

Difference Race Type

Object avoidance

For Object avoidance, add these keys manually

Key	Description	Possible values
NUMBER_OF_OBSTACLES	Number of objects on a track	int from 1 to 6
OBSTACLE_TYPE	Object Type (Only support box now)	BOX
RANDOMIZE_OBSTACLE_LOCATIONS	Obstacle location (Fixed or Random)	true (Random) or false (Fixed)

Head-to-head

For Head-to-head racing, add these keys manually

Key	Description	Possible values
NUMBER_OF_BOT_CARS	Number of bot vehicle(s)	int e.g. 4
BOT_CAR_SPEED	Speed for bot vehicles	number e.g. 0.5, 1.6, 4
IS_LANE_CHANGE	Enable lane changes for bot vehicle	true or false
UPPER_LANE_CHANGE_TIME	Maximum time interval for each bot to randomly change lanes	int e.g. 5
LOWER_LANE_CHANGE_TIME	Minimum time interval for each bot to randomly change lanes	int e.g. 1
RANDOMIZE_BOT_CAR_LOCATIONS	Use random locations for bot cars	true or false