This python projects is a DQN agent for google chrome dino.
For the interaction with the environment, it uses a modified
version of package gym-chrome-dino and it is built using
tensorflow 2.0.
The project also uses tensorboard logs that are saved in directory ./logs
The first step to use the repository is to clone it using git bash:
git clone https://github.com/margaeor/dino-dqn.git
If you don't have git bash, you can download the zip of the project and extract into a directory.
Next, you will need to install the appropriate packages.
For this project, I am using Anaconda and I have exported the environment into file environment.yaml.
So, open environment.yaml and change the field name to the name of the environment you want to create.
For instance, you can write name:dino_tf_env if you want to create an environment named dino_tf_env.
Next, change prefix to resemble the following format:
prefix: <anaconda-path>\env\<env_name>
For example, for environment dino_tf_env the path may be:
prefix: C:\Anaconda3\envs\dino_tf_env
The next step is to import the environment from file.
In order to do that, open an Anaconda prompt in the directory
containing environment.yaml and execute the following command:
conda env create -f=environment.yaml
After Anaconda finishes creating the environment, we will need to activate it using:
conda activate dino_tf_env
We are almost ready to go. The last thing we need is the chrome driver for selenium. You can try running the project using Anaconda prompt to see whether there are any problems with the chrome driver.
In order to make a sample run of the project, just execute the following command (parameters will be discussed later):
python main.py test_model --no-duck --no-logs
If the project runs successfully and you can see a chrome window with a dino, then everything is fine.
Otherwise, if you see an error regarding the version of chrome webdriver, go to http://chromedriver.chromium.org/downloads and download the chromeriver executable that corresponds to the version of chrome you have installed.
Then, place this executable in the same folder as main.py and everything should be fine.
Before doing anything, you first have to open Anaconda Prompt into the directory containing main.py.
Then you can execute main.py with the following parameters:
python main.py [-h] [--no-duck] [--no-logs] [--evaluate] [--headless]
[--no-acceleration] [--use-statistics] [--model MODEL]
[--episode EPISODE]
model_name
| Short | Long | Default | Description |
|---|---|---|---|
-h |
--help |
show this help message and exit | |
--no-duck |
Disable duck action | ||
--no-logs |
Disable tensorboard logs | ||
--evaluate |
Evaluate model instead of training | ||
--headless |
Run without browser window and show the result every some episodes | ||
--no-acceleration |
Disable environment acceleration | ||
--use-statistics |
Use statistics instead of images as input to the network. Those statistics include distance between dino and obstacle, dino height, bird height, obstacle gap e.t.c | ||
--model |
None |
Directory from which to restore saved model. This directory must contain saved_model.pb | |
--episode |
1 |
Number of episode where training starts from (default 1) |
Let's say we want to start a new training process from scratch with a model named new_model.
We also want to have logs, but we want to disable the duck action, so that our action space has just 2 actions: doing nothing and jumping.
Then we would run main.py as follows:
python main.py new_model --no-duck
Let's suppose now that we want to load the pretrained model located in ./models/no_duck/model1__400 and continue training from there.
This model doesn't use duck, so we must have flag --no-duck enabled.
In addition to that, this model was trained until episode 400 so we must resume training from episode 401.
Hence, we can run main.py as follows:
python main.py new_model --no-duck --model ./models/no_duck/model1_400 --episode 401
As a final example, lets suppose we want to load model located in ./models/duck/model1__2400 just for evaluation, but we don't want to enable logs.
Taking into account that this model uses duck we must not specify --no-duck.
So, we would execute our script as follows:
python main.py new_model --model ./models/duck/model2_2400 --evaluate --no-logs
If you choose to not disable logs, then the logs will be stored in folder ./logs in the main project directory.
Then, in order to see the logs, you have to run tensorboard from Anaconda.
To do that, open an Anaconda prompt at the directory of the project (where main.py) is located.
Then run the following command:
tensorboard --logdir=./logs/
This will open an http port in locahost that will allow you to view the logs. The address that you generally need to visit from any browser is http://localhost:6006 to see the detailed logs.
The name of a particular log contains the name of the model, the episode when training started (usually 1) and the timestamp.
Logs are stored every AGGREGATE_STATS_EVERY episodes, a constant that can be changed from broker.py