Description
GANTheftAuto is a fork of the Nvidia's GameGAN, which is research focused on emulating dynamic game environments. The early research done with GameGAN was with games like Pacman, and we aimed to try to emulate one of the most complex environments in games to date with Grand Theft Auto 5.
Video
(click to watch)
GANTheftAuto focuses mainly on the Grand Theft Auto 5 (GTA5) game, but contains other environments as well. In addition to the original project, we provide a set of improvements and fixes, with the most important ones being:
- ability to use the newest PyTorch version, which as of now is 1.8.1
- ability to use non-square images (16:8 in our case)
- larger generator and discriminator models
- ability to use more than 2 generators
- inference script (which is absent in the GameGAN repository)
- ability to use upsample model with inference
- ability to show generator outputs live during training (training preview) (soon with one of the future commits)
The work is still in progress as we know that our results can be greatly improved still.
GANTheftAuto output on the left, upscaled 4x for better visibility, and upsampled output (by a separate model)
Playable demo
You can instantly run the demo:
(you need a CUDA capable Nvidia GPU to run this demo)
-
Download and unzip or clone this repository:
git clone https://github.com/Sentdex/GANTheftAuto.git cd GANTheftAuto
-
Install requirements
Install (the highest) CUDA version of PyTorch following instructions at PyTorch's website (there is no universal command to do so). ROCm and CPU versions are currently not supported by the project.
pip3 install -r requirements.txt pip3 install tensorflow-gpu tensorflow_addons
-
Run inference:
./scripts/gtav_inference_demo.sh
or
scripts\gtav_inference_demo.bat
We are providing one of our trained models on GTA5 data as well as an 8x upsample model (part of a separate project). There's no GTA V running, this is the actual GAN output of a human playing within the GAN environment.
Example actual output of these demo models:
Trainable demo
(you need a CUDA capable Nvidia GPU to run this demo)
Since we cannot share out data collecting script, which involves a GTA5 mod and python code, we are sharing a sample dataset which you can use to train your model on. It's included within the data/gtav/gtagan_2_sample
folder.
To train your own model, follow the steps above, but run a training script instead.
- Run training:
or
./scripts/gtav_multi_demo.sh
scripts\gtav_inference_demo.bat
You'll need a GPU with at least 8 GB of VRAM.
Batch size in the demo scripts is currently set to 1
. If you have 16 GB of VRAM or more, try to find the biggest batch that you can fit in your GPU(s).
General
(you need a CUDA capable Nvidia GPU to run this code, but we are open for contribution to make it running on AMD GPUs as well)
Environment Setup
-
Download and unzip or clone the repository
git clone https://github.com/Sentdex/GANTheftAuto.git cd GANTheftAuto
-
Install dependencies
Install (the highest) CUDA version of PyTorch following instructions at PyTorch's website (there is no universal command to do so). ROCm and CPU versions are currently not supported by the project.
pip3 install -r requirements.txt
Dataset extraction
Currently, GTA V, Vroom and Cartpole are the only implemented data sources.
GTA V environment
This is an environment created using Grand Theft Auto V. We created our own GTA5 mod accompanied by a Python script to collect the data. It contains a simple driving AI (which we named DumbAI ;) ). We are pulling road nodes from the game and apply math transformations to create paths, then we are spawning multiple cars at the same time and alternate them to pull multiple streams at the same time (to speedup training). Game mod accepts steering commands from the Python script as well as limits the speed and sets other options like weather, traffic, etc. Python script analyzes current car position and nearest road nodes to drive using different paths to cover all possible actions and car positions as best as possible. This is important for further seamless experience with player "playing" the environment - it needs to output coherent and believable images.
Data collecting demo with visible road nodes (not included in the final training data):
(click to watch on YouTube)
As mentioned above, we can't share our data collecting scripts, but we are providing sample dataset. If you believe you have a model that has interesting results, feel free to reach out and we may try to train it on the full dataset.
You can also create your own dataset by recording frames and actions at 10 FPS. Save format is gzipped pickle file containing a dictionary of 'actions'
and 'observations'
. Actions are a single-dimensional NumPy array of 0
(left), 1
(straight) and 2
(right), while observations are a four-dimensional array where the first dimension are samples, and the other are (48, 80, 3)
- RGB image size. Ungzip and unpickle example sample from the sample dataset to learn more about the data structure. Each file should contain a single sequence length of at least 32 frames.
Example train script is available at scripts/gtav_multi.sh
(as well as its .bat
version).
Vroom environment
Vroom is our own environment based on the OpenmAI Gym's Box2D CarRacing environment, but this one does not require Gym to run. Its only dependencies are OpenCV and NumPy.
Example track with a slice of what's actually saved as a training data:
(blue, red and purple lines are shown for visualization purposes only and are not a part of the training data)
Example model output (we've never hunted for best possible output and switch to GTAV instead):
We are including the data collecting script - a simple AI (DumbAI) is playing the environment to collect the data. The car is "instructed" to follow the road, with additional constantly changing offset from the center of teh road, turns and u-turns to cover all possible scenarios.
To run the data collector:
- Install dependencies
cd data/vroom pip3 install - requirements.txt
- Edit
collect_data.py
if you need to change any defaults - Run the data extraction
python3 collect_data.py
NEAT-Cartpole
This environment is created with OpenAI Gym's Cartpole. However, the data collecting part is unattended as we are first training the NEAT algoritm to play it, then collect data generated this way.
Warning: recently we've discovered a possible issue with this environment causing actions to alternate between a direction and no action. As for now we have no fix for this environment, so your model results are highly likely to not be very useful. We'd recommend trying to build your own agent to play cartpole instead of a NEAT bot.
To run the data collector:
- Install dependencies
cd data/cartpole pip3 install - requirements.txt
- Edit
neat_cartpole.py
and update constants (at the bottom of the script) to your needs - Run the data extraction
python3 neat_cartpole.py
Training
We provide training scripts in './scripts'.
GTA V
- For training the full GameGAN model, run:
./scripts/gtav_multi.sh
Vroom
- For training the full GameGAN model, run:
./scripts/vroom_multi.sh
NEAT-Cartpole
- For training the full GameGAN model, run:
./scripts/cartpole_multi.sh
Monitoring
- You can monitor the training process with tensorboard:
tensorboard --logdir=./results
Inference
Inference is currently unfinished - can be ran, but actions are randomly generated instead of taken from the user input. We'll finish it up shortly.
Vroom
Edit scripts/gtav_inference.sh
and update the model filename, then run:
./scripts/gtav_inference.sh
Vroom
Edit scripts/cartpole_inference.sh
and update the model filename, then run:
./scripts/cartpole_inference.sh
NEAT-Cartpole
Edit scripts/cartpole_inference.sh
and update the model filename, then run:
./scripts/cartpole_inference.sh
Parts of the Original Nvidia's GameGAN readme
(head to the GameGAN for a full version)
This part describes the VidDom environment which we did not use in our work. The repository also contains Pac Man environment which have been never described and no data collection scrpts were provided.
Dataset extraction
Please clone and follow https://github.com/hardmaru/WorldModelsExperiments/tree/master/doomrnn to install the VizDoom environment.
- Copy extraction scripts and run
cp data/extract.py DOOMRNN_DIR/
cp data/extract_data.sh DOOMRNN_DIR/
cd DOOMRNN_DIR
./extract_data.sh
- Now, extracted data is saved in 'DOOMRNN_DIR/vizdoom_skip3'
cd GameGAN_code/data
python dataloader.py DOOMRNN_DIR/vizdoom_skip3 vizdoom
- You should now see .npy files extracted in 'data/vizdoom' directory.
For custom datasets, you can construct .npy files that contain a sequence of image and action pairs and define a dataloader similar to 'class vizdoom_dataset'. Please refer to 'data/dataloder.py'.
-- The above repository is deprecated and VizDoom environment might not run correctly in certain systems. In that case, you can use the docker installtaion of https://github.com/zacwellmer/WorldModels and copy the extraction scripts in the docker environment.
Training
We provide training scripts in './scripts'.
- For training the full GameGAN model, run:
./scripts/vizdoom_multi.sh
- For training the GameGAN model without the external memory module, run:
./scripts/vizdoom_single.sh
Monitoring
- You can monitor the training process with tensorboard:
tensorboard --port=PORT --logdir=./results
Tips
- Different environments might need different hyper-parameters. The most important hyper-parameter is 'recon_loss_multiplier' in 'config.py', which usually works well with 0.001 ~ 0.05.
- Environments that do not need long-term consistency usually works better without the external memory module