A deep neural network that learns to drive in video games
T.E.D.D. 1104 S: 138M Parameters. 1.6GB
control_mode: keyboard
cnn_model_name: efficientnet_v2_l
pretrained_cnn: true
encoder_type: transformer
embedded_size: 896
nhead: 8
num_layers_encoder: 4
learning_rate: 1e-05
optimizer_name: adafactor
scheduler_name: cosine
warmup_factor: 0.05
max_epochs: 20
batch_size: 32
mask_prob: 0.2
dropout_cnn_out: 0.3
dropout_encoder: 0.1
dropout_encoder_features: 0.3
positional_embeddings_dropout: 0.1
weight_decay: 1e-3
Accuracy in the test datasets:
Time | Weather | Micro-Acc K@1 | Micro-Acc k@3 | Macro-Acc K@1 | |
---|---|---|---|---|---|
City | :sun_with_face: | :sunny: | 53.2 | 84.4 | 46.2 |
City | :sun_with_face: | :umbrella: | 51.4 | 83.4 | 46.3 |
City | :first_quarter_moon_with_face: | :sunny: | 54.3 | 85.6 | 46.3 |
City | :first_quarter_moon_with_face: | :umbrella: | 47.3 | 82.3 | 49.9 |
Highway | :sun_with_face: | :sunny: | 72.7 | 97.7 | 40.6 |
Highway | :sun_with_face: | :umbrella: | 70.6 | 99.3 | 39.6 |
Highway | :first_quarter_moon_with_face: | :sunny: | 77.9 | 99.3 | 45.7 |
Highway | :first_quarter_moon_with_face: | :umbrella: | 70.9 | 97.6 | 30.8 |
The release includes the best epoch in the development set and the last epoch.
T.E.D.D. 1104 S: 68M Parameters. 685MB
control_mode: keyboard
cnn_model_name: efficientnet_v2_m
pretrained_cnn: true
encoder_type: transformer
embedded_size: 512
nhead: 8
num_layers_encoder: 4
learning_rate: 1e-05
optimizer_name: adafactor
scheduler_name: cosine
warmup_factor: 0.05
max_epochs: 20
batch_size: 32
mask_prob: 0.2
dropout_cnn_out: 0.3
dropout_encoder: 0.1
dropout_encoder_features: 0.3
positional_embeddings_dropout: 0.1
weight_decay: 1e-3
Accuracy in the test datasets:
Time | Weather | Micro-Acc K@1 | Micro-Acc k@3 | Macro-Acc K@1 | |
---|---|---|---|---|---|
City | :sun_with_face: | :sunny: | 52.9 | 84.1 | 43.1 |
City | :sun_with_face: | :umbrella: | 49.9 | 81.3 | 42.2 |
City | :first_quarter_moon_with_face: | :sunny: | 54.7 | 85.1 | 48.4 |
City | :first_quarter_moon_with_face: | :umbrella: | 49.5 | 81.1 | 41.1 |
Highway | :sun_with_face: | :sunny: | 62.5 | 99.2 | 43.1 |
Highway | :sun_with_face: | :umbrella: | 71.9 | 99.3 | 39.2 |
Highway | :first_quarter_moon_with_face: | :sunny: | 79.4 | 99.3 | 45.3 |
Highway | :first_quarter_moon_with_face: | :umbrella: | 63.0 | 97.2 | 47.2 |
The release includes the best epoch in the development set and the last epoch.
T.E.D.D. 1104 S: 26M Parameters. 260MB
control_mode: keyboard
cnn_model_name: efficientnet_v2_s
pretrained_cnn: true
encoder_type: transformer
embedded_size: 384
nhead: 8
num_layers_encoder: 2
learning_rate: 1e-05
optimizer_name: adafactor
scheduler_name: cosine
warmup_factor: 0.05
max_epochs: 20
batch_size: 32
mask_prob: 0.2
dropout_cnn_out: 0.3
dropout_encoder: 0.1
dropout_encoder_features: 0.3
positional_embeddings_dropout: 0.1
weight_decay: 1e-3
Accuracy in the test datasets:
Time | Weather | Micro-Acc K@1 | Micro-Acc k@3 | Macro-Acc K@1 | |
---|---|---|---|---|---|
City | :sun_with_face: | :sunny: | 51.0 | 83.0 | 46.3 |
City | :sun_with_face: | :umbrella: | 49.0 | 82.5 | 45.2 |
City | :first_quarter_moon_with_face: | :sunny: | 56.3 | 86.6 | 49.0 |
City | :first_quarter_moon_with_face: | :umbrella: | 49.4 | 81.4 | 42.5 |
Highway | :sun_with_face: | :sunny: | 70.3 | 100 | 68.5 |
Highway | :sun_with_face: | :umbrella: | 71.2 | 100 | 37.6 |
Highway | :first_quarter_moon_with_face: | :sunny: | 80.9 | 100 | 49.1 |
Highway | :first_quarter_moon_with_face: | :umbrella: | 69.3 | 100 | 61.1 |
The release includes the best epoch in the development set and the last epoch.
THIS MODEL WILL NOT WORK WITH THE CURRENT VERSION OF TEDD1104
T.E.D.D. 1104 Base: 34.6M Parameters. 415.9Mb
cnn_model_name: efficientnet_b4
control_mode: keyboard
dropout_cnn_out: 0.3
dropout_encoder: 0.1
dropout_encoder_features: 0.3
embedded_size: 512
encoder_type: transformer
learning_rate: 1.0e-05
lstm_hidden_size: 512
mask_prob: 0.2
nhead: 8
num_layers_encoder: 4
positional_embeddings_dropout: 0.1
pretrained_cnn: true
sequence_size: 5
weight_decay: 0.001
weights: null
num_epochs: 12
batch_size: 64
Time | Weather | Micro-Acc K@1 | Micro-Acc k@3 | Macro-Acc K@1 | |
---|---|---|---|---|---|
City | :sun_with_face: | :sunny: | 49.8 | 83.8 | 44.1 |
City | :sun_with_face: | :umbrella: | 52.1 | 84.7 | 46.1 |
City | :first_quarter_moon_with_face: | :sunny: | 54.5 | 86.9 | 48.0 |
City | :first_quarter_moon_with_face: | :umbrella: | 48.8 | 82.5 | 43.2 |
Highway | :sun_with_face: | :sunny: | 65.6 | 100.0 | 53.2 |
Highway | :sun_with_face: | :umbrella: | 70.6 | 98.0 | 54.2 |
Highway | :first_quarter_moon_with_face: | :sunny: | 71.3 | 100.0 | 52.3 |
Highway | :first_quarter_moon_with_face: | :umbrella: | 67.7 | 100.0 | 50.9 |
Trained using 20,479,120 examples (different vehicles and in different weather conditions). -"resnet": 18 -"pretrained_resnet": true -"sequence_size": 5 -"embedded_size": 256 -"hidden_size": 128 -"num_layers_lstm": 1 -"bidirectional_lstm": true -"layers_out": null -"dropout_cnn": 0.2 -"dropout_cnn_out": 0.1 -"dropout_lstm": 0.0 -"dropout_lstm_out": 0.1 -"fp16": true
HOWTO: Unzip the model somewhere on your computer. Use the command python run_TEDD1104.py --model_dir PATH where "PATH" is the path to the directory where you have unzipped the model (i.e D:\GTAV-AI\models), the directory should contain the files: "model.bin" and "model_hyperparameters.json". For more info see: https://github.com/ikergarcia1996/Self-Driving-Car-in-Video-Games#run-the-model
Trained using 1TB of data (100 hours of gameplay using different vehicles and in different weather conditions) for 4 epochs. The minimap was removed from the images with a probability of 0% in the first epoch, 20% in the second epoch and 50% in the following epochs. The model is trained to follow a route in the minimap, you need to mark points in the map and the model will drive to them.
Hyper-parameters: -"resnet": 18 -"pretrained_resnet": true -"sequence_size": 5, -"embedded_size": 256 -"hidden_size": 128 -"num_layers_lstm": 1 -"bidirectional_lstm": false -"layers_out": null, "dropout_cnn": 0.1 -"dropout_cnn_out": 0.1 -"dropout_lstm": 0.1 -"dropout_lstm_out": 0.1 -"fp16": true -"amp_opt_level": "O2"
HOWTO: Unzip the model somewhere on your computer. Use the command python run_TEDD1104.py --model_dir PATH where "PATH" is the path to the directory where you have unzipped the model (i.e D:\GTAV-AI\models), the directory should contain the files: "model.bin" and "model_hyperparameters.json". For more info see: https://github.com/ikergarcia1996/Self-Driving-Car-in-Video-Games#run-the-model
Trained using 658GB of data (65 hours of gameplay using different vehicles and in different weather conditions) for two epochs. During the 2º epoch, the minimap was removed from 20% of the images. The model is trained to follow a route in the minimap, you need to mark points in the map and the model will drive to them.
Parameters: -"resnet": 18 -"pretrained_resnet": true -"sequence_size": 5, -"embedded_size": 256 -"hidden_size": 128 -"num_layers_lstm": 1 -"bidirectional_lstm": false -"layers_out": null, "dropout_cnn": 0.1 -"dropout_cnn_out": 0.1 -"dropout_lstm": 0.1 -"dropout_lstm_out": 0.1 -"fp16": true -"amp_opt_level": "O2"
HOWTO: Unzip the model somewhere on your computer. Use the command python run_TEDD1104.py --model_dir PATH where "PATH" is the path to the directory where you have unzipped the model (i.e D:\GTAV-AI\models), the directory should contain the files: "model.bin" and "model_hyperparameters.json". For more info see: https://github.com/ikergarcia1996/Self-Driving-Car-in-Video-Games#run-the-model
[EN] This is a .exe file that runs the generate_data.py script. It does not require python or any dependency so anybody can execute it with just double-clicking it and help to develop a dataset to train the model. By default, the executable will save the generated dataset files in the same directory where the .exe is stored. Note: First time you launch the .exe it may take a while due to windows antivirus service. This executable is outdated uses the V2 dataset format, use the generate_data.py script.
[ES] Este es un archivo.exe que ejecuta el script generate_data.py. No requiere python ni ninguna otra dependencia por lo tanto cualquiera puede ejecutarlo simplemente dandole doble clic y ayudar a generar un dataset para entrenar el modelo. Por defecto, el ejecutable guardará los archivos generados en el mismo directorio donde guardes el .exe. Nota: La primera vez que lo ejecutas puede tardar un rato por el escaneo que realiza el antivirus de windows. Este executable está obsoleto, usa el formato V2, usa el script generate_data.py