Using Deep Learning to Predict Steering Angles. Udacity Challenge #2

Install all dependencies with:

$ pip install -r requirements.txt

This project is built on top of TensorKit which helps organizing experiments.

pip install https://github.com/nghiattran/tensorkit/archive/master.zip

Also, make sure you have Tensorflow above 1.0.

Check Hypes to see all options or to create your own ones.

You can train your own model by using.

$ tk-train path-to-your-hype

There are times that training process is interrupted due to unexpected reasons. In this case, you can resume training by using tk-continue.

$ tk-continue path-to-logdir

To evaluate trained model on valuation set, you can use tk-evaluate.

$ tk-evaluate path-to-logdir

To test your model, use submission.py to generate csv file. Format of this csv file would be similar to CH2_final_evaluation.csv.

$ python submission.py path-to-logdir path-image-folder

See submission.py for more detail.

Use visualize.py to create a demo video from csv files generated above.

$ python visualize.py path-to-input-csv path-groundtruth-csv path-to-image-folder

See visualize.py for more detail.

• model:
  • dataset_file: path to python file that handles input.
  • architecture_file: path to python file that constructs main graph.
  • objective_file: path to python file that handles losses.
  • optimizer_file: path to python file that handles updating variables.
  • evaluator_file: path to python file that handles validation.
• data:
  • train_file: path to training data file.
  • val_file: path to validation data file.
• logging:
  • display_iter: display frequency.
  • eval_iter: validation frequency.
  • save_iter: saving model frequency.
• solver:
  • opt: type of optimizer. Supported Adam, RMS, and SGD.
  • rnd_seed: seed number for random.
  • epsilon: value for epsilon (a small number that is used to avoid zero division).
  • learning_rate
  • max_steps
  • batch_size
• clip_norm: upper bound for gradients to avoid gradient exploding.
• image_height
• image_width
• reg_strength: regularization strength.
• color_space: image preprocessing color space. Supported rgb and yuv. Default rgb.
• crop: crop size. This value count from bottom to top.

To fine-tune, you only need to change learning_rate, opt, and max_steps in solver and other hyperparameters like reg_strength, color_space, crop,...

You can also create your own neural network architect by using layout in model/nvidia/architect.py and keep other parameter as is.

• nvidia: inspired by SullyChen's implementation of Nvidia's paper End to End Learning for Self-Driving Cars but our implementation runs way faster.

• rambo: inspired by rambo team's submission.

NOTE: All models in this project are similar but not identical to what they are based on.