Skip to content

Latest commit

 

History

History
47 lines (39 loc) · 2.84 KB

training_log.md

File metadata and controls

47 lines (39 loc) · 2.84 KB
Raw

CarND-Behavioral-Cloning

  • Three-image-input

    • Started with three images as input, one with positive steering angle, one with negative and the other one with 0. Started a complete cover of nvidia end-to-end net.
    • Dropout was added in, the loss didn't drop at all. After removing the dropouts, it started showing life. It makes sense, as dropout is to avoid overfitting. How can I draw "redundant" information from so limited (3 images!) data? :P
    • Started with SGD and grid search on LR. Found SGD is a bit not stable. Changing to Adam, it is much more stable and faster.

  • Data provided by Udacity

    • Moved on to training data provided by Udacity. It contains ~8k images.
    • It took me a while to be able to connect to simulator and try my net. Drive.py needs customization.
    • Training with ~20 epoches, the car can pass first 2-3 curves, then headed off track and into water. I'm still very excited already, :)

  • Track1 at 30 mph  - Steps:

      1. Training from scratch with left/center/right images augmented from udacity dataset. With this alone, car will crash into wall on black bridge on track1.
      1. Train based on step 1 result but with original udacity dataset, not the augmented one. That's it.

  • Track1 at 14, 24, 30 mph solid

    • Changed database from pickle to tables(HDF5)
    • Changed to use keras image generator, used random shear, zoom, rotation.
    • Alternating between local recorded track1 and udacity dataset, used 1-2 times of left/right augmented track1 dataset, achieved the stable state.
    • Still failed badly on track2. Has to deal with overfitting.

  • Track2 shows sign of life, crashed at a right turn where has light background (Choose fastest mode thus there's no shade in track2)

    • Recorded track2 training data
    • Trained based on previous best model/weight, using smaller learning rate, for two five epochs. Loss at ~0.45.
    • Checked still solid on track 1, speed 11, 20, 30 mph.
    • Track 2 failed at a spot rith right turn that background is bright sky.

  • Track2 completed at 30mph.

    • Recorded the failure scene (right turn with bright background) for several times of passing with right steering angle.
    • Trained with lower lr for 5 epoches, RMSprop optimizer.

  • Track2 completed at 20mph w/ shadow turned on.

    • steps:

      git checkout model.json model.h5 p3 model.py track2_recover.h5.warp_shadow 3

  • Track2 completed at 30mph w/ shadow on.

    • Model becomes super sensitive. Very easy to overfit either to track2 or track1. Eg. passing on track2 for 30mph but failed with Zigzag on black bridge on track1.
    • steps: git checkout model.h5 model.json p3 model.py track2_recover_2.h5.ori_shadow 3 (following is another git check in) p3 model.py data.h5.shadow_lr 1 (track1) - baby step p3 model.py track2_recover.h5.ori_shadow 1 - baby step p3 model.py track2_recover_2.h5.ori_shadow 1 - baby step