Add support for dropout layers to the MalpiConvNet class.

Seekable.update_line rewrites the file from the updated line to the end. Fine for single edits, not so good if you have many edits. Add a new method that goes through the file only once, replacing lines from a list of edits.

See what can be done to speed these up:

• Getting an image from the camera: 0.4 - 0.5 seconds
  Greatly improved by going with a separate thread for grabbing frames from the camera.
• Preprocessing an image: 0.1 seconds
• Forward pass (cnn+lstm): 0.1 - 0.2 seconds

Add support for reading v1 Tub files and writing them to a database.

https://github.com/Bleyddyn/malpi/blob/master/dk/scripts/tub2db.py

Add more data to the Experiment Meta class including:

• Framework versions: Keras, Tensorflow, numpy
• Github repo commit id (git rev-parse HEAD)
• Flag if there are any un-commited changes. Maybe with a list?
• Add a notes field

Use the Experiment class to record meta information when running train_vae.py.

Need some way to automatically calculate rewards, probably based on accelerometer data.

Sub task: add some example data to the repo.

Reward structure:

1. +1 for every time step actually moving forward
2. -10 for every time step that includes any crash or bumping into any object
3. 0 in all other cases

DKWMEnv.init should take the VAE and MDRNN parameters as objects not as paths to the weights files. That will help decouple the gym env from the implementations. Add examples of how to do it in the sample code.

The same should be done with the renderer. Just importing piglet requires jumping through hoops in Google's Colab. Letting the user skip that would be useful.

• Switch Pi to the 0.6A outlet on the battery
• See if the 2A outlet can even run the motors
• See if the battery has circuits to prevent motor load from shutting down the Pi
• If necessary, add such a circuit

If a tub record has non-zero, non-nan data for pilot/throttle or pilot/steering and the user versions are both zero, then use the pilot values. There was code to do that in the jupyter notebook at one point.

Add support for Batch Normalization layers to the model.

But first, make sure that BN and Dropout will work well together, especially once MaLPi switches to Reinforcement Learning.

Tub v2's don't handle user meta data correctly.

Meta: {'DONKEY_GYM_ENV_NAME': 'donkey-warehouse-v0', 'JOYSTICK_MAX_THROTTLE': 8.0, 'JOYSTICK_STEERING_SCALE': 1.0, 'location': 'sim', 'task': 'Train', 'driver': 'Andrew'}
Traceback (most recent call last):
File "scripts/collect_meta.py", line 110, in
vehicle = MyDriver(cfg, model_path=args['--model'], model_type=model_type, use_joystick=cfg.USE_JOYSTICK_AS_DEFAULT, meta=meta)
File "scripts/collect_meta.py", line 52, in init
super().init(cfg, model_path=model_path, use_joystick=use_joystick, model_type=model_type, camera_type=camera_type, meta=meta )
File "/home/andrew/malpi/malpi/dk/drive.py", line 111, in init
self.build(camera_type=camera_type)
File "/home/andrew/malpi/malpi/dk/drive.py", line 130, in build
self.build_recording()
File "/home/andrew/malpi/malpi/dk/drive.py", line 651, in build_recording
tub_writer = TubWriter(tub_path, inputs=inputs, types=types, metadata=self.meta)
File "/home/andrew/donkeycar/donkeycar/parts/tub_v2.py", line 114, in init
self.tub = Tub(base_path, inputs, types, metadata, max_catalog_len)
File "/home/andrew/donkeycar/donkeycar/parts/tub_v2.py", line 26, in init
self.manifest = Manifest(base_path, inputs=inputs, types=types,
File "/home/andrew/donkeycar/donkeycar/parts/datastore_v2.py", line 234, in init
self._read_metadata(metadata)
File "/home/andrew/donkeycar/donkeycar/parts/datastore_v2.py", line 320, in _read_metadata
for (key, value) in metadata:
ValueError: too many values to unpack (expected 2)