Draft: Tobi Dance #22
Draft
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…mass M to m_cart and cart pole mass m to m_pole added comments about meaning of "stage" in MPPI renamed TargetPositionGenerator to TargetTrajectoryGenerator, since we target a complete trajectory, not just cart position. renamed my_logger() to get_logger()
say which compiler we are using
…istency with other config files. added more docstrings for undocumented methods and constructors. added updated_attributes to barebones controller. improved formatting of logging output added dictdiffer to requirements.txt renamed target_position cost function to cartpole_trajectory_cost.py. replaced some load_config with the load_or_reload_config_if_modified, so that the configs are cached for change checking. added logger to Compile.py
…g and the cartpole is at least balancing itself and swinging up from down position. Big changes all over! Now anytime any of the config files is modified during runtime, the using class has its tf.variable assigned the new value. To get this work, the variable MUST NOT EXIST before it is set this way the first time. Otherwise the compiler just uses the field at compile time and the changes are never seen inside. The cartpole window now does not reset the initial position and angle sliders on each start.
…hanging the cost weights during runtime! Yay, finally! Trick was to make sure that all values are actually propagated to tf variables, e.g. numpy arrays, ints, string, not just float Changed logging format to start with level for better readability. added string type to computation_library.py
frehe
commented
Dec 14, 2022
Comment on lines
+61
to
+63
| if not computation_library is None and computation_library not in self.predictor.supported_computation_libraries: | ||
| raise ValueError( | ||
| f"Predictor {self.predictor.__class__.__name__} does not support {computation_library.__name__}") |
There was a problem hiding this comment.
Suggested change
| if not computation_library is None and computation_library not in self.predictor.supported_computation_libraries: | |
| raise ValueError( | |
| f"Predictor {self.predictor.__class__.__name__} does not support {computation_library.__name__}") | |
| if computation_library is not None and computation_library not in self.predictor.supported_computation_libraries: | |
| raise ValueError( | |
| f"Predictor {self.predictor.__class__.__name__} does not support {computation_library.__name__}" | |
| ) |
frehe
commented
Dec 14, 2022
| @@ -54,7 +54,7 @@ class ComputationLibrary: | |||
| gather: Callable[[TensorType, TensorType, int], TensorType] = None | |||
| gather_last: Callable[[TensorType, TensorType], TensorType] = None | |||
| arange: Callable[[Optional[NumericType], NumericType, Optional[NumericType]], TensorType] = None | |||
| zeros: Callable[["tuple[int]"], TensorType] = None | |||
| zeros: Callable[["tuple[int,...]"], TensorType] = None | |||
There was a problem hiding this comment.
Very nice! I had wrong assumptions about how to type subscript tuples. Learned something again.
frehe
commented
Dec 14, 2022
|
|
||
| class TensorFlowLibrary(ComputationLibrary): | ||
| lib = 'TF' | ||
| reshape = tf.reshape | ||
| permute = tf.transpose | ||
| newaxis = tf.newaxis | ||
| shape = lambda x: x.get_shape() # .as_list() | ||
| shape = tf.shape # tobi does not understand reason for this previous definition: # lambda x: x.get_shape() # .as_list() |
There was a problem hiding this comment.
No difference: If x is a tf.Tensor, then x.shape and x.get_shape() are identical https://www.tensorflow.org/api_docs/python/tf/Tensor#get_shape
Using tf.shape(x) seems to be the best choice here, which is a third option.
From https://www.tensorflow.org/api_docs/python/tf/shape:
tf.shape and Tensor.shape should be identical in eager mode. Within tf.function or within a compat.v1 context, not all dimensions may be known until execution time. Hence, when defining custom layers and models for graph mode, prefer the dynamic tf.shape(x) over the static x.shape.
|
I updated this branch with all recent changes of |
# Conflicts: # src/SI_Toolkit/computation_library.py
added prefs to cartpole gui for initial position and angle increased num_rollouts to 1000, decreased horizon to 25 time steps, changed cost function for spin to pure cart angle speed. Now spin, balance, and shimmy all work quite well. moved cartpole_trajectory_generator back out outside compiled TF code for ease of development and debugging, it is cheap so this is OK. added more logging of what gets compiled. added time to arguments for the base controller so that contained methods can access current time. added ruamel.yaml to requirements (this is the yaml parser that can handle scientific notation numbers)
# Conflicts: # Control_Toolkit_ASF_Template/Cost_Functions/EnvironmentName/cost_function_barebone.py # Control_Toolkit_ASF_Template/config_optimizers.yml # Controllers/__init__.py # Controllers/controller_neural_imitator_tf.py # Cost_Functions/__init__.py # Cost_Functions/cost_function_wrapper.py # Optimizers/optimizer_mppi.py # Optimizers/optimizer_rpgd_me_tf.py # Optimizers/optimizer_rpgd_ml_tf.py # Optimizers/optimizer_rpgd_particle_tf.py # Optimizers/optimizer_rpgd_tf.py # requirements.txt
# Conflicts: # src/SI_Toolkit/Predictors/predictor_wrapper.py
Started merging from remote/origin/Tobi_Dance
…cal-cartpole repo. Fix some cartpole parameter import. Fix rename of config_cost_functions.yml. Add timer.py and yes_or_no.py utility classes.
… seem to work. mppi parameters copied from development. cost weights adjusted. controller update interval inccreased to 25ms from 20ms to match simulator and actual achieved rate with dancer mppi. mppi rollouts reduced to 700 to speed up control. PhysicalCartPoleDriver.py has class self variable for reference from other classes to the instance. Compile.py raises exception when either option is undefined.
… for next timestep and recording this prediction along with the measurement in the PhysicalCartPoleDriver.py CSV output file. In this commit, control is enabled by default because the pycharm debugger hangs on interactive terminal keyboard intput, and the motor output is disabled for remote debugging.
commented out some noisy log.debug. made horizon same as physical cartpole
updated satisfaction CSV to use markers from audition and added more cartsheels
…or for compatibilty with RPGD gradient computation. update_attributes in globals_and_utils.py supports assignment of tf.Variable and now raises exception if we try to .assign an immutable Tensor. Balance step now uses int +1 or -1 for desired pole up or down, not 'up' or 'down'. computation_library.py now casts variale to numpy() from tensorflow only if it is a Tensor type.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.