- Updated SACPolicy to create critic heads using a list comprehension for better readability.
- Simplified the saving and loading of models using `save_model` and `load_model` functions from the safetensors library.
- Introduced `initialize_offline_replay_buffer` function in the learner server to streamline offline dataset handling and replay buffer initialization.
- Enhanced logging for dataset loading processes to improve traceability during training.
- Introduced `grad_clip_norm` parameter in SAC configuration for gradient clipping
- Updated SACPolicy to store temperature as an instance variable for consistent usage
- Modified loss calculations in SACPolicy to utilize the instance temperature
- Enhanced MLP and CriticHead to support a customizable final activation function
- Implemented gradient clipping in the learner server during training steps for both actor and critic
- Added tracking for gradient norms in training information
- Implement `_save_pretrained` method to handle TensorDict state saving
- Add `_from_pretrained` class method for loading SAC policy from files
- Create utility function `find_and_copy_params` to handle parameter copying
- Removed `@torch.no_grad` decorator from Unnormalize forward method
- Added TODO comment for optimizing next action prediction in SAC policy
- Minor formatting adjustment in NaN assertion for log standard deviation
Co-authored-by: Yoel Chornton <yoel.chornton@gmail.com>
- Introduce `storage_device` parameter in SAC configuration and training settings
- Update learner server to use configurable storage device for replay buffer
- Reduce online buffer capacity in ManiSkill configuration
- Modify replay buffer initialization to support custom storage device
- Introduced Ensemble and CriticHead classes for more efficient critic network handling
- Added support for multiple camera inputs in observation encoder
- Optimized image encoding by batching image processing
- Updated configuration for ManiSkill environment with reduced image size and action scaling
- Compiled critic networks for improved performance
- Simplified normalization and ensemble handling in critic networks
Co-authored-by: michel-aractingi <michel.aractingi@gmail.com>
- Fixed big issue in the normalization of the actions in the `forward` function of the critic -- remove the `torch.no_grad` decorator in `normalize.py` in the normalization function
- Fixed performance issue to boost the optimization frequency by setting the storage device to be the same as the device of learning.
Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>
- Modify logger to support multiple custom step keys
- Update logging method to handle custom step keys more flexibly
- Enhance logging of optimization step and frequency
Co-authored-by: michel-aractingi <michel.aractingi@gmail.com>
- Added Nan detection mechanisms in the actor, learner and gym_manipulator for the case where we encounter nans in the loop.
- changed the non-blocking in the `.to(device)` functions to only work for the case of cuda because they were causing nans when running the policy on mps
- Added some joint clipping and limits in the env, robot and policy configs. TODO clean this part and make the limits in one config file only.
Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>
- Added `lerobot/scripts/server/find_joint_limits.py` to test the min and max angles of the motion you wish the robot to explore during RL training.
- Added logic in `manipulator.py` to limit the maximum possible joint angles to allow motion within a predefined joint position range. The limits are specified in the yaml config for each robot. Checkout the so100.yaml.
Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>
- added `torch.compile` to the actor and learner servers.
- organized imports in `train_sac.py`
- optimized the parameters push by not sending the frozen pre-trained encoder.
Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>
Added functions for converting the replay buffer from and to LeRobotDataset. When we want to save the replay buffer, we convert it first to LeRobotDataset format and save it locally and vice-versa.
Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>
- Optimized critic design that improves the performance of the learner loop by a factor of 2
- Cleaned the code and fixed style issues
- Completed the config with actor_learner_config field that contains host-ip and port elemnts that are necessary for the actor-learner servers.
Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>
- Updated SACConfig to replace standard deviation parameterization with log_std_min and log_std_max for better control over action distributions.
- Modified SACPolicy to streamline action selection and log probability calculations, enhancing stochastic behavior.
- Removed deprecated TanhMultivariateNormalDiag class to simplify the codebase and improve maintainability.
These changes aim to enhance the robustness and performance of the SAC implementation during training and inference.
- Updated standard deviation parameterization in SACConfig to 'softplus' with defined min and max values for improved stability.
- Modified action sampling in SACPolicy to use reparameterized sampling, ensuring better gradient flow and log probability calculations.
- Cleaned up log probability calculations in TanhMultivariateNormalDiag for clarity and efficiency.
- Increased evaluation frequency in YAML configuration to 50000 for more efficient training cycles.
These changes aim to enhance the robustness and performance of the SAC implementation during training and inference.
- Updated action selection to use distribution sampling and log probabilities for better stochastic behavior.
- Enhanced standard deviation clamping to prevent extreme values, ensuring stability in policy outputs.
- Cleaned up code by removing unnecessary comments and improving readability.
These changes aim to refine the SAC implementation, enhancing its robustness and performance during training and inference.
