Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* added daemon=True for multi-process rollout, policy manager and inference * removed obsolete files * [REDO][PR#406]V0.2 rl refinement taskq (#408) * Add a usable task_queue * Rename some variables * 1. Add ; 2. Integrate related files; 3. Remove * merge `data_parallel` and `num_grad_workers` into `data_parallelism` * Fix bugs in docker_compose_yml.py and Simple/Multi-process mode. * Move `grad_worker` into marl/rl/workflows * 1.Merge data_parallel and num_workers into data_parallelism in config; 2.Assign recently used workers as possible in task_queue. * Refine code and update docs of `TaskQueue` * Support priority for tasks in `task_queue` * Update diagram of policy manager and task queue. * Add configurable `single_task_limit` and correct docstring about `data_parallelism` * Fix lint errors in `supply chain` * RL policy redesign (V2) (#405) * Drafi v2.0 for V2 * Polish models with more comments * Polish policies with more comments * Lint * Lint * Add developer doc for models. * Add developer doc for policies. * Remove policy manager V2 since it is not used and out-of-date * Lint * Lint * refined messy workflow code * merged 'scenario_dir' and 'scenario' in rl config * 1. refined env_sampler and agent_wrapper code; 2. added docstrings for env_sampler methods * 1. temporarily renamed RLPolicy from polivy_v2 to RLPolicyV2; 2. merged env_sampler and env_sampler_v2 * merged cim and cim_v2 * lint issue fix * refined logging logic * lint issue fix * reversed unwanted changes * . . . . ReplayMemory & IndexScheduler ReplayMemory & IndexScheduler . MultiReplayMemory get_actions_with_logps EnvSampler on the road EnvSampler Minor * LearnerManager * Use batch to transfer data & add SHAPE_CHECK_FLAG * Rename learner to trainer * Add property for policy._is_exploring * CIM test scenario for V3. Manual test passed. Next step: run it, make it works. * env_sampler.py could run * env_sampler refine on the way * First runnable version done * AC could run, but the result is bad. Need to check the logic * Refine abstract method & shape check error info. * Docs * Very detailed compare. Try again. * AC done * DQN check done * Minor * DDPG, not tested * Minors * A rough draft of MAAC * Cannot use CIM as the multi-agent scenario. * Minor * MAAC refinement on the way * Remove ActionWithAux * Refine batch & memory * MAAC example works * Reproduce-able fix. Policy share between env_sampler and trainer_manager. * Detail refinement * Simplify the user configed workflow * Minor * Refine example codes * Minor polishment * Migrate rollout_manager to V3 * Error on the way * Redesign torch.device management * Rl v3 maddpg (#418) * Add MADDPG trainer * Fit independent critics and shared critic modes. * Add a new property: num_policies * Lint * Fix a bug in `sum(rewards)` * Rename `MADDPG` to `DiscreteMADDPG` and fix type hint. * Rename maddpg in examples. * Preparation for data parallel (#420) * Preparation for data parallel * Minor refinement & lint fix * Lint * Lint * rename atomic_get_batch_grad to get_batch_grad * Fix a unexpected commit * distributed maddpg * Add critic worker * Minor * Data parallel related minorities * Refine code structure for trainers & add more doc strings * Revert a unwanted change * Use TrainWorker to do the actual calculations. * Some minor redesign of the worker's abstraction * Add set/get_policy_state_dict back * Refine set/get_policy_state_dict * Polish policy trainers move train_batch_size to abs trainer delete _train_step_impl() remove _record_impl remove unused methods a minor bug fix in maddpg * Rl v3 data parallel grad worker (#432) * Fit new `trainer_worker` in `grad_worker` and `task_queue`. * Add batch dispatch * Add `tensor_dict` for task submit interface * Move `_remote_learn` to `AbsTrainWorker`. * Complement docstring for task queue and trainer. * Rename train worker to train ops; add placeholder for abstract methods; * Lint Co-authored-by: GQ.Chen <v-guanchen@microsoft.com> * [DRAFT] distributed training pipeline based on RL Toolkit V3 (#450) * Preparation for data parallel * Minor refinement & lint fix * Lint * Lint * rename atomic_get_batch_grad to get_batch_grad * Fix a unexpected commit * distributed maddpg * Add critic worker * Minor * Data parallel related minorities * Refine code structure for trainers & add more doc strings * Revert a unwanted change * Use TrainWorker to do the actual calculations. * Some minor redesign of the worker's abstraction * Add set/get_policy_state_dict back * Refine set/get_policy_state_dict * Polish policy trainers move train_batch_size to abs trainer delete _train_step_impl() remove _record_impl remove unused methods a minor bug fix in maddpg * Rl v3 data parallel grad worker (#432) * Fit new `trainer_worker` in `grad_worker` and `task_queue`. * Add batch dispatch * Add `tensor_dict` for task submit interface * Move `_remote_learn` to `AbsTrainWorker`. * Complement docstring for task queue and trainer. * dsitributed training pipeline draft * added temporary test files for review purposes * Several code style refinements (#451) * Polish rl_v3/utils/ * Polish rl_v3/distributed/ * Polish rl_v3/policy_trainer/abs_trainer.py * fixed merge conflicts * unified sync and async interfaces * refactored rl_v3; refinement in progress * Finish the runnable pipeline under new design * Remove outdated files; refine class names; optimize imports; * Lint * Minor maddpg related refinement * Lint Co-authored-by: Default <huo53926@126.com> Co-authored-by: Huoran Li <huoranli@microsoft.com> Co-authored-by: GQ.Chen <v-guanchen@microsoft.com> Co-authored-by: ysqyang <v-yangqi@microsoft.com> * Miner bug fix * Coroutine-related bug fix ("get_policy_state") (#452) * fixed rebase conflicts * renamed get_policy_func_dict to policy_creator * deleted unwanted folder * removed unwanted changes * resolved PR452 comments Co-authored-by: ysqyang <v-yangqi@microsoft.com> * Quick fix * Redesign experience recording logic (#453) * Two not important fix * Temp draft. Prepare to WFH * Done * Lint * Lint * Calculating advantages / returns (#454) * V1.0 * Complete DDPG * Rl v3 hanging issue fix (#455) * fixed rebase conflicts * renamed get_policy_func_dict to policy_creator * unified worker interfaces * recovered some files * dist training + cli code move * fixed bugs * added retry logic to client * 1. refactored CIM with various algos; 2. lint * lint * added type hint * removed some logs * lint * Make main.py more IDE friendly * Make main.py more IDE friendly * Lint * Final test & format. Ready to merge. Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> Co-authored-by: Huoran Li <huoranli@microsoft.com> * Rl v3 parallel rollout (#457) * fixed rebase conflicts * renamed get_policy_func_dict to policy_creator * unified worker interfaces * recovered some files * dist training + cli code move * fixed bugs * added retry logic to client * 1. refactored CIM with various algos; 2. lint * lint * added type hint * removed some logs * lint * Make main.py more IDE friendly * Make main.py more IDE friendly * Lint * load balancing dispatcher * added parallel rollout * lint * Tracker variable type issue; rename to env_sampler_creator; * Rl v3 parallel rollout follow ups (#458) * AbsWorker & AbsDispatcher * Pass env idx to AbsTrainer.record() method, and let the trainer to decide how to record experiences sampled from different worlds. * Fix policy_creator reuse bug * Format code * Merge AbsTrainerManager & SimpleTrainerManager * AC test passed * Lint * Remove AbsTrainer.build() method. Put all initialization operations into __init__ * Redesign AC preprocess batches logic Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> Co-authored-by: Huoran Li <huoranli@microsoft.com> * MADDPG performance bug fix (#459) * Fix MARL (MADDPG) terminal recording bug; some other minor refinements; * Restore Trainer.build() method * Calculate latest action in the get_actor_grad method in MADDPG. * Share critic bug fix * Rl v3 example update (#461) * updated vm_scheduling example and cim notebook * fixed bugs in vm_scheduling * added local train method * bug fix * modified async client logic to fix hidden issue * reverted to default config * fixed PR comments and some bugs * removed hardcode Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> * Done (#462) * Rl v3 load save (#463) * added load/save feature * fixed some bugs * reverted unwanted changes * lint * fixed PR comments Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> * RL Toolkit data parallelism revamp & config utils (#464) * added load/save feature * fixed some bugs * reverted unwanted changes * lint * fixed PR comments * 1. fixed data parallelism issue; 2. added config validator; 3. refactored cli local * 1. fixed rollout exit issue; 2. refined config * removed config file from example * fixed lint issues * fixed lint issues * added main.py under examples/rl * fixed lint issues Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> * RL doc string (#465) * First rough draft * Minors * Reformat * Lint * Resolve PR comments * Rl type specific env getter (#466) * 1. type-sensitive env variable getter; 2. updated READMEs for examples * fixed bugs * fixed bugs * bug fixes * lint Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> * Example bug fix * Optimize parser.py * Resolve PR comments * Rl config doc (#467) * 1. type-sensitive env variable getter; 2. updated READMEs for examples * added detailed doc * lint * wording refined * resolved some PR comments * resolved more PR comments * typo fix Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: ysqyang <ysqyang@gmail.com> Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> * RL online doc (#469) * Model, policy, trainer * RL workflows and env sampler doc in RST (#468) * First rough draft * Minors * Reformat * Lint * Resolve PR comments * 1. type-sensitive env variable getter; 2. updated READMEs for examples * Rl type specific env getter (#466) * 1. type-sensitive env variable getter; 2. updated READMEs for examples * fixed bugs * fixed bugs * bug fixes * lint Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> * Example bug fix * Optimize parser.py * Resolve PR comments * added detailed doc * lint * wording refined * resolved some PR comments * rewriting rl toolkit rst * resolved more PR comments * typo fix * updated rst Co-authored-by: Huoran Li <huoranli@microsoft.com> Co-authored-by: Default <huo53926@126.com> Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> * Finish docs/source/key_components/rl_toolkit.rst * API doc * RL online doc image fix (#470) * resolved some PR comments * fix * fixed PR comments * added numfig=True setting in conf.py for sphinx Co-authored-by: ysqyang <v-yangqi@microsoft.com> * Resolve PR comments * Add example github link Co-authored-by: ysqyang <ysqyang@gmail.com> Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> * Rl v3 pr comment resolution (#474) * added load/save feature * 1. resolved pr comments; 2. reverted maro/cli/k8s * fixed some bugs Co-authored-by: ysqyang <v-yangqi@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> Co-authored-by: yaqiu <v-yaqiu@microsoft.com> Co-authored-by: GQ.Chen <v-guanchen@microsoft.com> Co-authored-by: ysqyang <ysqyang@gmail.com> Co-authored-by: ysqyang <v-yangqi@microsoft.com>
- Loading branch information
5 people
authored
Mar 7, 2022
1 parent
526627c
commit 696f5b5