a1a9bd6174f23d0f15b79539d716dcb613fcf74e
7 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
John Schulman
|
07cbf1e26a |
Grad clipping in MpiAdamOptimizer, transformer changes (#304)
* transformer mnist experiments * version that only builds one model * work on inverted mnist * Add grad clipping to MpiAdamOptimizer * various * transformer changes, loading * get rid of soft labels * transformer baseline * minor * experiments involving all possible training sets * vary training * minor * get ready for fine-tuning expers * lint * minor |
||
Karl Cobbe
|
5082e5d34b |
Workbench (#303)
* begin workbench * cleanup * begin procgen config integration * arg tweaks * more args * parameter saving * begin procgen enjoy * tweaks * more workbench * more args sync/restore * cleanup * merge in master * rework args priority * more workbench * more loggign * impala cnn * impala lstm * tweak * tweaks * rl19 time logging * misc fixes * faster pipeline * update local.py * sess and log config tweaks * num processes * logging tweaks * difficulty reward wrapper * logging fixes * gin tweaks * tweak * fix * task id * param loading * more variable loading * entrypoint * tweak * ksync * restore lstm * begin rl19 support * tweak * rl19 rnn * more rl19 integration * fix * cleanup * restore rl19 rnn * cleanup * cleanup * wrappers.get_log_info * cleanup * cleanup * directory cleanup * logging, num_experiments * fixes * cleanup * gin fixes * fix local max gpu * resid nx * num machines and download params * rename * cleanup * create workbench * more reorg * fix * more logging wrappers * lint fix * restore train procgen * restore train procgen * pylint fix * better wrapping * config sweep * args sweep * test workers * mpi_weight * train test comm and high difficulty fix * enjoy show returns * removing gin, procgen_parser * removing gin * procgen args * config fixes * cleanup * cleanup * procgen args fix * fix * rcall syncing * lint * rename mpi_weight * use username for sync * fixes * microbatch fix |
||
Peter Zhokhov
|
5bc6f53960 | merged master | ||
pzhokhov
|
74101a9f24 |
fix freeze of ppo2 (#849)
* fix freeze of ppo2 * unit test for freeze, updated docstring * more docstring update * set number of threads to 1 in the test |
||
|
pzhokhov
|
b875fb7b5e |
release Internal changes (#800)
* joshim5 changes (width and height to WarpFrame wrapper) * match network output with action distribution via a linear layer only if necessary (#167) * support color vs. grayscale option in WarpFrame wrapper (#166) * support color vs. grayscale option in WarpFrame wrapper * Support color in other wrappers * Updated per Peters suggestions * fixing test failures * ppo2 with microbatches (#168) * pass microbatch_size to the model during construction * microbatch fixes and test (#169) * microbatch fixes and test * tiny cleanup * added assertions to the test * vpg-related fix * Peterz joshim5 subclass ppo2 model (#170) * microbatch fixes and test * tiny cleanup * added assertions to the test * vpg-related fix * subclassing the model to make microbatched version of model WIP * made microbatched model a subclass of ppo2 Model * flake8 complaint * mpi-less ppo2 (resolving merge conflict) * flake8 and mpi4py imports in ppo2/model.py * more un-mpying * merge master * updates to the benchmark viewer code + autopep8 (#184) * viz docs and syntactic sugar wip * update viewer yaml to use persistent volume claims * move plot_util to baselines.common, update links * use 1Tb hard drive for results viewer * small updates to benchmark vizualizer code * autopep8 * autopep8 * any folder can be a benchmark * massage games image a little bit * fixed --preload option in app.py * remove preload from run_viewer.sh * remove pdb breakpoints * update bench-viewer.yaml * fixed bug (#185) * fixed bug it's wrong to do the else statement, because no other nodes would start. * changed the fix slightly * Refactor her phase 1 (#194) * add monitor to the rollout envs in her RUN BENCHMARKS her * Slice -> Slide in her benchmarks RUN BENCHMARKS her * run her benchmark for 200 epochs * dummy commit to RUN BENCHMARKS her * her benchmark for 500 epochs RUN BENCHMARKS her * add num_timesteps to her benchmark to be compatible with viewer RUN BENCHMARKS her * add num_timesteps to her benchmark to be compatible with viewer RUN BENCHMARKS her * add num_timesteps to her benchmark to be compatible with viewer RUN BENCHMARKS her * disable saving of policies in her benchmark RUN BENCHMARKS her * run fetch benchmarks with ppo2 and ddpg RUN BENCHMARKS Fetch * run fetch benchmarks with ppo2 and ddpg RUN BENCHMARKS Fetch * launcher refactor wip * wip * her works on FetchReach * her runner refactor RUN BENCHMARKS Fetch1M * unit test for her * fixing warnings in mpi_average in her, skip test_fetchreach if mujoco is not present * pickle-based serialization in her * remove extra import from subproc_vec_env.py * investigating differences in rollout.py * try with old rollout code RUN BENCHMARKS her * temporarily use DummyVecEnv in cmd_util.py RUN BENCHMARKS her * dummy commit to RUN BENCHMARKS her * set info_values in rollout worker in her RUN BENCHMARKS her * bug in rollout_new.py RUN BENCHMARKS her * fixed bug in rollout_new.py RUN BENCHMARKS her * do not use last step because vecenv calls reset and returns obs after reset RUN BENCHMARKS her * updated buffer sizes RUN BENCHMARKS her * fixed loading/saving via joblib * dust off learning from demonstrations in HER, docs, refactor * add deprecation notice on her play and plot files * address comments by Matthias * 1.5 months of codegen changes (#196) * play with resnet * feed_dict version * coinrun prob and more stats * fixes to get_choices_specs & hp search * minor prob fixes * minor fixes * minor * alternative version of rl_algo stuff * pylint fixes * fix bugs, move node_filters to soup * changed how get_algo works * change how get_algo works, probably broke all tests * continue previous refactor * get eval_agent running again * fixing tests * fix tests * fix more tests * clean up cma stuff * fix experiment * minor changes to eval_agent to make ppo_metal use gpu * make dict space work * modify mac makefile to use conda * recurrent layers * play with bn and resnets * minor hp changes * minor * got rid of use_fb argument and jtft (joint-train-fine-tune) functionality built test phase directly into AlgoProb * make new rl algos generateable * pylint; start fixing tests * fixing tests * more test fixes * pylint * fix search * work on search * hack around infinite loop caused by scan * algo search fixes * misc changes for search expt * enable annealing, overriding options of Op * pylint fixes * identity op * achieve use_last_output through masking so it automatically works in other distributions * fix tests * minor * discrete * use_last_output to be just a preference, not a hard constraint * pred delay, pruning * require nontrivial inputs * aliases for get_sm * add probname to probs * fixes * small fixes * fix tests * fix tests * fix tests * minor * test scripts * dualgru network improvements * minor * work on mysterious bugs * rcall gpu-usage command for kube * use cache dir that’s not in code folder, so that it doesn’t get removed by rcall code rsync * add power mode to gpu usage * make sure train/test actually different * remove VR for now * minor fixes * simplify soln_db * minor * big refactor of mpi eda * improve mpieda for multitask * - get rid of timelimit hack - add __del__ to cleanup SubprocVecEnv * get multitask working better * fixes * working on atari, various * annotate ops with whether they’re parametrized * minor * gym version * rand atari prob * minor * SolnDb bugfix and name change * pyspy script * switch conv layers * fix roboschool/bullet3 * nenvs assertion * fix rand atari * get rid of blanket exception catching fix soln_db bug * fix rand_atari * dynamic routing as cmdline arg * slight modifications to test_mpi_map and pyspy-all * max_tries argument for run_until_successs * dedup option in train_mle * simplify soln_db * increase atari horizon for 1 experiment * start implementing reward increment * ent multiplier * create cc dsl other misc fixes * cc ops * q_func -> qs in rl_algos_cc.py * fix PredictDistr * rl_ops_cc fixes, MakeAction op * augment algo agent to support cc stuff * work on ddpg experiments * fix blocking temporarily change logger * allow layer scaling * pylint fixes * spawn_method * isolate ddpg hacks * improve pruning * use spawn for subproc * remove use of python -c in rcall * fix pylint warning * fix static * maybe fix local backend * switch to DummyVecEnv * making some fixes via pylint * pylint fixes * fixing tests * fix tests * fix tests * write scaffolding for SSL in Codegen * logger fix * fix error * add EMA op to sl_ops * save many changes * save * add upsampler * add sl ops, enhance state machine * get ssl search working — some gross hacking * fix session/graph issue * fix importing * work on mle * - scale embeddings in gru model - better exception handling in sl_prob - use emas for test/val - use non-contrib batch_norm layer * improve logging * option to average before dumping in logger * default arguments, etc * new ddpg and identity test * concat fix * minor * move realistic ssl stuff to third-party (underscore to dash) * fixes * remove realistic_ssl_evaluation * pylint fixes * use gym master * try again * pass around args without gin * fix tests * separate line to install gym * rename failing tests that should be ignored * add data aug * ssl improvements * use fixed time limit * try to fix baselines tests * add score_floor, max_walltime, fiddle with lr decay * realistic_ssl * autopep8 * various ssl - enable blocking grad for simplification - kl - multiple final prediction * fix pruning * misc ssl stuff * bring back linear schedule, don’t use allgather for collecting stats (i’ve been getting nondeterministic errors from the old code) * save/load weights in SSL, big stepsize * cleanup SslProb * fix * get rid of kl coef * fix simplification, lower lr * search over hps * minor fixes * minor * static analysis * move files and rename things for improved consistency. still broken, and just saving before making nontrivial changes * various * make tests pass * move coinrun_train to codegen since it depends on codegen * fixes * pylint fixes * improve tests fix some things * improve tests * lint * fix up db_info.py, tests * mostly restore master version of envs directory, except for makefile changes * fix tests * improve printing * minor fixes * fix fixmes * pruning test * fixes * lint * write new test that makes tf graphs of random algos; fix some bugs it caught * add —delete flag to rcall upload-code command * lint * get cifar10 lazily for testing purposes * disable codegen ci tests for now * clean up rl_ops * rename spec classes * td3 with identity test * identity tests without gin files * remove gin.configurable from AlgoAgent * comments about reduction in rl_ops_cc * address @pzhokhov comments * fix tests * more linting * better tests * clean up filtering a bit * fix concat * delayed logger configuration (#208) * delayed logger configuration * fix typo * setters and getters for Logger.DEFAULT as well * do away with fancy property stuff - unable to get it to work with class level methods * grammar and spaces * spaces * use get_current function instead of reading Logger.CURRENT * autopep8 * disable mpi in subprocesses (#213) * lazy_mpi load * cleanups * more lazy mpi * don't pretend that class is a module, just use it as a class * mass-replace mpi4py imports * flake8 * fix previous lazy_mpi imports * silly recursion * try os.environ hack * better prefix test, work with mpich * restored MPI imports * removed commented import in test_with_mpi * restored codegen from master * remove lazy mpi * restored changes from rl-algs * remove extra files * address Chris' comments * use spawn for shmem vec env as well (#2) (#219) * lazy_mpi load * cleanups * more lazy mpi * don't pretend that class is a module, just use it as a class * mass-replace mpi4py imports * flake8 * fix previous lazy_mpi imports * silly recursion * try os.environ hack * better prefix test, work with mpich * restored MPI imports * removed commented import in test_with_mpi * restored codegen from master * remove lazy mpi * restored changes from rl-algs * remove extra files * port mpi fix to shmem vec env * increase the mpi test default timeout * change humanoid hyperparameters, get rid of clip_Frac annealing, as it's apparently dangerous * remove clip_frac schedule from ppo2 * more timesteps in humanoid run * whitespace + RUN BENCHMARKS * baselines: export vecenvs from folder (#221) * baselines: export vecenvs from folder * put missing function back in * add missing imports * more imports * longer mpi timeout? * make default logger configuration the same as call to logger.configure() (#222) * Vecenv refactor (#223) * update karl util * restore pvi flag * change rcall auto cpu behavior, move gin.configurable, add os.makedirs * vecenv refactor * aux buf index fix * add num aux obs * reset level with enter * restore high difficulty flag * bugfix * restore train_coinrun.py * tweaks * renaming * renaming * better arguments handling * more options * options cleanup * game data refactor * more options * args for train_procgen * add close handler to interactive base class * use debug build if debug=True, fix range on aux_obs * add ProcGenEnv to __init__.py, add missing imports to procgen.py * export RemoveDictWrapper and build, update train_procgen.py, move assets download into env creation and replace init_assets_and_build with just build * fix formatting issues * only call global init once * fix path in setup.py * revert part of makefile * ignore IDE files and folders * vec remove dict * export VecRemoveDictObs * remove RemoveDictWrapper * remove IDE files * move shared .h and .cpp files to common folder, update build to use those, dedupe env.cpp * fix missing header * try unified build function * remove old scripts dir * add comment on build * upload libenv with render fixes * tell qthreads to die when we unload the library * pyglet.app.run is garbage * static fixes * whoops * actually vsync is on * cleanup * cleanup * extern C for libenv interface * parse util rcall arg * high difficulty fix * game type enums * ProcGenEnv subclasses * game type cleanup * unrecognized key * unrecognized game type * parse util reorg * args management * typo fix * GinParser * arg tweaks * tweak * restore start_level/num_levels setting * fix create_procgen_env interface * build fix * procgen args in init signature * fix * build fix * fix logger usage in ppo_metal/run_retro * removed unnecessary OrderedDict requirement in subproc_vec_env * flake8 fix * allow for non-mpi tests * mpi test fixes * flake8; removed special logic for discrete spaces in dummy_vec_env * remove forked argument in front of tests - does not play nicely with subprocvecenv in spawned processes; analog of forked in ddpg/test_smoke * Everyrl initial commit & a few minor baselines changes (#226) * everyrl initial commit * add keep_buf argument to VecMonitor * logger changes: set_comm and fix to mpi_mean functionality * if filename not provided, don't create ResultsWriter * change variable syncing function to simplify its usage. now you should initialize from all mpi processes * everyrl coinrun changes * tf_distr changes, bugfix * get_one * bring back get_next to temporarily restore code * lint fixes * fix test * rename profile function * rename gaussian * fix coinrun training script * change random seeding to work with new gym version (#231) * change random seeding to work with new gym version * move seeding to seed() method * fix mnistenv * actually try some of the tests before pushing * more deterministic fixed seq * misc changes to vecenvs and run.py for benchmarks (#236) * misc changes to vecenvs and run.py for benchmarks * dont seed global gen * update more references to assert_venvs_equal * Rl19 (#232) * everyrl initial commit * add keep_buf argument to VecMonitor * logger changes: set_comm and fix to mpi_mean functionality * if filename not provided, don't create ResultsWriter * change variable syncing function to simplify its usage. now you should initialize from all mpi processes * everyrl coinrun changes * tf_distr changes, bugfix * get_one * bring back get_next to temporarily restore code * lint fixes * fix test * rename profile function * rename gaussian * fix coinrun training script * rl19 * remove everyrl dir which appeared in the merge for some reason * readme * fiddle with ddpg * make ddpg work * steps_total argument * gpu count * clean up hyperparams and shape math * logging + saving * configuration stuff * fixes, smoke tests * fix stats * make load_results return dicts -- easier to create the same kind of objects with some other mechanism for passing to downstream functions * benchmarks * fix tests * add dqn to tests, fix it * minor * turned annotated transformer (pytorch) into a script * more refactoring * jax stuff * cluster * minor * copy & paste alec code * sign error * add huber, rename some parameters, snapshotting off by default * remove jax stuff * minor * move maze env * minor * remove trailing spaces * remove trailing space * lint * fix test breakage due to gym update * rename function * move maze back to codegen * get recurrent ppo working * enable both lstm and gru * script to print table of benchmark results * various * fix dqn * add fixup initializer, remove lastrew * organize logging stats * fix silly bug * refactor models * fix mpi usage * check sync * minor * change vf coef, hps * clean up slicing in ppo * minor fixes * caching transformer * docstrings * xf fixes * get rid of 'B' and 'BT' arguments * minor * transformer example * remove output_kind from base class until we have a better idea how to use it * add comments, revert maze stuff * flake8 * codegen lint * fix codegen tests * responded to peter's comments * lint fixes * minor changes to baselines (#243) * minor changes to baselines * fix spaces reference * remove flake8 disable comments and fix import * okay maybe don't add spec to vec_env * Merge branch 'master' of github.com:openai/games the commit. * flake8 complaints in baselines/her |
||
John Schulman
|
fb6fd51fe6 |
Rl19 (#232)
* everyrl initial commit * add keep_buf argument to VecMonitor * logger changes: set_comm and fix to mpi_mean functionality * if filename not provided, don't create ResultsWriter * change variable syncing function to simplify its usage. now you should initialize from all mpi processes * everyrl coinrun changes * tf_distr changes, bugfix * get_one * bring back get_next to temporarily restore code * lint fixes * fix test * rename profile function * rename gaussian * fix coinrun training script * rl19 * remove everyrl dir which appeared in the merge for some reason * readme * fiddle with ddpg * make ddpg work * steps_total argument * gpu count * clean up hyperparams and shape math * logging + saving * configuration stuff * fixes, smoke tests * fix stats * make load_results return dicts -- easier to create the same kind of objects with some other mechanism for passing to downstream functions * benchmarks * fix tests * add dqn to tests, fix it * minor * turned annotated transformer (pytorch) into a script * more refactoring * jax stuff * cluster * minor * copy & paste alec code * sign error * add huber, rename some parameters, snapshotting off by default * remove jax stuff * minor * move maze env * minor * remove trailing spaces * remove trailing space * lint * fix test breakage due to gym update * rename function * move maze back to codegen * get recurrent ppo working * enable both lstm and gru * script to print table of benchmark results * various * fix dqn * add fixup initializer, remove lastrew * organize logging stats * fix silly bug * refactor models * fix mpi usage * check sync * minor * change vf coef, hps * clean up slicing in ppo * minor fixes * caching transformer * docstrings * xf fixes * get rid of 'B' and 'BT' arguments * minor * transformer example * remove output_kind from base class until we have a better idea how to use it * add comments, revert maze stuff * flake8 * codegen lint * fix codegen tests * responded to peter's comments * lint fixes |
||
|
pzhokhov
|
8c2aea2add |
refactor a2c, acer, acktr, ppo2, deepq, and trpo_mpi (#490)
* exported rl-algs * more stuff from rl-algs * run slow tests * re-exported rl_algs * re-exported rl_algs - fixed problems with serialization test and test_cartpole * replaced atari_arg_parser with common_arg_parser * run.py can run algos from both baselines and rl_algs * added approximate humanoid reward with ppo2 into the README for reference * dummy commit to RUN BENCHMARKS * dummy commit to RUN BENCHMARKS * dummy commit to RUN BENCHMARKS * dummy commit to RUN BENCHMARKS * very dummy commit to RUN BENCHMARKS * serialize variables as a dict, not as a list * running_mean_std uses tensorflow variables * fixed import in vec_normalize * dummy commit to RUN BENCHMARKS * dummy commit to RUN BENCHMARKS * flake8 complaints * save all variables to make sure we save the vec_normalize normalization * benchmarks on ppo2 only RUN BENCHMARKS * make_atari_env compatible with mpi * run ppo_mpi benchmarks only RUN BENCHMARKS * hardcode names of retro environments * add defaults * changed default ppo2 lr schedule to linear RUN BENCHMARKS * non-tf normalization benchmark RUN BENCHMARKS * use ncpu=1 for mujoco sessions - gives a bit of a performance speedup * reverted running_mean_std to user property decorators for mean, var, count * reverted VecNormalize to use RunningMeanStd (no tf) * reverted VecNormalize to use RunningMeanStd (no tf) * profiling wip * use VecNormalize with regular RunningMeanStd * added acer runner (missing import) * flake8 complaints * added a note in README about TfRunningMeanStd and serialization of VecNormalize * dummy commit to RUN BENCHMARKS * merged benchmarks branch |