Add scripts and configs for hyperparameter tuning (HumanCompatibleAI#675

)

* Merge py file changes from benchmark-algs

* Clean parallel script

* Undo the changes from HumanCompatibleAI#653 to the dagger benchmark config files.

This change just made some error messages go away indicating the missing imitation.algorithms.dagger.ExponentialBetaSchedule but it did not fix the root cause.

* Improve readability and interpretability of benchmarking tests.

* Add pxponential beta scheduler for dagger

* Ignore coverage for unknown algorithms.

* Cleanup and extend tests for beta schedules in dagger.

* Add optuna to dependencies

* Fix test case

* Clean up the scripts

* Remove reporter(done) since mean_return is reported by the runs

* Add beta_schedule parameter to dagger script

* Update config policy kwargs

* Changes from review

* Fix errors with some configs

* Updates based on review

* Change metric everywhere

* Separate tuning code from parallel.py

* Fix docstring

* Removing resume option as it is getting tricky to correctly implement

* Minor fixes

* Updates from review

* fix lint error

* Add documentation for using the tuning script

* Fix lint error

* Updates from the review

* Fix file name test errors

* Add tune_run_kwargs in parallel script

* Fix test errors

* Fix test

* Fix lint

* Updates from review

* Simplify few lines of code

* Updates from review

* Fix test

* Revert "Fix test"

This reverts commit 8b55134.

* Fix test

* Convert Dict to Mapping in input argument

* Ignore coverage in script configurations.

* Pin huggingface_sb3 version.

* Update to the newest seals environment versions.

* Push gymnasium dependency to 0.29 to ensure mujoco envs work.

* Incorporate review comments

* Fix test errors

* Move benchmarking/ to scripts/ and add named configs for tuned hyperparams

* Bump cache version & remove unnecessary files

* Include tuned hyperparam json files in package data

* Update storage hash

* Update search space of bc

* update benchmark and hyper parameter tuning readme

* Update README.md

* Incorporate reviewer's comments in benchmarking readme

* Update gymnasium version and render mode in eval policy

* Fix error

* Update commands.py hex strings

---------

Co-authored-by: Maximilian Ernestus <[email protected]>
Co-authored-by: ZiyueWang25 <[email protected]>

benchmarking/README.md

@@ -1,19 +1,42 @@
 # Benchmarking imitation
 
-This directory contains sacred configuration files for benchmarking imitation's algorithms. For v0.3.2, these correspond to the hyperparameters used in the paper [imitation: Clean Imitation Learning Implementations](https://www.rocamonde.com/publication/gleave-imitation-2022/).
+The `src/imitation/scripts/config/tuned_hps` directory provides the tuned hyperparameter configs for benchmarking imitation. For v0.4.0, these correspond to the hyperparameters used in the paper [imitation: Clean Imitation Learning Implementations](https://arxiv.org/abs/2211.11972).
 
-Configuration files can be loaded either from the CLI or from the Python API. The examples below assume that your current working directory is the root of the `imitation` repository. This is not necessarily the case and you should adjust your paths accordingly.
+Configuration files can be loaded either from the CLI or from the Python API.
 
 ## CLI
 
 ```bash
-python -m imitation.scripts.<train_script> <algo> with benchmarking/<config_name>.json
+python -m imitation.scripts.<train_script> <algo> with <algo>_<env>
 ```
-`train_script` can be either 1) `train_imitation` with `algo` as `bc` or `dagger` or 2) `train_adversarial`  with `algo` as `gail` or `airl`.
+`train_script` can be either 1) `train_imitation` with `algo` as `bc` or `dagger` or 2) `train_adversarial`  with `algo` as `gail` or `airl`. The `env` can be either of `seals_ant`, `seals_half_cheetah`, `seals_hopper`, `seals_swimmer`, or `seals_walker`. The hyperparameters for other environments are not tuned yet. You may be able to get reasonable performance by using hyperparameters tuned for a similar environment; alternatively, you can tune the hyperparameters using the `tuning` script.
 
 ## Python
 
 ```python
 ...
-ex.add_config('benchmarking/<config_name>.json')
+from imitation.scripts.<train_script> import <train_ex>
+<train_ex>.run(command_name="<algo>", named_configs=["<algo>_<env>"])
 ```
+
+# Tuning Hyperparameters
+
+The hyperparameters of any algorithm in imitation can be tuned using `src/imitation/scripts/tuning.py`.
+The benchmarking hyperparameter configs were generated by tuning the hyperparameters using
+the search space defined in the `scripts/config/tuning.py`.
+
+The tuning script proceeds in two phases:
+1. Tune the hyperparameters using the search space provided.
+2. Re-evaluate the best hyperparameter config found in the first phase based on the maximum mean return on a separate set of seeds. Report the mean and standard deviation of these trials.
+
+To use it with the default search space:
+```bash
+python -m imitation.scripts.tuning with <algo> 'parallel_run_config.base_named_configs=["<env>"]'
+```
+
+In this command:
+- `<algo>` provides the default search space and settings for the specific algorithm, which is defined in the `scripts/config/tuning.py`
+- `<env>` sets the environment to tune the algorithm in. They are defined in the algo-specifc `scripts/config/train_[adversarial|imitation|preference_comparisons|rl].py` files. For the already tuned environments, use the `<algo>_<env>` named configs here.
+
+See the documentation of `scripts/tuning.py` and `scripts/parallel.py` for many other arguments that can be
+provided through the command line to change the tuning behavior.

benchmarking/util.py

@@ -79,7 +79,7 @@ def clean_config_file(file: pathlib.Path, write_path: pathlib.Path, /) -> None:
 
     remove_empty_dicts(config)
     # files are of the format
-    # /path/to/file/example_<algo>_<env>_best_hp_eval/<other_info>/sacred/1/config.json
+    # /path/to/file/<algo>_<env>_best_hp_eval/<other_info>/sacred/1/config.json
     # we want to write to /<write_path>/<algo>_<env>.json
     with open(write_path / f"{file.parents[3].name}.json", "w") as f:
         json.dump(config, f, indent=4)

experiments/commands.py

@@ -12,23 +12,24 @@
 
 For example, we can run:
 
+TUNED_HPS_DIR=../src/imitation/scripts/config/tuned_hps
 python commands.py \
     --name=run0 \
-    --cfg_pattern=../benchmarking/*ai*_seals_walker_*.json \
+    --cfg_pattern=$TUNED_HPS_DIR/*ai*_seals_walker_*.json \
     --output_dir=output
 
 And get the following commands printed out:
 
 python -m imitation.scripts.train_adversarial airl \
     --capture=sys --name=run0 \
     --file_storage=output/sacred/$USER-cmd-run0-airl-0-a3531726 \
-    with ../benchmarking/example_airl_seals_walker_best_hp_eval.json \
+    with ../src/imitation/scripts/config/tuned_hps/airl_seals_walker_best_hp_eval.json \
     seed=0 logging.log_root=output
 
 python -m imitation.scripts.train_adversarial gail \
     --capture=sys --name=run0 \
     --file_storage=output/sacred/$USER-cmd-run0-gail-0-a1ec171b \
-    with ../benchmarking/example_gail_seals_walker_best_hp_eval.json \
+    with $TUNED_HPS_DIR/gail_seals_walker_best_hp_eval.json \
     seed=0 logging.log_root=output
 
 We can execute commands in parallel by piping them to GNU parallel:
@@ -40,9 +41,10 @@
 
 For example, we can run:
 
+TUNED_HPS_DIR=../src/imitation/scripts/config/tuned_hps
 python commands.py \
     --name=run0 \
-    --cfg_pattern=../benchmarking/example_bc_seals_half_cheetah_best_hp_eval.json \
+    --cfg_pattern=$TUNED_HPS_DIR/bc_seals_half_cheetah_best_hp_eval.json \
     --output_dir=/data/output \
     --remote
 
@@ -51,8 +53,9 @@
 ctl job run --name $USER-cmd-run0-bc-0-72cb1df3 \
     --command "python -m imitation.scripts.train_imitation bc \
     --capture=sys --name=run0 \
-    --file_storage=/data/output/sacred/$USER-cmd-run0-bc-0-72cb1df3 \
-    with /data/imitation/benchmarking/example_bc_seals_half_cheetah_best_hp_eval.json \
+    --file_storage=/data/output/sacred/$USER-cmd-run0-bc-0-72cb1df3 with \
+    /data/imitation/src/imitation/scripts/config/tuned_hps/
+    bc_seals_half_cheetah_best_hp_eval.json \
     seed=0 logging.log_root=/data/output" \
     --container hacobe/devbox:imitation \
     --login --force-pull --never-restart --gpu 0 --shared-host-dir-mount /data
@@ -85,7 +88,7 @@ def _get_algo_name(cfg_file: str) -> str:
     """Get the algorithm name from the given config filename."""
     algo_names = set()
     for key in _ALGO_NAME_TO_SCRIPT_NAME:
-        if cfg_file.find("_" + key + "_") != -1:
+        if cfg_file.find(key + "_") != -1:
             algo_names.add(key)
 
     if len(algo_names) == 0:
@@ -121,7 +124,7 @@ def main(args: argparse.Namespace) -> None:
         else:
             cfg_path = os.path.join(args.remote_cfg_dir, cfg_file)
 
-        cfg_id = _get_cfg_id(cfg_path)
+        cfg_id = _get_cfg_id(cfg_file)
 
         for seed in args.seeds:
             cmd_id = _CMD_ID_TEMPLATE.format(
@@ -177,19 +180,19 @@ def parse() -> argparse.Namespace:
     parser.add_argument(
         "--cfg_pattern",
         type=str,
-        default="example_bc_seals_half_cheetah_best_hp_eval.json",
+        default="bc_seals_half_cheetah_best_hp_eval.json",
         help="""Generate a command for every file that matches this glob pattern. \
 Each matching file should be a config file that has its algorithm name \
 (bc, dagger, airl or gail) bookended by underscores in the filename. \
 If the --remote flag is enabled, then generate a command for every file in the \
 --remote_cfg_dir directory that has the same filename as a file that matches this \
 glob pattern. E.g., suppose the current, local working directory is 'foo' and \
-the subdirectory 'foo/bar' contains the config files 'example_bc_best.json' and \
-'example_dagger_best.json'. If the pattern 'bar/*.json' is supplied, then globbing \
-will return ['bar/example_bc_best.json', 'bar/example_dagger_best.json']. \
+the subdirectory 'foo/bar' contains the config files 'bc_best.json' and \
+'dagger_best.json'. If the pattern 'bar/*.json' is supplied, then globbing \
+will return ['bar/bc_best.json', 'bar/dagger_best.json']. \
 If the --remote flag is enabled, 'bar' will be replaced with `remote_cfg_dir` and \
 commands will be created for the following configs: \
-[`remote_cfg_dir`/example_bc_best.json, `remote_cfg_dir`/example_dagger_best.json] \
+[`remote_cfg_dir`/bc_best.json, `remote_cfg_dir`/dagger_best.json] \
 Why not just supply the pattern '`remote_cfg_dir`/*.json' directly? \
 Because the `remote_cfg_dir` directory may not exist on the local machine.""",
     )
@@ -220,7 +223,7 @@ def parse() -> argparse.Namespace:
     parser.add_argument(
         "--remote_cfg_dir",
         type=str,
-        default="/data/imitation/benchmarking",
+        default="/data/imitation/src/imitation/scripts/config/tuned_hps",
         help="""Path to a directory storing config files \
 accessible from each container. """,
     )

setup.cfg

@@ -45,6 +45,8 @@ source = imitation
 include=
     src/*
     tests/*
+omit =
+    src/imitation/scripts/config/*
 
 [coverage:report]
 exclude_lines =

setup.py

@@ -182,7 +182,7 @@ def get_local_version(version: "ScmVersion", time_format="%Y%m%d") -> str:
     python_requires=">=3.8.0",
     packages=find_packages("src"),
     package_dir={"": "src"},
-    package_data={"imitation": ["py.typed", "envs/examples/airl_envs/assets/*.xml"]},
+    package_data={"imitation": ["py.typed", "scripts/config/tuned_hps/*.json"]},
     # Note: while we are strict with our test and doc requirement versions, we try to
     #   impose as little restrictions on the install requirements as possible. Try to
     #   encode only known incompatibilities here. This prevents nasty dependency issues
@@ -200,6 +200,7 @@ def get_local_version(version: "ScmVersion", time_format="%Y%m%d") -> str:
         "sacred>=0.8.4",
         "tensorboard>=1.14",
         "huggingface_sb3~=3.0",
+        "optuna>=3.0.1",
         "datasets>=2.8.0",
     ],
     tests_require=TESTS_REQUIRE,

src/imitation/scripts/analyze.py

@@ -262,38 +262,47 @@ def analyze_imitation(
         csv_output_path: If provided, then save a CSV output file to this path.
         tex_output_path: If provided, then save a LaTeX-format table to this path.
         print_table: If True, then print the dataframe to stdout.
-        table_verbosity: Increasing levels of verbosity, from 0 to 2, increase the
-            number of columns in the table.
+        table_verbosity: Increasing levels of verbosity, from 0 to 3, increase the
+            number of columns in the table. Level 3 prints all of the columns available.
 
     Returns:
         The DataFrame generated from the Sacred logs.
     """
-    table_entry_fns_subset = _get_table_entry_fns_subset(table_verbosity)
+    # Get column names for which we have get value using make_entry_fn
+    # These are same across Level 2 & 3. In Level 3, we additionally add remaining
+    #  config columns.
+    table_entry_fns_subset = _get_table_entry_fns_subset(min(table_verbosity, 2))
 
-    rows = []
+    output_table = pd.DataFrame()
     for sd in _gather_sacred_dicts():
-        row = {}
+        if table_verbosity == 3:
+            # gets all config columns
+            row = pd.json_normalize(sd.config)
+        else:
+            # create an empty dataframe with a single row
+            row = pd.DataFrame(index=[0])
+
         for col_name, make_entry_fn in table_entry_fns_subset.items():
             row[col_name] = make_entry_fn(sd)
-        rows.append(row)
 
-    df = pd.DataFrame(rows)
-    if len(df) > 0:
-        df.sort_values(by=["algo", "env_name"], inplace=True)
+        output_table = pd.concat([output_table, row])
+
+    if len(output_table) > 0:
+        output_table.sort_values(by=["algo", "env_name"], inplace=True)
 
-    display_options = dict(index=False)
+    display_options: Mapping[str, Any] = dict(index=False)
     if csv_output_path is not None:
-        df.to_csv(csv_output_path, **display_options)
+        output_table.to_csv(csv_output_path, **display_options)
         print(f"Wrote CSV file to {csv_output_path}")
     if tex_output_path is not None:
-        s: str = df.to_latex(**display_options)
+        s: str = output_table.to_latex(**display_options)
         with open(tex_output_path, "w") as f:
             f.write(s)
         print(f"Wrote TeX file to {tex_output_path}")
 
     if print_table:
-        print(df.to_string(**display_options))
-    return df
+        print(output_table.to_string(**display_options))
+    return output_table
 
 
 def _make_return_summary(stats: dict, prefix="") -> str:

src/imitation/scripts/config/analyze.py

@@ -18,7 +18,7 @@ def config():
     tex_output_path = None  # Write LaTex output to this path
     print_table = True  # Set to True to print analysis to stdout
     split_str = ","  # str used to split source_dir_str into multiple source dirs
-    table_verbosity = 1  # Choose from 0, 1, or 2
+    table_verbosity = 1  # Choose from 0, 1, 2 or 3
     source_dirs = None
 
 

src/imitation/scripts/config/parallel.py

@@ -5,7 +5,10 @@
 `@parallel_ex.named_config` to define a new parallel experiment.
 
 Adding custom named configs is necessary because the CLI interface can't add
-search spaces to the config like `"seed": tune.grid_search([0, 1, 2, 3])`.
+search spaces to the config like `"seed": tune.choice([0, 1, 2, 3])`.
+
+For tuning hyperparameters of an algorithm on a given environment,
+check out the imitation/scripts/tuning.py script.
 """
 
 import numpy as np
@@ -31,19 +34,10 @@ def config():
         "config_updates": {},
     }  # `config` argument to `ray.tune.run(trainable, config)`
 
-    local_dir = None  # `local_dir` arg for `ray.tune.run`
-    upload_dir = None  # `upload_dir` arg for `ray.tune.run`
-    n_seeds = 3  # Number of seeds to search over by default
-
-
-@parallel_ex.config
-def seeds(n_seeds):
-    search_space = {"config_updates": {"seed": tune.grid_search(list(range(n_seeds)))}}
-
-
-@parallel_ex.named_config
-def s3():
-    upload_dir = "s3://shwang-chai/private"
+    num_samples = 1  # Number of samples per grid search configuration
+    repeat = 1  # Number of times to repeat a sampled configuration
+    experiment_checkpoint_path = ""  # Path to checkpoint of experiment
+    tune_run_kwargs = {}  # Additional kwargs to pass to `tune.run`
 
 
 # Debug named configs
@@ -58,12 +52,12 @@ def generate_test_data():
     """
     sacred_ex_name = "train_rl"
     run_name = "TEST"
-    n_seeds = 1
+    repeat = 1
     search_space = {
         "config_updates": {
             "rl": {
                 "rl_kwargs": {
-                    "learning_rate": tune.grid_search(
+                    "learning_rate": tune.choice(
                         [3e-4 * x for x in (1 / 3, 1 / 2)],
                     ),
                 },
@@ -86,63 +80,16 @@ def generate_test_data():
 def example_cartpole_rl():
     sacred_ex_name = "train_rl"
     run_name = "example-cartpole"
-    n_seeds = 2
+    repeat = 2
     search_space = {
         "config_updates": {
             "rl": {
                 "rl_kwargs": {
-                    "learning_rate": tune.grid_search(np.logspace(3e-6, 1e-1, num=3)),
-                    "nminibatches": tune.grid_search([16, 32, 64]),
+                    "learning_rate": tune.choice(np.logspace(3e-6, 1e-1, num=3)),
+                    "nminibatches": tune.choice([16, 32, 64]),
                 },
             },
         },
     }
     base_named_configs = ["cartpole"]
     resources_per_trial = dict(cpu=4)
-
-
-EASY_ENVS = ["cartpole", "pendulum", "mountain_car"]
-
-
-@parallel_ex.named_config
-def example_rl_easy():
-    sacred_ex_name = "train_rl"
-    run_name = "example-rl-easy"
-    n_seeds = 2
-    search_space = {
-        "named_configs": tune.grid_search([[env] for env in EASY_ENVS]),
-        "config_updates": {
-            "rl": {
-                "rl_kwargs": {
-                    "learning_rate": tune.grid_search(np.logspace(3e-6, 1e-1, num=3)),
-                    "nminibatches": tune.grid_search([16, 32, 64]),
-                },
-            },
-        },
-    }
-    resources_per_trial = dict(cpu=4)
-
-
-@parallel_ex.named_config
-def example_gail_easy():
-    sacred_ex_name = "train_adversarial"
-    run_name = "example-gail-easy"
-    n_seeds = 1
-    search_space = {
-        "named_configs": tune.grid_search([[env] for env in EASY_ENVS]),
-        "config_updates": {
-            "init_trainer_kwargs": {
-                "rl": {
-                    "rl_kwargs": {
-                        "learning_rate": tune.grid_search(
-                            np.logspace(3e-6, 1e-1, num=3),
-                        ),
-                        "nminibatches": tune.grid_search([16, 32, 64]),
-                    },
-                },
-            },
-        },
-    }
-    search_space = {
-        "command_name": "gail",
-    }