PPO playing Acrobot-v1 from https://github.com/sgoodfriend/rl-algo-impls/tree/0511de345b17175b7cf1ea706c3e05981f11761c

README.md

@@ -10,7 +10,7 @@ model-index:
   results:
   - metrics:
     - type: mean_reward
-      value: -75.75 +/- 13.96
+      value: -70.5 +/- 9.68
       name: mean_reward
     task:
       type: reinforcement-learning
@@ -23,17 +23,17 @@ model-index:
 
 This is a trained model of a **PPO** agent playing **Acrobot-v1** using the [/sgoodfriend/rl-algo-impls](https://github.com/sgoodfriend/rl-algo-impls) repo.
 
-All models trained at this commit can be found at https://api.wandb.ai/links/sgoodfriend/09frjfcs.
+All models trained at this commit can be found at https://api.wandb.ai/links/sgoodfriend/7lx79bf0.
 
 ## Training Results
 
-This model was trained from 3 trainings of **PPO** agents using different initial seeds. These agents were trained by checking out [2067e21](https://github.com/sgoodfriend/rl-algo-impls/tree/2067e21d62fff5db60168687e7d9e89019a8bfc0). The best and last models were kept from each training. This submission has loaded the best models from each training, reevaluates them, and selects the best model from these latest evaluations (mean - std).
+This model was trained from 3 trainings of **PPO** agents using different initial seeds. These agents were trained by checking out [0511de3](https://github.com/sgoodfriend/rl-algo-impls/tree/0511de345b17175b7cf1ea706c3e05981f11761c). The best and last models were kept from each training. This submission has loaded the best models from each training, reevaluates them, and selects the best model from these latest evaluations (mean - std).
 
 | algo   | env        |   seed |   reward_mean |   reward_std |   eval_episodes | best   | wandb_url                                                                    |
 |:-------|:-----------|-------:|--------------:|-------------:|----------------:|:-------|:-----------------------------------------------------------------------------|
-| ppo    | Acrobot-v1 |      1 |      -75.75   |      13.9575 |              16 | *      | [wandb](https://wandb.ai/sgoodfriend/rl-algo-impls-benchmarks/runs/97ek81ma) |
-| ppo    | Acrobot-v1 |      2 |      -80.5625 |      32.3979 |              16 |        | [wandb](https://wandb.ai/sgoodfriend/rl-algo-impls-benchmarks/runs/zvau4y2s) |
-| ppo    | Acrobot-v1 |      3 |      -77.0625 |      19.2759 |              16 |        | [wandb](https://wandb.ai/sgoodfriend/rl-algo-impls-benchmarks/runs/furggpyg) |
+| ppo    | Acrobot-v1 |      1 |      -75.8125 |     14.2882  |              16 |        | [wandb](https://wandb.ai/sgoodfriend/rl-algo-impls-benchmarks/runs/mup0760u) |
+| ppo    | Acrobot-v1 |      2 |      -88.5    |     46.9281  |              16 |        | [wandb](https://wandb.ai/sgoodfriend/rl-algo-impls-benchmarks/runs/43r8cj6o) |
+| ppo    | Acrobot-v1 |      3 |      -70.5    |      9.68246 |              16 | *      | [wandb](https://wandb.ai/sgoodfriend/rl-algo-impls-benchmarks/runs/z5b5v48f) |
 
 
 ### Prerequisites: Weights & Biases (WandB)
@@ -53,10 +53,10 @@ login`.
 Note: While the model state dictionary and hyperaparameters are saved, the latest
 implementation could be sufficiently different to not be able to reproduce similar
 results. You might need to checkout the commit the agent was trained on:
-[2067e21](https://github.com/sgoodfriend/rl-algo-impls/tree/2067e21d62fff5db60168687e7d9e89019a8bfc0).
+[0511de3](https://github.com/sgoodfriend/rl-algo-impls/tree/0511de345b17175b7cf1ea706c3e05981f11761c).
 ```
 # Downloads the model, sets hyperparameters, and runs agent for 3 episodes
-python enjoy.py --wandb-run-path=sgoodfriend/rl-algo-impls-benchmarks/97ek81ma
+python enjoy.py --wandb-run-path=sgoodfriend/rl-algo-impls-benchmarks/z5b5v48f
 ```
 
 Setup hasn't been completely worked out yet, so you might be best served by using Google
@@ -68,11 +68,11 @@ notebook.
 
 ## Training
 If you want the highest chance to reproduce these results, you'll want to checkout the
-commit the agent was trained on: [2067e21](https://github.com/sgoodfriend/rl-algo-impls/tree/2067e21d62fff5db60168687e7d9e89019a8bfc0). While
+commit the agent was trained on: [0511de3](https://github.com/sgoodfriend/rl-algo-impls/tree/0511de345b17175b7cf1ea706c3e05981f11761c). While
 training is deterministic, different hardware will give different results.
 
 ```
-python train.py --algo ppo --env Acrobot-v1 --seed 1
+python train.py --algo ppo --env Acrobot-v1 --seed 3
 ```
 
 Setup hasn't been completely worked out yet, so you might be best served by using Google
@@ -83,7 +83,7 @@ notebook.
 
 
 ## Benchmarking (with Lambda Labs instance)
-This and other models from https://api.wandb.ai/links/sgoodfriend/09frjfcs were generated by running a script on a Lambda
+This and other models from https://api.wandb.ai/links/sgoodfriend/7lx79bf0 were generated by running a script on a Lambda
 Labs instance. In a Lambda Labs instance terminal:
 ```
 git clone [email protected]:sgoodfriend/rl-algo-impls.git
@@ -105,6 +105,7 @@ can be used. However, this requires a Google Colab Pro+ subscription and running
 This isn't exactly the format of hyperparams in hyperparams/ppo.yml, but instead the Wandb Run Config. However, it's very
 close and has some additional data:
 ```
+additional_keys_to_log: []
 algo: ppo
 algo_hyperparams:
   ent_coef: 0
@@ -121,13 +122,15 @@ env_id: null
 eval_params: {}
 n_timesteps: 1000000
 policy_hyperparams: {}
-seed: 1
+seed: 3
 use_deterministic_algorithms: true
 wandb_entity: null
 wandb_group: null
 wandb_project_name: rl-algo-impls-benchmarks
 wandb_tags:
-- benchmark_2067e21
-- host_155-248-199-228
+- benchmark_0511de3
+- host_152-67-249-42
+- branch_main
+- v0.0.8
 
 ```

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "rl_algo_impls"
-version = "0.0.4"
+version = "0.0.8"
 description = "Implementations of reinforcement learning algorithms"
 authors = [
     {name = "Scott Goodfriend", email = "[email protected]"},
@@ -35,6 +35,7 @@ dependencies = [
     "dash",
     "kaleido",
     "PyYAML",
+    "scikit-learn",
 ]
 
 [tool.setuptools]
@@ -55,10 +56,30 @@ procgen = [
     "glfw >= 1.12.0, < 1.13",
     "procgen; platform_machine=='x86_64'",
 ]
+microrts-old = [
+    "numpy < 1.24.0", # Support for gym-microrts < 0.6.0
+    "gym-microrts == 0.2.0", # Match ppo-implementation-details
+]
+microrts = [
+    "numpy < 1.24.0", # Support for gym-microrts < 0.6.0
+    "gym-microrts == 0.3.2",
+]
+jupyter = [
+    "jupyter",
+    "notebook"
+]
+all = [
+    "rl-algo-impls[test]",
+    "rl-algo-impls[procgen]",
+    "rl-algo-impls[microrts]",
+]
 
 [project.urls]
 "Homepage" = "https://github.com/sgoodfriend/rl-algo-impls"
 
 [build-system]
 requires = ["setuptools==65.5.0", "setuptools-scm"]
-build-backend = "setuptools.build_meta"
+build-backend = "setuptools.build_meta"
+
+[tool.isort]
+profile = "black" 

replay.meta.json

@@ -1 +1 @@
-{"content_type": "video/mp4", "encoder_version": {"backend": "ffmpeg", "version": "b'ffmpeg version 4.2.7-0ubuntu0.1 Copyright (c) 2000-2022 the FFmpeg developers\\nbuilt with gcc 9 (Ubuntu 9.4.0-1ubuntu1~20.04.1)\\nconfiguration: --prefix=/usr --extra-version=0ubuntu0.1 --toolchain=hardened --libdir=/usr/lib/x86_64-linux-gnu --incdir=/usr/include/x86_64-linux-gnu --arch=amd64 --enable-gpl --disable-stripping --enable-avresample --disable-filter=resample --enable-avisynth --enable-gnutls --enable-ladspa --enable-libaom --enable-libass --enable-libbluray --enable-libbs2b --enable-libcaca --enable-libcdio --enable-libcodec2 --enable-libflite --enable-libfontconfig --enable-libfreetype --enable-libfribidi --enable-libgme --enable-libgsm --enable-libjack --enable-libmp3lame --enable-libmysofa --enable-libopenjpeg --enable-libopenmpt --enable-libopus --enable-libpulse --enable-librsvg --enable-librubberband --enable-libshine --enable-libsnappy --enable-libsoxr --enable-libspeex --enable-libssh --enable-libtheora --enable-libtwolame --enable-libvidstab --enable-libvorbis --enable-libvpx --enable-libwavpack --enable-libwebp --enable-libx265 --enable-libxml2 --enable-libxvid --enable-libzmq --enable-libzvbi --enable-lv2 --enable-omx --enable-openal --enable-opencl --enable-opengl --enable-sdl2 --enable-libdc1394 --enable-libdrm --enable-libiec61883 --enable-nvenc --enable-chromaprint --enable-frei0r --enable-libx264 --enable-shared\\nlibavutil      56. 31.100 / 56. 31.100\\nlibavcodec     58. 54.100 / 58. 54.100\\nlibavformat    58. 29.100 / 58. 29.100\\nlibavdevice    58.  8.100 / 58.  8.100\\nlibavfilter     7. 57.100 /  7. 57.100\\nlibavresample   4.  0.  0 /  4.  0.  0\\nlibswscale      5.  5.100 /  5.  5.100\\nlibswresample   3.  5.100 /  3.  5.100\\nlibpostproc    55.  5.100 / 55.  5.100\\n'", "cmdline": ["ffmpeg", "-nostats", "-loglevel", "error", "-y", "-f", "rawvideo", "-s:v", "500x500", "-pix_fmt", "rgb24", "-framerate", "15", "-i", "-", "-vf", "scale=trunc(iw/2)*2:trunc(ih/2)*2", "-vcodec", "libx264", "-pix_fmt", "yuv420p", "-r", "15", "/tmp/tmpvgm8oa_n/ppo-Acrobot-v1/replay.mp4"]}, "episode": {"r": -63.0, "l": 64, "t": 1.256068}}
+{"content_type": "video/mp4", "encoder_version": {"backend": "ffmpeg", "version": "b'ffmpeg version 4.2.7-0ubuntu0.1 Copyright (c) 2000-2022 the FFmpeg developers\\nbuilt with gcc 9 (Ubuntu 9.4.0-1ubuntu1~20.04.1)\\nconfiguration: --prefix=/usr --extra-version=0ubuntu0.1 --toolchain=hardened --libdir=/usr/lib/x86_64-linux-gnu --incdir=/usr/include/x86_64-linux-gnu --arch=amd64 --enable-gpl --disable-stripping --enable-avresample --disable-filter=resample --enable-avisynth --enable-gnutls --enable-ladspa --enable-libaom --enable-libass --enable-libbluray --enable-libbs2b --enable-libcaca --enable-libcdio --enable-libcodec2 --enable-libflite --enable-libfontconfig --enable-libfreetype --enable-libfribidi --enable-libgme --enable-libgsm --enable-libjack --enable-libmp3lame --enable-libmysofa --enable-libopenjpeg --enable-libopenmpt --enable-libopus --enable-libpulse --enable-librsvg --enable-librubberband --enable-libshine --enable-libsnappy --enable-libsoxr --enable-libspeex --enable-libssh --enable-libtheora --enable-libtwolame --enable-libvidstab --enable-libvorbis --enable-libvpx --enable-libwavpack --enable-libwebp --enable-libx265 --enable-libxml2 --enable-libxvid --enable-libzmq --enable-libzvbi --enable-lv2 --enable-omx --enable-openal --enable-opencl --enable-opengl --enable-sdl2 --enable-libdc1394 --enable-libdrm --enable-libiec61883 --enable-nvenc --enable-chromaprint --enable-frei0r --enable-libx264 --enable-shared\\nlibavutil      56. 31.100 / 56. 31.100\\nlibavcodec     58. 54.100 / 58. 54.100\\nlibavformat    58. 29.100 / 58. 29.100\\nlibavdevice    58.  8.100 / 58.  8.100\\nlibavfilter     7. 57.100 /  7. 57.100\\nlibavresample   4.  0.  0 /  4.  0.  0\\nlibswscale      5.  5.100 /  5.  5.100\\nlibswresample   3.  5.100 /  3.  5.100\\nlibpostproc    55.  5.100 / 55.  5.100\\n'", "cmdline": ["ffmpeg", "-nostats", "-loglevel", "error", "-y", "-f", "rawvideo", "-s:v", "500x500", "-pix_fmt", "rgb24", "-framerate", "15", "-i", "-", "-vf", "scale=trunc(iw/2)*2:trunc(ih/2)*2", "-vcodec", "libx264", "-pix_fmt", "yuv420p", "-r", "15", "/tmp/tmpq_g6fglc/ppo-Acrobot-v1/replay.mp4"]}, "episode": {"r": -70.0, "l": 71, "t": 1.8153}}

rl_algo_impls/a2c/a2c.py

@@ -10,6 +10,7 @@ from typing import Optional, TypeVar
 
 from rl_algo_impls.shared.algorithm import Algorithm
 from rl_algo_impls.shared.callbacks.callback import Callback
+from rl_algo_impls.shared.gae import compute_advantages
 from rl_algo_impls.shared.policy.on_policy import ActorCritic
 from rl_algo_impls.shared.schedule import schedule, update_learning_rate
 from rl_algo_impls.shared.stats import log_scalars
@@ -84,12 +85,12 @@ class A2C(Algorithm):
         obs = np.zeros(epoch_dim + obs_space.shape, dtype=obs_space.dtype)
         actions = np.zeros(epoch_dim + act_space.shape, dtype=act_space.dtype)
         rewards = np.zeros(epoch_dim, dtype=np.float32)
-        episode_starts = np.zeros(epoch_dim, dtype=np.byte)
+        episode_starts = np.zeros(epoch_dim, dtype=np.bool8)
         values = np.zeros(epoch_dim, dtype=np.float32)
         logprobs = np.zeros(epoch_dim, dtype=np.float32)
 
         next_obs = self.env.reset()
-        next_episode_starts = np.ones(step_dim, dtype=np.byte)
+        next_episode_starts = np.full(step_dim, True, dtype=np.bool8)
 
         timesteps_elapsed = start_timesteps
         while timesteps_elapsed < start_timesteps + train_timesteps:
@@ -126,23 +127,16 @@ class A2C(Algorithm):
                     clamped_action
                 )
 
-            advantages = np.zeros(epoch_dim, dtype=np.float32)
-            last_gae_lam = 0
-            for t in reversed(range(self.n_steps)):
-                if t == self.n_steps - 1:
-                    next_nonterminal = 1.0 - next_episode_starts
-                    next_value = self.policy.value(next_obs)
-                else:
-                    next_nonterminal = 1.0 - episode_starts[t + 1]
-                    next_value = values[t + 1]
-                delta = (
-                    rewards[t] + self.gamma * next_value * next_nonterminal - values[t]
-                )
-                last_gae_lam = (
-                    delta
-                    + self.gamma * self.gae_lambda * next_nonterminal * last_gae_lam
-                )
-                advantages[t] = last_gae_lam
+            advantages = compute_advantages(
+                rewards,
+                values,
+                episode_starts,
+                next_episode_starts,
+                next_obs,
+                self.policy,
+                self.gamma,
+                self.gae_lambda,
+            )
             returns = advantages + values
 
             b_obs = torch.tensor(obs.reshape((-1,) + obs_space.shape)).to(self.device)

rl_algo_impls/a2c/optimize.py

@@ -3,7 +3,7 @@ import optuna
 from copy import deepcopy
 
 from rl_algo_impls.runner.config import Config, Hyperparams, EnvHyperparams
-from rl_algo_impls.runner.env import make_eval_env
+from rl_algo_impls.shared.vec_env import make_eval_env
 from rl_algo_impls.shared.policy.optimize_on_policy import sample_on_policy_hyperparams
 from rl_algo_impls.tuning.optimize_env import sample_env_hyperparams
 

rl_algo_impls/benchmark_publish.py

@@ -54,8 +54,8 @@ def benchmark_publish() -> None:
         "--virtual-display", action="store_true", help="Use headless virtual display"
     )
     # parser.set_defaults(
-    #     wandb_tags=["benchmark_2067e21", "host_155-248-199-228"],
-    #     wandb_report_url="https://api.wandb.ai/links/sgoodfriend/09frjfcs",
+    #     wandb_tags=["benchmark_e47a44c", "host_129-146-2-230"],
+    #     wandb_report_url="https://api.wandb.ai/links/sgoodfriend/v4wd7cp5",
     #     envs=[],
     #     exclude_envs=[],
     # )

rl_algo_impls/compare_runs.py

@@ -194,5 +194,6 @@ def compare_runs() -> None:
     df.loc["mean"] = df.mean(numeric_only=True)
     print(df.to_markdown())
 
+
 if __name__ == "__main__":
-    compare_runs()
+    compare_runs()

rl_algo_impls/dqn/policy.py

@@ -1,16 +1,16 @@
-import numpy as np
 import os
-import torch
-
 from typing import Optional, Sequence, TypeVar
 
+import numpy as np
+import torch
+
 from rl_algo_impls.dqn.q_net import QNetwork
 from rl_algo_impls.shared.policy.policy import Policy
 from rl_algo_impls.wrappers.vectorable_wrapper import (
     VecEnv,
     VecEnvObs,
-    single_observation_space,
     single_action_space,
+    single_observation_space,
 )
 
 DQNPolicySelf = TypeVar("DQNPolicySelf", bound="DQNPolicy")
@@ -21,7 +21,7 @@ class DQNPolicy(Policy):
         self,
         env: VecEnv,
         hidden_sizes: Sequence[int] = [],
-        cnn_feature_dim: int = 512,
+        cnn_flatten_dim: int = 512,
         cnn_style: str = "nature",
         cnn_layers_init_orthogonal: Optional[bool] = None,
         impala_channels: Sequence[int] = (16, 32, 32),
@@ -32,16 +32,23 @@ class DQNPolicy(Policy):
             single_observation_space(env),
             single_action_space(env),
             hidden_sizes,
-            cnn_feature_dim=cnn_feature_dim,
+            cnn_flatten_dim=cnn_flatten_dim,
             cnn_style=cnn_style,
             cnn_layers_init_orthogonal=cnn_layers_init_orthogonal,
             impala_channels=impala_channels,
         )
 
     def act(
-        self, obs: VecEnvObs, eps: float = 0, deterministic: bool = True
+        self,
+        obs: VecEnvObs,
+        eps: float = 0,
+        deterministic: bool = True,
+        action_masks: Optional[np.ndarray] = None,
     ) -> np.ndarray:
         assert eps == 0 if deterministic else eps >= 0
+        assert (
+            action_masks is None
+        ), f"action_masks not currently supported in {self.__class__.__name__}"
         if not deterministic and np.random.random() < eps:
             return np.array(
                 [

rl_algo_impls/dqn/q_net.py

@@ -1,11 +1,11 @@
+from typing import Optional, Sequence, Type
+
 import gym
 import torch as th
 import torch.nn as nn
-
 from gym.spaces import Discrete
-from typing import Optional, Sequence, Type
 
-from rl_algo_impls.shared.module.feature_extractor import FeatureExtractor
+from rl_algo_impls.shared.encoder import Encoder
 from rl_algo_impls.shared.module.module import mlp
 
 
@@ -16,17 +16,17 @@ class QNetwork(nn.Module):
         action_space: gym.Space,
         hidden_sizes: Sequence[int] = [],
         activation: Type[nn.Module] = nn.ReLU,  # Used by stable-baselines3
-        cnn_feature_dim: int = 512,
+        cnn_flatten_dim: int = 512,
         cnn_style: str = "nature",
         cnn_layers_init_orthogonal: Optional[bool] = None,
         impala_channels: Sequence[int] = (16, 32, 32),
     ) -> None:
         super().__init__()
         assert isinstance(action_space, Discrete)
-        self._feature_extractor = FeatureExtractor(
+        self._feature_extractor = Encoder(
             observation_space,
             activation,
-            cnn_feature_dim=cnn_feature_dim,
+            cnn_flatten_dim=cnn_flatten_dim,
             cnn_style=cnn_style,
             cnn_layers_init_orthogonal=cnn_layers_init_orthogonal,
             impala_channels=impala_channels,

rl_algo_impls/huggingface_publish.py

@@ -19,7 +19,7 @@ from pyvirtualdisplay.display import Display
 from rl_algo_impls.publish.markdown_format import EvalTableData, model_card_text
 from rl_algo_impls.runner.config import EnvHyperparams
 from rl_algo_impls.runner.evaluate import EvalArgs, evaluate_model
-from rl_algo_impls.runner.env import make_eval_env
+from rl_algo_impls.shared.vec_env import make_eval_env
 from rl_algo_impls.shared.callbacks.eval_callback import evaluate
 from rl_algo_impls.wrappers.vec_episode_recorder import VecEpisodeRecorder
 

rl_algo_impls/hyperparams/a2c.yml

@@ -97,31 +97,35 @@ Walker2DBulletEnv-v0:
 HopperBulletEnv-v0:
   <<: *pybullet-defaults
 
+# Tuned
 CarRacing-v0:
   n_timesteps: !!float 4e6
   env_hyperparams:
-    n_envs: 8
+    n_envs: 16
     frame_stack: 4
     normalize: true
     normalize_kwargs:
       norm_obs: false
       norm_reward: true
   policy_hyperparams:
-    use_sde: true
-    log_std_init: -2
-    init_layers_orthogonal: false
-    activation_fn: relu
+    use_sde: false
+    log_std_init: -1.3502584927786276
+    init_layers_orthogonal: true
+    activation_fn: tanh
     share_features_extractor: false
-    cnn_feature_dim: 256
+    cnn_flatten_dim: 256
     hidden_sizes: [256]
   algo_hyperparams:
-    n_steps: 512
-    learning_rate: !!float 1.62e-5
-    gamma: 0.997
-    gae_lambda: 0.975
-    ent_coef: 0
-    sde_sample_freq: 128
-    vf_coef: 0.64
+    n_steps: 16
+    learning_rate: 0.000025630993245026736
+    learning_rate_decay: linear
+    gamma: 0.99957617037542
+    gae_lambda: 0.949455676599436
+    ent_coef: !!float 1.707983205298309e-7
+    vf_coef: 0.10428178193833336
+    max_grad_norm: 0.5406643389792273
+    normalize_advantage: true
+    use_rms_prop: false
 
 _atari: &atari-defaults
   n_timesteps: !!float 1e7

rl_algo_impls/hyperparams/dqn.yml

@@ -108,7 +108,7 @@ _impala-atari: &impala-atari-defaults
   <<: *atari-defaults
   policy_hyperparams:
     cnn_style: impala
-    cnn_feature_dim: 256
+    cnn_flatten_dim: 256
     init_layers_orthogonal: true
     cnn_layers_init_orthogonal: false
 

rl_algo_impls/hyperparams/ppo.yml

@@ -112,7 +112,7 @@ CarRacing-v0: &carracing-defaults
     init_layers_orthogonal: false
     activation_fn: relu
     share_features_extractor: false
-    cnn_feature_dim: 256
+    cnn_flatten_dim: 256
     hidden_sizes: [256]
   algo_hyperparams:
     n_steps: 512
@@ -152,7 +152,7 @@ _atari: &atari-defaults
     vec_env_class: async
   policy_hyperparams: &atari-policy-defaults
     activation_fn: relu
-  algo_hyperparams:
+  algo_hyperparams: &atari-algo-defaults
     n_steps: 128
     batch_size: 256
     n_epochs: 4
@@ -192,7 +192,7 @@ _impala-atari: &impala-atari-defaults
   policy_hyperparams:
     <<: *atari-policy-defaults
     cnn_style: impala
-    cnn_feature_dim: 256
+    cnn_flatten_dim: 256
     init_layers_orthogonal: true
     cnn_layers_init_orthogonal: false
 
@@ -212,6 +212,126 @@ impala-QbertNoFrameskip-v4:
   <<: *impala-atari-defaults
   env_id: QbertNoFrameskip-v4
 
+_microrts: &microrts-defaults
+  <<: *atari-defaults
+  n_timesteps: !!float 2e6
+  env_hyperparams: &microrts-env-defaults
+    n_envs: 8
+    vec_env_class: sync
+    mask_actions: true
+  policy_hyperparams: &microrts-policy-defaults
+    <<: *atari-policy-defaults
+    cnn_style: microrts
+    cnn_flatten_dim: 128
+  algo_hyperparams: &microrts-algo-defaults
+    <<: *atari-algo-defaults
+    clip_range_decay: none
+    clip_range_vf: 0.1
+    ppo2_vf_coef_halving: true
+  eval_params:
+    deterministic: false # Good idea because MultiCategorical mode isn't great
+
+_no-mask-microrts: &no-mask-microrts-defaults
+  <<: *microrts-defaults
+  env_hyperparams:
+    <<: *microrts-env-defaults
+    mask_actions: false
+
+MicrortsMining-v1-NoMask:
+  <<: *no-mask-microrts-defaults
+  env_id: MicrortsMining-v1
+
+MicrortsAttackShapedReward-v1-NoMask:
+  <<: *no-mask-microrts-defaults
+  env_id: MicrortsAttackShapedReward-v1
+
+MicrortsRandomEnemyShapedReward3-v1-NoMask:
+  <<: *no-mask-microrts-defaults
+  env_id: MicrortsRandomEnemyShapedReward3-v1
+
+_microrts_ai: &microrts-ai-defaults
+  <<: *microrts-defaults
+  n_timesteps: !!float 100e6
+  additional_keys_to_log: ["microrts_stats"]
+  env_hyperparams: &microrts-ai-env-defaults
+    n_envs: 24
+    env_type: microrts
+    make_kwargs:
+      num_selfplay_envs: 0
+      max_steps: 2000
+      render_theme: 2
+      map_path: maps/16x16/basesWorkers16x16.xml
+      reward_weight: [10.0, 1.0, 1.0, 0.2, 1.0, 4.0]
+  policy_hyperparams: &microrts-ai-policy-defaults
+    <<: *microrts-policy-defaults
+    cnn_flatten_dim: 256
+    actor_head_style: gridnet
+  algo_hyperparams: &microrts-ai-algo-defaults
+    <<: *microrts-algo-defaults
+    learning_rate: !!float 2.5e-4
+    learning_rate_decay: linear
+    n_steps: 512
+    batch_size: 3072
+    n_epochs: 4
+    ent_coef: 0.01
+    vf_coef: 0.5
+    max_grad_norm: 0.5
+    clip_range: 0.1
+    clip_range_vf: 0.1
+
+MicrortsAttackPassiveEnemySparseReward-v3:
+  <<: *microrts-ai-defaults
+  n_timesteps: !!float 2e6
+  env_id: MicrortsAttackPassiveEnemySparseReward-v3 # Workaround to keep model name simple
+  env_hyperparams:
+    <<: *microrts-ai-env-defaults
+    bots:
+      passiveAI: 24
+
+MicrortsDefeatRandomEnemySparseReward-v3: &microrts-random-ai-defaults
+  <<: *microrts-ai-defaults
+  n_timesteps: !!float 2e6
+  env_id: MicrortsDefeatRandomEnemySparseReward-v3 # Workaround to keep model name simple
+  env_hyperparams:
+    <<: *microrts-ai-env-defaults
+    bots:
+      randomBiasedAI: 24
+
+enc-dec-MicrortsDefeatRandomEnemySparseReward-v3:
+  <<: *microrts-random-ai-defaults
+  policy_hyperparams:
+    <<: *microrts-ai-policy-defaults
+    cnn_style: gridnet_encoder
+    actor_head_style: gridnet_decoder
+    v_hidden_sizes: [128]
+
+MicrortsDefeatCoacAIShaped-v3: &microrts-coacai-defaults
+  <<: *microrts-ai-defaults
+  env_id: MicrortsDefeatCoacAIShaped-v3 # Workaround to keep model name simple
+  n_timesteps: !!float 300e6
+  env_hyperparams: &microrts-coacai-env-defaults
+    <<: *microrts-ai-env-defaults
+    bots:
+      coacAI: 24
+
+MicrortsDefeatCoacAIShaped-v3-diverseBots: &microrts-diverse-defaults
+  <<: *microrts-coacai-defaults
+  env_hyperparams:
+    <<: *microrts-coacai-env-defaults
+    bots:
+      coacAI: 18
+      randomBiasedAI: 2
+      lightRushAI: 2
+      workerRushAI: 2
+
+enc-dec-MicrortsDefeatCoacAIShaped-v3-diverseBots:
+  <<: *microrts-diverse-defaults
+  policy_hyperparams:
+    <<: *microrts-ai-policy-defaults
+    cnn_style: gridnet_encoder
+    actor_head_style: gridnet_decoder
+    v_hidden_sizes: [128]
+
 HalfCheetahBulletEnv-v0: &pybullet-defaults
   n_timesteps: !!float 2e6
   env_hyperparams: &pybullet-env-defaults
@@ -282,7 +402,7 @@ _procgen: &procgen-defaults
   policy_hyperparams: &procgen-policy-defaults
     activation_fn: relu
     cnn_style: impala
-    cnn_feature_dim: 256
+    cnn_flatten_dim: 256
     init_layers_orthogonal: true
     cnn_layers_init_orthogonal: false
   algo_hyperparams: &procgen-algo-defaults
@@ -368,7 +488,7 @@ procgen-starpilot-hard-2xIMPALA-fat:
   policy_hyperparams:
     <<: *procgen-policy-defaults
     impala_channels: [32, 64, 64]
-    cnn_feature_dim: 512
+    cnn_flatten_dim: 512
   algo_hyperparams:
     <<: *procgen-hard-algo-defaults
     learning_rate: !!float 2.5e-4

rl_algo_impls/hyperparams/vpg.yml

@@ -110,7 +110,7 @@ CarRacing-v0:
     log_std_init: -2
     init_layers_orthogonal: false
     activation_fn: relu
-    cnn_feature_dim: 256
+    cnn_flatten_dim: 256
     hidden_sizes: [256]
   algo_hyperparams:
     n_steps: 1000
@@ -175,9 +175,9 @@ FrozenLake-v1:
     save_best: true
 
 _atari: &atari-defaults
-  n_timesteps: !!float 25e6
+  n_timesteps: !!float 10e6
   env_hyperparams:
-    n_envs: 4
+    n_envs: 2
     frame_stack: 4
     no_reward_timeout_steps: 1000
     no_reward_fire_steps: 500
@@ -185,7 +185,7 @@ _atari: &atari-defaults
   policy_hyperparams:
     activation_fn: relu
   algo_hyperparams:
-    n_steps: 2048
+    n_steps: 3072
     pi_lr: !!float 5e-5
     gamma: 0.99
     gae_lambda: 0.95

rl_algo_impls/optimize.py

@@ -17,7 +17,7 @@ from typing import Callable, List, NamedTuple, Optional, Sequence, Union
 
 from rl_algo_impls.a2c.optimize import sample_params as a2c_sample_params
 from rl_algo_impls.runner.config import Config, EnvHyperparams, RunArgs
-from rl_algo_impls.runner.env import make_env, make_eval_env
+from rl_algo_impls.shared.vec_env import make_env, make_eval_env
 from rl_algo_impls.runner.running_utils import (
     base_parser,
     load_hyperparams,
@@ -194,7 +194,7 @@ def simple_optimize(trial: optuna.Trial, args: RunArgs, study_args: StudyArgs) -
     env = make_env(
         config, EnvHyperparams(**config.env_hyperparams), tb_writer=tb_writer
     )
-    device = get_device(config.device, env)
+    device = get_device(config, env)
     policy = make_policy(args.algo, env, device, **config.policy_hyperparams)
     algo = ALGOS[args.algo](policy, env, device, tb_writer, **config.algo_hyperparams)
 
@@ -274,7 +274,7 @@ def stepwise_optimize(
             project=study_args.wandb_project_name,
             entity=study_args.wandb_entity,
             config=asdict(hyperparams),
-            name=f"{study_args.study_name}-{str(trial.number)}",
+            name=f"{str(trial.number)}-S{base_config.seed()}",
             tags=study_args.wandb_tags,
             group=study_args.wandb_group,
             save_code=True,
@@ -298,7 +298,7 @@ def stepwise_optimize(
                 normalize_load_path=config.model_dir_path() if i > 0 else None,
                 tb_writer=tb_writer,
             )
-            device = get_device(config.device, env)
+            device = get_device(config, env)
             policy = make_policy(arg.algo, env, device, **config.policy_hyperparams)
             if i > 0:
                 policy.load(config.model_dir_path())
@@ -433,6 +433,7 @@ def optimize() -> None:
 
     fig1 = plot_optimization_history(study)
     fig1.write_image("opt_history.png")
+
     fig2 = plot_param_importances(study)
     fig2.write_image("param_importances.png")
 

rl_algo_impls/ppo/ppo.py

@@ -1,59 +1,26 @@
+import logging
+from dataclasses import asdict, dataclass
+from time import perf_counter
+from typing import List, NamedTuple, Optional, TypeVar
+
 import numpy as np
 import torch
 import torch.nn as nn
-
-from dataclasses import asdict, dataclass, field
-from time import perf_counter
 from torch.optim import Adam
 from torch.utils.tensorboard.writer import SummaryWriter
-from typing import List, Optional, NamedTuple, TypeVar
 
 from rl_algo_impls.shared.algorithm import Algorithm
 from rl_algo_impls.shared.callbacks.callback import Callback
-from rl_algo_impls.shared.gae import compute_advantage, compute_rtg_and_advantage
+from rl_algo_impls.shared.gae import compute_advantages
 from rl_algo_impls.shared.policy.on_policy import ActorCritic
-from rl_algo_impls.shared.schedule import (
-    constant_schedule,
-    linear_schedule,
-    update_learning_rate,
+from rl_algo_impls.shared.schedule import schedule, update_learning_rate
+from rl_algo_impls.shared.stats import log_scalars
+from rl_algo_impls.wrappers.action_mask_wrapper import find_action_masker
+from rl_algo_impls.wrappers.vectorable_wrapper import (
+    VecEnv,
+    single_action_space,
+    single_observation_space,
 )
-from rl_algo_impls.shared.trajectory import Trajectory, TrajectoryAccumulator
-from rl_algo_impls.wrappers.vectorable_wrapper import VecEnv, VecEnvObs
-
-
-@dataclass
-class PPOTrajectory(Trajectory):
-    logp_a: List[float] = field(default_factory=list)
-
-    def add(
-        self,
-        obs: np.ndarray,
-        act: np.ndarray,
-        next_obs: np.ndarray,
-        rew: float,
-        terminated: bool,
-        v: float,
-        logp_a: float,
-    ):
-        super().add(obs, act, next_obs, rew, terminated, v)
-        self.logp_a.append(logp_a)
-
-
-class PPOTrajectoryAccumulator(TrajectoryAccumulator):
-    def __init__(self, num_envs: int) -> None:
-        super().__init__(num_envs, PPOTrajectory)
-
-    def step(
-        self,
-        obs: VecEnvObs,
-        action: np.ndarray,
-        next_obs: VecEnvObs,
-        reward: np.ndarray,
-        done: np.ndarray,
-        val: np.ndarray,
-        logp_a: np.ndarray,
-    ) -> None:
-        super().step(obs, action, next_obs, reward, done, val, logp_a)
 
 
 class TrainStepStats(NamedTuple):
@@ -132,39 +99,31 @@ class PPO(Algorithm):
         vf_coef: float = 0.5,
         ppo2_vf_coef_halving: bool = False,
         max_grad_norm: float = 0.5,
-        update_rtg_between_epochs: bool = False,
         sde_sample_freq: int = -1,
+        update_advantage_between_epochs: bool = True,
+        update_returns_between_epochs: bool = False,
     ) -> None:
         super().__init__(policy, env, device, tb_writer)
         self.policy = policy
+        self.action_masker = find_action_masker(env)
 
         self.gamma = gamma
         self.gae_lambda = gae_lambda
         self.optimizer = Adam(self.policy.parameters(), lr=learning_rate, eps=1e-7)
-        self.lr_schedule = (
-            linear_schedule(learning_rate, 0)
-            if learning_rate_decay == "linear"
-            else constant_schedule(learning_rate)
-        )
+        self.lr_schedule = schedule(learning_rate_decay, learning_rate)
         self.max_grad_norm = max_grad_norm
-        self.clip_range_schedule = (
-            linear_schedule(clip_range, 0)
-            if clip_range_decay == "linear"
-            else constant_schedule(clip_range)
-        )
+        self.clip_range_schedule = schedule(clip_range_decay, clip_range)
         self.clip_range_vf_schedule = None
         if clip_range_vf:
-            self.clip_range_vf_schedule = (
-                linear_schedule(clip_range_vf, 0)
-                if clip_range_vf_decay == "linear"
-                else constant_schedule(clip_range_vf)
-            )
+            self.clip_range_vf_schedule = schedule(clip_range_vf_decay, clip_range_vf)
+
+        if normalize_advantage:
+            assert (
+                env.num_envs * n_steps > 1 and batch_size > 1
+            ), f"Each minibatch must be larger than 1 to support normalization"
         self.normalize_advantage = normalize_advantage
-        self.ent_coef_schedule = (
-            linear_schedule(ent_coef, 0)
-            if ent_coef_decay == "linear"
-            else constant_schedule(ent_coef)
-        )
+
+        self.ent_coef_schedule = schedule(ent_coef_decay, ent_coef)
         self.vf_coef = vf_coef
         self.ppo2_vf_coef_halving = ppo2_vf_coef_halving
 
@@ -173,181 +132,243 @@ class PPO(Algorithm):
         self.n_epochs = n_epochs
         self.sde_sample_freq = sde_sample_freq
 
-        self.update_rtg_between_epochs = update_rtg_between_epochs
+        self.update_advantage_between_epochs = update_advantage_between_epochs
+        self.update_returns_between_epochs = update_returns_between_epochs
 
     def learn(
         self: PPOSelf,
-        total_timesteps: int,
+        train_timesteps: int,
         callback: Optional[Callback] = None,
+        total_timesteps: Optional[int] = None,
+        start_timesteps: int = 0,
     ) -> PPOSelf:
-        obs = self.env.reset()
-        ts_elapsed = 0
-        while ts_elapsed < total_timesteps:
-            start_time = perf_counter()
-            accumulator = self._collect_trajectories(obs)
-            rollout_steps = self.n_steps * self.env.num_envs
-            ts_elapsed += rollout_steps
-            progress = ts_elapsed / total_timesteps
-            train_stats = self.train(accumulator.all_trajectories, progress, ts_elapsed)
-            train_stats.write_to_tensorboard(self.tb_writer, ts_elapsed)
-            end_time = perf_counter()
-            self.tb_writer.add_scalar(
-                "train/steps_per_second",
-                rollout_steps / (end_time - start_time),
-                ts_elapsed,
+        if total_timesteps is None:
+            total_timesteps = train_timesteps
+        assert start_timesteps + train_timesteps <= total_timesteps
+
+        epoch_dim = (self.n_steps, self.env.num_envs)
+        step_dim = (self.env.num_envs,)
+        obs_space = single_observation_space(self.env)
+        act_space = single_action_space(self.env)
+        act_shape = self.policy.action_shape
+
+        next_obs = self.env.reset()
+        next_action_masks = (
+            self.action_masker.action_masks() if self.action_masker else None
+        )
+        next_episode_starts = np.full(step_dim, True, dtype=np.bool8)
+
+        obs = np.zeros(epoch_dim + obs_space.shape, dtype=obs_space.dtype)  # type: ignore
+        actions = np.zeros(epoch_dim + act_shape, dtype=act_space.dtype)  # type: ignore
+        rewards = np.zeros(epoch_dim, dtype=np.float32)
+        episode_starts = np.zeros(epoch_dim, dtype=np.bool8)
+        values = np.zeros(epoch_dim, dtype=np.float32)
+        logprobs = np.zeros(epoch_dim, dtype=np.float32)
+        action_masks = (
+            np.zeros(
+                (self.n_steps,) + next_action_masks.shape, dtype=next_action_masks.dtype
             )
-            if callback:
-                callback.on_step(timesteps_elapsed=rollout_steps)
-
-        return self
-
-    def _collect_trajectories(self, obs: VecEnvObs) -> PPOTrajectoryAccumulator:
-        self.policy.eval()
-        accumulator = PPOTrajectoryAccumulator(self.env.num_envs)
-        self.policy.reset_noise()
-        for i in range(self.n_steps):
-            if self.sde_sample_freq > 0 and i > 0 and i % self.sde_sample_freq == 0:
-                self.policy.reset_noise()
-            action, value, logp_a, clamped_action = self.policy.step(obs)
-            next_obs, reward, done, _ = self.env.step(clamped_action)
-            accumulator.step(obs, action, next_obs, reward, done, value, logp_a)
-            obs = next_obs
-        return accumulator
-
-    def train(
-        self, trajectories: List[PPOTrajectory], progress: float, timesteps_elapsed: int
-    ) -> TrainStats:
-        self.policy.train()
-        learning_rate = self.lr_schedule(progress)
-        update_learning_rate(self.optimizer, learning_rate)
-        self.tb_writer.add_scalar(
-            "charts/learning_rate",
-            self.optimizer.param_groups[0]["lr"],
-            timesteps_elapsed,
+            if next_action_masks is not None
+            else None
         )
 
-        pi_clip = self.clip_range_schedule(progress)
-        self.tb_writer.add_scalar("charts/pi_clip", pi_clip, timesteps_elapsed)
-        if self.clip_range_vf_schedule:
-            v_clip = self.clip_range_vf_schedule(progress)
-            self.tb_writer.add_scalar("charts/v_clip", v_clip, timesteps_elapsed)
-        else:
-            v_clip = None
-        ent_coef = self.ent_coef_schedule(progress)
-        self.tb_writer.add_scalar("charts/ent_coef", ent_coef, timesteps_elapsed)
-
-        obs = torch.as_tensor(
-            np.concatenate([np.array(t.obs) for t in trajectories]), device=self.device
-        )
-        act = torch.as_tensor(
-            np.concatenate([np.array(t.act) for t in trajectories]), device=self.device
-        )
-        rtg, adv = compute_rtg_and_advantage(
-            trajectories, self.policy, self.gamma, self.gae_lambda, self.device
-        )
-        orig_v = torch.as_tensor(
-            np.concatenate([np.array(t.v) for t in trajectories]), device=self.device
-        )
-        orig_logp_a = torch.as_tensor(
-            np.concatenate([np.array(t.logp_a) for t in trajectories]),
-            device=self.device,
-        )
+        timesteps_elapsed = start_timesteps
+        while timesteps_elapsed < start_timesteps + train_timesteps:
+            start_time = perf_counter()
 
-        step_stats = []
-        for _ in range(self.n_epochs):
-            step_stats.clear()
-            if self.update_rtg_between_epochs:
-                rtg, adv = compute_rtg_and_advantage(
-                    trajectories, self.policy, self.gamma, self.gae_lambda, self.device
-                )
+            progress = timesteps_elapsed / total_timesteps
+            ent_coef = self.ent_coef_schedule(progress)
+            learning_rate = self.lr_schedule(progress)
+            update_learning_rate(self.optimizer, learning_rate)
+            pi_clip = self.clip_range_schedule(progress)
+            chart_scalars = {
+                "learning_rate": self.optimizer.param_groups[0]["lr"],
+                "ent_coef": ent_coef,
+                "pi_clip": pi_clip,
+            }
+            if self.clip_range_vf_schedule:
+                v_clip = self.clip_range_vf_schedule(progress)
+                chart_scalars["v_clip"] = v_clip
             else:
-                adv = compute_advantage(
-                    trajectories, self.policy, self.gamma, self.gae_lambda, self.device
+                v_clip = None
+            log_scalars(self.tb_writer, "charts", chart_scalars, timesteps_elapsed)
+
+            self.policy.eval()
+            self.policy.reset_noise()
+            for s in range(self.n_steps):
+                timesteps_elapsed += self.env.num_envs
+                if self.sde_sample_freq > 0 and s > 0 and s % self.sde_sample_freq == 0:
+                    self.policy.reset_noise()
+
+                obs[s] = next_obs
+                episode_starts[s] = next_episode_starts
+                if action_masks is not None:
+                    action_masks[s] = next_action_masks
+
+                (
+                    actions[s],
+                    values[s],
+                    logprobs[s],
+                    clamped_action,
+                ) = self.policy.step(next_obs, action_masks=next_action_masks)
+                next_obs, rewards[s], next_episode_starts, _ = self.env.step(
+                    clamped_action
                 )
-            idxs = torch.randperm(len(obs))
-            for i in range(0, len(obs), self.batch_size):
-                mb_idxs = idxs[i : i + self.batch_size]
-                mb_adv = adv[mb_idxs]
-                if self.normalize_advantage:
-                    mb_adv = (mb_adv - mb_adv.mean(-1)) / (mb_adv.std(-1) + 1e-8)
-                self.policy.reset_noise(self.batch_size)
-                step_stats.append(
-                    self._train_step(
-                        pi_clip,
-                        v_clip,
-                        ent_coef,
-                        obs[mb_idxs],
-                        act[mb_idxs],
-                        rtg[mb_idxs],
-                        mb_adv,
-                        orig_v[mb_idxs],
-                        orig_logp_a[mb_idxs],
-                    )
+                next_action_masks = (
+                    self.action_masker.action_masks() if self.action_masker else None
                 )
 
-        y_pred, y_true = orig_v.cpu().numpy(), rtg.cpu().numpy()
-        var_y = np.var(y_true).item()
-        explained_var = (
-            np.nan if var_y == 0 else 1 - np.var(y_true - y_pred).item() / var_y
-        )
+            self.policy.train()
+
+            b_obs = torch.tensor(obs.reshape((-1,) + obs_space.shape)).to(self.device)  # type: ignore
+            b_actions = torch.tensor(actions.reshape((-1,) + act_shape)).to(  # type: ignore
+                self.device
+            )
+            b_logprobs = torch.tensor(logprobs.reshape(-1)).to(self.device)
+            b_action_masks = (
+                torch.tensor(action_masks.reshape((-1,) + next_action_masks.shape[1:])).to(  # type: ignore
+                    self.device
+                )
+                if action_masks is not None
+                else None
+            )
+
+            y_pred = values.reshape(-1)
+            b_values = torch.tensor(y_pred).to(self.device)
+
+            step_stats = []
+            # Define variables that will definitely be set through the first epoch
+            advantages: np.ndarray = None  # type: ignore
+            b_advantages: torch.Tensor = None  # type: ignore
+            y_true: np.ndarray = None  # type: ignore
+            b_returns: torch.Tensor = None  # type: ignore
+            for e in range(self.n_epochs):
+                if e == 0 or self.update_advantage_between_epochs:
+                    advantages = compute_advantages(
+                        rewards,
+                        values,
+                        episode_starts,
+                        next_episode_starts,
+                        next_obs,
+                        self.policy,
+                        self.gamma,
+                        self.gae_lambda,
+                    )
+                    b_advantages = torch.tensor(advantages.reshape(-1)).to(self.device)
+                if e == 0 or self.update_returns_between_epochs:
+                    returns = advantages + values
+                    y_true = returns.reshape(-1)
+                    b_returns = torch.tensor(y_true).to(self.device)
+
+                b_idxs = torch.randperm(len(b_obs))
+                # Only record last epoch's stats
+                step_stats.clear()
+                for i in range(0, len(b_obs), self.batch_size):
+                    self.policy.reset_noise(self.batch_size)
+
+                    mb_idxs = b_idxs[i : i + self.batch_size]
+
+                    mb_obs = b_obs[mb_idxs]
+                    mb_actions = b_actions[mb_idxs]
+                    mb_values = b_values[mb_idxs]
+                    mb_logprobs = b_logprobs[mb_idxs]
+                    mb_action_masks = (
+                        b_action_masks[mb_idxs] if b_action_masks is not None else None
+                    )
 
-        return TrainStats(step_stats, explained_var)
+                    mb_adv = b_advantages[mb_idxs]
+                    if self.normalize_advantage:
+                        mb_adv = (mb_adv - mb_adv.mean()) / (mb_adv.std() + 1e-8)
+                    mb_returns = b_returns[mb_idxs]
 
-    def _train_step(
-        self,
-        pi_clip: float,
-        v_clip: Optional[float],
-        ent_coef: float,
-        obs: torch.Tensor,
-        act: torch.Tensor,
-        rtg: torch.Tensor,
-        adv: torch.Tensor,
-        orig_v: torch.Tensor,
-        orig_logp_a: torch.Tensor,
-    ) -> TrainStepStats:
-        logp_a, entropy, v = self.policy(obs, act)
-        logratio = logp_a - orig_logp_a
-        ratio = torch.exp(logratio)
-        clip_ratio = torch.clamp(ratio, min=1 - pi_clip, max=1 + pi_clip)
-        pi_loss = torch.maximum(-ratio * adv, -clip_ratio * adv).mean()
-
-        v_loss_unclipped = (v - rtg) ** 2
-        if v_clip:
-            v_loss_clipped = (
-                orig_v + torch.clamp(v - orig_v, -v_clip, v_clip) - rtg
-            ) ** 2
-            v_loss = torch.max(v_loss_unclipped, v_loss_clipped).mean()
-        else:
-            v_loss = v_loss_unclipped.mean()
-        if self.ppo2_vf_coef_halving:
-            v_loss *= 0.5
-
-        entropy_loss = -entropy.mean()
-
-        loss = pi_loss + ent_coef * entropy_loss + self.vf_coef * v_loss
-
-        self.optimizer.zero_grad()
-        loss.backward()
-        nn.utils.clip_grad_norm_(self.policy.parameters(), self.max_grad_norm)
-        self.optimizer.step()
-
-        with torch.no_grad():
-            approx_kl = ((ratio - 1) - logratio).mean().cpu().numpy().item()
-            clipped_frac = (
-                ((ratio - 1).abs() > pi_clip).float().mean().cpu().numpy().item()
+                    new_logprobs, entropy, new_values = self.policy(
+                        mb_obs, mb_actions, action_masks=mb_action_masks
+                    )
+
+                    logratio = new_logprobs - mb_logprobs
+                    ratio = torch.exp(logratio)
+                    clipped_ratio = torch.clamp(ratio, min=1 - pi_clip, max=1 + pi_clip)
+                    pi_loss = torch.max(-ratio * mb_adv, -clipped_ratio * mb_adv).mean()
+
+                    v_loss_unclipped = (new_values - mb_returns) ** 2
+                    if v_clip:
+                        v_loss_clipped = (
+                            mb_values
+                            + torch.clamp(new_values - mb_values, -v_clip, v_clip)
+                            - mb_returns
+                        ) ** 2
+                        v_loss = torch.max(v_loss_unclipped, v_loss_clipped).mean()
+                    else:
+                        v_loss = v_loss_unclipped.mean()
+
+                    if self.ppo2_vf_coef_halving:
+                        v_loss *= 0.5
+
+                    entropy_loss = -entropy.mean()
+
+                    loss = pi_loss + ent_coef * entropy_loss + self.vf_coef * v_loss
+
+                    self.optimizer.zero_grad()
+                    loss.backward()
+                    nn.utils.clip_grad_norm_(
+                        self.policy.parameters(), self.max_grad_norm
+                    )
+                    self.optimizer.step()
+
+                    with torch.no_grad():
+                        approx_kl = ((ratio - 1) - logratio).mean().cpu().numpy().item()
+                        clipped_frac = (
+                            ((ratio - 1).abs() > pi_clip)
+                            .float()
+                            .mean()
+                            .cpu()
+                            .numpy()
+                            .item()
+                        )
+                        val_clipped_frac = (
+                            ((new_values - mb_values).abs() > v_clip)
+                            .float()
+                            .mean()
+                            .cpu()
+                            .numpy()
+                            .item()
+                            if v_clip
+                            else 0
+                        )
+
+                    step_stats.append(
+                        TrainStepStats(
+                            loss.item(),
+                            pi_loss.item(),
+                            v_loss.item(),
+                            entropy_loss.item(),
+                            approx_kl,
+                            clipped_frac,
+                            val_clipped_frac,
+                        )
+                    )
+
+            var_y = np.var(y_true).item()
+            explained_var = (
+                np.nan if var_y == 0 else 1 - np.var(y_true - y_pred).item() / var_y
             )
-            val_clipped_frac = (
-                (((v - orig_v).abs() > v_clip).float().mean().cpu().numpy().item())
-                if v_clip
-                else 0
+            TrainStats(step_stats, explained_var).write_to_tensorboard(
+                self.tb_writer, timesteps_elapsed
             )
 
-        return TrainStepStats(
-            loss.item(),
-            pi_loss.item(),
-            v_loss.item(),
-            entropy_loss.item(),
-            approx_kl,
-            clipped_frac,
-            val_clipped_frac,
-        )
+            end_time = perf_counter()
+            rollout_steps = self.n_steps * self.env.num_envs
+            self.tb_writer.add_scalar(
+                "train/steps_per_second",
+                rollout_steps / (end_time - start_time),
+                timesteps_elapsed,
+            )
+
+            if callback:
+                if not callback.on_step(timesteps_elapsed=rollout_steps):
+                    logging.info(
+                        f"Callback terminated training at {timesteps_elapsed} timesteps"
+                    )
+                    break
+
+        return self

rl_algo_impls/runner/config.py

@@ -2,12 +2,10 @@ import dataclasses
 import inspect
 import itertools
 import os
-
-from datetime import datetime
 from dataclasses import dataclass
+from datetime import datetime
 from typing import Any, Dict, List, Optional, Type, TypeVar, Union
 
-
 RunArgsSelf = TypeVar("RunArgsSelf", bound="RunArgs")
 
 
@@ -50,6 +48,9 @@ class EnvHyperparams:
     video_step_interval: Union[int, float] = 1_000_000
     initial_steps_to_truncate: Optional[int] = None
     clip_atari_rewards: bool = True
+    normalize_type: Optional[str] = None
+    mask_actions: bool = False
+    bots: Optional[Dict[str, int]] = None
 
 
 HyperparamsSelf = TypeVar("HyperparamsSelf", bound="Hyperparams")
@@ -64,6 +65,7 @@ class Hyperparams:
     algo_hyperparams: Dict[str, Any] = dataclasses.field(default_factory=dict)
     eval_params: Dict[str, Any] = dataclasses.field(default_factory=dict)
     env_id: Optional[str] = None
+    additional_keys_to_log: List[str] = dataclasses.field(default_factory=list)
 
     @classmethod
     def from_dict_with_extra_fields(
@@ -119,6 +121,10 @@ class Config:
     def env_id(self) -> str:
         return self.hyperparams.env_id or self.args.env
 
+    @property
+    def additional_keys_to_log(self) -> List[str]:
+        return self.hyperparams.additional_keys_to_log
+
     def model_name(self, include_seed: bool = True) -> str:
         # Use arg env name instead of environment name
         parts = [self.algo, self.args.env]

rl_algo_impls/runner/evaluate.py

@@ -4,7 +4,7 @@ import shutil
 from dataclasses import dataclass
 from typing import NamedTuple, Optional
 
-from rl_algo_impls.runner.env import make_eval_env
+from rl_algo_impls.shared.vec_env import make_eval_env
 from rl_algo_impls.runner.config import Config, EnvHyperparams, Hyperparams, RunArgs
 from rl_algo_impls.runner.running_utils import (
     load_hyperparams,
@@ -75,7 +75,7 @@ def evaluate_model(args: EvalArgs, root_dir: str) -> Evaluation:
         render=args.render,
         normalize_load_path=model_path,
     )
-    device = get_device(config.device, env)
+    device = get_device(config, env)
     policy = make_policy(
         args.algo,
         env,

rl_algo_impls/runner/running_utils.py

@@ -1,32 +1,32 @@
 import argparse
-import gym
 import json
-import matplotlib.pyplot as plt
-import numpy as np
 import os
 import random
+from dataclasses import asdict
+from pathlib import Path
+from typing import Dict, Optional, Type, Union
+
+import gym
+import matplotlib.pyplot as plt
+import numpy as np
 import torch
 import torch.backends.cudnn
 import yaml
-
-from dataclasses import asdict
 from gym.spaces import Box, Discrete
-from pathlib import Path
 from torch.utils.tensorboard.writer import SummaryWriter
-from typing import Dict, Optional, Type, Union
-
-from rl_algo_impls.runner.config import Hyperparams
-from rl_algo_impls.shared.algorithm import Algorithm
-from rl_algo_impls.shared.callbacks.eval_callback import EvalCallback
-from rl_algo_impls.shared.policy.on_policy import ActorCritic
-from rl_algo_impls.shared.policy.policy import Policy
 
 from rl_algo_impls.a2c.a2c import A2C
 from rl_algo_impls.dqn.dqn import DQN
 from rl_algo_impls.dqn.policy import DQNPolicy
 from rl_algo_impls.ppo.ppo import PPO
-from rl_algo_impls.vpg.vpg import VanillaPolicyGradient
+from rl_algo_impls.runner.config import Config, Hyperparams
+from rl_algo_impls.shared.algorithm import Algorithm
+from rl_algo_impls.shared.callbacks.eval_callback import EvalCallback
+from rl_algo_impls.shared.policy.on_policy import ActorCritic
+from rl_algo_impls.shared.policy.policy import Policy
+from rl_algo_impls.shared.vec_env.utils import import_for_env_id, is_microrts
 from rl_algo_impls.vpg.policy import VPGActorCritic
+from rl_algo_impls.vpg.vpg import VanillaPolicyGradient
 from rl_algo_impls.wrappers.vectorable_wrapper import VecEnv, single_observation_space
 
 ALGOS: Dict[str, Type[Algorithm]] = {
@@ -81,16 +81,19 @@ def load_hyperparams(algo: str, env_id: str) -> Hyperparams:
     if env_id in hyperparams_dict:
         return Hyperparams(**hyperparams_dict[env_id])
 
-    if "BulletEnv" in env_id:
-        import pybullet_envs
+    import_for_env_id(env_id)
     spec = gym.spec(env_id)
-    if "AtariEnv" in str(spec.entry_point) and "_atari" in hyperparams_dict:
+    entry_point_name = str(spec.entry_point)  # type: ignore
+    if "AtariEnv" in entry_point_name and "_atari" in hyperparams_dict:
         return Hyperparams(**hyperparams_dict["_atari"])
+    elif "gym_microrts" in entry_point_name and "_microrts" in hyperparams_dict:
+        return Hyperparams(**hyperparams_dict["_microrts"])
     else:
         raise ValueError(f"{env_id} not specified in {algo} hyperparameters file")
 
 
-def get_device(device: str, env: VecEnv) -> torch.device:
+def get_device(config: Config, env: VecEnv) -> torch.device:
+    device = config.device
     # cuda by default
     if device == "auto":
         device = "cuda"
@@ -108,6 +111,16 @@ def get_device(device: str, env: VecEnv) -> torch.device:
             device = "cpu"
         elif isinstance(obs_space, Box) and len(obs_space.shape) == 1:
             device = "cpu"
+        if is_microrts(config):
+            try:
+                from gym_microrts.envs.vec_env import MicroRTSGridModeVecEnv
+
+                # Models that move more than one unit at a time should use mps
+                if not isinstance(env.unwrapped, MicroRTSGridModeVecEnv):
+                    device = "cpu"
+            except ModuleNotFoundError:
+                # Likely on gym_microrts v0.0.2 to match ppo-implementation-details
+                device = "cpu"
     print(f"Device: {device}")
     return torch.device(device)
 
@@ -187,6 +200,8 @@ def hparam_dict(
                     flattened[key] = str(sv)
                 else:
                     flattened[key] = sv
+        elif isinstance(v, list):
+            flattened[k] = json.dumps(v)
         else:
             flattened[k] = v  # type: ignore
     return flattened  # type: ignore

rl_algo_impls/runner/train.py

@@ -5,26 +5,26 @@ os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
 
 import dataclasses
 import shutil
-import wandb
-import yaml
-
 from dataclasses import asdict, dataclass
-from torch.utils.tensorboard.writer import SummaryWriter
 from typing import Any, Dict, Optional, Sequence
 
-from rl_algo_impls.shared.callbacks.eval_callback import EvalCallback
+import yaml
+from torch.utils.tensorboard.writer import SummaryWriter
+
+import wandb
 from rl_algo_impls.runner.config import Config, EnvHyperparams, RunArgs
-from rl_algo_impls.runner.env import make_env, make_eval_env
 from rl_algo_impls.runner.running_utils import (
     ALGOS,
-    load_hyperparams,
-    set_seeds,
     get_device,
+    hparam_dict,
+    load_hyperparams,
     make_policy,
     plot_eval_callback,
-    hparam_dict,
+    set_seeds,
 )
+from rl_algo_impls.shared.callbacks.eval_callback import EvalCallback
 from rl_algo_impls.shared.stats import EpisodesStats
+from rl_algo_impls.shared.vec_env import make_env, make_eval_env
 
 
 @dataclass
@@ -65,7 +65,7 @@ def train(args: TrainArgs):
     env = make_env(
         config, EnvHyperparams(**config.env_hyperparams), tb_writer=tb_writer
     )
-    device = get_device(config.device, env)
+    device = get_device(config, env)
     policy = make_policy(args.algo, env, device, **config.policy_hyperparams)
     algo = ALGOS[args.algo](policy, env, device, tb_writer, **config.algo_hyperparams)
 
@@ -94,6 +94,7 @@ def train(args: TrainArgs):
         if record_best_videos
         else None,
         best_video_dir=config.best_videos_dir,
+        additional_keys_to_log=config.additional_keys_to_log,
     )
     algo.learn(config.n_timesteps, callback=callback)
 

rl_algo_impls/shared/actor/__init__.py

@@ -0,0 +1,2 @@
+from rl_algo_impls.shared.actor.actor import Actor, PiForward
+from rl_algo_impls.shared.actor.make_actor import actor_head

rl_algo_impls/shared/actor/actor.py

@@ -0,0 +1,42 @@
+from abc import ABC, abstractmethod
+from typing import NamedTuple, Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+from torch.distributions import Distribution
+
+
+class PiForward(NamedTuple):
+    pi: Distribution
+    logp_a: Optional[torch.Tensor]
+    entropy: Optional[torch.Tensor]
+
+
+class Actor(nn.Module, ABC):
+    @abstractmethod
+    def forward(
+        self,
+        obs: torch.Tensor,
+        actions: Optional[torch.Tensor] = None,
+        action_masks: Optional[torch.Tensor] = None,
+    ) -> PiForward:
+        ...
+
+    def sample_weights(self, batch_size: int = 1) -> None:
+        pass
+
+    @property
+    @abstractmethod
+    def action_shape(self) -> Tuple[int, ...]:
+        ...
+
+    def pi_forward(
+        self, distribution: Distribution, actions: Optional[torch.Tensor] = None
+    ) -> PiForward:
+        logp_a = None
+        entropy = None
+        if actions is not None:
+            logp_a = distribution.log_prob(actions)
+            entropy = distribution.entropy()
+        return PiForward(distribution, logp_a, entropy)

rl_algo_impls/shared/actor/categorical.py

@@ -0,0 +1,64 @@
+from typing import Optional, Tuple, Type
+
+import torch
+import torch.nn as nn
+from torch.distributions import Categorical
+
+from rl_algo_impls.shared.actor import Actor, PiForward
+from rl_algo_impls.shared.module.module import mlp
+
+
+class MaskedCategorical(Categorical):
+    def __init__(
+        self,
+        probs=None,
+        logits=None,
+        validate_args=None,
+        mask: Optional[torch.Tensor] = None,
+    ):
+        if mask is not None:
+            assert logits is not None, "mask requires logits and not probs"
+            logits = torch.where(mask, logits, -1e8)
+        self.mask = mask
+        super().__init__(probs, logits, validate_args)
+
+    def entropy(self) -> torch.Tensor:
+        if self.mask is None:
+            return super().entropy()
+        # If mask set, then use approximation for entropy
+        p_log_p = self.logits * self.probs  # type: ignore
+        masked = torch.where(self.mask, p_log_p, 0)
+        return -masked.sum(-1)
+
+
+class CategoricalActorHead(Actor):
+    def __init__(
+        self,
+        act_dim: int,
+        in_dim: int,
+        hidden_sizes: Tuple[int, ...] = (32,),
+        activation: Type[nn.Module] = nn.Tanh,
+        init_layers_orthogonal: bool = True,
+    ) -> None:
+        super().__init__()
+        layer_sizes = (in_dim,) + hidden_sizes + (act_dim,)
+        self._fc = mlp(
+            layer_sizes,
+            activation,
+            init_layers_orthogonal=init_layers_orthogonal,
+            final_layer_gain=0.01,
+        )
+
+    def forward(
+        self,
+        obs: torch.Tensor,
+        actions: Optional[torch.Tensor] = None,
+        action_masks: Optional[torch.Tensor] = None,
+    ) -> PiForward:
+        logits = self._fc(obs)
+        pi = MaskedCategorical(logits=logits, mask=action_masks)
+        return self.pi_forward(pi, actions)
+
+    @property
+    def action_shape(self) -> Tuple[int, ...]:
+        return ()

rl_algo_impls/shared/actor/gaussian.py

@@ -0,0 +1,61 @@
+from typing import Optional, Tuple, Type
+
+import torch
+import torch.nn as nn
+from torch.distributions import Distribution, Normal
+
+from rl_algo_impls.shared.actor.actor import Actor, PiForward
+from rl_algo_impls.shared.module.module import mlp
+
+
+class GaussianDistribution(Normal):
+    def log_prob(self, a: torch.Tensor) -> torch.Tensor:
+        return super().log_prob(a).sum(axis=-1)
+
+    def sample(self) -> torch.Tensor:
+        return self.rsample()
+
+
+class GaussianActorHead(Actor):
+    def __init__(
+        self,
+        act_dim: int,
+        in_dim: int,
+        hidden_sizes: Tuple[int, ...] = (32,),
+        activation: Type[nn.Module] = nn.Tanh,
+        init_layers_orthogonal: bool = True,
+        log_std_init: float = -0.5,
+    ) -> None:
+        super().__init__()
+        self.act_dim = act_dim
+        layer_sizes = (in_dim,) + hidden_sizes + (act_dim,)
+        self.mu_net = mlp(
+            layer_sizes,
+            activation,
+            init_layers_orthogonal=init_layers_orthogonal,
+            final_layer_gain=0.01,
+        )
+        self.log_std = nn.Parameter(
+            torch.ones(act_dim, dtype=torch.float32) * log_std_init
+        )
+
+    def _distribution(self, obs: torch.Tensor) -> Distribution:
+        mu = self.mu_net(obs)
+        std = torch.exp(self.log_std)
+        return GaussianDistribution(mu, std)
+
+    def forward(
+        self,
+        obs: torch.Tensor,
+        actions: Optional[torch.Tensor] = None,
+        action_masks: Optional[torch.Tensor] = None,
+    ) -> PiForward:
+        assert (
+            not action_masks
+        ), f"{self.__class__.__name__} does not support action_masks"
+        pi = self._distribution(obs)
+        return self.pi_forward(pi, actions)
+
+    @property
+    def action_shape(self) -> Tuple[int, ...]:
+        return (self.act_dim,)

rl_algo_impls/shared/actor/gridnet.py

@@ -0,0 +1,108 @@
+from typing import Dict, Optional, Tuple, Type
+
+import numpy as np
+import torch
+import torch.nn as nn
+from numpy.typing import NDArray
+from torch.distributions import Distribution, constraints
+
+from rl_algo_impls.shared.actor import Actor, PiForward
+from rl_algo_impls.shared.actor.categorical import MaskedCategorical
+from rl_algo_impls.shared.encoder import EncoderOutDim
+from rl_algo_impls.shared.module.module import mlp
+
+
+class GridnetDistribution(Distribution):
+    def __init__(
+        self,
+        map_size: int,
+        action_vec: NDArray[np.int64],
+        logits: torch.Tensor,
+        masks: torch.Tensor,
+        validate_args: Optional[bool] = None,
+    ) -> None:
+        self.map_size = map_size
+        self.action_vec = action_vec
+
+        masks = masks.view(-1, masks.shape[-1])
+        split_masks = torch.split(masks[:, 1:], action_vec.tolist(), dim=1)
+
+        grid_logits = logits.reshape(-1, action_vec.sum())
+        split_logits = torch.split(grid_logits, action_vec.tolist(), dim=1)
+        self.categoricals = [
+            MaskedCategorical(logits=lg, validate_args=validate_args, mask=m)
+            for lg, m in zip(split_logits, split_masks)
+        ]
+
+        batch_shape = logits.size()[:-1] if logits.ndimension() > 1 else torch.Size()
+        super().__init__(batch_shape=batch_shape, validate_args=validate_args)
+
+    def log_prob(self, action: torch.Tensor) -> torch.Tensor:
+        prob_stack = torch.stack(
+            [
+                c.log_prob(a)
+                for a, c in zip(action.view(-1, action.shape[-1]).T, self.categoricals)
+            ],
+            dim=-1,
+        )
+        logprob = prob_stack.view(-1, self.map_size, len(self.action_vec))
+        return logprob.sum(dim=(1, 2))
+
+    def entropy(self) -> torch.Tensor:
+        ent = torch.stack([c.entropy() for c in self.categoricals], dim=-1)
+        ent = ent.view(-1, self.map_size, len(self.action_vec))
+        return ent.sum(dim=(1, 2))
+
+    def sample(self, sample_shape: torch.Size = torch.Size()) -> torch.Tensor:
+        s = torch.stack([c.sample(sample_shape) for c in self.categoricals], dim=-1)
+        return s.view(-1, self.map_size, len(self.action_vec))
+
+    @property
+    def mode(self) -> torch.Tensor:
+        m = torch.stack([c.mode for c in self.categoricals], dim=-1)
+        return m.view(-1, self.map_size, len(self.action_vec))
+
+    @property
+    def arg_constraints(self) -> Dict[str, constraints.Constraint]:
+        # Constraints handled by child distributions in dist
+        return {}
+
+
+class GridnetActorHead(Actor):
+    def __init__(
+        self,
+        map_size: int,
+        action_vec: NDArray[np.int64],
+        in_dim: EncoderOutDim,
+        hidden_sizes: Tuple[int, ...] = (32,),
+        activation: Type[nn.Module] = nn.ReLU,
+        init_layers_orthogonal: bool = True,
+    ) -> None:
+        super().__init__()
+        self.map_size = map_size
+        self.action_vec = action_vec
+        assert isinstance(in_dim, int)
+        layer_sizes = (in_dim,) + hidden_sizes + (map_size * action_vec.sum(),)
+        self._fc = mlp(
+            layer_sizes,
+            activation,
+            init_layers_orthogonal=init_layers_orthogonal,
+            final_layer_gain=0.01,
+        )
+
+    def forward(
+        self,
+        obs: torch.Tensor,
+        actions: Optional[torch.Tensor] = None,
+        action_masks: Optional[torch.Tensor] = None,
+    ) -> PiForward:
+        assert (
+            action_masks is not None
+        ), f"No mask case unhandled in {self.__class__.__name__}"
+        logits = self._fc(obs)
+        pi = GridnetDistribution(self.map_size, self.action_vec, logits, action_masks)
+        return self.pi_forward(pi, actions)
+
+    @property
+    def action_shape(self) -> Tuple[int, ...]:
+        return (self.map_size, len(self.action_vec))

rl_algo_impls/shared/actor/gridnet_decoder.py

@@ -0,0 +1,80 @@
+from typing import Optional, Tuple, Type
+
+import numpy as np
+import torch
+import torch.nn as nn
+from numpy.typing import NDArray
+
+from rl_algo_impls.shared.actor import Actor, PiForward
+from rl_algo_impls.shared.actor.categorical import MaskedCategorical
+from rl_algo_impls.shared.actor.gridnet import GridnetDistribution
+from rl_algo_impls.shared.encoder import EncoderOutDim
+from rl_algo_impls.shared.module.module import layer_init
+
+
+class Transpose(nn.Module):
+    def __init__(self, permutation: Tuple[int, ...]) -> None:
+        super().__init__()
+        self.permutation = permutation
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x.permute(self.permutation)
+
+
+class GridnetDecoder(Actor):
+    def __init__(
+        self,
+        map_size: int,
+        action_vec: NDArray[np.int64],
+        in_dim: EncoderOutDim,
+        activation: Type[nn.Module] = nn.ReLU,
+        init_layers_orthogonal: bool = True,
+    ) -> None:
+        super().__init__()
+        self.map_size = map_size
+        self.action_vec = action_vec
+        assert isinstance(in_dim, tuple)
+        self.deconv = nn.Sequential(
+            layer_init(
+                nn.ConvTranspose2d(
+                    in_dim[0], 128, 3, stride=2, padding=1, output_padding=1
+                ),
+                init_layers_orthogonal=init_layers_orthogonal,
+            ),
+            activation(),
+            layer_init(
+                nn.ConvTranspose2d(128, 64, 3, stride=2, padding=1, output_padding=1),
+                init_layers_orthogonal=init_layers_orthogonal,
+            ),
+            activation(),
+            layer_init(
+                nn.ConvTranspose2d(64, 32, 3, stride=2, padding=1, output_padding=1),
+                init_layers_orthogonal=init_layers_orthogonal,
+            ),
+            activation(),
+            layer_init(
+                nn.ConvTranspose2d(
+                    32, action_vec.sum(), 3, stride=2, padding=1, output_padding=1
+                ),
+                init_layers_orthogonal=init_layers_orthogonal,
+                std=0.01,
+            ),
+            Transpose((0, 2, 3, 1)),
+        )
+
+    def forward(
+        self,
+        obs: torch.Tensor,
+        actions: Optional[torch.Tensor] = None,
+        action_masks: Optional[torch.Tensor] = None,
+    ) -> PiForward:
+        assert (
+            action_masks is not None
+        ), f"No mask case unhandled in {self.__class__.__name__}"
+        logits = self.deconv(obs)
+        pi = GridnetDistribution(self.map_size, self.action_vec, logits, action_masks)
+        return self.pi_forward(pi, actions)
+
+    @property
+    def action_shape(self) -> Tuple[int, ...]:
+        return (self.map_size, len(self.action_vec))

rl_algo_impls/shared/actor/make_actor.py

@@ -0,0 +1,95 @@
+from typing import Tuple, Type
+
+import gym
+import torch.nn as nn
+from gym.spaces import Box, Discrete, MultiDiscrete
+
+from rl_algo_impls.shared.actor.actor import Actor
+from rl_algo_impls.shared.actor.categorical import CategoricalActorHead
+from rl_algo_impls.shared.actor.gaussian import GaussianActorHead
+from rl_algo_impls.shared.actor.gridnet import GridnetActorHead
+from rl_algo_impls.shared.actor.gridnet_decoder import GridnetDecoder
+from rl_algo_impls.shared.actor.multi_discrete import MultiDiscreteActorHead
+from rl_algo_impls.shared.actor.state_dependent_noise import (
+    StateDependentNoiseActorHead,
+)
+from rl_algo_impls.shared.encoder import EncoderOutDim
+
+
+def actor_head(
+    action_space: gym.Space,
+    in_dim: EncoderOutDim,
+    hidden_sizes: Tuple[int, ...],
+    init_layers_orthogonal: bool,
+    activation: Type[nn.Module],
+    log_std_init: float = -0.5,
+    use_sde: bool = False,
+    full_std: bool = True,
+    squash_output: bool = False,
+    actor_head_style: str = "single",
+) -> Actor:
+    assert not use_sde or isinstance(
+        action_space, Box
+    ), "use_sde only valid if Box action_space"
+    assert not squash_output or use_sde, "squash_output only valid if use_sde"
+    if isinstance(action_space, Discrete):
+        assert isinstance(in_dim, int)
+        return CategoricalActorHead(
+            action_space.n,  # type: ignore
+            in_dim=in_dim,
+            hidden_sizes=hidden_sizes,
+            activation=activation,
+            init_layers_orthogonal=init_layers_orthogonal,
+        )
+    elif isinstance(action_space, Box):
+        assert isinstance(in_dim, int)
+        if use_sde:
+            return StateDependentNoiseActorHead(
+                action_space.shape[0],  # type: ignore
+                in_dim=in_dim,
+                hidden_sizes=hidden_sizes,
+                activation=activation,
+                init_layers_orthogonal=init_layers_orthogonal,
+                log_std_init=log_std_init,
+                full_std=full_std,
+                squash_output=squash_output,
+            )
+        else:
+            return GaussianActorHead(
+                action_space.shape[0],  # type: ignore
+                in_dim=in_dim,
+                hidden_sizes=hidden_sizes,
+                activation=activation,
+                init_layers_orthogonal=init_layers_orthogonal,
+                log_std_init=log_std_init,
+            )
+    elif isinstance(action_space, MultiDiscrete):
+        if actor_head_style == "single":
+            return MultiDiscreteActorHead(
+                action_space.nvec,  # type: ignore
+                in_dim=in_dim,
+                hidden_sizes=hidden_sizes,
+                activation=activation,
+                init_layers_orthogonal=init_layers_orthogonal,
+            )
+        elif actor_head_style == "gridnet":
+            return GridnetActorHead(
+                action_space.nvec[0],  # type: ignore
+                action_space.nvec[1:],  # type: ignore
+                in_dim=in_dim,
+                hidden_sizes=hidden_sizes,
+                activation=activation,
+                init_layers_orthogonal=init_layers_orthogonal,
+            )
+        elif actor_head_style == "gridnet_decoder":
+            return GridnetDecoder(
+                action_space.nvec[0],  # type: ignore
+                action_space.nvec[1:],  # type: ignore
+                in_dim=in_dim,
+                activation=activation,
+                init_layers_orthogonal=init_layers_orthogonal,
+            )
+        else:
+            raise ValueError(f"Doesn't support actor_head_style {actor_head_style}")
+    else:
+        raise ValueError(f"Unsupported action space: {action_space}")

rl_algo_impls/shared/actor/multi_discrete.py

@@ -0,0 +1,101 @@
+from typing import Dict, Optional, Tuple, Type
+
+import numpy as np
+import torch
+import torch.nn as nn
+from numpy.typing import NDArray
+from torch.distributions import Distribution, constraints
+
+from rl_algo_impls.shared.actor.actor import Actor, PiForward
+from rl_algo_impls.shared.actor.categorical import MaskedCategorical
+from rl_algo_impls.shared.encoder import EncoderOutDim
+from rl_algo_impls.shared.module.module import mlp
+
+
+class MultiCategorical(Distribution):
+    def __init__(
+        self,
+        nvec: NDArray[np.int64],
+        probs=None,
+        logits=None,
+        validate_args=None,
+        masks: Optional[torch.Tensor] = None,
+    ):
+        # Either probs or logits should be set
+        assert (probs is None) != (logits is None)
+        masks_split = (
+            torch.split(masks, nvec.tolist(), dim=1)
+            if masks is not None
+            else [None] * len(nvec)
+        )
+        if probs:
+            self.dists = [
+                MaskedCategorical(probs=p, validate_args=validate_args, mask=m)
+                for p, m in zip(torch.split(probs, nvec.tolist(), dim=1), masks_split)
+            ]
+            param = probs
+        else:
+            assert logits is not None
+            self.dists = [
+                MaskedCategorical(logits=lg, validate_args=validate_args, mask=m)
+                for lg, m in zip(torch.split(logits, nvec.tolist(), dim=1), masks_split)
+            ]
+            param = logits
+        batch_shape = param.size()[:-1] if param.ndimension() > 1 else torch.Size()
+        super().__init__(batch_shape=batch_shape, validate_args=validate_args)
+
+    def log_prob(self, action: torch.Tensor) -> torch.Tensor:
+        prob_stack = torch.stack(
+            [c.log_prob(a) for a, c in zip(action.T, self.dists)], dim=-1
+        )
+        return prob_stack.sum(dim=-1)
+
+    def entropy(self) -> torch.Tensor:
+        return torch.stack([c.entropy() for c in self.dists], dim=-1).sum(dim=-1)
+
+    def sample(self, sample_shape: torch.Size = torch.Size()) -> torch.Tensor:
+        return torch.stack([c.sample(sample_shape) for c in self.dists], dim=-1)
+
+    @property
+    def mode(self) -> torch.Tensor:
+        return torch.stack([c.mode for c in self.dists], dim=-1)
+
+    @property
+    def arg_constraints(self) -> Dict[str, constraints.Constraint]:
+        # Constraints handled by child distributions in dist
+        return {}
+
+
+class MultiDiscreteActorHead(Actor):
+    def __init__(
+        self,
+        nvec: NDArray[np.int64],
+        in_dim: EncoderOutDim,
+        hidden_sizes: Tuple[int, ...] = (32,),
+        activation: Type[nn.Module] = nn.ReLU,
+        init_layers_orthogonal: bool = True,
+    ) -> None:
+        super().__init__()
+        self.nvec = nvec
+        assert isinstance(in_dim, int)
+        layer_sizes = (in_dim,) + hidden_sizes + (nvec.sum(),)
+        self._fc = mlp(
+            layer_sizes,
+            activation,
+            init_layers_orthogonal=init_layers_orthogonal,
+            final_layer_gain=0.01,
+        )
+
+    def forward(
+        self,
+        obs: torch.Tensor,
+        actions: Optional[torch.Tensor] = None,
+        action_masks: Optional[torch.Tensor] = None,
+    ) -> PiForward:
+        logits = self._fc(obs)
+        pi = MultiCategorical(self.nvec, logits=logits, masks=action_masks)
+        return self.pi_forward(pi, actions)
+
+    @property
+    def action_shape(self) -> Tuple[int, ...]:
+        return (len(self.nvec),)

rl_algo_impls/shared/{policy/actor.py → actor/state_dependent_noise.py}

@@ -1,99 +1,13 @@
-import gym
+from typing import Optional, Tuple, Type, TypeVar, Union
+
 import torch
 import torch.nn as nn
+from torch.distributions import Distribution, Normal
 
-from abc import ABC, abstractmethod
-from gym.spaces import Box, Discrete
-from torch.distributions import Categorical, Distribution, Normal
-from typing import NamedTuple, Optional, Sequence, Type, TypeVar, Union
-
+from rl_algo_impls.shared.actor.actor import Actor, PiForward
 from rl_algo_impls.shared.module.module import mlp
 
 
-class PiForward(NamedTuple):
-    pi: Distribution
-    logp_a: Optional[torch.Tensor]
-    entropy: Optional[torch.Tensor]
-
-
-class Actor(nn.Module, ABC):
-    @abstractmethod
-    def forward(self, obs: torch.Tensor, a: Optional[torch.Tensor] = None) -> PiForward:
-        ...
-
-
-class CategoricalActorHead(Actor):
-    def __init__(
-        self,
-        act_dim: int,
-        hidden_sizes: Sequence[int] = (32,),
-        activation: Type[nn.Module] = nn.Tanh,
-        init_layers_orthogonal: bool = True,
-    ) -> None:
-        super().__init__()
-        layer_sizes = tuple(hidden_sizes) + (act_dim,)
-        self._fc = mlp(
-            layer_sizes,
-            activation,
-            init_layers_orthogonal=init_layers_orthogonal,
-            final_layer_gain=0.01,
-        )
-
-    def forward(self, obs: torch.Tensor, a: Optional[torch.Tensor] = None) -> PiForward:
-        logits = self._fc(obs)
-        pi = Categorical(logits=logits)
-        logp_a = None
-        entropy = None
-        if a is not None:
-            logp_a = pi.log_prob(a)
-            entropy = pi.entropy()
-        return PiForward(pi, logp_a, entropy)
-
-
-class GaussianDistribution(Normal):
-    def log_prob(self, a: torch.Tensor) -> torch.Tensor:
-        return super().log_prob(a).sum(axis=-1)
-
-    def sample(self) -> torch.Tensor:
-        return self.rsample()
-
-
-class GaussianActorHead(Actor):
-    def __init__(
-        self,
-        act_dim: int,
-        hidden_sizes: Sequence[int] = (32,),
-        activation: Type[nn.Module] = nn.Tanh,
-        init_layers_orthogonal: bool = True,
-        log_std_init: float = -0.5,
-    ) -> None:
-        super().__init__()
-        layer_sizes = tuple(hidden_sizes) + (act_dim,)
-        self.mu_net = mlp(
-            layer_sizes,
-            activation,
-            init_layers_orthogonal=init_layers_orthogonal,
-            final_layer_gain=0.01,
-        )
-        self.log_std = nn.Parameter(
-            torch.ones(act_dim, dtype=torch.float32) * log_std_init
-        )
-
-    def _distribution(self, obs: torch.Tensor) -> Distribution:
-        mu = self.mu_net(obs)
-        std = torch.exp(self.log_std)
-        return GaussianDistribution(mu, std)
-
-    def forward(self, obs: torch.Tensor, a: Optional[torch.Tensor] = None) -> PiForward:
-        pi = self._distribution(obs)
-        logp_a = None
-        entropy = None
-        if a is not None:
-            logp_a = pi.log_prob(a)
-            entropy = pi.entropy()
-        return PiForward(pi, logp_a, entropy)
-
-
 class TanhBijector:
     def __init__(self, epsilon: float = 1e-6) -> None:
         self.epsilon = epsilon
@@ -173,7 +87,8 @@ class StateDependentNoiseActorHead(Actor):
     def __init__(
         self,
         act_dim: int,
-        hidden_sizes: Sequence[int] = (32,),
+        in_dim: int,
+        hidden_sizes: Tuple[int, ...] = (32,),
         activation: Type[nn.Module] = nn.Tanh,
         init_layers_orthogonal: bool = True,
         log_std_init: float = -0.5,
@@ -183,7 +98,7 @@ class StateDependentNoiseActorHead(Actor):
     ) -> None:
         super().__init__()
         self.act_dim = act_dim
-        layer_sizes = tuple(hidden_sizes) + (self.act_dim,)
+        layer_sizes = (in_dim,) + hidden_sizes + (act_dim,)
         if len(layer_sizes) == 2:
             self.latent_net = nn.Identity()
         elif len(layer_sizes) > 2:
@@ -193,8 +108,6 @@ class StateDependentNoiseActorHead(Actor):
                 output_activation=activation,
                 init_layers_orthogonal=init_layers_orthogonal,
             )
-        else:
-            raise ValueError("hidden_sizes must be of at least length 1")
         self.mu_net = mlp(
             layer_sizes[-2:],
             activation,
@@ -202,7 +115,7 @@ class StateDependentNoiseActorHead(Actor):
             final_layer_gain=0.01,
         )
         self.full_std = full_std
-        std_dim = (hidden_sizes[-1], act_dim if self.full_std else 1)
+        std_dim = (layer_sizes[-2], act_dim if self.full_std else 1)
         self.log_std = nn.Parameter(
             torch.ones(std_dim, dtype=torch.float32) * log_std_init
         )
@@ -249,14 +162,17 @@ class StateDependentNoiseActorHead(Actor):
             ones = ones.to(self.device)
         return ones * std
 
-    def forward(self, obs: torch.Tensor, a: Optional[torch.Tensor] = None) -> PiForward:
+    def forward(
+        self,
+        obs: torch.Tensor,
+        actions: Optional[torch.Tensor] = None,
+        action_masks: Optional[torch.Tensor] = None,
+    ) -> PiForward:
+        assert (
+            not action_masks
+        ), f"{self.__class__.__name__} does not support action_masks"
         pi = self._distribution(obs)
-        logp_a = None
-        entropy = None
-        if a is not None:
-            logp_a = pi.log_prob(a)
-            entropy = -logp_a if self.bijector else sum_independent_dims(pi.entropy())
-        return PiForward(pi, logp_a, entropy)
+        return self.pi_forward(pi, actions)
 
     def sample_weights(self, batch_size: int = 1) -> None:
         std = self._get_std()
@@ -265,46 +181,20 @@ class StateDependentNoiseActorHead(Actor):
         self.exploration_mat = weights_dist.rsample()
         self.exploration_matrices = weights_dist.rsample(torch.Size((batch_size,)))
 
+    @property
+    def action_shape(self) -> Tuple[int, ...]:
+        return (self.act_dim,)
 
-def actor_head(
-    action_space: gym.Space,
-    hidden_sizes: Sequence[int],
-    init_layers_orthogonal: bool,
-    activation: Type[nn.Module],
-    log_std_init: float = -0.5,
-    use_sde: bool = False,
-    full_std: bool = True,
-    squash_output: bool = False,
-) -> Actor:
-    assert not use_sde or isinstance(
-        action_space, Box
-    ), "use_sde only valid if Box action_space"
-    assert not squash_output or use_sde, "squash_output only valid if use_sde"
-    if isinstance(action_space, Discrete):
-        return CategoricalActorHead(
-            action_space.n,
-            hidden_sizes=hidden_sizes,
-            activation=activation,
-            init_layers_orthogonal=init_layers_orthogonal,
-        )
-    elif isinstance(action_space, Box):
-        if use_sde:
-            return StateDependentNoiseActorHead(
-                action_space.shape[0],
-                hidden_sizes=hidden_sizes,
-                activation=activation,
-                init_layers_orthogonal=init_layers_orthogonal,
-                log_std_init=log_std_init,
-                full_std=full_std,
-                squash_output=squash_output,
-            )
-        else:
-            return GaussianActorHead(
-                action_space.shape[0],
-                hidden_sizes=hidden_sizes,
-                activation=activation,
-                init_layers_orthogonal=init_layers_orthogonal,
-                log_std_init=log_std_init,
+    def pi_forward(
+        self, distribution: Distribution, actions: Optional[torch.Tensor] = None
+    ) -> PiForward:
+        logp_a = None
+        entropy = None
+        if actions is not None:
+            logp_a = distribution.log_prob(actions)
+            entropy = (
+                -logp_a
+                if self.bijector
+                else sum_independent_dims(distribution.entropy())
             )
-    else:
-        raise ValueError(f"Unsupported action space: {action_space}")
+        return PiForward(distribution, logp_a, entropy)

rl_algo_impls/shared/callbacks/eval_callback.py

@@ -1,14 +1,15 @@
 import itertools
-import numpy as np
 import os
-
 from time import perf_counter
+from typing import Dict, List, Optional, Union
+
+import numpy as np
 from torch.utils.tensorboard.writer import SummaryWriter
-from typing import List, Optional, Union
 
 from rl_algo_impls.shared.callbacks.callback import Callback
 from rl_algo_impls.shared.policy.policy import Policy
 from rl_algo_impls.shared.stats import Episode, EpisodeAccumulator, EpisodesStats
+from rl_algo_impls.wrappers.action_mask_wrapper import find_action_masker
 from rl_algo_impls.wrappers.vec_episode_recorder import VecEpisodeRecorder
 from rl_algo_impls.wrappers.vectorable_wrapper import VecEnv
 
@@ -20,6 +21,7 @@ class EvaluateAccumulator(EpisodeAccumulator):
         goal_episodes: int,
         print_returns: bool = True,
         ignore_first_episode: bool = False,
+        additional_keys_to_log: Optional[List[str]] = None,
     ):
         super().__init__(num_envs)
         self.completed_episodes_by_env_idx = [[] for _ in range(num_envs)]
@@ -36,8 +38,11 @@ class EvaluateAccumulator(EpisodeAccumulator):
             self.should_record_done = should_record_done
         else:
             self.should_record_done = lambda idx: True
+        self.additional_keys_to_log = additional_keys_to_log
 
-    def on_done(self, ep_idx: int, episode: Episode) -> None:
+    def on_done(self, ep_idx: int, episode: Episode, info: Dict) -> None:
+        if self.additional_keys_to_log:
+            episode.info = {k: info[k] for k in self.additional_keys_to_log}
         if (
             self.should_record_done(ep_idx)
             and len(self.completed_episodes_by_env_idx[ep_idx])
@@ -74,19 +79,29 @@ def evaluate(
     deterministic: bool = True,
     print_returns: bool = True,
     ignore_first_episode: bool = False,
+    additional_keys_to_log: Optional[List[str]] = None,
 ) -> EpisodesStats:
     policy.sync_normalization(env)
     policy.eval()
 
     episodes = EvaluateAccumulator(
-        env.num_envs, n_episodes, print_returns, ignore_first_episode
+        env.num_envs,
+        n_episodes,
+        print_returns,
+        ignore_first_episode,
+        additional_keys_to_log=additional_keys_to_log,
     )
 
     obs = env.reset()
+    action_masker = find_action_masker(env)
     while not episodes.is_done():
-        act = policy.act(obs, deterministic=deterministic)
-        obs, rew, done, _ = env.step(act)
-        episodes.step(rew, done)
+        act = policy.act(
+            obs,
+            deterministic=deterministic,
+            action_masks=action_masker.action_masks() if action_masker else None,
+        )
+        obs, rew, done, info = env.step(act)
+        episodes.step(rew, done, info)
         if render:
             env.render()
     stats = EpisodesStats(episodes.episodes)
@@ -111,6 +126,7 @@ class EvalCallback(Callback):
         best_video_dir: Optional[str] = None,
         max_video_length: int = 3600,
         ignore_first_episode: bool = False,
+        additional_keys_to_log: Optional[List[str]] = None,
     ) -> None:
         super().__init__()
         self.policy = policy
@@ -133,8 +149,8 @@ class EvalCallback(Callback):
             os.makedirs(best_video_dir, exist_ok=True)
         self.max_video_length = max_video_length
         self.best_video_base_path = None
-
         self.ignore_first_episode = ignore_first_episode
+        self.additional_keys_to_log = additional_keys_to_log
 
     def on_step(self, timesteps_elapsed: int = 1) -> bool:
         super().on_step(timesteps_elapsed)
@@ -153,6 +169,7 @@ class EvalCallback(Callback):
             deterministic=self.deterministic,
             print_returns=print_returns or False,
             ignore_first_episode=self.ignore_first_episode,
+            additional_keys_to_log=self.additional_keys_to_log,
         )
         end_time = perf_counter()
         self.tb_writer.add_scalar(

rl_algo_impls/shared/encoder/__init__.py

@@ -0,0 +1,2 @@
+from rl_algo_impls.shared.encoder.cnn import EncoderOutDim
+from rl_algo_impls.shared.encoder.encoder import Encoder

rl_algo_impls/shared/encoder/cnn.py

@@ -0,0 +1,72 @@
+from abc import ABC, abstractmethod
+from typing import Optional, Tuple, Type, Union
+
+import gym
+import numpy as np
+import torch
+import torch.nn as nn
+
+from rl_algo_impls.shared.module.module import layer_init
+
+EncoderOutDim = Union[int, Tuple[int, ...]]
+
+
+class CnnEncoder(nn.Module, ABC):
+    @abstractmethod
+    def __init__(
+        self,
+        obs_space: gym.Space,
+        **kwargs,
+    ) -> None:
+        super().__init__()
+        self.range_size = np.max(obs_space.high) - np.min(obs_space.low)  # type: ignore
+
+    def preprocess(self, obs: torch.Tensor) -> torch.Tensor:
+        if len(obs.shape) == 3:
+            obs = obs.unsqueeze(0)
+        return obs.float() / self.range_size
+
+    def forward(self, obs: torch.Tensor) -> torch.Tensor:
+        return self.preprocess(obs)
+
+    @property
+    @abstractmethod
+    def out_dim(self) -> EncoderOutDim:
+        ...
+
+
+class FlattenedCnnEncoder(CnnEncoder):
+    def __init__(
+        self,
+        obs_space: gym.Space,
+        activation: Type[nn.Module],
+        linear_init_layers_orthogonal: bool,
+        cnn_flatten_dim: int,
+        cnn: nn.Module,
+        **kwargs,
+    ) -> None:
+        super().__init__(obs_space, **kwargs)
+        self.cnn = cnn
+        self.flattened_dim = cnn_flatten_dim
+        with torch.no_grad():
+            cnn_out = torch.flatten(
+                cnn(self.preprocess(torch.as_tensor(obs_space.sample()))), start_dim=1
+            )
+        self.fc = nn.Sequential(
+            nn.Flatten(),
+            layer_init(
+                nn.Linear(cnn_out.shape[1], cnn_flatten_dim),
+                linear_init_layers_orthogonal,
+            ),
+            activation(),
+        )
+
+    def forward(self, obs: torch.Tensor) -> torch.Tensor:
+        x = super().forward(obs)
+        x = self.cnn(x)
+        x = self.fc(x)
+        return x
+
+    @property
+    def out_dim(self) -> EncoderOutDim:
+        return self.flattened_dim

rl_algo_impls/shared/encoder/encoder.py

@@ -0,0 +1,73 @@
+from typing import Dict, Optional, Sequence, Type
+
+import gym
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from gym.spaces import Box, Discrete
+from stable_baselines3.common.preprocessing import get_flattened_obs_dim
+
+from rl_algo_impls.shared.encoder.cnn import CnnEncoder
+from rl_algo_impls.shared.encoder.gridnet_encoder import GridnetEncoder
+from rl_algo_impls.shared.encoder.impala_cnn import ImpalaCnn
+from rl_algo_impls.shared.encoder.microrts_cnn import MicrortsCnn
+from rl_algo_impls.shared.encoder.nature_cnn import NatureCnn
+from rl_algo_impls.shared.module.module import layer_init
+
+CNN_EXTRACTORS_BY_STYLE: Dict[str, Type[CnnEncoder]] = {
+    "nature": NatureCnn,
+    "impala": ImpalaCnn,
+    "microrts": MicrortsCnn,
+    "gridnet_encoder": GridnetEncoder,
+}
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        obs_space: gym.Space,
+        activation: Type[nn.Module],
+        init_layers_orthogonal: bool = False,
+        cnn_flatten_dim: int = 512,
+        cnn_style: str = "nature",
+        cnn_layers_init_orthogonal: Optional[bool] = None,
+        impala_channels: Sequence[int] = (16, 32, 32),
+    ) -> None:
+        super().__init__()
+        if isinstance(obs_space, Box):
+            # Conv2D: (channels, height, width)
+            if len(obs_space.shape) == 3:  # type: ignore
+                self.preprocess = None
+                cnn = CNN_EXTRACTORS_BY_STYLE[cnn_style](
+                    obs_space,
+                    activation=activation,
+                    cnn_init_layers_orthogonal=cnn_layers_init_orthogonal,
+                    linear_init_layers_orthogonal=init_layers_orthogonal,
+                    cnn_flatten_dim=cnn_flatten_dim,
+                    impala_channels=impala_channels,
+                )
+                self.feature_extractor = cnn
+                self.out_dim = cnn.out_dim
+            elif len(obs_space.shape) == 1:  # type: ignore
+
+                def preprocess(obs: torch.Tensor) -> torch.Tensor:
+                    if len(obs.shape) == 1:
+                        obs = obs.unsqueeze(0)
+                    return obs.float()
+
+                self.preprocess = preprocess
+                self.feature_extractor = nn.Flatten()
+                self.out_dim = get_flattened_obs_dim(obs_space)
+            else:
+                raise ValueError(f"Unsupported observation space: {obs_space}")
+        elif isinstance(obs_space, Discrete):
+            self.preprocess = lambda x: F.one_hot(x, obs_space.n).float()
+            self.feature_extractor = nn.Flatten()
+            self.out_dim = obs_space.n  # type: ignore
+        else:
+            raise NotImplementedError
+
+    def forward(self, obs: torch.Tensor) -> torch.Tensor:
+        if self.preprocess:
+            obs = self.preprocess(obs)
+        return self.feature_extractor(obs)

rl_algo_impls/shared/encoder/gridnet_encoder.py

@@ -0,0 +1,64 @@
+from typing import Optional, Tuple, Type, Union
+
+import gym
+import torch
+import torch.nn as nn
+
+from rl_algo_impls.shared.encoder.cnn import CnnEncoder, EncoderOutDim
+from rl_algo_impls.shared.module.module import layer_init
+
+
+class GridnetEncoder(CnnEncoder):
+    """
+    Encoder for encoder-decoder for Gym-MicroRTS
+    """
+
+    def __init__(
+        self,
+        obs_space: gym.Space,
+        activation: Type[nn.Module] = nn.ReLU,
+        cnn_init_layers_orthogonal: Optional[bool] = None,
+        **kwargs
+    ) -> None:
+        if cnn_init_layers_orthogonal is None:
+            cnn_init_layers_orthogonal = True
+        super().__init__(obs_space, **kwargs)
+        in_channels = obs_space.shape[0]  # type: ignore
+        self.encoder = nn.Sequential(
+            layer_init(
+                nn.Conv2d(in_channels, 32, kernel_size=3, padding=1),
+                cnn_init_layers_orthogonal,
+            ),
+            nn.MaxPool2d(3, stride=2, padding=1),
+            activation(),
+            layer_init(
+                nn.Conv2d(32, 64, kernel_size=3, padding=1),
+                cnn_init_layers_orthogonal,
+            ),
+            nn.MaxPool2d(3, stride=2, padding=1),
+            activation(),
+            layer_init(
+                nn.Conv2d(64, 128, kernel_size=3, padding=1),
+                cnn_init_layers_orthogonal,
+            ),
+            nn.MaxPool2d(3, stride=2, padding=1),
+            activation(),
+            layer_init(
+                nn.Conv2d(128, 256, kernel_size=3, padding=1),
+                cnn_init_layers_orthogonal,
+            ),
+            nn.MaxPool2d(3, stride=2, padding=1),
+            activation(),
+        )
+        with torch.no_grad():
+            encoder_out = self.encoder(
+                self.preprocess(torch.as_tensor(obs_space.sample()))  # type: ignore
+            )
+            self._out_dim = encoder_out.shape[1:]
+
+    def forward(self, obs: torch.Tensor) -> torch.Tensor:
+        return self.encoder(super().forward(obs))
+
+    @property
+    def out_dim(self) -> EncoderOutDim:
+        return self._out_dim

rl_algo_impls/shared/encoder/impala_cnn.py

@@ -0,0 +1,92 @@
+from typing import Optional, Sequence, Type
+
+import gym
+import torch
+import torch.nn as nn
+
+from rl_algo_impls.shared.encoder.cnn import FlattenedCnnEncoder
+from rl_algo_impls.shared.module.module import layer_init
+
+
+class ResidualBlock(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        init_layers_orthogonal: bool = False,
+    ) -> None:
+        super().__init__()
+        self.residual = nn.Sequential(
+            activation(),
+            layer_init(
+                nn.Conv2d(channels, channels, 3, padding=1), init_layers_orthogonal
+            ),
+            activation(),
+            layer_init(
+                nn.Conv2d(channels, channels, 3, padding=1), init_layers_orthogonal
+            ),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.residual(x)
+
+
+class ConvSequence(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        init_layers_orthogonal: bool = False,
+    ) -> None:
+        super().__init__()
+        self.seq = nn.Sequential(
+            layer_init(
+                nn.Conv2d(in_channels, out_channels, 3, padding=1),
+                init_layers_orthogonal,
+            ),
+            nn.MaxPool2d(3, stride=2, padding=1),
+            ResidualBlock(out_channels, activation, init_layers_orthogonal),
+            ResidualBlock(out_channels, activation, init_layers_orthogonal),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.seq(x)
+
+
+class ImpalaCnn(FlattenedCnnEncoder):
+    """
+    IMPALA-style CNN architecture
+    """
+
+    def __init__(
+        self,
+        obs_space: gym.Space,
+        activation: Type[nn.Module],
+        cnn_init_layers_orthogonal: Optional[bool],
+        linear_init_layers_orthogonal: bool,
+        cnn_flatten_dim: int,
+        impala_channels: Sequence[int] = (16, 32, 32),
+        **kwargs,
+    ) -> None:
+        if cnn_init_layers_orthogonal is None:
+            cnn_init_layers_orthogonal = False
+        in_channels = obs_space.shape[0]  # type: ignore
+        sequences = []
+        for out_channels in impala_channels:
+            sequences.append(
+                ConvSequence(
+                    in_channels, out_channels, activation, cnn_init_layers_orthogonal
+                )
+            )
+            in_channels = out_channels
+        sequences.append(activation())
+        cnn = nn.Sequential(*sequences)
+        super().__init__(
+            obs_space,
+            activation,
+            linear_init_layers_orthogonal,
+            cnn_flatten_dim,
+            cnn,
+            **kwargs,
+        )

rl_algo_impls/shared/encoder/microrts_cnn.py

@@ -0,0 +1,45 @@
+from typing import Optional, Type
+
+import gym
+import torch
+import torch.nn as nn
+
+from rl_algo_impls.shared.encoder.cnn import FlattenedCnnEncoder
+from rl_algo_impls.shared.module.module import layer_init
+
+
+class MicrortsCnn(FlattenedCnnEncoder):
+    """
+    Base CNN architecture for Gym-MicroRTS
+    """
+
+    def __init__(
+        self,
+        obs_space: gym.Space,
+        activation: Type[nn.Module],
+        cnn_init_layers_orthogonal: Optional[bool],
+        linear_init_layers_orthogonal: bool,
+        cnn_flatten_dim: int,
+        **kwargs,
+    ) -> None:
+        if cnn_init_layers_orthogonal is None:
+            cnn_init_layers_orthogonal = True
+        in_channels = obs_space.shape[0]  # type: ignore
+        cnn = nn.Sequential(
+            layer_init(
+                nn.Conv2d(in_channels, 16, kernel_size=3, stride=2),
+                cnn_init_layers_orthogonal,
+            ),
+            activation(),
+            layer_init(nn.Conv2d(16, 32, kernel_size=2), cnn_init_layers_orthogonal),
+            activation(),
+            nn.Flatten(),
+        )
+        super().__init__(
+            obs_space,
+            activation,
+            linear_init_layers_orthogonal,
+            cnn_flatten_dim,
+            cnn,
+            **kwargs,
+        )

rl_algo_impls/shared/encoder/nature_cnn.py

@@ -0,0 +1,53 @@
+from typing import Optional, Type
+
+import gym
+import torch.nn as nn
+
+from rl_algo_impls.shared.encoder.cnn import FlattenedCnnEncoder
+from rl_algo_impls.shared.module.module import layer_init
+
+
+class NatureCnn(FlattenedCnnEncoder):
+    """
+    CNN from DQN Nature paper: Mnih, Volodymyr, et al.
+    "Human-level control through deep reinforcement learning."
+    Nature 518.7540 (2015): 529-533.
+    """
+
+    def __init__(
+        self,
+        obs_space: gym.Space,
+        activation: Type[nn.Module],
+        cnn_init_layers_orthogonal: Optional[bool],
+        linear_init_layers_orthogonal: bool,
+        cnn_flatten_dim: int,
+        **kwargs,
+    ) -> None:
+        if cnn_init_layers_orthogonal is None:
+            cnn_init_layers_orthogonal = True
+        in_channels = obs_space.shape[0]  # type: ignore
+        cnn = nn.Sequential(
+            layer_init(
+                nn.Conv2d(in_channels, 32, kernel_size=8, stride=4),
+                cnn_init_layers_orthogonal,
+            ),
+            activation(),
+            layer_init(
+                nn.Conv2d(32, 64, kernel_size=4, stride=2),
+                cnn_init_layers_orthogonal,
+            ),
+            activation(),
+            layer_init(
+                nn.Conv2d(64, 64, kernel_size=3, stride=1),
+                cnn_init_layers_orthogonal,
+            ),
+            activation(),
+        )
+        super().__init__(
+            obs_space,
+            activation,
+            linear_init_layers_orthogonal,
+            cnn_flatten_dim,
+            cnn,
+            **kwargs,
+        )

rl_algo_impls/shared/gae.py

@@ -5,6 +5,7 @@ from typing import NamedTuple, Sequence
 
 from rl_algo_impls.shared.policy.on_policy import OnPolicy
 from rl_algo_impls.shared.trajectory import Trajectory
+from rl_algo_impls.wrappers.vectorable_wrapper import VecEnvObs
 
 
 class RtgAdvantage(NamedTuple):
@@ -19,7 +20,7 @@ def discounted_cumsum(x: np.ndarray, gamma: float) -> np.ndarray:
     return dc
 
 
-def compute_advantage(
+def compute_advantage_from_trajectories(
     trajectories: Sequence[Trajectory],
     policy: OnPolicy,
     gamma: float,
@@ -40,7 +41,7 @@ def compute_advantage(
     )
 
 
-def compute_rtg_and_advantage(
+def compute_rtg_and_advantage_from_trajectories(
     trajectories: Sequence[Trajectory],
     policy: OnPolicy,
     gamma: float,
@@ -65,3 +66,29 @@ def compute_rtg_and_advantage(
         ),
         torch.as_tensor(np.concatenate(advantages), dtype=torch.float32, device=device),
     )
+
+
+def compute_advantages(
+    rewards: np.ndarray,
+    values: np.ndarray,
+    episode_starts: np.ndarray,
+    next_episode_starts: np.ndarray,
+    next_obs: VecEnvObs,
+    policy: OnPolicy,
+    gamma: float,
+    gae_lambda: float,
+) -> np.ndarray:
+    advantages = np.zeros_like(rewards)
+    last_gae_lam = 0
+    n_steps = advantages.shape[0]
+    for t in reversed(range(n_steps)):
+        if t == n_steps - 1:
+            next_nonterminal = 1.0 - next_episode_starts
+            next_value = policy.value(next_obs)
+        else:
+            next_nonterminal = 1.0 - episode_starts[t + 1]
+            next_value = values[t + 1]
+        delta = rewards[t] + gamma * next_value * next_nonterminal - values[t]
+        last_gae_lam = delta + gamma * gae_lambda * next_nonterminal * last_gae_lam
+        advantages[t] = last_gae_lam
+    return advantages

rl_algo_impls/shared/module/feature_extractor.py

@@ -1,215 +0,0 @@
-import gym
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-from abc import ABC, abstractmethod
-from gym.spaces import Box, Discrete
-from stable_baselines3.common.preprocessing import get_flattened_obs_dim
-from typing import Dict, Optional, Sequence, Type
-
-from rl_algo_impls.shared.module.module import layer_init
-
-
-class CnnFeatureExtractor(nn.Module, ABC):
-    @abstractmethod
-    def __init__(
-        self,
-        in_channels: int,
-        activation: Type[nn.Module] = nn.ReLU,
-        init_layers_orthogonal: Optional[bool] = None,
-        **kwargs,
-    ) -> None:
-        super().__init__()
-
-
-class NatureCnn(CnnFeatureExtractor):
-    """
-    CNN from DQN Nature paper: Mnih, Volodymyr, et al.
-    "Human-level control through deep reinforcement learning."
-    Nature 518.7540 (2015): 529-533.
-    """
-
-    def __init__(
-        self,
-        in_channels: int,
-        activation: Type[nn.Module] = nn.ReLU,
-        init_layers_orthogonal: Optional[bool] = None,
-        **kwargs,
-    ) -> None:
-        if init_layers_orthogonal is None:
-            init_layers_orthogonal = True
-        super().__init__(in_channels, activation, init_layers_orthogonal)
-        self.cnn = nn.Sequential(
-            layer_init(
-                nn.Conv2d(in_channels, 32, kernel_size=8, stride=4),
-                init_layers_orthogonal,
-            ),
-            activation(),
-            layer_init(
-                nn.Conv2d(32, 64, kernel_size=4, stride=2),
-                init_layers_orthogonal,
-            ),
-            activation(),
-            layer_init(
-                nn.Conv2d(64, 64, kernel_size=3, stride=1),
-                init_layers_orthogonal,
-            ),
-            activation(),
-            nn.Flatten(),
-        )
-
-    def forward(self, obs: torch.Tensor) -> torch.Tensor:
-        return self.cnn(obs)
-
-
-class ResidualBlock(nn.Module):
-    def __init__(
-        self,
-        channels: int,
-        activation: Type[nn.Module] = nn.ReLU,
-        init_layers_orthogonal: bool = False,
-    ) -> None:
-        super().__init__()
-        self.residual = nn.Sequential(
-            activation(),
-            layer_init(
-                nn.Conv2d(channels, channels, 3, padding=1), init_layers_orthogonal
-            ),
-            activation(),
-            layer_init(
-                nn.Conv2d(channels, channels, 3, padding=1), init_layers_orthogonal
-            ),
-        )
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        return x + self.residual(x)
-
-
-class ConvSequence(nn.Module):
-    def __init__(
-        self,
-        in_channels: int,
-        out_channels: int,
-        activation: Type[nn.Module] = nn.ReLU,
-        init_layers_orthogonal: bool = False,
-    ) -> None:
-        super().__init__()
-        self.seq = nn.Sequential(
-            layer_init(
-                nn.Conv2d(in_channels, out_channels, 3, padding=1),
-                init_layers_orthogonal,
-            ),
-            nn.MaxPool2d(3, stride=2, padding=1),
-            ResidualBlock(out_channels, activation, init_layers_orthogonal),
-            ResidualBlock(out_channels, activation, init_layers_orthogonal),
-        )
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        return self.seq(x)
-
-
-class ImpalaCnn(CnnFeatureExtractor):
-    """
-    IMPALA-style CNN architecture
-    """
-
-    def __init__(
-        self,
-        in_channels: int,
-        activation: Type[nn.Module] = nn.ReLU,
-        init_layers_orthogonal: Optional[bool] = None,
-        impala_channels: Sequence[int] = (16, 32, 32),
-        **kwargs,
-    ) -> None:
-        if init_layers_orthogonal is None:
-            init_layers_orthogonal = False
-        super().__init__(in_channels, activation, init_layers_orthogonal)
-        sequences = []
-        for out_channels in impala_channels:
-            sequences.append(
-                ConvSequence(
-                    in_channels, out_channels, activation, init_layers_orthogonal
-                )
-            )
-            in_channels = out_channels
-        sequences.extend(
-            [
-                activation(),
-                nn.Flatten(),
-            ]
-        )
-        self.seq = nn.Sequential(*sequences)
-
-    def forward(self, obs: torch.Tensor) -> torch.Tensor:
-        return self.seq(obs)
-
-
-CNN_EXTRACTORS_BY_STYLE: Dict[str, Type[CnnFeatureExtractor]] = {
-    "nature": NatureCnn,
-    "impala": ImpalaCnn,
-}
-
-
-class FeatureExtractor(nn.Module):
-    def __init__(
-        self,
-        obs_space: gym.Space,
-        activation: Type[nn.Module],
-        init_layers_orthogonal: bool = False,
-        cnn_feature_dim: int = 512,
-        cnn_style: str = "nature",
-        cnn_layers_init_orthogonal: Optional[bool] = None,
-        impala_channels: Sequence[int] = (16, 32, 32),
-    ) -> None:
-        super().__init__()
-        if isinstance(obs_space, Box):
-            # Conv2D: (channels, height, width)
-            if len(obs_space.shape) == 3:
-                cnn = CNN_EXTRACTORS_BY_STYLE[cnn_style](
-                    obs_space.shape[0],
-                    activation,
-                    init_layers_orthogonal=cnn_layers_init_orthogonal,
-                    impala_channels=impala_channels,
-                )
-
-                def preprocess(obs: torch.Tensor) -> torch.Tensor:
-                    if len(obs.shape) == 3:
-                        obs = obs.unsqueeze(0)
-                    return obs.float() / 255.0
-
-                with torch.no_grad():
-                    cnn_out = cnn(preprocess(torch.as_tensor(obs_space.sample())))
-                self.preprocess = preprocess
-                self.feature_extractor = nn.Sequential(
-                    cnn,
-                    layer_init(
-                        nn.Linear(cnn_out.shape[1], cnn_feature_dim),
-                        init_layers_orthogonal,
-                    ),
-                    activation(),
-                )
-                self.out_dim = cnn_feature_dim
-            elif len(obs_space.shape) == 1:
-
-                def preprocess(obs: torch.Tensor) -> torch.Tensor:
-                    if len(obs.shape) == 1:
-                        obs = obs.unsqueeze(0)
-                    return obs.float()
-
-                self.preprocess = preprocess
-                self.feature_extractor = nn.Flatten()
-                self.out_dim = get_flattened_obs_dim(obs_space)
-            else:
-                raise ValueError(f"Unsupported observation space: {obs_space}")
-        elif isinstance(obs_space, Discrete):
-            self.preprocess = lambda x: F.one_hot(x, obs_space.n).float()
-            self.feature_extractor = nn.Flatten()
-            self.out_dim = obs_space.n
-        else:
-            raise NotImplementedError
-
-    def forward(self, obs: torch.Tensor) -> torch.Tensor:
-        if self.preprocess:
-            obs = self.preprocess(obs)
-        return self.feature_extractor(obs)

rl_algo_impls/shared/module/module.py

@@ -1,8 +1,8 @@
+from typing import Sequence, Type
+
 import numpy as np
 import torch.nn as nn
 
-from typing import Sequence, Type
-
 
 def mlp(
     layer_sizes: Sequence[int],
@@ -10,12 +10,15 @@ def mlp(
     output_activation: Type[nn.Module] = nn.Identity,
     init_layers_orthogonal: bool = False,
     final_layer_gain: float = np.sqrt(2),
+    hidden_layer_gain: float = np.sqrt(2),
 ) -> nn.Module:
     layers = []
     for i in range(len(layer_sizes) - 2):
         layers.append(
             layer_init(
-                nn.Linear(layer_sizes[i], layer_sizes[i + 1]), init_layers_orthogonal
+                nn.Linear(layer_sizes[i], layer_sizes[i + 1]),
+                init_layers_orthogonal,
+                std=hidden_layer_gain,
             )
         )
         layers.append(activation())

rl_algo_impls/shared/policy/critic.py

@@ -1,27 +1,39 @@
-import gym
+from typing import Sequence, Type
+
+import numpy as np
 import torch
 import torch.nn as nn
 
-from typing import Sequence, Type
-
+from rl_algo_impls.shared.encoder import EncoderOutDim
 from rl_algo_impls.shared.module.module import mlp
 
 
 class CriticHead(nn.Module):
     def __init__(
         self,
-        hidden_sizes: Sequence[int] = (32,),
+        in_dim: EncoderOutDim,
+        hidden_sizes: Sequence[int] = (),
         activation: Type[nn.Module] = nn.Tanh,
         init_layers_orthogonal: bool = True,
     ) -> None:
         super().__init__()
-        layer_sizes = tuple(hidden_sizes) + (1,)
-        self._fc = mlp(
-            layer_sizes,
-            activation,
-            init_layers_orthogonal=init_layers_orthogonal,
-            final_layer_gain=1.0,
+        seq = []
+        if isinstance(in_dim, tuple):
+            seq.append(nn.Flatten())
+            in_channels = int(np.prod(in_dim))
+        else:
+            in_channels = in_dim
+        layer_sizes = (in_channels,) + tuple(hidden_sizes) + (1,)
+        seq.append(
+            mlp(
+                layer_sizes,
+                activation,
+                init_layers_orthogonal=init_layers_orthogonal,
+                final_layer_gain=1.0,
+                hidden_layer_gain=1.0,
+            )
         )
+        self._fc = nn.Sequential(*seq)
 
     def forward(self, obs: torch.Tensor) -> torch.Tensor:
         v = self._fc(obs)

rl_algo_impls/shared/policy/on_policy.py

@@ -1,24 +1,20 @@
+from abc import abstractmethod
+from typing import NamedTuple, Optional, Sequence, Tuple, TypeVar
+
 import gym
 import numpy as np
 import torch
-
-from abc import abstractmethod
 from gym.spaces import Box, Discrete, Space
-from typing import NamedTuple, Optional, Sequence, Tuple, TypeVar
 
-from rl_algo_impls.shared.module.feature_extractor import FeatureExtractor
-from rl_algo_impls.shared.policy.actor import (
-    PiForward,
-    StateDependentNoiseActorHead,
-    actor_head,
-)
+from rl_algo_impls.shared.actor import PiForward, actor_head
+from rl_algo_impls.shared.encoder import Encoder
 from rl_algo_impls.shared.policy.critic import CriticHead
 from rl_algo_impls.shared.policy.policy import ACTIVATION, Policy
 from rl_algo_impls.wrappers.vectorable_wrapper import (
     VecEnv,
     VecEnvObs,
-    single_observation_space,
     single_action_space,
+    single_observation_space,
 )
 
 
@@ -77,7 +73,12 @@ class OnPolicy(Policy):
         ...
 
     @abstractmethod
-    def step(self, obs: VecEnvObs) -> Step:
+    def step(self, obs: VecEnvObs, action_masks: Optional[np.ndarray] = None) -> Step:
+        ...
+
+    @property
+    @abstractmethod
+    def action_shape(self) -> Tuple[int, ...]:
         ...
 
 
@@ -94,10 +95,11 @@ class ActorCritic(OnPolicy):
         full_std: bool = True,
         squash_output: bool = False,
         share_features_extractor: bool = True,
-        cnn_feature_dim: int = 512,
+        cnn_flatten_dim: int = 512,
         cnn_style: str = "nature",
         cnn_layers_init_orthogonal: Optional[bool] = None,
         impala_channels: Sequence[int] = (16, 32, 32),
+        actor_head_style: str = "single",
         **kwargs,
     ) -> None:
         super().__init__(env, **kwargs)
@@ -120,52 +122,56 @@ class ActorCritic(OnPolicy):
         self.action_space = action_space
         self.squash_output = squash_output
         self.share_features_extractor = share_features_extractor
-        self._feature_extractor = FeatureExtractor(
+        self._feature_extractor = Encoder(
             observation_space,
             activation,
             init_layers_orthogonal=init_layers_orthogonal,
-            cnn_feature_dim=cnn_feature_dim,
+            cnn_flatten_dim=cnn_flatten_dim,
             cnn_style=cnn_style,
             cnn_layers_init_orthogonal=cnn_layers_init_orthogonal,
             impala_channels=impala_channels,
         )
         self._pi = actor_head(
             self.action_space,
-            (self._feature_extractor.out_dim,) + tuple(pi_hidden_sizes),
+            self._feature_extractor.out_dim,
+            tuple(pi_hidden_sizes),
             init_layers_orthogonal,
             activation,
             log_std_init=log_std_init,
             use_sde=use_sde,
             full_std=full_std,
             squash_output=squash_output,
+            actor_head_style=actor_head_style,
         )
 
         if not share_features_extractor:
-            self._v_feature_extractor = FeatureExtractor(
+            self._v_feature_extractor = Encoder(
                 observation_space,
                 activation,
                 init_layers_orthogonal=init_layers_orthogonal,
-                cnn_feature_dim=cnn_feature_dim,
+                cnn_flatten_dim=cnn_flatten_dim,
                 cnn_style=cnn_style,
                 cnn_layers_init_orthogonal=cnn_layers_init_orthogonal,
             )
-            v_hidden_sizes = (self._v_feature_extractor.out_dim,) + tuple(
-                v_hidden_sizes
-            )
+            critic_in_dim = self._v_feature_extractor.out_dim
         else:
             self._v_feature_extractor = None
-            v_hidden_sizes = (self._feature_extractor.out_dim,) + tuple(v_hidden_sizes)
+            critic_in_dim = self._feature_extractor.out_dim
         self._v = CriticHead(
+            in_dim=critic_in_dim,
             hidden_sizes=v_hidden_sizes,
             activation=activation,
             init_layers_orthogonal=init_layers_orthogonal,
         )
 
     def _pi_forward(
-        self, obs: torch.Tensor, action: Optional[torch.Tensor] = None
+        self,
+        obs: torch.Tensor,
+        action_masks: Optional[torch.Tensor],
+        action: Optional[torch.Tensor] = None,
     ) -> Tuple[PiForward, torch.Tensor]:
         p_fe = self._feature_extractor(obs)
-        pi_forward = self._pi(p_fe, action)
+        pi_forward = self._pi(p_fe, actions=action, action_masks=action_masks)
 
         return pi_forward, p_fe
 
@@ -173,8 +179,13 @@ class ActorCritic(OnPolicy):
         v_fe = self._v_feature_extractor(obs) if self._v_feature_extractor else p_fc
         return self._v(v_fe)
 
-    def forward(self, obs: torch.Tensor, action: torch.Tensor) -> ACForward:
-        (_, logp_a, entropy), p_fc = self._pi_forward(obs, action)
+    def forward(
+        self,
+        obs: torch.Tensor,
+        action: torch.Tensor,
+        action_masks: Optional[torch.Tensor] = None,
+    ) -> ACForward:
+        (_, logp_a, entropy), p_fc = self._pi_forward(obs, action_masks, action=action)
         v = self._v_forward(obs, p_fc)
 
         assert logp_a is not None
@@ -192,10 +203,11 @@ class ActorCritic(OnPolicy):
             v = self._v(fe)
         return v.cpu().numpy()
 
-    def step(self, obs: VecEnvObs) -> Step:
+    def step(self, obs: VecEnvObs, action_masks: Optional[np.ndarray] = None) -> Step:
         o = self._as_tensor(obs)
+        a_masks = self._as_tensor(action_masks) if action_masks is not None else None
         with torch.no_grad():
-            (pi, _, _), p_fc = self._pi_forward(o)
+            (pi, _, _), p_fc = self._pi_forward(o, action_masks=a_masks)
             a = pi.sample()
             logp_a = pi.log_prob(a)
 
@@ -205,13 +217,21 @@ class ActorCritic(OnPolicy):
         clamped_a_np = clamp_actions(a_np, self.action_space, self.squash_output)
         return Step(a_np, v.cpu().numpy(), logp_a.cpu().numpy(), clamped_a_np)
 
-    def act(self, obs: np.ndarray, deterministic: bool = True) -> np.ndarray:
+    def act(
+        self,
+        obs: np.ndarray,
+        deterministic: bool = True,
+        action_masks: Optional[np.ndarray] = None,
+    ) -> np.ndarray:
         if not deterministic:
-            return self.step(obs).clamped_a
+            return self.step(obs, action_masks=action_masks).clamped_a
         else:
             o = self._as_tensor(obs)
+            a_masks = (
+                self._as_tensor(action_masks) if action_masks is not None else None
+            )
             with torch.no_grad():
-                (pi, _, _), _ = self._pi_forward(o)
+                (pi, _, _), _ = self._pi_forward(o, action_masks=a_masks)
                 a = pi.mode
             return clamp_actions(a.cpu().numpy(), self.action_space, self.squash_output)
 
@@ -220,7 +240,10 @@ class ActorCritic(OnPolicy):
         self.reset_noise()
 
     def reset_noise(self, batch_size: Optional[int] = None) -> None:
-        if isinstance(self._pi, StateDependentNoiseActorHead):
-            self._pi.sample_weights(
-                batch_size=batch_size if batch_size else self.env.num_envs
-            )
+        self._pi.sample_weights(
+            batch_size=batch_size if batch_size else self.env.num_envs
+        )
+
+    @property
+    def action_shape(self) -> Tuple[int, ...]:
+        return self._pi.action_shape

rl_algo_impls/shared/policy/policy.py

@@ -46,7 +46,12 @@ class Policy(nn.Module, ABC):
         return self
 
     @abstractmethod
-    def act(self, obs: VecEnvObs, deterministic: bool = True) -> np.ndarray:
+    def act(
+        self,
+        obs: VecEnvObs,
+        deterministic: bool = True,
+        action_masks: Optional[np.ndarray] = None,
+    ) -> np.ndarray:
         ...
 
     def save(self, path: str) -> None:

rl_algo_impls/shared/schedule.py

@@ -20,10 +20,38 @@ def constant_schedule(val: float) -> Schedule:
     return lambda f: val
 
 
+def spike_schedule(
+    max_value: float,
+    start_fraction: float = 1e-2,
+    end_fraction: float = 1e-4,
+    peak_progress: float = 0.1,
+) -> Schedule:
+    assert 0 < peak_progress < 1
+
+    def func(progress_fraction: float) -> float:
+        if progress_fraction < peak_progress:
+            fraction = (
+                start_fraction
+                + (1 - start_fraction) * progress_fraction / peak_progress
+            )
+        else:
+            fraction = 1 + (end_fraction - 1) * (progress_fraction - peak_progress) / (
+                1 - peak_progress
+            )
+        return max_value * fraction
+
+    return func
+
+
 def schedule(name: str, start_val: float) -> Schedule:
     if name == "linear":
         return linear_schedule(start_val, 0)
-    return constant_schedule(start_val)
+    elif name == "none":
+        return constant_schedule(start_val)
+    elif name == "spike":
+        return spike_schedule(start_val)
+    else:
+        raise ValueError(f"Schedule {name} not supported")
 
 
 def update_learning_rate(optimizer: Optimizer, learning_rate: float) -> None:

rl_algo_impls/shared/stats.py

@@ -1,14 +1,17 @@
-import numpy as np
-
+import dataclasses
+from collections import defaultdict
 from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Sequence, TypeVar, Union
+
+import numpy as np
 from torch.utils.tensorboard.writer import SummaryWriter
-from typing import Dict, List, Optional, Sequence, Union, TypeVar
 
 
 @dataclass
 class Episode:
     score: float = 0
     length: int = 0
+    info: Dict[str, Dict[str, Any]] = dataclasses.field(default_factory=dict)
 
 
 StatisticSelf = TypeVar("StatisticSelf", bound="Statistic")
@@ -75,12 +78,25 @@ class EpisodesStats:
     simple: bool
     score: Statistic
     length: Statistic
+    additional_stats: Dict[str, Statistic]
 
     def __init__(self, episodes: Sequence[Episode], simple: bool = False) -> None:
         self.episodes = episodes
         self.simple = simple
         self.score = Statistic(np.array([e.score for e in episodes]))
         self.length = Statistic(np.array([e.length for e in episodes]), round_digits=0)
+        additional_values = defaultdict(list)
+        for e in self.episodes:
+            if e.info:
+                for k, v in e.info.items():
+                    if isinstance(v, dict):
+                        for k2, v2 in v.items():
+                            additional_values[f"{k}_{k2}"].append(v2)
+                    else:
+                        additional_values[k].append(v)
+        self.additional_stats = {
+            k: Statistic(np.array(values)) for k, values in additional_values.items()
+        }
 
     def __gt__(self: EpisodesStatsSelf, o: EpisodesStatsSelf) -> bool:
         return self.score > o.score
@@ -118,6 +134,8 @@ class EpisodesStats:
                     "length": self.length.mean,
                 }
             )
+            for k, addl_stats in self.additional_stats.items():
+                stats[k] = addl_stats.mean
         for name, value in stats.items():
             tb_writer.add_scalar(f"{main_tag}/{name}", value, global_step=global_step)
 
@@ -131,19 +149,19 @@ class EpisodeAccumulator:
     def episodes(self) -> List[Episode]:
         return self._episodes
 
-    def step(self, reward: np.ndarray, done: np.ndarray) -> None:
+    def step(self, reward: np.ndarray, done: np.ndarray, info: List[Dict]) -> None:
         for idx, current in enumerate(self.current_episodes):
             current.score += reward[idx]
             current.length += 1
             if done[idx]:
                 self._episodes.append(current)
                 self.current_episodes[idx] = Episode()
-                self.on_done(idx, current)
+                self.on_done(idx, current, info[idx])
 
     def __len__(self) -> int:
         return len(self.episodes)
 
-    def on_done(self, ep_idx: int, episode: Episode) -> None:
+    def on_done(self, ep_idx: int, episode: Episode, info: Dict) -> None:
         pass
 
     def stats(self) -> EpisodesStats:

rl_algo_impls/shared/vec_env/__init__.py

@@ -0,0 +1 @@
+from rl_algo_impls.shared.vec_env.make_env import make_env, make_eval_env

rl_algo_impls/shared/vec_env/make_env.py

@@ -0,0 +1,66 @@
+from dataclasses import asdict
+from typing import Optional
+
+from torch.utils.tensorboard.writer import SummaryWriter
+
+from rl_algo_impls.runner.config import Config, EnvHyperparams
+from rl_algo_impls.shared.vec_env.microrts import make_microrts_env
+from rl_algo_impls.shared.vec_env.procgen import make_procgen_env
+from rl_algo_impls.shared.vec_env.vec_env import make_vec_env
+from rl_algo_impls.wrappers.vectorable_wrapper import VecEnv
+
+
+def make_env(
+    config: Config,
+    hparams: EnvHyperparams,
+    training: bool = True,
+    render: bool = False,
+    normalize_load_path: Optional[str] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+) -> VecEnv:
+    if hparams.env_type == "procgen":
+        return make_procgen_env(
+            config,
+            hparams,
+            training=training,
+            render=render,
+            normalize_load_path=normalize_load_path,
+            tb_writer=tb_writer,
+        )
+    elif hparams.env_type in {"sb3vec", "gymvec"}:
+        return make_vec_env(
+            config,
+            hparams,
+            training=training,
+            render=render,
+            normalize_load_path=normalize_load_path,
+            tb_writer=tb_writer,
+        )
+    elif hparams.env_type == "microrts":
+        return make_microrts_env(
+            config,
+            hparams,
+            training=training,
+            render=render,
+            normalize_load_path=normalize_load_path,
+            tb_writer=tb_writer,
+        )
+    else:
+        raise ValueError(f"env_type {hparams.env_type} not supported")
+
+
+def make_eval_env(
+    config: Config,
+    hparams: EnvHyperparams,
+    override_n_envs: Optional[int] = None,
+    **kwargs,
+) -> VecEnv:
+    kwargs = kwargs.copy()
+    kwargs["training"] = False
+    if override_n_envs is not None:
+        hparams_kwargs = asdict(hparams)
+        hparams_kwargs["n_envs"] = override_n_envs
+        if override_n_envs == 1:
+            hparams_kwargs["vec_env_class"] = "sync"
+        hparams = EnvHyperparams(**hparams_kwargs)
+    return make_env(config, hparams, **kwargs)

rl_algo_impls/shared/vec_env/microrts.py

@@ -0,0 +1,94 @@
+from dataclasses import astuple
+from typing import Optional
+
+import gym
+import numpy as np
+from torch.utils.tensorboard.writer import SummaryWriter
+
+from rl_algo_impls.runner.config import Config, EnvHyperparams
+from rl_algo_impls.wrappers.action_mask_wrapper import MicrortsMaskWrapper
+from rl_algo_impls.wrappers.episode_stats_writer import EpisodeStatsWriter
+from rl_algo_impls.wrappers.hwc_to_chw_observation import HwcToChwObservation
+from rl_algo_impls.wrappers.is_vector_env import IsVectorEnv
+from rl_algo_impls.wrappers.microrts_stats_recorder import MicrortsStatsRecorder
+from rl_algo_impls.wrappers.vectorable_wrapper import VecEnv
+
+
+def make_microrts_env(
+    config: Config,
+    hparams: EnvHyperparams,
+    training: bool = True,
+    render: bool = False,
+    normalize_load_path: Optional[str] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+) -> VecEnv:
+    import gym_microrts
+    from gym_microrts import microrts_ai
+
+    from rl_algo_impls.shared.vec_env.microrts_compat import (
+        MicroRTSGridModeVecEnvCompat,
+    )
+
+    (
+        _,  # env_type
+        n_envs,
+        _,  # frame_stack
+        make_kwargs,
+        _,  # no_reward_timeout_steps
+        _,  # no_reward_fire_steps
+        _,  # vec_env_class
+        _,  # normalize
+        _,  # normalize_kwargs,
+        rolling_length,
+        _,  # train_record_video
+        _,  # video_step_interval
+        _,  # initial_steps_to_truncate
+        _,  # clip_atari_rewards
+        _,  # normalize_type
+        _,  # mask_actions
+        bots,
+    ) = astuple(hparams)
+
+    seed = config.seed(training=training)
+
+    make_kwargs = make_kwargs or {}
+    if "num_selfplay_envs" not in make_kwargs:
+        make_kwargs["num_selfplay_envs"] = 0
+    if "num_bot_envs" not in make_kwargs:
+        make_kwargs["num_bot_envs"] = n_envs - make_kwargs["num_selfplay_envs"]
+    if "reward_weight" in make_kwargs:
+        make_kwargs["reward_weight"] = np.array(make_kwargs["reward_weight"])
+    if bots:
+        ai2s = []
+        for ai_name, n in bots.items():
+            for _ in range(n):
+                if len(ai2s) >= make_kwargs["num_bot_envs"]:
+                    break
+                ai = getattr(microrts_ai, ai_name)
+                assert ai, f"{ai_name} not in microrts_ai"
+                ai2s.append(ai)
+    else:
+        ai2s = [microrts_ai.randomAI for _ in make_kwargs["num_bot_envs"]]
+    make_kwargs["ai2s"] = ai2s
+    envs = MicroRTSGridModeVecEnvCompat(**make_kwargs)
+    envs = HwcToChwObservation(envs)
+    envs = IsVectorEnv(envs)
+    envs = MicrortsMaskWrapper(envs)
+
+    if seed is not None:
+        envs.action_space.seed(seed)
+        envs.observation_space.seed(seed)
+
+    envs = gym.wrappers.RecordEpisodeStatistics(envs)
+    envs = MicrortsStatsRecorder(envs, config.algo_hyperparams.get("gamma", 0.99))
+    if training:
+        assert tb_writer
+        envs = EpisodeStatsWriter(
+            envs,
+            tb_writer,
+            training=training,
+            rolling_length=rolling_length,
+            additional_keys_to_log=config.additional_keys_to_log,
+        )
+
+    return envs

rl_algo_impls/shared/vec_env/microrts_compat.py

@@ -0,0 +1,49 @@
+from typing import TypeVar
+
+import numpy as np
+from gym_microrts.envs.vec_env import MicroRTSGridModeVecEnv
+from jpype.types import JArray, JInt
+
+from rl_algo_impls.wrappers.vectorable_wrapper import VecEnvStepReturn
+
+MicroRTSGridModeVecEnvCompatSelf = TypeVar(
+    "MicroRTSGridModeVecEnvCompatSelf", bound="MicroRTSGridModeVecEnvCompat"
+)
+
+
+class MicroRTSGridModeVecEnvCompat(MicroRTSGridModeVecEnv):
+    def step(self, action: np.ndarray) -> VecEnvStepReturn:
+        indexed_actions = np.concatenate(
+            [
+                np.expand_dims(
+                    np.stack(
+                        [np.arange(0, action.shape[1]) for i in range(self.num_envs)]
+                    ),
+                    axis=2,
+                ),
+                action,
+            ],
+            axis=2,
+        )
+        action_mask = np.array(self.vec_client.getMasks(0), dtype=np.bool8).reshape(
+            indexed_actions.shape[:-1] + (-1,)
+        )
+        valid_action_mask = action_mask[:, :, 0]
+        valid_actions_counts = valid_action_mask.sum(1)
+        valid_actions = indexed_actions[valid_action_mask]
+        valid_actions_idx = 0
+
+        all_valid_actions = []
+        for env_act_cnt in valid_actions_counts:
+            env_valid_actions = []
+            for _ in range(env_act_cnt):
+                env_valid_actions.append(JArray(JInt)(valid_actions[valid_actions_idx]))
+                valid_actions_idx += 1
+            all_valid_actions.append(JArray(JArray(JInt))(env_valid_actions))
+        return super().step(JArray(JArray(JArray(JInt)))(all_valid_actions))  # type: ignore
+
+    @property
+    def unwrapped(
+        self: MicroRTSGridModeVecEnvCompatSelf,
+    ) -> MicroRTSGridModeVecEnvCompatSelf:
+        return self