Quentin Gallouédec commited on
Commit
a64aa9f
1 Parent(s): 9c6c190

initial_agent

Browse files
README.md ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - QbertNoFrameskip-v4
4
+ - deep-reinforcement-learning
5
+ - reinforcement-learning
6
+ - custom-implementation
7
+ library_name: cleanrl
8
+ model-index:
9
+ - name: DQN
10
+ results:
11
+ - task:
12
+ type: reinforcement-learning
13
+ name: reinforcement-learning
14
+ dataset:
15
+ name: QbertNoFrameskip-v4
16
+ type: QbertNoFrameskip-v4
17
+ metrics:
18
+ - type: mean_reward
19
+ value: 10695.00 +/- 4551.15
20
+ name: mean_reward
21
+ verified: false
22
+ ---
23
+
24
+ # (CleanRL) **DQN** Agent Playing **QbertNoFrameskip-v4**
25
+
26
+ This is a trained model of a DQN agent playing QbertNoFrameskip-v4.
27
+ The model was trained by using [CleanRL](https://github.com/vwxyzjn/cleanrl) and the most up-to-date training code can be
28
+ found [here](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/dqn_atari.py).
29
+
30
+ ## Get Started
31
+
32
+ To use this model, please install the `cleanrl` package with the following command:
33
+
34
+ ```
35
+ pip install "cleanrl[dqn_atari]"
36
+ python -m cleanrl_utils.enjoy --exp-name dqn_atari --env-id QbertNoFrameskip-v4
37
+ ```
38
+
39
+ Please refer to the [documentation](https://docs.cleanrl.dev/get-started/zoo/) for more detail.
40
+
41
+
42
+ ## Command to reproduce the training
43
+
44
+ ```bash
45
+ curl -OL https://huggingface.co/cleanrl/QbertNoFrameskip-v4-dqn_atari-seed1/raw/main/dqn_atari.py
46
+ curl -OL https://huggingface.co/cleanrl/QbertNoFrameskip-v4-dqn_atari-seed1/raw/main/pyproject.toml
47
+ curl -OL https://huggingface.co/cleanrl/QbertNoFrameskip-v4-dqn_atari-seed1/raw/main/poetry.lock
48
+ poetry install --all-extras
49
+ python dqn_atari.py --capture-video --env-id QbertNoFrameskip-v4 --seed 1 --save-model --upload-model --hf-entity cleanrl --track
50
+ ```
51
+
52
+ # Hyperparameters
53
+ ```python
54
+ {'batch_size': 32,
55
+ 'buffer_size': 1000000,
56
+ 'capture_video': True,
57
+ 'cuda': True,
58
+ 'end_e': 0.01,
59
+ 'env_id': 'QbertNoFrameskip-v4',
60
+ 'exp_name': 'dqn_atari',
61
+ 'exploration_fraction': 0.1,
62
+ 'gamma': 0.99,
63
+ 'hf_entity': 'cleanrl',
64
+ 'learning_rate': 0.0001,
65
+ 'learning_starts': 80000,
66
+ 'num_envs': 1,
67
+ 'save_model': True,
68
+ 'seed': 1,
69
+ 'start_e': 1,
70
+ 'target_network_frequency': 1000,
71
+ 'tau': 1.0,
72
+ 'torch_deterministic': True,
73
+ 'total_timesteps': 10000000,
74
+ 'track': True,
75
+ 'train_frequency': 4,
76
+ 'upload_model': True,
77
+ 'wandb_entity': None,
78
+ 'wandb_project_name': 'cleanRL'}
79
+ ```
80
+
dqn_atari.cleanrl_model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0ac213d2aa94d98a34c79b25e40f7d09bd3126fb58ea488b2e6495a44f9add37
3
+ size 6752415
dqn_atari.py ADDED
@@ -0,0 +1,286 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/dqn/#dqn_ataripy
2
+ import argparse
3
+ import os
4
+ import random
5
+ import time
6
+ from distutils.util import strtobool
7
+
8
+ import gymnasium as gym
9
+ import numpy as np
10
+ import torch
11
+ import torch.nn as nn
12
+ import torch.nn.functional as F
13
+ import torch.optim as optim
14
+ from stable_baselines3.common.atari_wrappers import (
15
+ ClipRewardEnv,
16
+ EpisodicLifeEnv,
17
+ FireResetEnv,
18
+ MaxAndSkipEnv,
19
+ NoopResetEnv,
20
+ )
21
+ from stable_baselines3.common.buffers import ReplayBuffer
22
+ from torch.utils.tensorboard import SummaryWriter
23
+
24
+
25
+ def parse_args():
26
+ # fmt: off
27
+ parser = argparse.ArgumentParser()
28
+ parser.add_argument("--exp-name", type=str, default=os.path.basename(__file__).rstrip(".py"),
29
+ help="the name of this experiment")
30
+ parser.add_argument("--seed", type=int, default=1,
31
+ help="seed of the experiment")
32
+ parser.add_argument("--torch-deterministic", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
33
+ help="if toggled, `torch.backends.cudnn.deterministic=False`")
34
+ parser.add_argument("--cuda", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
35
+ help="if toggled, cuda will be enabled by default")
36
+ parser.add_argument("--track", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
37
+ help="if toggled, this experiment will be tracked with Weights and Biases")
38
+ parser.add_argument("--wandb-project-name", type=str, default="cleanRL",
39
+ help="the wandb's project name")
40
+ parser.add_argument("--wandb-entity", type=str, default=None,
41
+ help="the entity (team) of wandb's project")
42
+ parser.add_argument("--capture-video", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
43
+ help="whether to capture videos of the agent performances (check out `videos` folder)")
44
+ parser.add_argument("--save-model", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
45
+ help="whether to save model into the `runs/{run_name}` folder")
46
+ parser.add_argument("--upload-model", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
47
+ help="whether to upload the saved model to huggingface")
48
+ parser.add_argument("--hf-entity", type=str, default="",
49
+ help="the user or org name of the model repository from the Hugging Face Hub")
50
+
51
+ # Algorithm specific arguments
52
+ parser.add_argument("--env-id", type=str, default="BreakoutNoFrameskip-v4",
53
+ help="the id of the environment")
54
+ parser.add_argument("--total-timesteps", type=int, default=10000000,
55
+ help="total timesteps of the experiments")
56
+ parser.add_argument("--learning-rate", type=float, default=1e-4,
57
+ help="the learning rate of the optimizer")
58
+ parser.add_argument("--num-envs", type=int, default=1,
59
+ help="the number of parallel game environments")
60
+ parser.add_argument("--buffer-size", type=int, default=1000000,
61
+ help="the replay memory buffer size")
62
+ parser.add_argument("--gamma", type=float, default=0.99,
63
+ help="the discount factor gamma")
64
+ parser.add_argument("--tau", type=float, default=1.,
65
+ help="the target network update rate")
66
+ parser.add_argument("--target-network-frequency", type=int, default=1000,
67
+ help="the timesteps it takes to update the target network")
68
+ parser.add_argument("--batch-size", type=int, default=32,
69
+ help="the batch size of sample from the reply memory")
70
+ parser.add_argument("--start-e", type=float, default=1,
71
+ help="the starting epsilon for exploration")
72
+ parser.add_argument("--end-e", type=float, default=0.01,
73
+ help="the ending epsilon for exploration")
74
+ parser.add_argument("--exploration-fraction", type=float, default=0.10,
75
+ help="the fraction of `total-timesteps` it takes from start-e to go end-e")
76
+ parser.add_argument("--learning-starts", type=int, default=80000,
77
+ help="timestep to start learning")
78
+ parser.add_argument("--train-frequency", type=int, default=4,
79
+ help="the frequency of training")
80
+ args = parser.parse_args()
81
+ # fmt: on
82
+ assert args.num_envs == 1, "vectorized envs are not supported at the moment"
83
+
84
+ return args
85
+
86
+
87
+ def make_env(env_id, seed, idx, capture_video, run_name):
88
+ def thunk():
89
+ if capture_video and idx == 0:
90
+ env = gym.make(env_id, render_mode="rgb_array")
91
+ env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
92
+ else:
93
+ env = gym.make(env_id)
94
+ env = gym.wrappers.RecordEpisodeStatistics(env)
95
+ env = NoopResetEnv(env, noop_max=30)
96
+ env = MaxAndSkipEnv(env, skip=4)
97
+ env = EpisodicLifeEnv(env)
98
+ if "FIRE" in env.unwrapped.get_action_meanings():
99
+ env = FireResetEnv(env)
100
+ env = ClipRewardEnv(env)
101
+ env = gym.wrappers.ResizeObservation(env, (84, 84))
102
+ env = gym.wrappers.GrayScaleObservation(env)
103
+ env = gym.wrappers.FrameStack(env, 4)
104
+ env.action_space.seed(seed)
105
+
106
+ return env
107
+
108
+ return thunk
109
+
110
+
111
+ # ALGO LOGIC: initialize agent here:
112
+ class QNetwork(nn.Module):
113
+ def __init__(self, env):
114
+ super().__init__()
115
+ self.network = nn.Sequential(
116
+ nn.Conv2d(4, 32, 8, stride=4),
117
+ nn.ReLU(),
118
+ nn.Conv2d(32, 64, 4, stride=2),
119
+ nn.ReLU(),
120
+ nn.Conv2d(64, 64, 3, stride=1),
121
+ nn.ReLU(),
122
+ nn.Flatten(),
123
+ nn.Linear(3136, 512),
124
+ nn.ReLU(),
125
+ nn.Linear(512, env.single_action_space.n),
126
+ )
127
+
128
+ def forward(self, x):
129
+ return self.network(x / 255.0)
130
+
131
+
132
+ def linear_schedule(start_e: float, end_e: float, duration: int, t: int):
133
+ slope = (end_e - start_e) / duration
134
+ return max(slope * t + start_e, end_e)
135
+
136
+
137
+ if __name__ == "__main__":
138
+ import stable_baselines3 as sb3
139
+
140
+ if sb3.__version__ < "2.0":
141
+ raise ValueError(
142
+ """Ongoing migration: run the following command to install the new dependencies:
143
+
144
+ poetry run pip install "stable_baselines3==2.0.0a1" "gymnasium[atari,accept-rom-license]==0.28.1" "ale-py==0.8.1"
145
+ """
146
+ )
147
+ args = parse_args()
148
+ run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
149
+ if args.track:
150
+ import wandb
151
+
152
+ wandb.init(
153
+ project=args.wandb_project_name,
154
+ entity=args.wandb_entity,
155
+ sync_tensorboard=True,
156
+ config=vars(args),
157
+ name=run_name,
158
+ monitor_gym=True,
159
+ save_code=True,
160
+ )
161
+ writer = SummaryWriter(f"runs/{run_name}")
162
+ writer.add_text(
163
+ "hyperparameters",
164
+ "|param|value|\n|-|-|\n%s" % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])),
165
+ )
166
+
167
+ # TRY NOT TO MODIFY: seeding
168
+ random.seed(args.seed)
169
+ np.random.seed(args.seed)
170
+ torch.manual_seed(args.seed)
171
+ torch.backends.cudnn.deterministic = args.torch_deterministic
172
+
173
+ device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
174
+
175
+ # env setup
176
+ envs = gym.vector.SyncVectorEnv(
177
+ [make_env(args.env_id, args.seed + i, i, args.capture_video, run_name) for i in range(args.num_envs)]
178
+ )
179
+ assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported"
180
+
181
+ q_network = QNetwork(envs).to(device)
182
+ optimizer = optim.Adam(q_network.parameters(), lr=args.learning_rate)
183
+ target_network = QNetwork(envs).to(device)
184
+ target_network.load_state_dict(q_network.state_dict())
185
+
186
+ rb = ReplayBuffer(
187
+ args.buffer_size,
188
+ envs.single_observation_space,
189
+ envs.single_action_space,
190
+ device,
191
+ optimize_memory_usage=True,
192
+ handle_timeout_termination=False,
193
+ )
194
+ start_time = time.time()
195
+
196
+ # TRY NOT TO MODIFY: start the game
197
+ obs, _ = envs.reset(seed=args.seed)
198
+ for global_step in range(args.total_timesteps):
199
+ # ALGO LOGIC: put action logic here
200
+ epsilon = linear_schedule(args.start_e, args.end_e, args.exploration_fraction * args.total_timesteps, global_step)
201
+ if random.random() < epsilon:
202
+ actions = np.array([envs.single_action_space.sample() for _ in range(envs.num_envs)])
203
+ else:
204
+ q_values = q_network(torch.Tensor(obs).to(device))
205
+ actions = torch.argmax(q_values, dim=1).cpu().numpy()
206
+
207
+ # TRY NOT TO MODIFY: execute the game and log data.
208
+ next_obs, rewards, terminated, truncated, infos = envs.step(actions)
209
+
210
+ # TRY NOT TO MODIFY: record rewards for plotting purposes
211
+ if "final_info" in infos:
212
+ for info in infos["final_info"]:
213
+ # Skip the envs that are not done
214
+ if "episode" not in info:
215
+ continue
216
+ print(f"global_step={global_step}, episodic_return={info['episode']['r']}")
217
+ writer.add_scalar("charts/episodic_return", info["episode"]["r"], global_step)
218
+ writer.add_scalar("charts/episodic_length", info["episode"]["l"], global_step)
219
+ writer.add_scalar("charts/epsilon", epsilon, global_step)
220
+
221
+ # TRY NOT TO MODIFY: save data to reply buffer; handle `final_observation`
222
+ real_next_obs = next_obs.copy()
223
+ for idx, d in enumerate(truncated):
224
+ if d:
225
+ real_next_obs[idx] = infos["final_observation"][idx]
226
+ rb.add(obs, real_next_obs, actions, rewards, terminated, infos)
227
+
228
+ # TRY NOT TO MODIFY: CRUCIAL step easy to overlook
229
+ obs = next_obs
230
+
231
+ # ALGO LOGIC: training.
232
+ if global_step > args.learning_starts:
233
+ if global_step % args.train_frequency == 0:
234
+ data = rb.sample(args.batch_size)
235
+ with torch.no_grad():
236
+ target_max, _ = target_network(data.next_observations).max(dim=1)
237
+ td_target = data.rewards.flatten() + args.gamma * target_max * (1 - data.dones.flatten())
238
+ old_val = q_network(data.observations).gather(1, data.actions).squeeze()
239
+ loss = F.mse_loss(td_target, old_val)
240
+
241
+ if global_step % 100 == 0:
242
+ writer.add_scalar("losses/td_loss", loss, global_step)
243
+ writer.add_scalar("losses/q_values", old_val.mean().item(), global_step)
244
+ print("SPS:", int(global_step / (time.time() - start_time)))
245
+ writer.add_scalar("charts/SPS", int(global_step / (time.time() - start_time)), global_step)
246
+
247
+ # optimize the model
248
+ optimizer.zero_grad()
249
+ loss.backward()
250
+ optimizer.step()
251
+
252
+ # update target network
253
+ if global_step % args.target_network_frequency == 0:
254
+ for target_network_param, q_network_param in zip(target_network.parameters(), q_network.parameters()):
255
+ target_network_param.data.copy_(
256
+ args.tau * q_network_param.data + (1.0 - args.tau) * target_network_param.data
257
+ )
258
+
259
+ if args.save_model:
260
+ model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
261
+ torch.save(q_network.state_dict(), model_path)
262
+ print(f"model saved to {model_path}")
263
+ from cleanrl_utils.evals.dqn_eval import evaluate
264
+
265
+ episodic_returns = evaluate(
266
+ model_path,
267
+ make_env,
268
+ args.env_id,
269
+ eval_episodes=10,
270
+ run_name=f"{run_name}-eval",
271
+ Model=QNetwork,
272
+ device=device,
273
+ epsilon=0.05,
274
+ )
275
+ for idx, episodic_return in enumerate(episodic_returns):
276
+ writer.add_scalar("eval/episodic_return", episodic_return, idx)
277
+
278
+ if args.upload_model:
279
+ from cleanrl_utils.huggingface import push_to_hub
280
+
281
+ repo_name = f"{args.env_id}-{args.exp_name}-seed{args.seed}"
282
+ repo_id = f"{args.hf_entity}/{repo_name}" if args.hf_entity else repo_name
283
+ push_to_hub(args, episodic_returns, repo_id, "DQN", f"runs/{run_name}", f"videos/{run_name}-eval")
284
+
285
+ envs.close()
286
+ writer.close()
events.out.tfevents.1696742846.3090-172.370956.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:771961046df8e7c1a0b83f5061d81ace9ef9522d0686e07a103c4df7f49afdbc
3
+ size 17908865
poetry.lock ADDED
The diff for this file is too large to render. See raw diff
 
pyproject.toml ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [tool.poetry]
2
+ name = "cleanrl"
3
+ version = "1.1.0"
4
+ description = "High-quality single file implementation of Deep Reinforcement Learning algorithms with research-friendly features"
5
+ authors = ["Costa Huang <[email protected]>"]
6
+ packages = [
7
+ { include = "cleanrl" },
8
+ { include = "cleanrl_utils" },
9
+ ]
10
+ keywords = ["reinforcement", "machine", "learning", "research"]
11
+ license="MIT"
12
+ readme = "README.md"
13
+
14
+ [tool.poetry.dependencies]
15
+ python = ">=3.7.1,<3.11"
16
+ tensorboard = "^2.10.0"
17
+ wandb = "^0.13.11"
18
+ gym = "0.23.1"
19
+ torch = ">=1.12.1"
20
+ stable-baselines3 = "1.2.0"
21
+ gymnasium = ">=0.28.1"
22
+ moviepy = "^1.0.3"
23
+ pygame = "2.1.0"
24
+ huggingface-hub = "^0.11.1"
25
+ rich = "<12.0"
26
+ tenacity = "^8.2.2"
27
+
28
+ ale-py = {version = "0.7.4", optional = true}
29
+ AutoROM = {extras = ["accept-rom-license"], version = "^0.4.2", optional = true}
30
+ opencv-python = {version = "^4.6.0.66", optional = true}
31
+ procgen = {version = "^0.10.7", optional = true}
32
+ pytest = {version = "^7.1.3", optional = true}
33
+ mujoco = {version = "<=2.3.3", optional = true}
34
+ imageio = {version = "^2.14.1", optional = true}
35
+ free-mujoco-py = {version = "^2.1.6", optional = true}
36
+ mkdocs-material = {version = "^8.4.3", optional = true}
37
+ markdown-include = {version = "^0.7.0", optional = true}
38
+ openrlbenchmark = {version = "^0.1.1b4", optional = true}
39
+ jax = {version = "^0.3.17", optional = true}
40
+ jaxlib = {version = "^0.3.15", optional = true}
41
+ flax = {version = "^0.6.0", optional = true}
42
+ optuna = {version = "^3.0.1", optional = true}
43
+ optuna-dashboard = {version = "^0.7.2", optional = true}
44
+ envpool = {version = "^0.6.4", optional = true}
45
+ PettingZoo = {version = "1.18.1", optional = true}
46
+ SuperSuit = {version = "3.4.0", optional = true}
47
+ multi-agent-ale-py = {version = "0.1.11", optional = true}
48
+ boto3 = {version = "^1.24.70", optional = true}
49
+ awscli = {version = "^1.25.71", optional = true}
50
+ shimmy = {version = ">=1.0.0", extras = ["dm-control"], optional = true}
51
+
52
+ [tool.poetry.group.dev.dependencies]
53
+ pre-commit = "^2.20.0"
54
+
55
+
56
+ [tool.poetry.group.isaacgym]
57
+ optional = true
58
+ [tool.poetry.group.isaacgym.dependencies]
59
+ isaacgymenvs = {git = "https://github.com/vwxyzjn/IsaacGymEnvs.git", rev = "poetry", python = ">=3.7.1,<3.10"}
60
+ isaacgym = {path = "cleanrl/ppo_continuous_action_isaacgym/isaacgym", develop = true}
61
+
62
+
63
+ [build-system]
64
+ requires = ["poetry-core"]
65
+ build-backend = "poetry.core.masonry.api"
66
+
67
+ [tool.poetry.extras]
68
+ atari = ["ale-py", "AutoROM", "opencv-python"]
69
+ procgen = ["procgen"]
70
+ plot = ["pandas", "seaborn"]
71
+ pytest = ["pytest"]
72
+ mujoco = ["mujoco", "imageio"]
73
+ mujoco_py = ["free-mujoco-py"]
74
+ jax = ["jax", "jaxlib", "flax"]
75
+ docs = ["mkdocs-material", "markdown-include", "openrlbenchmark"]
76
+ envpool = ["envpool"]
77
+ optuna = ["optuna", "optuna-dashboard"]
78
+ pettingzoo = ["PettingZoo", "SuperSuit", "multi-agent-ale-py"]
79
+ cloud = ["boto3", "awscli"]
80
+ dm_control = ["shimmy", "mujoco"]
81
+
82
+ # dependencies for algorithm variant (useful when you want to run a specific algorithm)
83
+ dqn = []
84
+ dqn_atari = ["ale-py", "AutoROM", "opencv-python"]
85
+ dqn_jax = ["jax", "jaxlib", "flax"]
86
+ dqn_atari_jax = [
87
+ "ale-py", "AutoROM", "opencv-python", # atari
88
+ "jax", "jaxlib", "flax" # jax
89
+ ]
90
+ c51 = []
91
+ c51_atari = ["ale-py", "AutoROM", "opencv-python"]
92
+ c51_jax = ["jax", "jaxlib", "flax"]
93
+ c51_atari_jax = [
94
+ "ale-py", "AutoROM", "opencv-python", # atari
95
+ "jax", "jaxlib", "flax" # jax
96
+ ]
97
+ ppo_atari_envpool_xla_jax_scan = [
98
+ "ale-py", "AutoROM", "opencv-python", # atari
99
+ "jax", "jaxlib", "flax", # jax
100
+ "envpool", # envpool
101
+ ]
102
+ qdagger_dqn_atari_impalacnn = [
103
+ "ale-py", "AutoROM", "opencv-python"
104
+ ]
105
+ qdagger_dqn_atari_jax_impalacnn = [
106
+ "ale-py", "AutoROM", "opencv-python", # atari
107
+ "jax", "jaxlib", "flax", # jax
108
+ ]