mmorales34 commited on
Commit
c930d43
1 Parent(s): a686b6e

pushing model

Browse files
DQPN_freq_5000_0.99.cleanrl_model ADDED
Binary file (276 kB). View file
 
README.md ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - CartPole-v1
4
+ - deep-reinforcement-learning
5
+ - reinforcement-learning
6
+ - custom-implementation
7
+ library_name: cleanrl
8
+ model-index:
9
+ - name: DQPN_freq
10
+ results:
11
+ - task:
12
+ type: reinforcement-learning
13
+ name: reinforcement-learning
14
+ dataset:
15
+ name: CartPole-v1
16
+ type: CartPole-v1
17
+ metrics:
18
+ - type: mean_reward
19
+ value: 343.71 +/- 0.00
20
+ name: mean_reward
21
+ verified: false
22
+ ---
23
+
24
+ # (CleanRL) **DQPN_freq** Agent Playing **CartPole-v1**
25
+
26
+ This is a trained model of a DQPN_freq agent playing CartPole-v1.
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/DQPN_freq_5000_0.99.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[DQPN_freq_5000_0.99]"
36
+ python -m cleanrl_utils.enjoy --exp-name DQPN_freq_5000_0.99 --env-id CartPole-v1
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/pfunk/CartPole-v1-DQPN_freq_5000_0.99-seed3/raw/main/dqpn_freq.py
46
+ curl -OL https://huggingface.co/pfunk/CartPole-v1-DQPN_freq_5000_0.99-seed3/raw/main/pyproject.toml
47
+ curl -OL https://huggingface.co/pfunk/CartPole-v1-DQPN_freq_5000_0.99-seed3/raw/main/poetry.lock
48
+ poetry install --all-extras
49
+ python dqpn_freq.py --track --wandb-entity pfunk --wandb-project-name dqpn --capture-video true --save-model true --upload-model true --hf-entity pfunk --exp-name DQPN_freq_5000_0.99 --gamma 0.99 --policy-network-frequency 5000 --seed 3
50
+ ```
51
+
52
+ # Hyperparameters
53
+ ```python
54
+ {'alg_type': 'dqpn_freq.py',
55
+ 'batch_size': 256,
56
+ 'buffer_size': 300000,
57
+ 'capture_video': True,
58
+ 'cuda': True,
59
+ 'end_e': 0.1,
60
+ 'env_id': 'CartPole-v1',
61
+ 'exp_name': 'DQPN_freq_5000_0.99',
62
+ 'exploration_fraction': 0.2,
63
+ 'gamma': 0.99,
64
+ 'hf_entity': 'pfunk',
65
+ 'learning_rate': 0.0001,
66
+ 'learning_starts': 1000,
67
+ 'policy_network_frequency': 5000,
68
+ 'policy_tau': 1.0,
69
+ 'save_model': True,
70
+ 'seed': 3,
71
+ 'start_e': 1.0,
72
+ 'target_network_frequency': 20,
73
+ 'target_tau': 1.0,
74
+ 'torch_deterministic': True,
75
+ 'total_timesteps': 500000,
76
+ 'track': True,
77
+ 'train_frequency': 1,
78
+ 'upload_model': True,
79
+ 'wandb_entity': 'pfunk',
80
+ 'wandb_project_name': 'dqpn'}
81
+ ```
82
+
dqpn_freq.py ADDED
@@ -0,0 +1,318 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/dqn/#dqnpy
2
+ import argparse
3
+ import os
4
+ import random
5
+ import time
6
+ from distutils.util import strtobool
7
+
8
+ import 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.buffers import ReplayBuffer
15
+ from torch.utils.tensorboard import SummaryWriter
16
+
17
+
18
+ def parse_args():
19
+ # fmt: off
20
+ parser = argparse.ArgumentParser()
21
+ parser.add_argument("--exp-name", type=str, default=os.path.basename(__file__).rstrip(".py"),
22
+ help="the name of this experiment")
23
+ parser.add_argument("--seed", type=int, default=1,
24
+ help="seed of the experiment")
25
+ parser.add_argument("--torch-deterministic", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
26
+ help="if toggled, `torch.backends.cudnn.deterministic=False`")
27
+ parser.add_argument("--cuda", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
28
+ help="if toggled, cuda will be enabled by default")
29
+ parser.add_argument("--track", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
30
+ help="if toggled, this experiment will be tracked with Weights and Biases")
31
+ parser.add_argument("--wandb-project-name", type=str, default="cleanRL",
32
+ help="the wandb's project name")
33
+ parser.add_argument("--wandb-entity", type=str, default=None,
34
+ help="the entity (team) of wandb's project")
35
+ parser.add_argument("--capture-video", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
36
+ help="whether to capture videos of the agent performances (check out `videos` folder)")
37
+ parser.add_argument("--save-model", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
38
+ help="whether to save model into the `runs/{run_name}` folder")
39
+ parser.add_argument("--upload-model", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
40
+ help="whether to upload the saved model to huggingface")
41
+ parser.add_argument("--hf-entity", type=str, default="",
42
+ help="the user or org name of the model repository from the Hugging Face Hub")
43
+
44
+ # Algorithm specific arguments
45
+ parser.add_argument("--env-id", type=str, default="CartPole-v1",
46
+ help="the id of the environment")
47
+ parser.add_argument("--total-timesteps", type=int, default=500000,
48
+ help="total timesteps of the experiments")
49
+ parser.add_argument("--learning-rate", type=float, default=0.0001,
50
+ help="the learning rate of the optimizer")
51
+ parser.add_argument("--buffer-size", type=int, default=300000,
52
+ help="the replay memory buffer size")
53
+ parser.add_argument("--gamma", type=float, default=1.0,
54
+ help="the discount factor gamma")
55
+ parser.add_argument("--target-tau", type=float, default=1.0,
56
+ help="the target network update rate")
57
+ parser.add_argument("--policy-tau", type=float, default=1.0,
58
+ help="the target network update rate")
59
+ parser.add_argument("--target-network-frequency", type=int, default=20,
60
+ help="the timesteps it takes to update the target network")
61
+ parser.add_argument("--policy-network-frequency", type=int, default=1000,
62
+ help="the timesteps it takes to update the policy network")
63
+ parser.add_argument("--batch-size", type=int, default=256,
64
+ help="the batch size of sample from the reply memory")
65
+ parser.add_argument("--start-e", type=float, default=1.0,
66
+ help="the starting epsilon for exploration")
67
+ parser.add_argument("--end-e", type=float, default=0.1,
68
+ help="the ending epsilon for exploration")
69
+ parser.add_argument("--exploration-fraction", type=float, default=0.2,
70
+ help="the fraction of `total-timesteps` it takes from start-e to go end-e")
71
+ parser.add_argument("--learning-starts", type=int, default=1000,
72
+ help="timestep to start learning")
73
+ parser.add_argument("--train-frequency", type=int, default=1,
74
+ help="the frequency of training")
75
+ args = parser.parse_args()
76
+ # fmt: on
77
+ return args
78
+
79
+
80
+ def make_env(env_id, seed, idx, capture_video, run_name):
81
+ def thunk():
82
+ env = gym.make(env_id)
83
+ env = gym.wrappers.RecordEpisodeStatistics(env)
84
+ if capture_video:
85
+ if idx == 0:
86
+ env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
87
+ env.seed(seed)
88
+ env.action_space.seed(seed)
89
+ env.observation_space.seed(seed)
90
+ return env
91
+
92
+ return thunk
93
+
94
+
95
+ # ALGO LOGIC: initialize agent here:
96
+ class QNetwork(nn.Module):
97
+ def __init__(self, env):
98
+ super().__init__()
99
+ self.network = nn.Sequential(
100
+ nn.Linear(np.array(env.single_observation_space.shape).prod(), 512),
101
+ nn.ReLU(),
102
+ nn.Linear(512, 128),
103
+ nn.ReLU(),
104
+ nn.Linear(128, env.single_action_space.n),
105
+ )
106
+
107
+ def forward(self, x):
108
+ return self.network(x)
109
+
110
+
111
+ def linear_schedule(start_e: float, end_e: float, duration: int, t: int):
112
+ slope = (end_e - start_e) / duration
113
+ return max(slope * t + start_e, end_e)
114
+
115
+
116
+ if __name__ == "__main__":
117
+ args = parse_args()
118
+ run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
119
+ if args.track:
120
+ import wandb
121
+
122
+ args.alg_type = os.path.basename(__file__)
123
+ wandb_sess = wandb.init(
124
+ project=args.wandb_project_name,
125
+ entity=args.wandb_entity,
126
+ config=vars(args),
127
+ save_code=True,
128
+ # group='string',
129
+ name=run_name,
130
+ sync_tensorboard=False,
131
+ monitor_gym=True,
132
+ )
133
+ writer = SummaryWriter(f"runs/{run_name}")
134
+ writer.add_text(
135
+ "hyperparameters",
136
+ "|param|value|\n|-|-|\n%s" % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])),
137
+ )
138
+
139
+ def log_value(name: str, x: float, y: int):
140
+ # writer.add_scalar(name, x, y)
141
+ wandb.log({name: x, "global_step": y})
142
+
143
+ # TRY NOT TO MODIFY: seeding
144
+ random.seed(args.seed)
145
+ np.random.seed(args.seed)
146
+ torch.manual_seed(args.seed)
147
+ torch.backends.cudnn.deterministic = args.torch_deterministic
148
+
149
+ device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
150
+
151
+ # env setup
152
+ envs = gym.vector.SyncVectorEnv([make_env(args.env_id, args.seed, 0, args.capture_video, run_name)])
153
+ assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported"
154
+ envs.seed(args.seed)
155
+
156
+ q_network = QNetwork(envs).to(device)
157
+ optimizer = optim.RMSprop(q_network.parameters(), lr=args.learning_rate)
158
+ target_network = QNetwork(envs).to(device)
159
+ policy_network = QNetwork(envs).to(device)
160
+ target_network.load_state_dict(q_network.state_dict())
161
+ policy_network.load_state_dict(q_network.state_dict())
162
+
163
+ rb = ReplayBuffer(
164
+ args.buffer_size,
165
+ envs.single_observation_space,
166
+ envs.single_action_space,
167
+ device,
168
+ optimize_memory_usage=True,
169
+ handle_timeout_termination=True,
170
+ )
171
+ start_time = time.time()
172
+ policy_update_counter = 0
173
+ episode_returns = []
174
+
175
+ # TRY NOT TO MODIFY: start the game
176
+ obs = envs.reset()
177
+ for global_step in range(args.total_timesteps):
178
+ # ALGO LOGIC: put action logic here
179
+ epsilon = linear_schedule(args.start_e, args.end_e, args.exploration_fraction * args.total_timesteps, global_step)
180
+
181
+ if random.random() < epsilon:
182
+ actions = np.array([envs.single_action_space.sample() for _ in range(envs.num_envs)])
183
+ else:
184
+ q_values = policy_network(torch.Tensor(obs).to(device))
185
+ actions = torch.argmax(q_values, dim=1).cpu().numpy()
186
+
187
+ # TRY NOT TO MODIFY: execute the game and log data.
188
+ next_obs, rewards, dones, infos = envs.step(actions)
189
+
190
+ # TRY NOT TO MODIFY: record rewards for plotting purposes
191
+ for info in infos:
192
+ if "episode" in info.keys():
193
+ episode_returns.append(info['episode']['r'])
194
+ episode_returns = episode_returns[-100:]
195
+ print(f"step={global_step}, return={info['episode']['r']}, sps={int(global_step / (time.time() - start_time))}")
196
+ log_value("perf/episodic_return", info["episode"]["r"], global_step)
197
+ log_value("perf/episodic_return_mean_100", np.mean(episode_returns), global_step)
198
+ log_value("perf/episodic_return_std_100", np.std(episode_returns), global_step)
199
+ log_value("debug/episodic_length", info["episode"]["l"], global_step)
200
+ log_value("ex2/epsilon", epsilon, global_step)
201
+ break
202
+
203
+ # TRY NOT TO MODIFY: save data to reply buffer; handle `terminal_observation`
204
+ real_next_obs = next_obs.copy()
205
+ for idx, d in enumerate(dones):
206
+ if d:
207
+ real_next_obs[idx] = infos[idx]["terminal_observation"]
208
+ rb.add(obs, real_next_obs, actions, rewards, dones, infos)
209
+
210
+ # TRY NOT TO MODIFY: CRUCIAL step easy to overlook
211
+ obs = next_obs
212
+
213
+ # ALGO LOGIC: training.
214
+ if global_step > args.learning_starts:
215
+ # NOTE: Current code does not work with train_frequency != 1
216
+ if global_step % args.train_frequency == 0:
217
+ data = rb.sample(args.batch_size)
218
+ with torch.no_grad():
219
+ target_max, _ = target_network(data.next_observations).max(dim=1)
220
+ td_target = data.rewards.flatten() + args.gamma * target_max * (1 - data.dones.flatten())
221
+ old_val = q_network(data.observations).gather(1, data.actions).squeeze()
222
+ loss = F.mse_loss(td_target, old_val)
223
+
224
+ if global_step % 100 == 0:
225
+
226
+ prev = old_val.detach().cpu().numpy()
227
+ new = td_target.detach().cpu().numpy()
228
+ diff, a_diff = new-prev, np.abs(new-prev)
229
+
230
+ mean, a_mean = np.mean(diff), np.mean(a_diff)
231
+ median, a_median = np.median(diff), np.median(a_diff)
232
+ maximum, a_maximum = np.max(diff), np.max(a_diff)
233
+ minimum, a_minimum = np.min(diff), np.min(a_diff)
234
+ std, a_std = np.std(diff), np.std(a_diff)
235
+ below, a_below = mean - std, a_mean - a_std
236
+ above, a_above = mean + std, a_mean + a_std
237
+ pu_scalar, a_pu_scalar = 2 * mean / maximum, 2 * a_mean / a_maximum
238
+ policy_frequency_scalar_ratio = args.policy_network_frequency * pu_scalar
239
+ a_policy_frequency_scalar_ratio = args.policy_network_frequency * a_pu_scalar
240
+
241
+ log_value("losses/td_loss", loss, global_step)
242
+ log_value("losses/q_values", old_val.mean().item(), global_step)
243
+ log_value("td/mean", mean, global_step)
244
+ log_value("td/a_mean", a_mean, global_step)
245
+ log_value("td/median", median, global_step)
246
+ log_value("td/a_median", a_median, global_step)
247
+ log_value("td/max", maximum, global_step)
248
+ log_value("td/a_max", a_maximum, global_step)
249
+ log_value("td/min", minimum, global_step)
250
+ log_value("td/a_min", a_minimum, global_step)
251
+ log_value("td/std", std, global_step)
252
+ log_value("td/a_std", a_std, global_step)
253
+ log_value("td/below", below, global_step)
254
+ log_value("td/a_below", a_below, global_step)
255
+ log_value("td/above", above, global_step)
256
+ log_value("td/a_above", a_above, global_step)
257
+ log_value("pu/pu_scalar", pu_scalar, global_step)
258
+ log_value("pu/a_pu_scalar", a_pu_scalar, global_step)
259
+ log_value("pu/policy_frequency_scalar_ratio", policy_frequency_scalar_ratio, global_step)
260
+ log_value("pu/a_policy_frequency_scalar_ratio", a_policy_frequency_scalar_ratio, global_step)
261
+ log_value("debug/steps_per_second", int(global_step / (time.time() - start_time)), global_step)
262
+
263
+ # optimize the model
264
+ optimizer.zero_grad()
265
+ loss.backward()
266
+ optimizer.step()
267
+
268
+ # update target network
269
+ if global_step % args.target_network_frequency == 0:
270
+ for target_network_param, q_network_param in zip(target_network.parameters(), q_network.parameters()):
271
+ target_network_param.data.copy_(
272
+ args.target_tau * q_network_param.data + (1.0 - args.target_tau) * target_network_param.data
273
+ )
274
+
275
+ # update policy network
276
+ if global_step % args.policy_network_frequency == 0:
277
+ policy_update_counter += 1
278
+ for policy_network_param, q_network_param in zip(policy_network.parameters(), q_network.parameters()):
279
+ policy_network_param.data.copy_(
280
+ args.policy_tau * q_network_param.data + (1.0 - args.policy_tau) * policy_network_param.data
281
+ )
282
+ for target_network_param, q_network_param in zip(target_network.parameters(), q_network.parameters()):
283
+ target_network_param.data.copy_(
284
+ args.target_tau * q_network_param.data + (1.0 - args.target_tau) * target_network_param.data
285
+ )
286
+ if global_step % 100 == 0:
287
+ log_value("pu/n_policy_update", policy_update_counter, global_step)
288
+
289
+ if args.save_model:
290
+ model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
291
+ torch.save(policy_network.state_dict(), model_path)
292
+ print(f"model saved to {model_path}")
293
+ from cleanrl_utils.evals.dqn_eval import evaluate
294
+
295
+ episodic_returns = evaluate(
296
+ model_path,
297
+ make_env,
298
+ args.env_id,
299
+ eval_episodes=10,
300
+ run_name=f"{run_name}-eval",
301
+ Model=QNetwork,
302
+ device=device,
303
+ epsilon=0.05,
304
+ )
305
+ for idx, episodic_return in enumerate(episodic_returns):
306
+ log_value("eval/episodic_return", episodic_return, idx)
307
+
308
+
309
+ if args.upload_model:
310
+ from cleanrl_utils.huggingface import push_to_hub
311
+
312
+ repo_name = f"{args.env_id}-{args.exp_name}-seed{args.seed}"
313
+ repo_id = f"{args.hf_entity}/{repo_name}" if args.hf_entity else repo_name
314
+ push_to_hub(args, np.mean(episode_returns), repo_id, "DQPN_freq", f"runs/{run_name}", f"videos/{run_name}-eval")
315
+
316
+ wandb_sess.finish()
317
+ envs.close()
318
+ writer.close()
events.out.tfevents.1679115745.rhea.136041.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da6960acc23cf4e659d472a0dec1f49036fe0e3963491fe4abd44cf10d54e946
3
+ size 696
poetry.lock ADDED
The diff for this file is too large to render. See raw diff
 
pyproject.toml ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 <costa.huang@outlook.com>"]
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,<3.10"
16
+ tensorboard = "^2.10.0"
17
+ wandb = "^0.13.6"
18
+ gym = "0.23.1"
19
+ torch = ">=1.12.1"
20
+ stable-baselines3 = "1.2.0"
21
+ gymnasium = "^0.26.3"
22
+ moviepy = "^1.0.3"
23
+ pygame = "2.1.0"
24
+ huggingface-hub = "^0.11.1"
25
+
26
+ ale-py = {version = "0.7.4", optional = true}
27
+ AutoROM = {extras = ["accept-rom-license"], version = "^0.4.2"}
28
+ opencv-python = {version = "^4.6.0.66", optional = true}
29
+ pybullet = {version = "3.1.8", optional = true}
30
+ procgen = {version = "^0.10.7", optional = true}
31
+ pytest = {version = "^7.1.3", optional = true}
32
+ mujoco = {version = "^2.2", optional = true}
33
+ imageio = {version = "^2.14.1", optional = true}
34
+ free-mujoco-py = {version = "^2.1.6", optional = true}
35
+ mkdocs-material = {version = "^8.4.3", optional = true}
36
+ markdown-include = {version = "^0.7.0", optional = true}
37
+ jax = {version = "^0.3.17", optional = true}
38
+ jaxlib = {version = "^0.3.15", optional = true}
39
+ flax = {version = "^0.6.0", optional = true}
40
+ optuna = {version = "^3.0.1", optional = true}
41
+ optuna-dashboard = {version = "^0.7.2", optional = true}
42
+ rich = {version = "<12.0", optional = true}
43
+ envpool = {version = "^0.6.4", optional = true}
44
+ PettingZoo = {version = "1.18.1", optional = true}
45
+ SuperSuit = {version = "3.4.0", optional = true}
46
+ multi-agent-ale-py = {version = "0.1.11", optional = true}
47
+ boto3 = {version = "^1.24.70", optional = true}
48
+ awscli = {version = "^1.25.71", optional = true}
49
+ shimmy = {version = "^0.1.0", optional = true}
50
+ dm-control = {version = "^1.0.8", optional = true}
51
+
52
+ [tool.poetry.group.dev.dependencies]
53
+ pre-commit = "^2.20.0"
54
+
55
+ [tool.poetry.group.atari]
56
+ optional = true
57
+ [tool.poetry.group.atari.dependencies]
58
+ ale-py = "0.7.4"
59
+ AutoROM = {extras = ["accept-rom-license"], version = "^0.4.2"}
60
+ opencv-python = "^4.6.0.66"
61
+
62
+ [tool.poetry.group.pybullet]
63
+ optional = true
64
+ [tool.poetry.group.pybullet.dependencies]
65
+ pybullet = "3.1.8"
66
+
67
+ [tool.poetry.group.procgen]
68
+ optional = true
69
+ [tool.poetry.group.procgen.dependencies]
70
+ procgen = "^0.10.7"
71
+
72
+ [tool.poetry.group.pytest]
73
+ optional = true
74
+ [tool.poetry.group.pytest.dependencies]
75
+ pytest = "^7.1.3"
76
+
77
+ [tool.poetry.group.mujoco]
78
+ optional = true
79
+ [tool.poetry.group.mujoco.dependencies]
80
+ mujoco = "^2.2"
81
+ imageio = "^2.14.1"
82
+
83
+ [tool.poetry.group.mujoco_py]
84
+ optional = true
85
+ [tool.poetry.group.mujoco_py.dependencies]
86
+ free-mujoco-py = "^2.1.6"
87
+
88
+ [tool.poetry.group.docs]
89
+ optional = true
90
+ [tool.poetry.group.docs.dependencies]
91
+ mkdocs-material = "^8.4.3"
92
+ markdown-include = "^0.7.0"
93
+
94
+ [tool.poetry.group.jax]
95
+ optional = true
96
+ [tool.poetry.group.jax.dependencies]
97
+ jax = "^0.3.17"
98
+ jaxlib = "^0.3.15"
99
+ flax = "^0.6.0"
100
+
101
+ [tool.poetry.group.optuna]
102
+ optional = true
103
+ [tool.poetry.group.optuna.dependencies]
104
+ optuna = "^3.0.1"
105
+ optuna-dashboard = "^0.7.2"
106
+ rich = "<12.0"
107
+
108
+ [tool.poetry.group.envpool]
109
+ optional = true
110
+ [tool.poetry.group.envpool.dependencies]
111
+ envpool = "^0.6.4"
112
+
113
+ [tool.poetry.group.pettingzoo]
114
+ optional = true
115
+ [tool.poetry.group.pettingzoo.dependencies]
116
+ PettingZoo = "1.18.1"
117
+ SuperSuit = "3.4.0"
118
+ multi-agent-ale-py = "0.1.11"
119
+
120
+ [tool.poetry.group.cloud]
121
+ optional = true
122
+ [tool.poetry.group.cloud.dependencies]
123
+ boto3 = "^1.24.70"
124
+ awscli = "^1.25.71"
125
+
126
+ [tool.poetry.group.isaacgym]
127
+ optional = true
128
+ [tool.poetry.group.isaacgym.dependencies]
129
+ isaacgymenvs = {git = "https://github.com/vwxyzjn/IsaacGymEnvs.git", rev = "poetry"}
130
+ isaacgym = {path = "cleanrl/ppo_continuous_action_isaacgym/isaacgym", develop = true}
131
+
132
+ [tool.poetry.group.dm_control]
133
+ optional = true
134
+ [tool.poetry.group.dm_control.dependencies]
135
+ shimmy = "^0.1.0"
136
+ dm-control = "^1.0.8"
137
+ mujoco = "^2.2"
138
+
139
+ [build-system]
140
+ requires = ["poetry-core"]
141
+ build-backend = "poetry.core.masonry.api"
142
+
143
+ [tool.poetry.extras]
144
+ atari = ["ale-py", "AutoROM", "opencv-python"]
145
+ pybullet = ["pybullet"]
146
+ procgen = ["procgen"]
147
+ plot = ["pandas", "seaborn"]
148
+ pytest = ["pytest"]
149
+ mujoco = ["mujoco", "imageio"]
150
+ mujoco_py = ["free-mujoco-py"]
151
+ jax = ["jax", "jaxlib", "flax"]
152
+ docs = ["mkdocs-material", "markdown-include"]
153
+ envpool = ["envpool"]
154
+ optuna = ["optuna", "optuna-dashboard", "rich"]
155
+ pettingzoo = ["PettingZoo", "SuperSuit", "multi-agent-ale-py"]
156
+ cloud = ["boto3", "awscli"]
157
+ dm_control = ["shimmy", "dm-control", "mujoco"]
158
+
159
+ # dependencies for algorithm variant (useful when you want to run a specific algorithm)
160
+ dqn = []
161
+ dqn_atari = ["ale-py", "AutoROM", "opencv-python"]
162
+ dqn_jax = ["jax", "jaxlib", "flax"]
163
+ dqn_atari_jax = [
164
+ "ale-py", "AutoROM", "opencv-python", # atari
165
+ "jax", "jaxlib", "flax" # jax
166
+ ]
167
+ c51 = []
168
+ c51_atari = ["ale-py", "AutoROM", "opencv-python"]
169
+ c51_jax = ["jax", "jaxlib", "flax"]
170
+ c51_atari_jax = [
171
+ "ale-py", "AutoROM", "opencv-python", # atari
172
+ "jax", "jaxlib", "flax" # jax
173
+ ]
174
+ ppo_atari_envpool_xla_jax_scan = [
175
+ "ale-py", "AutoROM", "opencv-python", # atari
176
+ "jax", "jaxlib", "flax", # jax
177
+ "envpool", # envpool
178
+ ]
replay.mp4 ADDED
Binary file (53.7 kB). View file
 
videos/CartPole-v1__DQPN_freq_5000_0.99__3__1679115741-eval/rl-video-episode-0.mp4 ADDED
Binary file (64.3 kB). View file
 
videos/CartPole-v1__DQPN_freq_5000_0.99__3__1679115741-eval/rl-video-episode-1.mp4 ADDED
Binary file (57 kB). View file
 
videos/CartPole-v1__DQPN_freq_5000_0.99__3__1679115741-eval/rl-video-episode-8.mp4 ADDED
Binary file (53.7 kB). View file