mmorales34 commited on
Commit
212030e
1 Parent(s): d1afa1a

pushing model

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ DQPN_x5.cleanrl_model filter=lfs diff=lfs merge=lfs -text
DQPN_x5.cleanrl_model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6758fd0967d4b0a49ce08ef51aafb49f913cbe3792ad32619f38c63630867a50
3
+ size 6747719
README.md ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - BreakoutNoFrameskip-v4
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: BreakoutNoFrameskip-v4
16
+ type: BreakoutNoFrameskip-v4
17
+ metrics:
18
+ - type: mean_reward
19
+ value: 3.97 +/- 0.00
20
+ name: mean_reward
21
+ verified: false
22
+ ---
23
+
24
+ # (CleanRL) **DQPN_freq** Agent Playing **BreakoutNoFrameskip-v4**
25
+
26
+ This is a trained model of a DQPN_freq agent playing BreakoutNoFrameskip-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/DQPN_x5.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_x5]"
36
+ python -m cleanrl_utils.enjoy --exp-name DQPN_x5 --env-id BreakoutNoFrameskip-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/pfunk/BreakoutNoFrameskip-v4-DQPN_x5-seed3/raw/main/dqpn_freq_atari.py
46
+ curl -OL https://huggingface.co/pfunk/BreakoutNoFrameskip-v4-DQPN_x5-seed3/raw/main/pyproject.toml
47
+ curl -OL https://huggingface.co/pfunk/BreakoutNoFrameskip-v4-DQPN_x5-seed3/raw/main/poetry.lock
48
+ poetry install --all-extras
49
+ python dqpn_freq_atari.py --track --wandb-entity pfunk --wandb-project-name dqpn --capture-video true --save-model true --upload-model true --hf-entity pfunk --exp-name DQPN_x5 --target-network-frequency 1000 --policy-network-frequency 5000 --seed 3 --env-id BreakoutNoFrameskip-v4
50
+ ```
51
+
52
+ # Hyperparameters
53
+ ```python
54
+ {'alg_type': 'dqpn_freq_atari.py',
55
+ 'batch_size': 32,
56
+ 'buffer_size': 1000000,
57
+ 'capture_video': True,
58
+ 'cuda': True,
59
+ 'double_learning': False,
60
+ 'end_e': 0.05,
61
+ 'env_id': 'BreakoutNoFrameskip-v4',
62
+ 'exp_name': 'DQPN_x5',
63
+ 'exploration_fraction': 0.2,
64
+ 'gamma': 0.99,
65
+ 'hf_entity': 'pfunk',
66
+ 'learning_rate': 0.0001,
67
+ 'learning_starts': 10000,
68
+ 'max_gradient_norm': inf,
69
+ 'policy_network_frequency': 5000,
70
+ 'policy_tau': 1.0,
71
+ 'save_model': True,
72
+ 'seed': 3,
73
+ 'start_e': 1.0,
74
+ 'target_network_frequency': 1000,
75
+ 'target_tau': 1.0,
76
+ 'torch_deterministic': True,
77
+ 'total_timesteps': 10000000,
78
+ 'track': True,
79
+ 'train_frequency': 1,
80
+ 'upload_model': True,
81
+ 'wandb_entity': 'pfunk',
82
+ 'wandb_project_name': 'dqpn'}
83
+ ```
84
+
dqpn_freq_atari.py ADDED
@@ -0,0 +1,350 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 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="PongNoFrameskip-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=0.0001,
57
+ help="the learning rate of the optimizer")
58
+ parser.add_argument("--max-gradient-norm", type=float, default=float('inf'),
59
+ help="gradient clipping value")
60
+ parser.add_argument("--double-learning", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
61
+ help="enable double learning DDQN")
62
+ parser.add_argument("--buffer-size", type=int, default=1000000,
63
+ help="the replay memory buffer size")
64
+ parser.add_argument("--gamma", type=float, default=0.99,
65
+ help="the discount factor gamma")
66
+ parser.add_argument("--target-tau", type=float, default=1.0,
67
+ help="the target network update rate")
68
+ parser.add_argument("--policy-tau", type=float, default=1.0,
69
+ help="the target network update rate")
70
+ parser.add_argument("--target-network-frequency", type=int, default=1000,
71
+ help="the timesteps it takes to update the target network")
72
+ parser.add_argument("--policy-network-frequency", type=int, default=5000,
73
+ help="the timesteps it takes to update the policy network")
74
+ parser.add_argument("--batch-size", type=int, default=32,
75
+ help="the batch size of sample from the reply memory")
76
+ parser.add_argument("--start-e", type=float, default=1.0,
77
+ help="the starting epsilon for exploration")
78
+ parser.add_argument("--end-e", type=float, default=0.05,
79
+ help="the ending epsilon for exploration")
80
+ parser.add_argument("--exploration-fraction", type=float, default=0.2,
81
+ help="the fraction of `total-timesteps` it takes from start-e to go end-e")
82
+ parser.add_argument("--learning-starts", type=int, default=10000,
83
+ help="timestep to start learning")
84
+ parser.add_argument("--train-frequency", type=int, default=1,
85
+ help="the frequency of training")
86
+ args = parser.parse_args()
87
+ # fmt: on
88
+ return args
89
+
90
+
91
+ def make_env(env_id, seed, idx, capture_video, run_name):
92
+ def thunk():
93
+ env = gym.make(env_id)
94
+ env = gym.wrappers.RecordEpisodeStatistics(env)
95
+ if capture_video:
96
+ if idx == 0:
97
+ env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
98
+ env = NoopResetEnv(env, noop_max=30)
99
+ env = MaxAndSkipEnv(env, skip=4)
100
+ env = EpisodicLifeEnv(env)
101
+ if "FIRE" in env.unwrapped.get_action_meanings():
102
+ env = FireResetEnv(env)
103
+ env = ClipRewardEnv(env)
104
+ env = gym.wrappers.ResizeObservation(env, (84, 84))
105
+ env = gym.wrappers.GrayScaleObservation(env)
106
+ env = gym.wrappers.FrameStack(env, 4)
107
+ env.seed(seed)
108
+ env.action_space.seed(seed)
109
+ env.observation_space.seed(seed)
110
+ return env
111
+
112
+ return thunk
113
+
114
+
115
+ # ALGO LOGIC: initialize agent here:
116
+ class QNetwork(nn.Module):
117
+ def __init__(self, env):
118
+ super().__init__()
119
+ self.network = nn.Sequential(
120
+ nn.Conv2d(4, 32, 8, stride=4),
121
+ nn.ReLU(),
122
+ nn.Conv2d(32, 64, 4, stride=2),
123
+ nn.ReLU(),
124
+ nn.Conv2d(64, 64, 3, stride=1),
125
+ nn.ReLU(),
126
+ nn.Flatten(),
127
+ nn.Linear(3136, 512),
128
+ nn.ReLU(),
129
+ nn.Linear(512, env.single_action_space.n),
130
+ )
131
+
132
+ def forward(self, x):
133
+ return self.network(x / 255.0)
134
+
135
+
136
+ def linear_schedule(start_e: float, end_e: float, duration: int, t: int):
137
+ slope = (end_e - start_e) / duration
138
+ return max(slope * t + start_e, end_e)
139
+
140
+
141
+ if __name__ == "__main__":
142
+ args = parse_args()
143
+ run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}"
144
+ if args.track:
145
+ import wandb
146
+
147
+ args.alg_type = os.path.basename(__file__)
148
+ wandb_sess = wandb.init(
149
+ project=args.wandb_project_name,
150
+ entity=args.wandb_entity,
151
+ config=vars(args),
152
+ save_code=True,
153
+ # group='string',
154
+ name=run_name,
155
+ sync_tensorboard=False,
156
+ monitor_gym=True,
157
+ )
158
+ writer = SummaryWriter(f"runs/{run_name}")
159
+ writer.add_text(
160
+ "hyperparameters",
161
+ "|param|value|\n|-|-|\n%s" % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])),
162
+ )
163
+
164
+ def log_value(name: str, x: float, y: int):
165
+ # writer.add_scalar(name, x, y)
166
+ wandb.log({name: x, "global_step": y})
167
+
168
+ # TRY NOT TO MODIFY: seeding
169
+ random.seed(args.seed)
170
+ np.random.seed(args.seed)
171
+ torch.manual_seed(args.seed)
172
+ torch.backends.cudnn.deterministic = args.torch_deterministic
173
+
174
+ device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
175
+
176
+ # env setup
177
+ envs = gym.vector.SyncVectorEnv([make_env(args.env_id, args.seed, 0, args.capture_video, run_name)])
178
+ assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported"
179
+
180
+ q_network = QNetwork(envs).to(device)
181
+ optimizer = optim.RMSprop(q_network.parameters(), lr=args.learning_rate)
182
+ target_network = QNetwork(envs).to(device)
183
+ policy_network = QNetwork(envs).to(device)
184
+ target_network.load_state_dict(q_network.state_dict())
185
+ policy_network.load_state_dict(q_network.state_dict())
186
+
187
+ rb = ReplayBuffer(
188
+ args.buffer_size,
189
+ envs.single_observation_space,
190
+ envs.single_action_space,
191
+ device,
192
+ optimize_memory_usage=True,
193
+ handle_timeout_termination=True,
194
+ )
195
+ start_time = time.time()
196
+ target_update_counter = 0
197
+ policy_update_counter = 0
198
+ episode_returns = []
199
+
200
+ # TRY NOT TO MODIFY: start the game
201
+ obs = envs.reset()
202
+ for global_step in range(args.total_timesteps):
203
+ # ALGO LOGIC: put action logic here
204
+ epsilon = linear_schedule(args.start_e, args.end_e, args.exploration_fraction * args.total_timesteps, global_step)
205
+
206
+ if random.random() < epsilon:
207
+ actions = np.array([envs.single_action_space.sample() for _ in range(envs.num_envs)])
208
+ else:
209
+ q_values = policy_network(torch.Tensor(obs).to(device))
210
+ actions = torch.argmax(q_values, dim=1).cpu().numpy()
211
+
212
+ # TRY NOT TO MODIFY: execute the game and log data.
213
+ next_obs, rewards, dones, infos = envs.step(actions)
214
+
215
+ # TRY NOT TO MODIFY: record rewards for plotting purposes
216
+ for info in infos:
217
+ if "episode" in info.keys():
218
+ episode_returns.append(info['episode']['r'])
219
+ episode_returns = episode_returns[-100:]
220
+ print(f"step={global_step}, return={info['episode']['r']}, sps={int(global_step / (time.time() - start_time))}")
221
+ log_value("perf/episodic_return", info["episode"]["r"], global_step)
222
+ log_value("perf/episodic_return_mean_100", np.mean(episode_returns), global_step)
223
+ log_value("perf/episodic_return_std_100", np.std(episode_returns), global_step)
224
+ log_value("debug/episodic_length", info["episode"]["l"], global_step)
225
+ log_value("ex2/epsilon", epsilon, global_step)
226
+ break
227
+
228
+ # TRY NOT TO MODIFY: save data to reply buffer; handle `terminal_observation`
229
+ real_next_obs = next_obs.copy()
230
+ for idx, d in enumerate(dones):
231
+ if d:
232
+ real_next_obs[idx] = infos[idx]["terminal_observation"]
233
+ rb.add(obs, real_next_obs, actions, rewards, dones, infos)
234
+
235
+ # TRY NOT TO MODIFY: CRUCIAL step easy to overlook
236
+ obs = next_obs
237
+
238
+ # ALGO LOGIC: training.
239
+ if global_step > args.learning_starts:
240
+ # NOTE: Current code does not work with train_frequency != 1
241
+ if global_step % args.train_frequency == 0:
242
+ data = rb.sample(args.batch_size)
243
+ with torch.no_grad():
244
+ if args.double_learning:
245
+ argmax_a = q_network(data.next_observations).max(1)[1].unsqueeze(1)
246
+ else:
247
+ argmax_a = target_network(data.next_observations).max(1)[1].unsqueeze(1)
248
+
249
+ target_max = target_network(data.next_observations).gather(1, argmax_a).squeeze()
250
+ td_target = data.rewards.flatten() + args.gamma * target_max * (1 - data.dones.flatten())
251
+
252
+ old_val = q_network(data.observations).gather(1, data.actions).squeeze()
253
+ loss = F.mse_loss(td_target, old_val)
254
+
255
+ if global_step % 100 == 0:
256
+
257
+ prev = old_val.detach().cpu().numpy()
258
+ new = td_target.detach().cpu().numpy()
259
+ diff, a_diff = new-prev, np.abs(new-prev)
260
+
261
+ mean, a_mean = np.mean(diff), np.mean(a_diff)
262
+ median, a_median = np.median(diff), np.median(a_diff)
263
+ maximum, a_maximum = np.max(diff), np.max(a_diff)
264
+ minimum, a_minimum = np.min(diff), np.min(a_diff)
265
+ std, a_std = np.std(diff), np.std(a_diff)
266
+ below, a_below = mean - std, a_mean - a_std
267
+ above, a_above = mean + std, a_mean + a_std
268
+ pu_scalar, a_pu_scalar = 2 * mean / maximum, 2 * a_mean / a_maximum
269
+ policy_frequency_scalar_ratio = args.policy_network_frequency * pu_scalar
270
+ a_policy_frequency_scalar_ratio = args.policy_network_frequency * a_pu_scalar
271
+
272
+ log_value("losses/td_loss", loss, global_step)
273
+ log_value("losses/q_values", old_val.mean().item(), global_step)
274
+ log_value("td/mean", mean, global_step)
275
+ log_value("td/a_mean", a_mean, global_step)
276
+ log_value("td/median", median, global_step)
277
+ log_value("td/a_median", a_median, global_step)
278
+ log_value("td/max", maximum, global_step)
279
+ log_value("td/a_max", a_maximum, global_step)
280
+ log_value("td/min", minimum, global_step)
281
+ log_value("td/a_min", a_minimum, global_step)
282
+ log_value("td/std", std, global_step)
283
+ log_value("td/a_std", a_std, global_step)
284
+ log_value("td/below", below, global_step)
285
+ log_value("td/a_below", a_below, global_step)
286
+ log_value("td/above", above, global_step)
287
+ log_value("td/a_above", a_above, global_step)
288
+ log_value("alg/pu_scalar", pu_scalar, global_step)
289
+ log_value("alg/a_pu_scalar", a_pu_scalar, global_step)
290
+ log_value("alg/policy_frequency_scalar_ratio", policy_frequency_scalar_ratio, global_step)
291
+ log_value("alg/a_policy_frequency_scalar_ratio", a_policy_frequency_scalar_ratio, global_step)
292
+ log_value("debug/steps_per_second", int(global_step / (time.time() - start_time)), global_step)
293
+
294
+ # optimize the model
295
+ optimizer.zero_grad()
296
+ loss.backward()
297
+ torch.nn.utils.clip_grad_norm_(q_network.parameters(),
298
+ args.max_gradient_norm)
299
+ optimizer.step()
300
+
301
+ # update target network
302
+ if global_step % args.target_network_frequency == 0:
303
+ target_update_counter += 1
304
+ for target_network_param, q_network_param in zip(target_network.parameters(), q_network.parameters()):
305
+ target_network_param.data.copy_(
306
+ args.target_tau * q_network_param.data + (1.0 - args.target_tau) * target_network_param.data
307
+ )
308
+
309
+ # update policy network
310
+ if global_step % args.policy_network_frequency == 0:
311
+ policy_update_counter += 1
312
+ for policy_network_param, q_network_param in zip(policy_network.parameters(), q_network.parameters()):
313
+ policy_network_param.data.copy_(
314
+ args.policy_tau * q_network_param.data + (1.0 - args.policy_tau) * policy_network_param.data
315
+ )
316
+
317
+ if global_step % 100 == 0:
318
+ log_value("alg/n_target_update", target_update_counter, global_step)
319
+ log_value("alg/n_policy_update", policy_update_counter, global_step)
320
+
321
+ if args.save_model:
322
+ model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
323
+ torch.save(policy_network.state_dict(), model_path)
324
+ print(f"model saved to {model_path}")
325
+ from cleanrl_utils.evals.dqn_eval import evaluate
326
+
327
+ episodic_returns = evaluate(
328
+ model_path,
329
+ make_env,
330
+ args.env_id,
331
+ eval_episodes=10,
332
+ run_name=f"{run_name}-eval",
333
+ Model=QNetwork,
334
+ device=device,
335
+ epsilon=0.05,
336
+ )
337
+ for idx, episodic_return in enumerate(episodic_returns):
338
+ log_value("eval/episodic_return", episodic_return, idx)
339
+
340
+
341
+ if args.upload_model:
342
+ from cleanrl_utils.huggingface import push_to_hub
343
+
344
+ repo_name = f"{args.env_id}-{args.exp_name}-seed{args.seed}"
345
+ repo_id = f"{args.hf_entity}/{repo_name}" if args.hf_entity else repo_name
346
+ push_to_hub(args, np.mean(episode_returns), repo_id, "DQPN_freq", f"runs/{run_name}", f"videos/{run_name}-eval")
347
+
348
+ wandb_sess.finish()
349
+ envs.close()
350
+ writer.close()
events.out.tfevents.1697580073.redi.9014.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e2f0f232e9eea2e01ef94604980315ac3c3a9b92ad79b82cefd304fa33a1e7be
3
+ size 755
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 (23.4 kB). View file
 
videos/BreakoutNoFrameskip-v4__DQPN_x5__3__1697580065-eval/rl-video-episode-0.mp4 ADDED
Binary file (23.9 kB). View file
 
videos/BreakoutNoFrameskip-v4__DQPN_x5__3__1697580065-eval/rl-video-episode-1.mp4 ADDED
Binary file (34.5 kB). View file
 
videos/BreakoutNoFrameskip-v4__DQPN_x5__3__1697580065-eval/rl-video-episode-8.mp4 ADDED
Binary file (23.4 kB). View file