Commit
·
2900a24
1
Parent(s):
1c62517
Update README.md
Browse files
README.md
CHANGED
@@ -4,4 +4,129 @@ tags:
|
|
4 |
- reinforcement-learning
|
5 |
- stable-baselines3
|
6 |
---
|
7 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
- reinforcement-learning
|
5 |
- stable-baselines3
|
6 |
---
|
7 |
+
# PPO Agent playing QbertNoFrameskip-v4
|
8 |
+
This is a trained model of a **PPO agent playing QbertNoFrameskip-v4 using the [stable-baselines3 library](https://stable-baselines3.readthedocs.io/en/master/index.html)**.
|
9 |
+
|
10 |
+
<video src="https://huggingface.co/ThomasSimonini/ppo-QbertNoFrameskip-v4/resolve/main/output.mp4" controls autoplay loop></video>
|
11 |
+
|
12 |
+
## Evaluation Results
|
13 |
+
Mean_reward: `15685.00 +/- 115.217`
|
14 |
+
|
15 |
+
# Usage (with Stable-baselines3)
|
16 |
+
- You need to use `gym==0.19` since it **includes Atari Roms**.
|
17 |
+
- The Action Space is 6 since we use only **possible actions in this game**.
|
18 |
+
|
19 |
+
|
20 |
+
Watch your agent interacts :
|
21 |
+
|
22 |
+
```python
|
23 |
+
# Import the libraries
|
24 |
+
import os
|
25 |
+
|
26 |
+
import gym
|
27 |
+
|
28 |
+
from stable_baselines3 import PPO
|
29 |
+
from stable_baselines3.common.vec_env import VecNormalize
|
30 |
+
|
31 |
+
from stable_baselines3.common.env_util import make_atari_env
|
32 |
+
from stable_baselines3.common.vec_env import VecFrameStack
|
33 |
+
|
34 |
+
from huggingface_sb3 import load_from_hub, push_to_hub
|
35 |
+
|
36 |
+
# Load the model
|
37 |
+
checkpoint = load_from_hub("ThomasSimonini/ppo-QbertNoFrameskip-v4", "ppo-QbertNoFrameskip-v4.zip")
|
38 |
+
|
39 |
+
# Because we using 3.7 on Colab and this agent was trained with 3.8 to avoid Pickle errors:
|
40 |
+
custom_objects = {
|
41 |
+
"learning_rate": 0.0,
|
42 |
+
"lr_schedule": lambda _: 0.0,
|
43 |
+
"clip_range": lambda _: 0.0,
|
44 |
+
}
|
45 |
+
|
46 |
+
model= PPO.load(checkpoint, custom_objects=custom_objects)
|
47 |
+
|
48 |
+
env = make_atari_env('QbertNoFrameskip-v4', n_envs=1)
|
49 |
+
env = VecFrameStack(env, n_stack=4)
|
50 |
+
|
51 |
+
obs = env.reset()
|
52 |
+
while True:
|
53 |
+
action, _states = model.predict(obs)
|
54 |
+
obs, rewards, dones, info = env.step(action)
|
55 |
+
env.render()
|
56 |
+
```
|
57 |
+
|
58 |
+
|
59 |
+
## Training Code
|
60 |
+
```python
|
61 |
+
import wandb
|
62 |
+
import gym
|
63 |
+
|
64 |
+
from stable_baselines3 import PPO
|
65 |
+
from stable_baselines3.common.env_util import make_atari_env
|
66 |
+
from stable_baselines3.common.vec_env import VecFrameStack, VecVideoRecorder
|
67 |
+
from stable_baselines3.common.callbacks import CheckpointCallback
|
68 |
+
|
69 |
+
from wandb.integration.sb3 import WandbCallback
|
70 |
+
|
71 |
+
from huggingface_sb3 import load_from_hub, push_to_hub
|
72 |
+
|
73 |
+
config = {
|
74 |
+
"env_name": "QbertNoFrameskip-v4",
|
75 |
+
"num_envs": 8,
|
76 |
+
"total_timesteps": int(10e6),
|
77 |
+
"seed": 1194709219,
|
78 |
+
}
|
79 |
+
|
80 |
+
run = wandb.init(
|
81 |
+
project="HFxSB3",
|
82 |
+
config = config,
|
83 |
+
sync_tensorboard = True, # Auto-upload sb3's tensorboard metrics
|
84 |
+
monitor_gym = True, # Auto-upload the videos of agents playing the game
|
85 |
+
save_code = True, # Save the code to W&B
|
86 |
+
)
|
87 |
+
|
88 |
+
# There already exists an environment generator
|
89 |
+
# that will make and wrap atari environments correctly.
|
90 |
+
# Here we are also multi-worker training (n_envs=8 => 8 environments)
|
91 |
+
env = make_atari_env(config["env_name"], n_envs=config["num_envs"], seed=config["seed"]) #QbertNoFrameskip-v4
|
92 |
+
|
93 |
+
print("ENV ACTION SPACE: ", env.action_space.n)
|
94 |
+
|
95 |
+
# Frame-stacking with 4 frames
|
96 |
+
env = VecFrameStack(env, n_stack=4)
|
97 |
+
# Video recorder
|
98 |
+
env = VecVideoRecorder(env, "videos", record_video_trigger=lambda x: x % 100000 == 0, video_length=2000)
|
99 |
+
|
100 |
+
model = PPO(policy = "CnnPolicy",
|
101 |
+
env = env,
|
102 |
+
batch_size = 256,
|
103 |
+
clip_range = 0.1,
|
104 |
+
ent_coef = 0.01,
|
105 |
+
gae_lambda = 0.9,
|
106 |
+
gamma = 0.99,
|
107 |
+
learning_rate = 2.5e-4,
|
108 |
+
max_grad_norm = 0.5,
|
109 |
+
n_epochs = 4,
|
110 |
+
n_steps = 128,
|
111 |
+
vf_coef = 0.5,
|
112 |
+
tensorboard_log = f"runs",
|
113 |
+
verbose=1,
|
114 |
+
)
|
115 |
+
|
116 |
+
model.learn(
|
117 |
+
total_timesteps = config["total_timesteps"],
|
118 |
+
callback = [
|
119 |
+
WandbCallback(
|
120 |
+
gradient_save_freq = 1000,
|
121 |
+
model_save_path = f"models/{run.id}",
|
122 |
+
),
|
123 |
+
CheckpointCallback(save_freq=10000, save_path='./qbert',
|
124 |
+
name_prefix=config["env_name"]),
|
125 |
+
]
|
126 |
+
)
|
127 |
+
|
128 |
+
model.save("ppo-QbertNoFrameskip-v4.zip")
|
129 |
+
push_to_hub(repo_id="ThomasSimonini/ppo-QbertNoFrameskip-v4",
|
130 |
+
filename="ppo-QbertNoFrameskip-v4.zip",
|
131 |
+
commit_message="Added Qbert trained agent")
|
132 |
+
```
|