diff --git a/README.md b/README.md
index bedab60c91db49a191ad723ec4413142f095d35b..87b6c05ac09fdf77eae88f1c41f278ca9114d37f 100644
--- a/README.md
+++ b/README.md
@@ -22,7 +22,8 @@ Now that you have trained your model, it is time to evaluate its performance. Ru
> 🛠**To be handed in**
> Implement a script which loads your saved model and use it to solve the cartpole enviroment. Run 100 evaluations and share the final success rate across all evaluations in the `README.md`. Share the code in `evaluate_reinforce_cartpole.py`.
-From the openai gym wiki we know that the environment counts as solved when the average reward is greater or equal to 195 for over 100 consecutuve trials.
+Remark: To me it was not clear what "Success rate" in this context means (Therefore i made some assumptions!). The openai gym wiki says that the environment is solved when the average is over 100 trials is greater than 195. To translate this into a successrate i assumed that it is just necessary to count how often out of 100 trials, the summed reward was bigger than 195.
+
From the evaluation script i used the success rate is 1.0 when we allow the maximum number of steps the environment offers.
## Familiarization with a complete RL pipeline: Application to training a robotic arm
diff --git a/a2c_sb3_cartpole.py b/a2c_sb3_cartpole.py
index 7c94adcbbdf8cb5b3af5c7def98c89e4e7b4bd90..701652c6f414fb742a920f904052bcb51c383be3 100644
--- a/a2c_sb3_cartpole.py
+++ b/a2c_sb3_cartpole.py
@@ -20,12 +20,12 @@ run = wandb.init(
save_code=True,
)
+
+# from documentation
env = gym.make("CartPole-v1", render_mode="rgb_array")
model = A2C("MlpPolicy", env, verbose=1, tensorboard_log=f"runs/{run.id}")
-#model = A2C("MlpPolicy", env, )
model.learn(total_timesteps=10_000, callback=WandbCallback(gradient_save_freq=100,model_save_path=f"models/{run.id}",verbose=2,),)
-#model.learn(total_timesteps=10_000)
vec_env = model.get_env()
obs = vec_env.reset()
for i in range(1000):
@@ -34,6 +34,8 @@ for i in range(1000):
run.finish()
+
+# from documentation of huggingface
package_to_hub(model=model,
model_name="CartPole-v1",
model_architecture="A2C",
diff --git a/a2c_sb3_panda_reach.py b/a2c_sb3_panda_reach.py
index 5df3b95d2996f3081a887ecad2ac85198e978a1a..b30b75381803bc59f5f809f4f79fa00e51888869 100644
--- a/a2c_sb3_panda_reach.py
+++ b/a2c_sb3_panda_reach.py
@@ -21,12 +21,11 @@ run = wandb.init(
save_code=True,
)
+# from docs
env = gym.make("PandaReachJointsDense-v3")
model = A2C("MultiInputPolicy", env, verbose=1, tensorboard_log=f"runs/{run.id}")
-#model = A2C("MlpPolicy", env, )
model.learn(total_timesteps=500_000, callback=WandbCallback(gradient_save_freq=100,model_save_path=f"models/{run.id}",verbose=2,),)
-#model.learn(total_timesteps=10_000)
vec_env = model.get_env()
obs = vec_env.reset()
for i in range(1000):
@@ -35,6 +34,7 @@ for i in range(1000):
run.finish()
+# from documentation of huggingface
package_to_hub(model=model,
model_name="PandaReachJointsDense-v3",
model_architecture="A2C",
diff --git a/evaluate_reinforce_cartpole.py b/evaluate_reinforce_cartpole.py
index 1021baf702d003289c61a334d91e7d3d87d02e2a..9b18de60a225c8d025b0bbc564749f23e553f743 100644
--- a/evaluate_reinforce_cartpole.py
+++ b/evaluate_reinforce_cartpole.py
@@ -2,8 +2,9 @@ import gymnasium as gym
import torch
from reinforce_cartpole import Policy
+# function to evaluate the policy
def eval_policy(eval_length, policy, env):
- # Reset the environment and get the initial observation
+ # reset the environment and get the initial observation
observation = env.reset()[0]
rewards = []
@@ -13,12 +14,11 @@ def eval_policy(eval_length, policy, env):
action = torch.distributions.Categorical(action_probs).sample()
observation, reward, terminated, truncated, info = env.step(action.numpy())
rewards.append(reward)
- # visualize agent behavio
- #env.render()
if terminated or truncated:
break
return sum(rewards)
-# Create the environment
+
+# create env
env = gym.make("CartPole-v1")
policy = Policy()
@@ -26,15 +26,18 @@ policy = Policy()
policy.load_state_dict(torch.load('reinforce_cartpole.pth', weights_only=True))
policy.eval()
+# max stesps
eval_length = env.spec.max_episode_steps
num_evals = 100
number_of_solves = 0
+
for eval in range(num_evals):
sum_reward = eval_policy(eval_length, policy, env)
- print(f"Average reward: {sum_reward}")
+ # check if reward is greater than 195
if sum_reward >= 195:
number_of_solves += 1
-
+
+# compute successrate
success_rate = number_of_solves / num_evals
print(f"Success rate: {success_rate}")
diff --git a/reinforce_cartpole.py b/reinforce_cartpole.py
index dc01fd7bc5aad5c71bf1a8115c311dc94b5095e5..32da2bc0bb3f8f9fd2eb2a403de4bebcc2ff65dc 100644
--- a/reinforce_cartpole.py
+++ b/reinforce_cartpole.py
@@ -3,8 +3,10 @@ import torch
import numpy as np
import matplotlib.pyplot as plt
+# modify the dropout rate here
DROPOUT_RATE = 0.5
+# the policy class
class Policy(torch.nn.Module):
def __init__(self, input_size=4, output_size=2):
super(Policy, self).__init__()
@@ -19,34 +21,34 @@ class Policy(torch.nn.Module):
x = self.relu(x)
x = self.dropout(x)
x = self.fc2(x)
- #print(x)
x = self.softmax(x)
- #print(x)
return x
def main():
+
+ # load policy and optimizer
policy = Policy()
optimizer = torch.optim.Adam(policy.parameters(), lr=5e-3)
- # Create the environment
+ # create the environment
env = gym.make("CartPole-v1")
- # Reset the environment and get the initial observation
-
gamma = 0.99
+ # track total loss and reward
total_reward = []
total_loss = []
epochs = 500
+ # max possible steps in the envorinment
max_steps = env.spec.max_episode_steps
for ep in range(epochs):
print(ep)
observation = env.reset()[0]
- # rewards = torch.zeros(max_steps)
- # log_probs = torch.zeros(max_steps)
+
+ # allocate enough spaces for the rewards and log_probs
rewards = torch.zeros(max_steps)
log_probs = torch.zeros(max_steps)
for step in range(max_steps):
@@ -55,6 +57,8 @@ def main():
action = torch.distributions.Categorical(action_probs).sample()
observation, reward, terminated, truncated, info = env.step(action.numpy())
+
+ # set rewards and log_probs
rewards[step] = reward
log_probs[step] = torch.log(action_probs[action])
@@ -79,8 +83,6 @@ def main():
loss.backward()
optimizer.step()
total_loss.append(loss.detach().numpy())
- # Render the environment to visualize the agent's behavior
- #env.render()
# save the model weights
torch.save(policy.state_dict(), "reinforce_cartpole.pth")