DiscretePPO/agent.py at main · naivoder/DiscretePPO · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
import torch
from memory import ReplayBuffer
import numpy as np

class ActorCritic(torch.nn.Module):
    def __init__(self, input_dims, n_actions, alpha=1e-4, chkpt_dir="weights/network.pt"):
        super(ActorCritic, self).__init__()
        self.input_dims = input_dims
        self.n_actions = n_actions
        self.chkpt_dir = chkpt_dir

        self.conv1 = self._init_weights(torch.nn.Conv2d(input_dims[0], 32, kernel_size=8, stride=4))
        self.conv2 = self._init_weights(torch.nn.Conv2d(32, 64, kernel_size=4, stride=2))
        self.conv3 = self._init_weights(torch.nn.Conv2d(64, 64, kernel_size=3, stride=1))

        self.fc1_input_dim = self._calculate_fc1_input_dim(input_dims)

        self.fc1 = self._init_weights(torch.nn.Linear(self.fc1_input_dim, 512))
        self.critic = self._init_weights(torch.nn.Linear(512, 1), std=1.0)
        self.actor = self._init_weights(torch.nn.Linear(512, n_actions), std=0.01)

        self.optimizer = torch.optim.AdamW(self.parameters(), alpha)
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.to(self.device)

    def _init_weights(self, layer, std=np.sqrt(2), scale=False):
        # taken from cleanrl implementation
        torch.nn.init.orthogonal_(layer.weight, std)
        torch.nn.init.constant_(layer.bias, 0)
        return layer

    def _calculate_fc1_input_dim(self, input_shape):
        dummy_input = torch.zeros(1, *input_shape)
        x = self.conv1(dummy_input)
        x = self.conv2(x)
        x = self.conv3(x)
        return x.numel()  # count flattened elements

    def forward(self, x):
        x = torch.nn.functional.relu(self.conv1(x))
        x = torch.nn.functional.relu(self.conv2(x))
        x = torch.nn.functional.relu(self.conv3(x))
        x = x.view(x.size()[0], -1)
        x = torch.nn.functional.relu(self.fc1(x))
        value = self.critic(x)
        x = torch.nn.functional.softmax(self.actor(x), dim=1)
        action = torch.distributions.Categorical(logits=x)
        return action, value

    def save_checkpoint(self):
        torch.save(self.state_dict(), self.chkpt_dir)

    def load_checkpoint(self):
        self.load_state_dict(torch.load(self.chkpt_dir))

class DiscretePPOAgent:
    def __init__(
        self,
        env_name,
        input_dims,
        n_actions,
        gamma=0.99,
        alpha=3e-4,
        gae_lambda=0.95,
        policy_clip=0.1,
        batch_size=64,
        n_epochs=5,
        max_grad_norm=0.5,
        entropy_coefficient=0.01,
        clip_value=True,
    ):
        self.env_name = env_name.split("/")[-1]
        self.gamma = gamma
        self.policy_clip = policy_clip
        self.n_epochs = n_epochs
        self.gae_lambda = gae_lambda
        self.entropy_coefficient = entropy_coefficient
        self.max_grad_norm = max_grad_norm
        self.clip_value = clip_value

        self.network = ActorCritic(input_dims, n_actions, alpha, f"weights/{env_name}.pt")
        self.memory = ReplayBuffer(batch_size)

    def remember(self, state, value, action, probs, reward, done):
        self.memory.store_transition(state, value, action, probs, reward, done)

    def save_checkpoints(self):
        self.network.save_checkpoint()

    def load_checkpoints(self):
        self.network.load_checkpoint()

    def choose_action(self, state):
        state = torch.FloatTensor(state).to(self.network.device).unsqueeze(0)
        dist, value = self.network(state)
        action = dist.sample()
        prob = dist.log_prob(action)
        return action.item(), prob.item(), value.item()

    def learn(self):
        state_arr, value_arr, action_arr, prob_arr, reward_arr, dones_arr = (
            self.memory.sample()
        )

        # https://github.com/philtabor/.../ppo_torch.py
        advantage = np.zeros(len(reward_arr), dtype=np.float32)
        for t in range(len(reward_arr)-1):
            discount = 1
            a_t = 0
            for k in range(t, len(reward_arr)-1):
                a_t += discount*(reward_arr[k] + self.gamma*value_arr[k+1]*\
                        (1-int(dones_arr[k])) - value_arr[k])
                discount *= self.gamma*self.gae_lambda
            advantage[t] = a_t

        advantage_arr = torch.FloatTensor(advantage).to(self.network.device)
        state_arr = torch.FloatTensor(state_arr).to(self.network.device)
        action_arr = torch.FloatTensor(action_arr).to(self.network.device)
        prob_arr = torch.FloatTensor(prob_arr).to(self.network.device)
        value_arr = torch.FloatTensor(value_arr).to(self.network.device)
        reward_arr = torch.FloatTensor(reward_arr).to(self.network.device)
        dones_arr = torch.BoolTensor(dones_arr).to(self.network.device)

        for _ in range(self.n_epochs):
            batches = self.memory.generate_batches()

            for batch in batches:
                states = state_arr[batch]
                actions = action_arr[batch]
                old_values = value_arr[batch]
                old_probs = prob_arr[batch]
                advantages = advantage_arr[batch]

                dist, new_values = self.network(states)

                new_probs = dist.log_prob(actions)
                prob_ratio = new_probs.exp() / old_probs.exp()

                weighted_probs = advantages * prob_ratio
                weighted_clipped_probs = advantages * torch.clamp(prob_ratio, 1-self.policy_clip, 1+self.policy_clip)

                actor_loss = torch.min(weighted_probs, weighted_clipped_probs).mean()
                actor_loss -= self.entropy_coefficient * dist.entropy().mean()

                returns = advantages + old_values

                if self.clip_value:
                    unclipped_critic_loss = (new_values - returns).pow(2)
                    clipped_values = old_values + torch.clamp(new_values - old_values, -self.policy_clip, self.policy_clip)
                    clipped_critic_loss = (clipped_values - returns).pow(2)
                    critic_loss = torch.max(unclipped_critic_loss, clipped_critic_loss).mean()
                else:
                    critic_loss = (returns-new_values).pow(2).mean()

                loss = actor_loss + 0.5 * critic_loss

                self.network.optimizer.zero_grad()
                loss.backward()
                torch.nn.utils.clip_grad_norm_(self.network.parameters(), self.max_grad_norm)
                self.network.optimizer.step()

        self.memory.clear_memory()