Skip to content

Sac update

mighty.mighty_update.sac_update #

SACUpdate #

SACUpdate(
    model: SACModel,
    policy_lr: float = 0.001,
    q_lr: float = 0.001,
    tau: float = 0.005,
    alpha: float = 0.2,
    gamma: float = 0.99,
    target_entropy: float = None,
    auto_alpha: bool = True,
    alpha_lr: float = 0.0003,
    policy_frequency: int = 2,
    target_network_frequency: int = 1,
)
Source code in mighty/mighty_update/sac_update.py 
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
def __init__(
    self,
    model: SACModel,
    policy_lr: float = 0.001,
    q_lr: float = 0.001,
    tau: float = 0.005,
    alpha: float = 0.2,
    gamma: float = 0.99,
    target_entropy: float = None,
    auto_alpha: bool = True,
    alpha_lr: float = 3e-4,
    policy_frequency: int = 2,
    target_network_frequency: int = 1,
):
    self.model = model

    self.policy_optimizer = optim.Adam(model.policy_net.parameters(), lr=policy_lr)
    self.q_optimizer = optim.Adam(
        list(model.q_net1.parameters()) + list(model.q_net2.parameters()), lr=q_lr
    )

    self.alpha = alpha
    self.gamma = gamma
    self.tau = tau
    self.auto_alpha = auto_alpha
    self.action_dim = self.model.action_size
    self.policy_frequency = policy_frequency
    self.target_network_frequency = target_network_frequency
    self.update_step = 0

    if self.auto_alpha:
        self.log_alpha = torch.zeros(1, requires_grad=True)
        self.alpha_optimizer = optim.Adam([self.log_alpha], lr=alpha_lr or q_lr)
        self.target_entropy = (
            -float(self.action_dim)
            if target_entropy is None
            else float(target_entropy)
        )
    else:
        self.alpha = alpha

calculate_td_error #

calculate_td_error(transition: TransitionBatch) -> Tuple

Calculate the TD error for a given transition.

Source code in mighty/mighty_update/sac_update.py 
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
def calculate_td_error(self, transition: TransitionBatch) -> Tuple:
    """Calculate the TD error for a given transition."""
    with torch.no_grad():
        a_next, z_next, mean_next, log_std_next = self.model(
            torch.as_tensor(transition.next_obs, dtype=torch.float32)
        )
        logp_next = self.model.policy_log_prob(z_next, mean_next, log_std_next)
        sa_next = torch.cat(
            [
                torch.as_tensor(transition.next_obs, dtype=torch.float32),
                a_next,
            ],
            dim=-1,
        )
        q1_t = self.model.target_q_net1(sa_next)
        q2_t = self.model.target_q_net2(sa_next)
        alpha = self.log_alpha.exp() if self.auto_alpha else self.alpha
        q_target = torch.as_tensor(
            transition.rewards, dtype=torch.float32
        ).unsqueeze(-1) + (
            1 - torch.as_tensor(transition.dones, dtype=torch.float32).unsqueeze(-1)
        ) * self.gamma * (torch.min(q1_t, q2_t) - alpha * logp_next)
    sa = torch.cat(
        [
            torch.as_tensor(transition.observations, dtype=torch.float32),
            torch.as_tensor(transition.actions, dtype=torch.float32),
        ],
        dim=-1,
    )
    q1_curr = self.model.q_net1(sa)
    q2_curr = self.model.q_net2(sa)
    td_error1 = q1_curr - q_target
    td_error2 = q2_curr - q_target
    return td_error1, td_error2

update #

update(batch: TransitionBatch) -> Dict

Perform an update of the SAC model using a batch of experience.

Source code in mighty/mighty_update/sac_update.py 
 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
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
def update(self, batch: TransitionBatch) -> Dict:
    """Perform an update of the SAC model using a batch of experience."""
    self.update_step += 1

    states = torch.as_tensor(batch.observations, dtype=torch.float32)
    actions = torch.as_tensor(batch.actions, dtype=torch.float32)
    rewards = torch.as_tensor(batch.rewards, dtype=torch.float32).unsqueeze(-1)
    dones = torch.as_tensor(batch.dones, dtype=torch.float32).unsqueeze(-1)
    next_states = torch.as_tensor(batch.next_obs, dtype=torch.float32)

    # --- Q-network update ---
    with torch.no_grad():
        a_next, z_next, mean_next, log_std_next = self.model(next_states)
        logp_next = self.model.policy_log_prob(z_next, mean_next, log_std_next)
        sa_next = torch.cat([next_states, a_next], dim=-1)
        q1_t = self.model.target_q_net1(sa_next)
        q2_t = self.model.target_q_net2(sa_next)
        current_alpha = (
            self.log_alpha.exp().detach() if self.auto_alpha else self.alpha
        )
        q_target = rewards + (1 - dones) * self.gamma * (
            torch.min(q1_t, q2_t) - current_alpha * logp_next
        )

    sa = torch.cat([states, actions], dim=-1)
    q1 = self.model.q_net1(sa)
    q2 = self.model.q_net2(sa)
    q_loss1 = F.mse_loss(q1, q_target)
    q_loss2 = F.mse_loss(q2, q_target)
    q_loss = q_loss1 + q_loss2

    # use combined optimizer for both Q-networks
    self.q_optimizer.zero_grad()
    q_loss.backward()
    self.q_optimizer.step()

    # --- Policy update (delayed) ---
    policy_loss = torch.tensor(0.0)
    alpha_loss = torch.tensor(0.0)
    if self.update_step % self.policy_frequency == 0:
        # do multiple policy updates to compensate for delay
        for _ in range(self.policy_frequency):
            # recompute alpha after q update
            current_alpha = (
                self.log_alpha.exp().detach() if self.auto_alpha else self.alpha
            )

            # Sample fresh actions for each policy update iteration
            # This ensures stochasticity across iterations
            a, z, mean, log_std = self.model(states)
            logp = self.model.policy_log_prob(z, mean, log_std)
            sa_pi = torch.cat([states, a], dim=-1)

            q1_pi = self.model.q_net1(sa_pi)
            q2_pi = self.model.q_net2(sa_pi)
            q_pi = torch.min(q1_pi, q2_pi)
            policy_loss = (current_alpha * logp - q_pi).mean()

            self.policy_optimizer.zero_grad()
            policy_loss.backward()
            self.policy_optimizer.step()

            # --- Entropy coefficient (alpha) update ---
            if self.auto_alpha:
                # Get fresh sample for alpha update
                with torch.no_grad():
                    _, z_alpha, mean_alpha, log_std_alpha = self.model(states)
                    logp_alpha = self.model.policy_log_prob(z_alpha, mean_alpha, log_std_alpha)

                alpha_loss = -(
                    self.log_alpha.exp() * (logp_alpha.detach() + self.target_entropy)
                ).mean()
                self.alpha_optimizer.zero_grad()
                alpha_loss.backward()
                self.alpha_optimizer.step()
                self.alpha = self.log_alpha.exp().item()

    # --- Soft update targets ---
    if self.update_step % self.target_network_frequency == 0:
        polyak_update(
            self.model.q_net1.parameters(),
            self.model.target_q_net1.parameters(),
            self.tau,
        )
        polyak_update(
            self.model.q_net2.parameters(),
            self.model.target_q_net2.parameters(),
            self.tau,
        )

    # --- Logging metrics ---
    td1, td2 = self.calculate_td_error(batch)
    return {
        "Update/q_loss1": q_loss1.item(),
        "Update/q_loss2": q_loss2.item(),
        "Update/policy_loss": policy_loss.item(),
        "Update/alpha_loss": alpha_loss.item(),
        "Update/td_error1": td1.mean().item(),
        "Update/td_error2": td2.mean().item(),
    }