Skip to content

Ppo update

mighty.mighty_update.ppo_update #

PPOUpdate #

PPOUpdate(
    model: PPOModel,
    policy_lr: float = 0.001,
    value_lr: float = 0.001,
    epsilon: float = 0.2,
    ent_coef: float = 0.0,
    vf_coef: float = 0.5,
    max_grad_norm: float = 0.5,
    n_epochs: int = 10,
    minibatch_size: int = 32,
    kl_target: float = 0.001,
    use_value_clip: bool = True,
    value_clip_eps: float = 0.2,
    total_timesteps: int = 1000000,
)
Source code in mighty/mighty_update/ppo_update.py 
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
def __init__(
    self,
    model: PPOModel,
    policy_lr: float = 0.001,
    value_lr: float = 0.001,
    epsilon: float = 0.2,
    ent_coef: float = 0.0,
    vf_coef: float = 0.5,
    max_grad_norm: float = 0.5,
    n_epochs: int = 10,
    minibatch_size: int = 32,
    kl_target: float = 0.001,
    use_value_clip: bool = True,
    value_clip_eps: float = 0.2,
    total_timesteps: int = 1_000_000,
):
    """Initialize PPO update mechanism."""
    self.model = model
    self.epsilon = epsilon
    self.ent_coef = ent_coef
    self.vf_coef = vf_coef
    self.max_grad_norm = max_grad_norm

    self.n_epochs = n_epochs
    self.minibatch_size = minibatch_size
    self.kl_target = kl_target
    self.use_value_clip = use_value_clip
    self.value_clip_eps = value_clip_eps

    self.total_steps = total_timesteps

    # Optimizers

    policy_params = list(
        model.policy_head.parameters()
    )  # includes feature_extractor_policy
    value_params = list(
        model.value_head.parameters()
    )  # includes feature_extractor_value

    extra_params = []
    if getattr(model, "continuous_action", False) and hasattr(model, "log_std"):
        extra_params.append(model.log_std)

    self.optimizer = optim.Adam(
        [
            {
                "params": policy_params,
                "lr": policy_lr,
            },  # policy net (feat-extractor + head)
            {
                "params": value_params,
                "lr": value_lr,
            },  # value net  (feat-extractor + head)
            *([{"params": extra_params, "lr": policy_lr}] if extra_params else []),
        ],
        eps=1e-5,
    )

    self.scheduler = optim.lr_scheduler.LambdaLR(
        self.optimizer,
        lr_lambda=lambda step: 1 - step / float(self.total_steps),
    )

update #

update(batch: MaxiBatch) -> Dict[str, float]

PPO update with advantage normalisation, epoch-level KL-stop, value-clipping.

Source code in mighty/mighty_update/ppo_update.py 
 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
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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
def update(self, batch: MaxiBatch) -> Dict[str, float]:
    """PPO update with advantage normalisation, epoch-level KL-stop, value-clipping."""

    # cache old values / log-probs with *one* forward pass per minibatch
    with torch.no_grad():
        old_values = [
            self.model.forward_value(mb.observations) for mb in batch.minibatches
        ]
        old_log_probs = [mb.log_probs.clone() for mb in batch.minibatches]

    # global advantage normalisation (once!) and detach from graph
    flat_adv = batch.advantages.squeeze(0)
    adv_mean = flat_adv.mean().detach()
    adv_std = (flat_adv.std() + 1e-8).detach()

    metrics = {
        "policy_loss": 0.0,
        "value_loss": 0.0,
        "entropy": 0.0,
        "approx_kl": 0.0,
    }
    mb_updates = 0  # mini-batch counter
    epoch_counts = 0  # how many epochs actually executed

    # main PPO loop
    for epoch in range(self.n_epochs):
        epoch_kls = []

        for i, mb in enumerate(batch.minibatches):
            # 2a) use the globally normalised advantages
            adv = ((mb.advantages - adv_mean) / adv_std).detach()

            # 2b) value loss (with optional clipping)
            values = self.model.forward_value(mb.observations)
            if self.use_value_clip:
                clipped = old_values[i] + torch.clamp(
                    values - old_values[i],
                    -self.value_clip_eps,
                    self.value_clip_eps,
                )
                v_loss = (
                    0.5
                    * torch.max(
                        (values - mb.returns).pow(2),
                        (clipped - mb.returns).pow(2),
                    ).mean()
                )
            else:
                v_loss = 0.5 * (mb.returns - values).pow(2).mean()

            # 2c) new policy log-probs & entropy
            if self.model.continuous_action:
                means, raw_std = self.model(mb.observations)
                # derive a safe std
                log_std = raw_std.clamp(-20, 2)
                std = torch.exp(log_std).clamp(min=1e-3)
                dist = torch.distributions.Normal(means, std)
                log_probs = dist.log_prob(mb.actions).sum(-1)
                entropy = dist.entropy().sum(-1).mean()
            else:
                logits = self.model(mb.observations)
                dist = torch.distributions.Categorical(logits=logits)
                log_probs = dist.log_prob(mb.actions)
                entropy = dist.entropy().mean()

            # 2d) PPO surrogate loss
            ratios = torch.exp((log_probs - old_log_probs[i]).clamp(-20, 20))
            surr1 = ratios * adv
            surr2 = torch.clamp(ratios, 1 - self.epsilon, 1 + self.epsilon) * adv
            p_loss = -torch.min(surr1, surr2).mean()

            # 2e) total loss, backprop, step
            loss = p_loss + self.vf_coef * v_loss - self.ent_coef * entropy
            self.optimizer.zero_grad()
            loss.backward()
            torch.nn.utils.clip_grad_norm_(
                self.model.parameters(), self.max_grad_norm
            )
            self.optimizer.step()

            # 2f) recompute log-probs *after* the step for a correct KL

            with torch.no_grad():
                if self.model.continuous_action:
                    means, stds = self.model(mb.observations)
                    new_lp = (
                        torch.distributions.Normal(means, stds)
                        .log_prob(mb.actions)
                        .sum(-1)
                    )
                else:
                    logits = self.model(mb.observations)
                    new_lp = torch.distributions.Categorical(
                        logits=logits
                    ).log_prob(mb.actions)
            kl = (old_log_probs[i] - new_lp).mean()
            epoch_kls.append(kl)

            # 2g) log accumulators
            metrics["policy_loss"] += p_loss.item()
            metrics["value_loss"] += v_loss.item()
            metrics["entropy"] += entropy.item()
            mb_updates += 1

        # epoch finished
        mean_kl = torch.mean(torch.stack(epoch_kls))
        metrics["approx_kl"] += mean_kl.item()
        epoch_counts += 1

        if mean_kl > self.kl_target:
            break  # early-stop epochs

    # LR decay (once per update call)

    self.scheduler.step()

    # 4) final metric averages

    if mb_updates > 0:
        metrics["policy_loss"] /= mb_updates
        metrics["value_loss"] /= mb_updates
        metrics["entropy"] /= mb_updates
    if epoch_counts > 0:
        metrics["approx_kl"] /= epoch_counts

    return {
        "Update/policy_loss": metrics["policy_loss"],
        "Update/value_loss": metrics["value_loss"],
        "Update/entropy": metrics["entropy"],
        "Update/approx_kl": metrics["approx_kl"],
    }