Add rlpd tricks

- Introduced target critic networks in SACPolicy to enhance stability during training.
- Updated TD target calculation to incorporate entropy adjustments, improving robustness.
- Increased online buffer capacity in configuration from 10,000 to 40,000 for better data handling.
- Adjusted learning rates for critic, actor, and temperature to 3e-4 for optimized training performance.

These changes aim to refine the SAC implementation, enhancing its robustness and performance during training and inference.

lerobot/common/datasets/lerobot_dataset.py

@@ -611,11 +611,24 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return query_timestamps
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
-        return {
+        # Step 1: Combine all unique indices
-            key: torch.stack(self.hf_dataset.select(q_idx)[key])
+        all_indices = sorted({idx for indices in query_indices.values() for idx in indices})
-            for key, q_idx in query_indices.items()
+
-            if key not in self.meta.video_keys
+        # Step 2: Select all required data at once
-        }
+        selected_dataset = self.hf_dataset.select(all_indices).to_dict()
+        selected_dataset = {key: torch.tensor(values) for key, values in selected_dataset.items()}
+
+        # Step 3: Map original indices to their positions in the selected dataset
+        index_map = {original_idx: i for i, original_idx in enumerate(all_indices)}
+
+        # Step 4: Build the result for each key
+        results = {}
+        for key, q_indices in query_indices.items():
+            if key not in self.meta.video_keys:
+                mapped_indices = [index_map[idx] for idx in q_indices]
+                results[key] = torch.stack([selected_dataset[key][i] for i in mapped_indices])
+
+        return results
 
     def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict:
         """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function

lerobot/common/policies/sac/configuration_sac.py

@@ -28,12 +28,18 @@ class SACConfig:
     )
     output_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
-            "action": [4],
+            "action": [2],
         }
     )
 
     # Normalization / Unnormalization
-    input_normalization_modes: dict[str, str] | None = None
+    input_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {
+            "observation.image": "mean_std",
+            "observation.state": "min_max",
+            "observation.environment_state": "min_max",
+        }
+    )
     output_normalization_modes: dict[str, str] = field(
         default_factory=lambda: {"action": "min_max"},
     )
@@ -48,9 +54,10 @@ class SACConfig:
     critic_target_update_weight = 0.005
     utd_ratio = 2
     state_encoder_hidden_dim = 256
-    latent_dim = 128
+    latent_dim = 256
     target_entropy = None
-    backup_entropy = True
+    # backup_entropy = False
+    use_backup_entropy = True
     critic_network_kwargs = {
         "hidden_dims": [256, 256],
         "activate_final": True,

lerobot/common/policies/sac/modeling_sac.py

@@ -18,8 +18,7 @@
 # TODO: (1) better device management
 
 from collections import deque
-from copy import deepcopy
+from typing import Callable, Optional, Sequence, Tuple, Union
-from typing import Callable, Optional, Sequence, Tuple
 
 import einops
 import numpy as np
@@ -57,12 +56,20 @@ class SACPolicy(
             )
         else:
             self.normalize_inputs = nn.Identity()
+        # HACK: we need to pass the dataset_stats to the normalization functions
+        dataset_stats = dataset_stats or {
+            "action": {
+                "min": torch.tensor([-1.0, -1.0, -1.0, -1.0]),
+                "max": torch.tensor([1.0, 1.0, 1.0, 1.0]),
+            }
+        }
         self.normalize_targets = Normalize(
             config.output_shapes, config.output_normalization_modes, dataset_stats
         )
         self.unnormalize_outputs = Unnormalize(
             config.output_shapes, config.output_normalization_modes, dataset_stats
         )
+
         encoder_critic = SACObservationEncoder(config)
         encoder_actor = SACObservationEncoder(config)
         # Define networks
@@ -72,26 +79,38 @@ class SACPolicy(
                 encoder=encoder_critic,
                 network=MLP(
                     input_dim=encoder_critic.output_dim + config.output_shapes["action"][0],
-                    **config.critic_network_kwargs
+                    **config.critic_network_kwargs,
-                )
+                ),
             )
             critic_nets.append(critic_net)
 
+        target_critic_nets = []
+        for _ in range(config.num_critics):
+            target_critic_net = Critic(
+                encoder=encoder_critic,
+                network=MLP(
+                    input_dim=encoder_critic.output_dim + config.output_shapes["action"][0],
+                    **config.critic_network_kwargs,
+                ),
+            )
+            target_critic_nets.append(target_critic_net)
+
         self.critic_ensemble = create_critic_ensemble(critic_nets, config.num_critics)
-        self.critic_target = deepcopy(self.critic_ensemble)
+        self.critic_target = create_critic_ensemble(target_critic_nets, config.num_critics)
+        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
 
         self.actor = Policy(
             encoder=encoder_actor,
-            network=MLP(
+            network=MLP(input_dim=encoder_actor.output_dim, **config.actor_network_kwargs),
-                input_dim=encoder_actor.output_dim,
-                **config.actor_network_kwargs
-            ),
             action_dim=config.output_shapes["action"][0],
-            **config.policy_kwargs
+            **config.policy_kwargs,
         )
         if config.target_entropy is None:
-            config.target_entropy = -np.prod(config.output_shapes["action"][0]) #  (-dim(A))
+            config.target_entropy = -np.prod(config.output_shapes["action"][0]) / 2  # (-dim(A)/2)
-        self.temperature = LagrangeMultiplier(init_value=config.temperature_init)    
+        # TODO: fix later device
+        # TODO: Handle the case where the temparameter is a fixed
+        self.log_alpha = torch.zeros(1, requires_grad=True, device="cpu")
+        self.temperature = self.log_alpha.exp().item()
 
     def reset(self):
         """
@@ -111,11 +130,13 @@ class SACPolicy(
     @torch.no_grad()
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select action for inference/evaluation"""
-        actions, _ = self.actor(batch)
+        actions, _, _ = self.actor(batch)
         actions = self.unnormalize_outputs({"action": actions})["action"]
         return actions
 
-    def critic_forward(self, observations: dict[str, Tensor], actions: Tensor, use_target: bool = False) -> Tensor:
+    def critic_forward(
+        self, observations: dict[str, Tensor], actions: Tensor, use_target: bool = False
+    ) -> Tensor:
         """Forward pass through a critic network ensemble
 
         Args:
@@ -130,16 +151,20 @@ class SACPolicy(
         q_values = torch.stack([critic(observations, actions) for critic in critics])
         return q_values
 
-
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor | float]:
         """Run the batch through the model and compute the loss.
 
         Returns a dictionary with loss as a tensor, and other information as native floats.
         """
+        # We have to actualize the value of the temperature because in the previous
+        self.temperature = self.log_alpha.exp().item()
+        temperature = self.temperature
+
         batch = self.normalize_inputs(batch)
         # batch shape is (b, 2, ...) where index 1 returns the current observation and
         # the next observation for calculating the right td index.
-        actions = batch["action"][:, 0]
+        # actions = batch["action"][:, 0]
+        actions = batch["action"]
         rewards = batch["next.reward"][:, 0]
         observations = {}
         next_observations = {}
@@ -147,18 +172,13 @@ class SACPolicy(
             if k.startswith("observation."):
                 observations[k] = batch[k][:, 0]
                 next_observations[k] = batch[k][:, 1]
+        done = batch["next.done"]
 
-        # perform image augmentation
+        with torch.no_grad():
-
+            next_action_preds, next_log_probs, _ = self.actor(next_observations)
-        # reward bias from HIL-SERL code base 
-        # add_or_replace={"rewards": batch["rewards"] + self.config["reward_bias"]} in reward_batch
-        
-        # calculate critics loss
-        # 1- compute actions from policy
-        action_preds, log_probs = self.actor(next_observations)
 
             # 2- compute q targets
-        q_targets = self.critic_forward(next_observations, action_preds, use_target=True)
+            q_targets = self.critic_forward(next_observations, next_action_preds, use_target=True)
 
             # subsample critics to prevent overfitting if use high UTD (update to date)
             if self.config.num_subsample_critics is not None:
@@ -168,84 +188,125 @@ class SACPolicy(
 
             # critics subsample size
             min_q, _ = q_targets.min(dim=0)  # Get values from min operation
-
+            if self.config.use_backup_entropy:
-        # compute td target
+                min_q -= self.temperature * next_log_probs
-        td_target = rewards + self.config.discount * min_q #+ self.config.discount * self.temperature() * log_probs # add entropy term
+            td_target = rewards + self.config.discount * min_q * ~done
 
         # 3- compute predicted qs
         q_preds = self.critic_forward(observations, actions, use_target=False)
 
         # 4- Calculate loss
         # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
-        critics_loss = F.mse_loss(
+        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
-            q_preds,  # shape: [num_critics, batch_size]
+        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
-            einops.repeat(td_target, "b -> e b", e=q_preds.shape[0]), # expand td_target to match q_preds shape
+        critics_loss = (
-            reduction="none"
+            F.mse_loss(
-        ).sum(0).mean()
+                input=q_preds,
+                target=td_target_duplicate,
+                reduction="none",
+            ).mean(1)
+        ).sum()
 
-        # critics_loss = (   
+        actions_pi, log_probs, _ = self.actor(observations)
-        #     F.mse_loss(
-        #             q_preds,
-        #             einops.repeat(td_target, "b -> e b", e=q_preds.shape[0]),
-        #             reduction="none",
-        #         ).sum(0)  # sum over ensemble
-        #         # `q_preds_ensemble` depends on the first observation and the actions.
-        #         * ~batch["observation.state_is_pad"][0]
-        #         * ~batch["action_is_pad"]
-        #         # q_targets depends on the reward and the next observations.
-        #         * ~batch["next.reward_is_pad"]
-        #         * ~batch["observation.state_is_pad"][1:]
-        #     ).sum(0).mean()
-        
-        # calculate actors loss
-        # 1- temperature
-        temperature = self.temperature()
-        # 2- get actions (batch_size, action_dim) and log probs (batch_size,)
-        actions, log_probs = self.actor(observations)
-        # 3- get q-value predictions
         with torch.inference_mode():
-            q_preds = self.critic_forward(observations, actions, use_target=False)
+            q_preds = self.critic_forward(observations, actions_pi, use_target=False)
-        actor_loss = (
+        min_q_preds = q_preds.min(dim=0)[0]
-            -(q_preds - temperature * log_probs).mean()
-            # * ~batch["observation.state_is_pad"][0]
-            # * ~batch["action_is_pad"]
-        ).mean()
 
+        actor_loss = ((temperature * log_probs) - min_q_preds).mean()
 
         # calculate temperature loss
-        # 1- calculate entropy
+        with torch.no_grad():
-        entropy = -log_probs.mean()
+            _, log_probs, _ = self.actor(observations)
-        temperature_loss = self.temperature(
+        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.config.target_entropy)).mean()
-            lhs=entropy,
-            rhs=self.config.target_entropy
-        )
 
         loss = critics_loss + actor_loss + temperature_loss
 
         return {
             "critics_loss": critics_loss.item(),
             "actor_loss": actor_loss.item(),
+            "mean_q_predicts": min_q_preds.mean().item(),
+            "min_q_predicts": min_q_preds.min().item(),
+            "max_q_predicts": min_q_preds.max().item(),
             "temperature_loss": temperature_loss.item(),
-                "temperature": temperature.item(),
+            "temperature": temperature,
-                "entropy": entropy.item(),
+            "mean_log_probs": log_probs.mean().item(),
+            "min_log_probs": log_probs.min().item(),
+            "max_log_probs": log_probs.max().item(),
+            "td_target_mean": td_target.mean().item(),
+            "td_target_max": td_target.max().item(),
+            "action_mean": actions.mean().item(),
+            "entropy": log_probs.mean().item(),
             "loss": loss,
         }
 
-    def update(self):
+    def update_target_networks(self):
-        # TODO: implement UTD update
-        # First update only critics for utd_ratio-1 times
-        #for critic_step in range(self.config.utd_ratio - 1):
-            # only update critic and critic target
-        # Then update critic, critic target, actor and temperature
         """Update target networks with exponential moving average"""
-        with torch.no_grad():
         for target_critic, critic in zip(self.critic_target, self.critic_ensemble, strict=False):
             for target_param, param in zip(target_critic.parameters(), critic.parameters(), strict=False):
                 target_param.data.copy_(
-                        target_param.data * self.config.critic_target_update_weight + 
+                    param.data * self.config.critic_target_update_weight
-                        param.data * (1.0 - self.config.critic_target_update_weight)
+                    + target_param.data * (1.0 - self.config.critic_target_update_weight)
                 )
 
+    def compute_loss_critic(self, observations, actions, rewards, next_observations, done) -> Tensor:
+        temperature = self.log_alpha.exp().item()
+        with torch.no_grad():
+            next_action_preds, next_log_probs, _ = self.actor(next_observations)
+
+            # 2- compute q targets
+            q_targets = self.critic_forward(
+                observations=next_observations, actions=next_action_preds, use_target=True
+            )
+
+            # subsample critics to prevent overfitting if use high UTD (update to date)
+            if self.config.num_subsample_critics is not None:
+                indices = torch.randperm(self.config.num_critics)
+                indices = indices[: self.config.num_subsample_critics]
+                q_targets = q_targets[indices]
+
+            # critics subsample size
+            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
+            if self.config.use_backup_entropy:
+                min_q = min_q - (temperature * next_log_probs)
+
+            td_target = rewards + (1 - done) * self.config.discount * min_q
+
+        # 3- compute predicted qs
+        q_preds = self.critic_forward(observations, actions, use_target=False)
+
+        # 4- Calculate loss
+        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
+        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
+        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
+        critics_loss = (
+            F.mse_loss(
+                input=q_preds,
+                target=td_target_duplicate,
+                reduction="none",
+            ).mean(1)
+        ).sum()
+        return critics_loss
+
+    def compute_loss_temperature(self, observations) -> Tensor:
+        """Compute the temperature loss"""
+        # calculate temperature loss
+        with torch.no_grad():
+            _, log_probs, _ = self.actor(observations)
+        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.config.target_entropy)).mean()
+        return temperature_loss
+
+    def compute_loss_actor(self, observations) -> Tensor:
+        temperature = self.log_alpha.exp().item()
+
+        actions_pi, log_probs, _ = self.actor(observations)
+
+        q_preds = self.critic_forward(observations, actions_pi, use_target=False)
+        min_q_preds = q_preds.min(dim=0)[0]
+
+        actor_loss = ((temperature * log_probs) - min_q_preds).mean()
+        return actor_loss
+
+
 class MLP(nn.Module):
     def __init__(
         self,
@@ -276,7 +337,9 @@ class MLP(nn.Module):
                 if dropout_rate is not None and dropout_rate > 0:
                     layers.append(nn.Dropout(p=dropout_rate))
                 layers.append(nn.LayerNorm(hidden_dims[i]))
-                layers.append(activations if isinstance(activations, nn.Module) else getattr(nn, activations)())
+                layers.append(
+                    activations if isinstance(activations, nn.Module) else getattr(nn, activations)()
+                )
 
         self.net = nn.Sequential(*layers)
 
@@ -290,7 +353,7 @@ class Critic(nn.Module):
         encoder: Optional[nn.Module],
         network: nn.Module,
         init_final: Optional[float] = None,
-        device: str = "cuda"
+        device: str = "cpu",
     ):
         super().__init__()
         self.device = torch.device(device)
@@ -321,9 +384,7 @@ class Critic(nn.Module):
         actions: torch.Tensor,
     ) -> torch.Tensor:
         # Move each tensor in observations to device
-        observations = {
+        observations = {k: v.to(self.device) for k, v in observations.items()}
-            k: v.to(self.device) for k, v in observations.items()
-        }
         actions = actions.to(self.device)
 
         obs_enc = observations if self.encoder is None else self.encoder(observations)
@@ -345,7 +406,7 @@ class Policy(nn.Module):
         fixed_std: Optional[torch.Tensor] = None,
         init_final: Optional[float] = None,
         use_tanh_squash: bool = False,
-        device: str = "cuda"
+        device: str = "cpu",
     ):
         super().__init__()
         self.device = torch.device(device)
@@ -386,7 +447,6 @@ class Policy(nn.Module):
         self,
         observations: torch.Tensor,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-                
         # Encode observations if encoder exists
         obs_enc = observations if self.encoder is None else self.encoder(observations)
 
@@ -397,22 +457,30 @@ class Policy(nn.Module):
         # Compute standard deviations
         if self.fixed_std is None:
             log_std = self.std_layer(outputs)
+            assert not torch.isnan(log_std).any(), "[ERROR] log_std became NaN after std_layer!"
+
             if self.use_tanh_squash:
                 log_std = torch.tanh(log_std)
+                log_std = self.log_std_min + 0.5 * (self.log_std_max - self.log_std_min) * (log_std + 1.0)
+            else:
                 log_std = torch.clamp(log_std, self.log_std_min, self.log_std_max)
         else:
             log_std = self.fixed_std.expand_as(means)
 
-        # uses tahn activation function to squash the action to be in the range of [-1, 1]
+        # uses tanh activation function to squash the action to be in the range of [-1, 1]
         normal = torch.distributions.Normal(means, torch.exp(log_std))
-        x_t = normal.rsample()  # for reparameterization trick (mean + std * N(0,1)) 
+        x_t = normal.rsample()  # Reparameterization trick (mean + std * N(0,1))
-        log_probs = normal.log_prob(x_t)
+        log_probs = normal.log_prob(x_t)  # Base log probability before Tanh
+
         if self.use_tanh_squash:
             actions = torch.tanh(x_t)
-            log_probs -= torch.log((1 - actions.pow(2)) + 1e-6)
+            log_probs -= torch.log((1 - actions.pow(2)) + 1e-6)  # Adjust log-probs for Tanh
-        log_probs = log_probs.sum(-1) # sum over action dim
+        else:
+            actions = x_t  # No Tanh; raw Gaussian sample
 
-        return actions, log_probs
+        log_probs = log_probs.sum(-1)  # Sum over action dimensions
+        means = torch.tanh(means) if self.use_tanh_squash else means
+        return actions, log_probs, means
 
     def get_features(self, observations: torch.Tensor) -> torch.Tensor:
         """Get encoded features from observations"""
@@ -461,19 +529,13 @@ class SACObservationEncoder(nn.Module):
             )
         if "observation.state" in config.input_shapes:
             self.state_enc_layers = nn.Sequential(
-                nn.Linear(config.input_shapes["observation.state"][0], config.state_encoder_hidden_dim),
+                nn.Linear(config.input_shapes["observation.state"][0], config.latent_dim),
-                nn.ELU(),
-                nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
                 nn.LayerNorm(config.latent_dim),
                 nn.Tanh(),
             )
         if "observation.environment_state" in config.input_shapes:
             self.env_state_enc_layers = nn.Sequential(
-                nn.Linear(
+                nn.Linear(config.input_shapes["observation.environment_state"][0], config.latent_dim),
-                    config.input_shapes["observation.environment_state"][0], config.state_encoder_hidden_dim
-                ),
-                nn.ELU(),
-                nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
                 nn.LayerNorm(config.latent_dim),
                 nn.Tanh(),
             )
@@ -502,59 +564,16 @@ class SACObservationEncoder(nn.Module):
         return self.config.latent_dim
 
 
-class LagrangeMultiplier(nn.Module):
-    def __init__(
-        self,
-        init_value: float = 1.0,
-        constraint_shape: Sequence[int] = (),
-        device: str = "cuda"
-    ):
-        super().__init__()
-        self.device = torch.device(device)
-        init_value = torch.log(torch.exp(torch.tensor(init_value, device=self.device)) - 1)
-            
-        # Initialize the Lagrange multiplier as a parameter
-        self.lagrange = nn.Parameter(
-            torch.full(constraint_shape, init_value, dtype=torch.float32, device=self.device)
-        )
-        
-        self.to(self.device)
-
-    def forward(
-        self, 
-        lhs: Optional[torch.Tensor | float | int] = None, 
-        rhs: Optional[torch.Tensor | float | int] = None
-    ) -> torch.Tensor:
-        # Get the multiplier value based on parameterization        
-        multiplier = torch.nn.functional.softplus(self.lagrange)
-                
-        # Return the raw multiplier if no constraint values provided
-        if lhs is None:
-            return multiplier
-            
-        # Convert inputs to tensors and move to device
-        lhs = torch.tensor(lhs, device=self.device) if not isinstance(lhs, torch.Tensor) else lhs.to(self.device)
-        if rhs is not None:
-            rhs = torch.tensor(rhs, device=self.device) if not isinstance(rhs, torch.Tensor) else rhs.to(self.device)
-        else:
-            rhs = torch.zeros_like(lhs, device=self.device)
-            
-        diff = lhs - rhs
-        
-        assert diff.shape == multiplier.shape, f"Shape mismatch: {diff.shape} vs {multiplier.shape}"
-        
-        return multiplier * diff
-
-
 def orthogonal_init():
     return lambda x: torch.nn.init.orthogonal_(x, gain=1.0)
 
 
-def create_critic_ensemble(critics: list[nn.Module], num_critics: int, device: str = "cuda") -> nn.ModuleList:
+def create_critic_ensemble(critics: list[nn.Module], num_critics: int, device: str = "cpu") -> nn.ModuleList:
     """Creates an ensemble of critic networks"""
     assert len(critics) == num_critics, f"Expected {num_critics} critics, got {len(critics)}"
     return nn.ModuleList(critics).to(device)
 
+
 # borrowed from tdmpc
 def flatten_forward_unflatten(fn: Callable[[Tensor], Tensor], image_tensor: Tensor) -> Tensor:
     """Helper to temporarily flatten extra dims at the start of the image tensor.

lerobot/configs/policy/sac_pusht_keypoints.yaml

@@ -15,11 +15,11 @@ training:
   # Offline training dataloader
   num_workers: 4
 
-  batch_size: 128
+  batch_size: 256
   grad_clip_norm: 10.0
   lr: 3e-4
 
-  eval_freq: 50000
+  eval_freq: 2500
   log_freq: 500
   save_freq: 50000
 
@@ -46,8 +46,8 @@ policy:
 
   # Input / output structure.
   n_action_repeats: 1
-  horizon: 5
+  horizon: 2
-  n_action_steps: 5
+  n_action_steps: 2
 
   input_shapes:
     # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?

lerobot/scripts/train.py

@@ -99,7 +99,8 @@ def make_optimizer_and_scheduler(cfg, policy):
             [
                 {"params": policy.actor.parameters(), "lr": policy.config.actor_lr},
                 {"params": policy.critic_ensemble.parameters(), "lr": policy.config.critic_lr},
-                {"params": policy.temperature.parameters(), "lr": policy.config.temperature_lr},
+                # We wrap policy log temperature in list because this is a torch tensor and not a nn.Module
+                {"params": [policy.log_alpha], "lr": policy.config.temperature_lr},
             ]
         )
         lr_scheduler = None