tangger/lerobot
1
0
Fork 1
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Potential fixes for SAC instability and NAN bug

Browse Source
This commit is contained in:
joeclinton1
2025-01-03 21:12:59 +00:00
committed by Ke-Wang1017
parent f99e670976
commit db3925df28
1 changed files with 149 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

131
lerobot/common/policies/sac/modeling_sac.py
View File

@@ -18,8 +18,7 @@
# TODO: (1) better device management # TODO: (1) better device management
from collections import deque 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 einops
import numpy as np import numpy as np
@@ -72,8 +71,8 @@ class SACPolicy(
encoder=encoder_critic, encoder=encoder_critic,
network=MLP( network=MLP(
input_dim=encoder_critic.output_dim + config.output_shapes["action"][0], input_dim=encoder_critic.output_dim + config.output_shapes["action"][0],
**config.critic_network_kwargs **config.critic_network_kwargs,
) ),
) )
critic_nets.append(critic_net) critic_nets.append(critic_net)
@@ -83,8 +82,8 @@ class SACPolicy(
encoder=encoder_critic, encoder=encoder_critic,
network=MLP( network=MLP(
input_dim=encoder_critic.output_dim + config.output_shapes["action"][0], input_dim=encoder_critic.output_dim + config.output_shapes["action"][0],
**config.critic_network_kwargs **config.critic_network_kwargs,
) ),
) )
target_critic_nets.append(target_critic_net) target_critic_nets.append(target_critic_net)
@@ -93,12 +92,9 @@ class SACPolicy(
self.actor = Policy( self.actor = Policy(
encoder=encoder_actor, 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], action_dim=config.output_shapes["action"][0],
**config.policy_kwargs **config.policy_kwargs,
) )
if config.target_entropy is None: if config.target_entropy is None:
config.target_entropy = -np.prod(config.output_shapes["action"][0]) / 2 # (-dim(A)/2) config.target_entropy = -np.prod(config.output_shapes["action"][0]) / 2 # (-dim(A)/2)
@@ -126,7 +122,9 @@ class SACPolicy(
actions = self.unnormalize_outputs({"action": actions})["action"] actions = self.unnormalize_outputs({"action": actions})["action"]
return actions 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 """Forward pass through a critic network ensemble
Args: Args:
@@ -141,7 +139,6 @@ class SACPolicy(
q_values = torch.stack([critic(observations, actions) for critic in critics]) q_values = torch.stack([critic(observations, actions) for critic in critics])
return q_values return q_values
def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor | float]: def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor | float]:
"""Run the batch through the model and compute the loss. """Run the batch through the model and compute the loss.
@@ -182,7 +179,12 @@ class SACPolicy(
min_q, _ = q_targets.min(dim=0) # Get values from min operation min_q, _ = q_targets.min(dim=0) # Get values from min operation
# breakpoint() # breakpoint()
if self.config.use_backup_entropy: if self.config.use_backup_entropy:
min_q -= self.temperature() * log_probs * ~batch["observation.state_is_pad"][:,0] * ~batch["action_is_pad"][:,0] # shape: [batch_size, horizon] min_q -= (
self.temperature()
* log_probs
* ~batch["observation.state_is_pad"][:, 0]
* ~batch["action_is_pad"][:, 0]
) # shape: [batch_size, horizon]
td_target = rewards + self.config.discount * min_q * ~batch["next.done"] td_target = rewards + self.config.discount * min_q * ~batch["next.done"]
# td_target -= self.config.discount * self.temperature() * log_probs \ # td_target -= self.config.discount * self.temperature() * log_probs \
# * ~batch["observation.state_is_pad"][:,0] * ~batch["action_is_pad"][:,0] # shape: [batch_size, horizon] # * ~batch["observation.state_is_pad"][:,0] * ~batch["action_is_pad"][:,0] # shape: [batch_size, horizon]
@@ -213,7 +215,7 @@ class SACPolicy(
# 2- get actions (batch_size, action_dim) and log probs (batch_size,) # 2- get actions (batch_size, action_dim) and log probs (batch_size,)
actions, log_probs, _ = self.actor(observations) actions, log_probs, _ = self.actor(observations)
# 3- get q-value predictions # 3- get q-value predictions
# with torch.inference_mode(): with torch.inference_mode():
q_preds = self.critic_forward(observations, actions, use_target=False) q_preds = self.critic_forward(observations, actions, use_target=False)
# q_preds_min = torch.min(q_preds, axis=0) # q_preds_min = torch.min(q_preds, axis=0)
min_q_preds = q_preds.min(dim=0)[0] min_q_preds = q_preds.min(dim=0)[0]
@@ -226,16 +228,12 @@ class SACPolicy(
* ~batch["action_is_pad"][:, 0] # shape: [batch_size, horizon] * ~batch["action_is_pad"][:, 0] # shape: [batch_size, horizon]
).mean() ).mean()
# calculate temperature loss # calculate temperature loss
# 1- calculate entropy # 1- calculate entropy
with torch.no_grad(): with torch.no_grad():
actions, log_probs, _ = self.actor(observations) actions, log_probs, _ = self.actor(observations)
entropy = -log_probs.mean() entropy = -log_probs.mean()
temperature_loss = self.temperature( temperature_loss = self.temperature(lhs=entropy, rhs=self.config.target_entropy)
lhs=entropy,
rhs=self.config.target_entropy
)
loss = critics_loss + actor_loss + temperature_loss loss = critics_loss + actor_loss + temperature_loss
@@ -251,7 +249,7 @@ class SACPolicy(
"min_log_probs": log_probs.min().item(), "min_log_probs": log_probs.min().item(),
"max_log_probs": log_probs.max().item(), "max_log_probs": log_probs.max().item(),
"td_target_mean": td_target.mean().item(), "td_target_mean": td_target.mean().item(),
"td_target_mean": td_target.max().item(), "td_target_max": td_target.max().item(),
"action_mean": actions.mean().item(), "action_mean": actions.mean().item(),
"entropy": entropy.item(), "entropy": entropy.item(),
"loss": loss, "loss": loss,
@@ -268,10 +266,11 @@ class SACPolicy(
for target_critic, critic in zip(self.critic_target, self.critic_ensemble, strict=False): 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): for target_param, param in zip(target_critic.parameters(), critic.parameters(), strict=False):
target_param.data.copy_( target_param.data.copy_(
param.data * self.config.critic_target_update_weight + param.data * self.config.critic_target_update_weight
target_param.data * (1.0 - self.config.critic_target_update_weight) + target_param.data * (1.0 - self.config.critic_target_update_weight)
) )
class MLP(nn.Module): class MLP(nn.Module):
def __init__( def __init__(
self, self,
@@ -302,7 +301,9 @@ class MLP(nn.Module):
if dropout_rate is not None and dropout_rate > 0: if dropout_rate is not None and dropout_rate > 0:
layers.append(nn.Dropout(p=dropout_rate)) layers.append(nn.Dropout(p=dropout_rate))
layers.append(nn.LayerNorm(hidden_dims[i])) 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) self.net = nn.Sequential(*layers)
@@ -316,7 +317,7 @@ class Critic(nn.Module):
encoder: Optional[nn.Module], encoder: Optional[nn.Module],
network: nn.Module, network: nn.Module,
init_final: Optional[float] = None, init_final: Optional[float] = None,
device: str = "cuda" device: str = "cuda",
): ):
super().__init__() super().__init__()
self.device = torch.device(device) self.device = torch.device(device)
@@ -347,9 +348,7 @@ class Critic(nn.Module):
actions: torch.Tensor, actions: torch.Tensor,
) -> torch.Tensor: ) -> torch.Tensor:
# Move each tensor in observations to device # 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) actions = actions.to(self.device)
obs_enc = observations if self.encoder is None else self.encoder(observations) obs_enc = observations if self.encoder is None else self.encoder(observations)
@@ -371,7 +370,7 @@ class Policy(nn.Module):
fixed_std: Optional[torch.Tensor] = None, fixed_std: Optional[torch.Tensor] = None,
init_final: Optional[float] = None, init_final: Optional[float] = None,
use_tanh_squash: bool = False, use_tanh_squash: bool = False,
device: str = "cuda" device: str = "cuda",
): ):
super().__init__() super().__init__()
self.device = torch.device(device) self.device = torch.device(device)
@@ -412,7 +411,6 @@ class Policy(nn.Module):
self, self,
observations: torch.Tensor, observations: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]: ) -> Tuple[torch.Tensor, torch.Tensor]:
# Encode observations if encoder exists # Encode observations if encoder exists
obs_enc = observations if self.encoder is None else self.encoder(observations) obs_enc = observations if self.encoder is None else self.encoder(observations)
@@ -423,23 +421,28 @@ class Policy(nn.Module):
# Compute standard deviations # Compute standard deviations
if self.fixed_std is None: if self.fixed_std is None:
log_std = self.std_layer(outputs) log_std = self.std_layer(outputs)
log_std = torch.clamp(log_std, self.log_std_min, self.log_std_max) assert not torch.isnan(log_std).any(), "[ERROR] log_std became NaN after std_layer!"
if self.use_tanh_squash: if self.use_tanh_squash:
log_std = torch.tanh(log_std) 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: else:
log_std = self.fixed_std.expand_as(means) 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)) 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))
x_t = torch.clamp(x_t, -2.0, 2.0) log_probs = normal.log_prob(x_t) # Base log probability before Tanh
log_probs = normal.log_prob(x_t)
if self.use_tanh_squash: if self.use_tanh_squash:
actions = torch.tanh(x_t) 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:
means = torch.tanh(means) actions = x_t # No Tanh; raw Gaussian sample
log_probs = log_probs.sum(-1) # Sum over action dimensions
return actions, log_probs, means return actions, log_probs, means
def get_features(self, observations: torch.Tensor) -> torch.Tensor: def get_features(self, observations: torch.Tensor) -> torch.Tensor:
@@ -495,9 +498,7 @@ class SACObservationEncoder(nn.Module):
) )
if "observation.environment_state" in config.input_shapes: if "observation.environment_state" in config.input_shapes:
self.env_state_enc_layers = nn.Sequential( 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.latent_dim
),
nn.LayerNorm(config.latent_dim), nn.LayerNorm(config.latent_dim),
nn.Tanh(), nn.Tanh(),
) )
@@ -527,48 +528,47 @@ class SACObservationEncoder(nn.Module):
class LagrangeMultiplier(nn.Module): class LagrangeMultiplier(nn.Module):
def __init__( def __init__(self, init_value: float = 1.0, constraint_shape: Sequence[int] = (), device: str = "cuda"):
self,
init_value: float = 1.0,
constraint_shape: Sequence[int] = (),
device: str = "cuda"
):
super().__init__() super().__init__()
self.device = torch.device(device) self.device = torch.device(device)
# init_value = torch.log(torch.exp(torch.tensor(init_value, device=self.device)) - 1)
init_value = torch.tensor(init_value, device=self.device)
# Parameterize log(alpha) directly to ensure positivity
# Initialize the Lagrange multiplier as a parameter log_alpha = torch.log(torch.tensor(init_value, dtype=torch.float32, device=self.device))
self.lagrange = nn.Parameter( self.log_alpha = nn.Parameter(torch.full(constraint_shape, log_alpha))
torch.full(constraint_shape, init_value, dtype=torch.float32, device=self.device)
)
def forward( def forward(
self, self,
lhs: Optional[torch.Tensor | float | int] = None, lhs: Optional[Union[torch.Tensor, float, int]] = None,
rhs: Optional[torch.Tensor | float | int] = None rhs: Optional[Union[torch.Tensor, float, int]] = None,
) -> torch.Tensor: ) -> torch.Tensor:
# Get the multiplier value based on parameterization # Compute alpha = exp(log_alpha)
# multiplier = torch.nn.functional.softplus(self.lagrange) alpha = self.log_alpha.exp()
log_multiplier = torch.log(self.lagrange)
# Return the raw multiplier if no constraint values provided # Return alpha directly if no constraints provided
if lhs is None: if lhs is None:
return log_multiplier.exp() return alpha
# Convert inputs to tensors and move to device # 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) lhs = (
torch.tensor(lhs, device=self.device)
if not isinstance(lhs, torch.Tensor)
else lhs.to(self.device)
)
if rhs is not None: if rhs is not None:
rhs = torch.tensor(rhs, device=self.device) if not isinstance(rhs, torch.Tensor) else rhs.to(self.device) rhs = (
torch.tensor(rhs, device=self.device)
if not isinstance(rhs, torch.Tensor)
else rhs.to(self.device)
)
else: else:
rhs = torch.zeros_like(lhs, device=self.device) rhs = torch.zeros_like(lhs, device=self.device)
# Compute the difference and apply the multiplier
diff = lhs - rhs diff = lhs - rhs
assert diff.shape == log_multiplier.shape, f"Shape mismatch: {diff.shape} vs {log_multiplier.shape}" assert diff.shape == alpha.shape, f"Shape mismatch: {diff.shape} vs {alpha.shape}"
return log_multiplier.exp() * diff # numerically better return alpha * diff
def orthogonal_init(): def orthogonal_init():
@@ -580,6 +580,7 @@ def create_critic_ensemble(critics: list[nn.Module], num_critics: int, device: s
assert len(critics) == num_critics, f"Expected {num_critics} critics, got {len(critics)}" assert len(critics) == num_critics, f"Expected {num_critics} critics, got {len(critics)}"
return nn.ModuleList(critics).to(device) return nn.ModuleList(critics).to(device)
# borrowed from tdmpc # borrowed from tdmpc
def flatten_forward_unflatten(fn: Callable[[Tensor], Tensor], image_tensor: Tensor) -> Tensor: 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. """Helper to temporarily flatten extra dims at the start of the image tensor.