File size: 6,333 Bytes
1641a08 | 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 |
"""
losses.py
"""
from dataclasses import dataclass
from typing import Dict, Tuple
import torch
import torch.nn.functional as F
# ============================================================
# 1) CONFIG
# ============================================================
@dataclass
class GeneralistLossConfig:
w_action: float = 1.0
w_physics: float = 20.0
w_value: float = 100.0
label_smoothing: float = 0.0
use_rtg_weighting: bool = True
rtg_weight_mode: str = "exp"
rtg_weight_beta: float = 2.0
min_token_weight: float = 0.05
# ============================================================
# 2) HELPERS
# ============================================================
def _expand_rtg_to_tokens(rtg_bt: torch.Tensor, K: int) -> torch.Tensor:
return rtg_bt.unsqueeze(-1).expand(-1, -1, K)
def _rtg_to_weights(rtg_input: torch.Tensor, mode: str, beta: float) -> torch.Tensor:
if mode == "none":
return torch.ones(rtg_input.shape[:2], device=rtg_input.device)
if rtg_input.dim() == 3:
mu = rtg_input.mean(dim=1, keepdim=True)
sig = rtg_input.std(dim=1, keepdim=True, unbiased=False).clamp_min(1e-5)
rtg_norm = (rtg_input - mu) / sig
scalar_rtg = rtg_norm.sum(dim=-1)
else:
scalar_rtg = rtg_input
mu_s = scalar_rtg.mean(dim=1, keepdim=True)
sig_s = scalar_rtg.std(dim=1, keepdim=True, unbiased=False).clamp_min(1e-5)
z = (scalar_rtg - mu_s) / sig_s
z = torch.clamp(z, -5.0, 5.0)
if mode == "clamp01":
w = torch.sigmoid(beta * z)
elif mode == "softplus":
w = F.softplus(beta * z)
elif mode == "exp":
w = torch.exp(beta * z)
else:
raise ValueError(f"Unknown rtg_weight_mode={mode}")
w = torch.clamp(w, min=0.01, max=50.0)
return w
# return total, metrics
def compute_generalist_loss(
model_out: Dict[str, torch.Tensor],
batch: Dict[str, torch.Tensor],
config: GeneralistLossConfig
) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
"""
Computes Physics loss and Rescaled Value loss.
"""
action_logits = model_out["action_logits"] # [B, T, K, n_bins]
state_preds = model_out["state_preds"] # [B, T, K]
state_preds_4h = model_out["state_preds_4h"] # [B, T, K]
return_preds = model_out["return_preds"] # [B, T, 2]
target_tokens = batch["target_action_tokens"]
target_mask = batch["target_mask"].float()
attn_mask = batch["attention_mask"].float()
target_rtg = batch["rtg"].float()
time_mask = batch.get("time_mask", torch.ones(target_rtg.shape[:2], device=target_rtg.device)).float()
B, T, K, n_bins = action_logits.shape
is_state = (1.0 - target_mask)
valid_phys = attn_mask * is_state
# 1) Stitching
if config.use_rtg_weighting:
w_bt = _rtg_to_weights(target_rtg, config.rtg_weight_mode, config.rtg_weight_beta)
w_btk = _expand_rtg_to_tokens(w_bt, K)
norm_factor = (target_mask * attn_mask).sum().clamp_min(1e-6) / (w_btk * target_mask * attn_mask).sum().clamp_min(1e-6)
token_importance = w_btk * norm_factor
else:
w_bt = torch.ones((B, T), device=action_logits.device)
token_importance = torch.ones((B, T, K), device=action_logits.device)
# 2) ACTION LOSS (CE)
flat_logits = action_logits.reshape(-1, n_bins)
flat_targets = target_tokens.reshape(-1)
flat_mask = (target_mask * attn_mask).reshape(-1)
flat_importance = token_importance.reshape(-1)
with torch.no_grad():
valid_t = flat_targets[flat_mask > 0.5]
if valid_t.numel() > 0:
counts = torch.bincount(valid_t, minlength=n_bins).float()
class_weights = (1.0 / (counts + 10.0)) / (1.0 / (counts + 10.0)).mean()
else:
class_weights = torch.ones(n_bins, device=flat_logits.device)
ce_per_token = F.cross_entropy(flat_logits, flat_targets, weight=class_weights, reduction="none", ignore_index=-100)
loss_action = (ce_per_token * flat_mask * flat_importance).sum() / flat_mask.sum().clamp_min(1e-6)
# ============================================================
# 3) PHYSICS LOSS (The Delta Fix)
# ============================================================
# Ground Truth from Dataloader
# next_obs is [B, T, 21]
# feature_values is [B, T, 64] (Padded tokens)
true_next = batch["next_obs"].float()
target_delta_4h = batch["target_4h_delta"].float()
K_limit = true_next.shape[2]
true_vals_sliced = batch["feature_values"].float().narrow(2, 0, K_limit)
s_pred_valid = state_preds.narrow(2, 0, K_limit)
s_pred_4h_valid = state_preds_4h.narrow(2, 0, K_limit)
v_phys_mask = valid_phys.narrow(2, 0, K_limit)
target_delta_1s = true_next - true_vals_sliced
mse_1s = (s_pred_valid - target_delta_1s) ** 2
mse_4h = (s_pred_4h_valid - target_delta_4h) ** 2
with torch.no_grad():
act_diff = torch.zeros((B, T), device=true_next.device)
if T > 1:
act_diff[:, 1:] = torch.abs(true_vals_sliced[:, 1:] - true_vals_sliced[:, :-1]).sum(dim=-1)
excitation = (1.0 + 5.0 * act_diff).unsqueeze(-1)
denom = (v_phys_mask * excitation).sum().clamp_min(1e-6)
loss_phys_1s = (mse_1s * v_phys_mask * excitation).sum() / denom
loss_phys_4h = (mse_4h * v_phys_mask * excitation).sum() / denom
loss_physics = loss_phys_1s + 0.5 * loss_phys_4h
val_mse = ((return_preds - target_rtg) ** 2).sum(dim=-1)
loss_value = (val_mse * w_bt * time_mask).sum() / time_mask.sum().clamp_min(1e-6)
loss_value = loss_value * 500.0
total = (config.w_action * loss_action) + \
(config.w_physics * loss_physics) + \
(config.w_value * loss_value)
with torch.no_grad():
acc = ((torch.argmax(flat_logits, -1) == flat_targets).float() * flat_mask).sum() / flat_mask.sum().clamp_min(1e-6)
if torch.rand(1) < 0.001:
print(f"[Loss Debug] Action: {loss_action.item():.3f} | Phys: {loss_physics.item():.3f} | Val: {loss_value.item():.3f}")
metrics = {
"loss_action": loss_action.detach(),
"loss_physics": loss_physics.detach(),
"loss_value": loss_value.detach(),
"accuracy": acc.detach(),
"total_loss": total.detach(),
}
return total, metrics |