TD3B / td3b /td3b_losses.py
chq1155's picture
Upload TD3B code (inference, training, baselines)
ee6da62 verified
raw
history blame
20.7 kB
"""
TD3B Loss Functions
Implements contrastive loss for separating agonist/antagonist embeddings.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple
class ContrastiveLoss(nn.Module):
"""
Margin-based contrastive loss for separating agonist and antagonist embeddings.
For a pair of sequences (y_i, y_j):
- If both are agonists OR both are antagonists (similar): minimize distance
- If one is agonist and one is antagonist (dissimilar): maximize distance
Loss formula:
L_ctr = (1 - y_ij) * 0.5 * d²
+ y_ij * 0.5 * max(0, margin - d)²
where:
- d = ||emb_i - emb_j||_2 (Euclidean distance)
- y_ij = 0 if similar, 1 if dissimilar
- margin = minimum distance between dissimilar pairs
"""
def __init__(self, margin: float = 1.0, distance_metric: str = 'euclidean', adaptive_margin: bool = False):
"""
Args:
margin: Minimum distance between dissimilar pairs (base margin)
distance_metric: 'euclidean' or 'cosine'
adaptive_margin: If True, adjust margin based on actual dissimilar distances
"""
super().__init__()
self.base_margin = margin
self.distance_metric = distance_metric
self.adaptive_margin = adaptive_margin
def compute_distance(self, emb1: torch.Tensor, emb2: torch.Tensor) -> torch.Tensor:
"""
Compute pairwise distance between embeddings.
Args:
emb1: (batch_size, embedding_dim)
emb2: (batch_size, embedding_dim)
Returns:
distances: (batch_size,)
"""
if self.distance_metric == 'euclidean':
# L2 distance
distances = torch.norm(emb1 - emb2, p=2, dim=-1) # (B,)
elif self.distance_metric == 'cosine':
# Cosine distance = 1 - cosine_similarity
cos_sim = F.cosine_similarity(emb1, emb2, dim=-1) # (B,)
distances = 1.0 - cos_sim
else:
raise ValueError(f"Unknown distance metric: {self.distance_metric}")
return distances
def forward(
self,
embeddings: torch.Tensor,
labels: torch.Tensor,
confidences: Optional[torch.Tensor] = None,
debug: bool = False
) -> torch.Tensor:
"""
Compute contrastive loss for a batch.
Args:
embeddings: (batch_size, embedding_dim) sequence embeddings
labels: (batch_size,) directional labels in {-1, +1}
+1 = agonist, -1 = antagonist
confidences: (batch_size,) oracle confidence scores; pairs with product <= 0 are masked out
debug: If True, print detailed debugging information
Returns:
loss: scalar contrastive loss
"""
batch_size = embeddings.size(0)
if batch_size < 2:
if debug:
print(f"[ContrastiveLoss DEBUG] Batch size {batch_size} < 2, returning 0 loss")
return torch.tensor(0.0, device=embeddings.device)
if confidences is not None:
if not torch.is_tensor(confidences):
confidences = torch.as_tensor(confidences, device=embeddings.device)
else:
confidences = confidences.to(embeddings.device)
confidences = confidences.view(-1)
if confidences.numel() != batch_size:
raise ValueError(
f"Confidences size {confidences.numel()} does not match batch size {batch_size}"
)
# Compute pairwise distances (all pairs)
if self.distance_metric == 'euclidean':
distances = torch.cdist(embeddings, embeddings, p=2) # (B, B)
elif self.distance_metric == 'cosine':
emb_norm = F.normalize(embeddings, p=2, dim=-1)
distances = 1.0 - torch.matmul(emb_norm, emb_norm.T) # (B, B)
else:
raise ValueError(f"Unknown distance metric: {self.distance_metric}")
# Compute pairwise similarity labels
# y_ij = 0 if same class (both agonist or both antagonist)
# y_ij = 1 if different class
labels = labels.view(-1)
labels_expanded = labels.unsqueeze(1) # (B, 1)
label_product = labels_expanded * labels_expanded.T # (B, B)
# label_product > 0 means same class (both +1 or both -1)
# label_product < 0 means different class
dissimilar_mask = (label_product < 0) # (B, B) bool
# Exclude diagonal
eye_mask = torch.eye(batch_size, device=embeddings.device, dtype=torch.bool)
pos_mask = (~dissimilar_mask) & ~eye_mask
neg_mask = dissimilar_mask & ~eye_mask
# Apply confidence mask: remove pairs with confidence product <= 0
conf_mask = None
if confidences is not None:
conf_product = confidences.unsqueeze(0) * confidences.unsqueeze(1)
conf_mask = conf_product > 0
pos_mask = pos_mask & conf_mask
neg_mask = neg_mask & conf_mask
# Adaptive margin: set margin based on actual dissimilar distances
if self.adaptive_margin and neg_mask.any():
# Get all dissimilar distances
dissimilar_distances = distances[neg_mask]
# Set margin to 150% of mean dissimilar distance
# This ensures there's always room for optimization
adaptive_margin = 1.5 * dissimilar_distances.mean().item()
# Use max of base_margin and adaptive_margin
margin = max(self.base_margin, adaptive_margin)
else:
margin = self.base_margin
pos_count = pos_mask.sum()
neg_count = neg_mask.sum()
total_pairs = pos_count + neg_count
if total_pairs.item() == 0:
if debug:
print("[ContrastiveLoss DEBUG] No valid pairs after filtering, returning 0 loss")
return torch.tensor(0.0, device=embeddings.device)
# Contrastive loss
# For similar pairs: minimize squared distance
# For dissimilar pairs: squared hinge loss with margin
pos_loss = distances[pos_mask].pow(2).sum() / (pos_count + 1e-8)
neg_loss = torch.clamp(margin - distances[neg_mask], min=0.0).pow(2).sum() / (neg_count + 1e-8)
loss = pos_loss + neg_loss
if debug:
print(f"\n[ContrastiveLoss DEBUG]")
print(f" Batch size: {batch_size}")
print(f" Labels: {labels.cpu().tolist()}")
print(f" Unique labels: {torch.unique(labels).cpu().tolist()}")
print(f" Embedding shape: {embeddings.shape}")
print(f" Embedding norm (mean): {embeddings.norm(dim=-1).mean().item():.4f}")
print(f" Embedding norm (std): {embeddings.norm(dim=-1).std().item():.4f}")
valid_mask = pos_mask | neg_mask
if valid_mask.any():
valid_dists = distances[valid_mask]
print(f" Distance stats (valid pairs): mean={valid_dists.mean().item():.4f} "
f"min={valid_dists.min().item():.4f} max={valid_dists.max().item():.4f}")
if self.adaptive_margin and neg_mask.any():
print(f" Margin: {margin:.4f} (adaptive, base={self.base_margin})")
else:
print(f" Margin: {margin:.4f} (fixed)")
print(f" Num similar pairs: {pos_count.item():.0f}")
print(f" Num dissimilar pairs: {neg_count.item():.0f}")
if conf_mask is not None:
print(f" Confidence-passing pairs: {conf_mask.sum().item():.0f}")
print(f" Similar loss (mean): {pos_loss.item():.4f}")
print(f" Dissimilar loss (mean): {neg_loss.item():.4f}")
print(f" Total loss: {loss.item():.4f}")
# Show which dissimilar pairs have margin violations
margin_violations = (distances < margin) & neg_mask
if margin_violations.sum() > 0:
print(f" Margin violations: {margin_violations.sum().item():.0f} dissimilar pairs have distance < margin")
else:
print(f" Margin violations: 0 (all dissimilar pairs are already separated)")
return loss
class InfoNCELoss(nn.Module):
"""
Alternative: InfoNCE contrastive loss (used in SimCLR, CLIP).
Treats agonists as positive class, antagonists as negative class.
For each agonist, pull it close to other agonists and push away from antagonists.
For each antagonist, pull it close to other antagonists and push away from agonists.
"""
def __init__(self, temperature: float = 0.1):
"""
Args:
temperature: Temperature parameter for softmax
"""
super().__init__()
self.temperature = temperature
def forward(
self,
embeddings: torch.Tensor,
labels: torch.Tensor,
confidences: Optional[torch.Tensor] = None,
debug: bool = False
) -> torch.Tensor:
"""
Compute InfoNCE loss.
Args:
embeddings: (batch_size, embedding_dim)
labels: (batch_size,) in {-1, +1}
confidences: (batch_size,) oracle confidence scores; pairs with product <= 0 are masked out
debug: Unused (kept for API compatibility)
Returns:
loss: scalar
"""
batch_size = embeddings.size(0)
if confidences is not None:
if not torch.is_tensor(confidences):
confidences = torch.as_tensor(confidences, device=embeddings.device)
else:
confidences = confidences.to(embeddings.device)
confidences = confidences.view(-1)
if confidences.numel() != batch_size:
raise ValueError(
f"Confidences size {confidences.numel()} does not match batch size {batch_size}"
)
if batch_size < 2:
return torch.tensor(0.0, device=embeddings.device)
# Normalize embeddings
embeddings = F.normalize(embeddings, p=2, dim=-1) # (B, D)
# Compute similarity matrix
similarity = torch.matmul(embeddings, embeddings.T) / self.temperature # (B, B)
# Create positive/negative masks
labels_expanded = labels.unsqueeze(1) # (B, 1)
label_product = labels_expanded * labels_expanded.T # (B, B)
positive_mask = (label_product > 0) # Same class
negative_mask = (label_product < 0) # Different class
# Remove self-similarity
positive_mask.fill_diagonal_(0)
if confidences is not None:
conf_product = confidences.unsqueeze(0) * confidences.unsqueeze(1)
conf_mask = conf_product > 0
positive_mask = positive_mask & conf_mask
negative_mask = negative_mask & conf_mask
# For each sample, compute InfoNCE loss
# log( exp(sim_pos) / (exp(sim_pos) + sum(exp(sim_neg))) )
losses = []
for i in range(batch_size):
# Positive samples
pos_sims = similarity[i][positive_mask[i]] # (num_pos,)
# Negative samples
neg_sims = similarity[i][negative_mask[i]] # (num_neg,)
# Check if there are positive samples
if pos_sims.numel() == 0:
continue
# LogSumExp for numerical stability
pos_exp = torch.exp(pos_sims) # (num_pos,)
neg_exp = torch.exp(neg_sims) # (num_neg,)
if neg_exp.numel() == 0:
continue
# Average over positive samples
denominator = pos_exp.sum() + neg_exp.sum()
loss_i = -torch.log(pos_exp.sum() / (denominator + 1e-8))
losses.append(loss_i)
if len(losses) == 0:
return torch.tensor(0.0, device=embeddings.device)
return torch.stack(losses).mean()
class TD3BTotalLoss:
"""
Combined TD3B loss: L_total = L_WDCE + λ * L_ctr + β * L_KL
Components:
- L_WDCE: Weighted Denoising Cross-Entropy (from TR2-D2)
- L_ctr: Contrastive loss for agonist/antagonist separation
- L_KL: KL divergence regularization between policy and reference model
"""
def __init__(
self,
contrastive_weight: float = 0.1,
contrastive_margin: float = 1.0,
contrastive_type: str = 'margin', # 'margin' or 'infonce'
kl_beta: float = 0.1, # β coefficient for KL divergence
reference_model: Optional[nn.Module] = None,
adaptive_margin: bool = True # Enable adaptive margin by default
):
"""
Args:
contrastive_weight: λ coefficient for contrastive loss
contrastive_margin: Margin for margin-based contrastive loss (base margin if adaptive)
contrastive_type: Type of contrastive loss ('margin' or 'infonce')
kl_beta: β coefficient for KL divergence regularization
reference_model: Frozen reference model for KL divergence (deepcopy of pretrained)
adaptive_margin: If True, automatically adjust margin based on dissimilar distances
"""
self.contrastive_weight = contrastive_weight
self.kl_beta = kl_beta
self.reference_model = reference_model
# Freeze reference model if provided
if self.reference_model is not None:
self.reference_model.eval()
for param in self.reference_model.parameters():
param.requires_grad = False
# Verify all parameters are frozen
assert all(not p.requires_grad for p in self.reference_model.parameters()), \
"ERROR: Reference model has parameters with requires_grad=True!"
if contrastive_type == 'margin':
self.contrastive_loss = ContrastiveLoss(
margin=contrastive_margin,
distance_metric='euclidean',
adaptive_margin=adaptive_margin
)
elif contrastive_type == 'infonce':
self.contrastive_loss = InfoNCELoss(temperature=0.1)
else:
raise ValueError(f"Unknown contrastive type: {contrastive_type}")
def compute_kl_categorical(
self,
log_p: torch.Tensor,
log_ref_p: torch.Tensor
) -> torch.Tensor:
"""
Compute KL divergence between categorical distributions.
KL(P || Q) = Σ P(x) * log(P(x) / Q(x))
= Σ P(x) * (log P(x) - log Q(x))
Args:
log_p: (B, L, Vocab) log-probabilities from policy model
log_ref_p: (B, L, Vocab) log-probabilities from reference model
Returns:
kl: (B, L) KL divergence per position
"""
# Convert log-probs to probabilities
p = torch.exp(log_p) # (B, L, Vocab)
# KL divergence element-wise
kl_elementwise = p * (log_p - log_ref_p) # (B, L, Vocab)
# Handle numerical issues: 0 * log(0) should be 0
# Replace NaNs or Infs that occur at -inf locations with 0
kl_elementwise = torch.where(
torch.isfinite(kl_elementwise),
kl_elementwise,
torch.zeros_like(kl_elementwise)
)
# Sum over vocabulary dimension
kl = kl_elementwise.sum(dim=-1) # (B, L)
return kl
def compute_kl_loss(
self,
policy_model: nn.Module,
sequences: torch.Tensor,
attn_mask: torch.Tensor,
sigma: torch.Tensor
) -> torch.Tensor:
"""
Compute KL divergence loss between policy model and reference model.
Args:
policy_model: Current policy model (being trained)
sequences: (B, L) input sequences
attn_mask: (B, L) attention mask
sigma: (B,) noise schedule
Returns:
kl_loss: Scalar KL divergence loss
"""
if self.reference_model is None:
return torch.tensor(0.0, device=sequences.device)
# Ensure reference model is in eval mode
assert not self.reference_model.training, \
"ERROR: Reference model is in training mode! It should always be in eval mode."
# Forward through policy model (already computed in WDCE, but need logits)
policy_logits = policy_model(sequences, attn_mask=attn_mask, sigma=sigma) # (B, L, Vocab)
# Forward through reference model (frozen, no gradients)
with torch.no_grad():
ref_logits = self.reference_model(sequences, attn_mask=attn_mask, sigma=sigma) # (B, L, Vocab)
# Convert to log-probabilities
log_p = F.log_softmax(policy_logits, dim=-1) # (B, L, Vocab)
log_ref_p = F.log_softmax(ref_logits, dim=-1) # (B, L, Vocab)
# Compute KL divergence
kl_per_position = self.compute_kl_categorical(log_p, log_ref_p) # (B, L)
# Mask out padding positions
kl_masked = kl_per_position * attn_mask.float() # (B, L)
# Average over all non-padding positions
num_valid = attn_mask.float().sum()
kl_loss = kl_masked.sum() / (num_valid + 1e-8)
return kl_loss
def compute_loss(
self,
wdce_loss: torch.Tensor,
embeddings: torch.Tensor,
directional_labels: torch.Tensor,
confidences: Optional[torch.Tensor] = None,
kl_loss: Optional[torch.Tensor] = None,
debug: bool = False
) -> Tuple[torch.Tensor, dict]:
"""
Compute total TD3B loss.
Args:
wdce_loss: Precomputed WDCE loss (scalar)
embeddings: (batch_size, embedding_dim) sequence embeddings from MDLM
directional_labels: (batch_size,) labels in {-1, +1}
confidences: (batch_size,) oracle confidence scores; pairs with product <= 0 are masked out
kl_loss: Precomputed KL divergence loss (optional)
debug: If True, enable debugging output in contrastive loss
Returns:
total_loss: Combined loss
loss_dict: Dictionary with individual loss components
"""
# Contrastive loss (pass debug flag)
contrastive_loss = self.contrastive_loss(
embeddings,
directional_labels,
confidences=confidences,
debug=debug
)
# KL divergence loss
if kl_loss is None:
kl_loss = torch.tensor(0.0, device=embeddings.device)
# Total loss: L_total = L_WDCE + λ * L_ctr + β * L_KL
total_loss = wdce_loss + self.contrastive_weight * contrastive_loss + self.kl_beta * kl_loss
loss_dict = {
'total_loss': total_loss.item(),
'wdce_loss': wdce_loss.item(),
'contrastive_loss': contrastive_loss.item(),
'kl_loss': kl_loss.item() if isinstance(kl_loss, torch.Tensor) else kl_loss
}
return total_loss, loss_dict
def extract_embeddings_from_mdlm(
model,
sequences: torch.Tensor,
pool_method: str = 'mean'
) -> torch.Tensor:
"""
Extract sequence-level embeddings from MDLM backbone.
Args:
model: MDLM model with backbone (Roformer)
sequences: (batch_size, seq_len) token sequences
pool_method: 'mean', 'max', or 'cls'
Returns:
embeddings: (batch_size, hidden_dim)
"""
# Create attention mask (1 for real tokens, 0 for padding)
attn_mask = (sequences != 0).long() # (B, L)
# Forward through Roformer backbone to get hidden states
# IMPORTANT: DO NOT use torch.no_grad() here - we need gradients for backprop!
# Access the underlying RoFormerForMaskedLM model and request hidden states
outputs = model.backbone.model(
input_ids=sequences,
attention_mask=attn_mask,
output_hidden_states=True,
return_dict=True
)
# Extract last hidden state from outputs
# outputs.hidden_states is a tuple of (embedding_output, layer1, layer2, ..., layerN)
# We want the last layer's hidden states
hidden_states = outputs.hidden_states[-1] # (B, L, D)
# Pool to get sequence-level embeddings
if pool_method == 'mean':
# Mean pooling (ignore padding)
mask = attn_mask.float().unsqueeze(-1) # (B, L, 1)
pooled = (hidden_states * mask).sum(dim=1) / (mask.sum(dim=1) + 1e-8) # (B, D)
elif pool_method == 'max':
# Max pooling
pooled = hidden_states.max(dim=1)[0] # (B, D)
elif pool_method == 'cls':
# Use first token (CLS-style)
pooled = hidden_states[:, 0, :] # (B, D)
else:
raise ValueError(f"Unknown pool method: {pool_method}")
return pooled