File size: 25,534 Bytes
ee6da62 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 | """
TD3B Finetuning Loop
Extends TR2-D2 training with contrastive loss and directional rewards.
"""
import numpy as np
import torch
import wandb
import os
from finetune_utils import loss_wdce
from .td3b_losses import TD3BTotalLoss, extract_embeddings_from_mdlm
from tqdm import tqdm
import pandas as pd
from plotting import plot_data_with_distribution_seaborn, plot_data
def td3b_finetune(
args,
cfg,
policy_model,
reward_model,
mcts=None,
pretrained=None,
filename=None,
prot_name=None,
eps=1e-5,
# TD3B-specific arguments
contrastive_weight=0.1,
contrastive_margin=1.0,
contrastive_type='margin',
embedding_pool_method='mean',
kl_beta=0.1
):
"""
TD3B finetuning with combined WDCE + contrastive loss + KL regularization.
Args:
args: Configuration arguments
cfg: Hydra config
policy_model: Policy model (MDLM)
reward_model: Reward scoring functions (TD3BRewardFunction)
mcts: TD3B_MCTS instance
pretrained: Pretrained model (for no-MCTS mode)
filename: Output filename
prot_name: Target protein name
eps: Small epsilon
contrastive_weight: λ for contrastive loss
contrastive_margin: Margin for margin-based contrastive loss
contrastive_type: 'margin' or 'infonce'
embedding_pool_method: 'mean', 'max', or 'cls'
kl_beta: β coefficient for KL divergence regularization
Returns:
batch_losses: List of training losses
"""
base_path = args.base_path
dt = (1 - eps) / args.total_num_steps
if args.no_mcts:
assert pretrained is not None, "pretrained model is required for no mcts"
else:
assert mcts is not None, "mcts is required for mcts"
# Create reference model (frozen copy of policy model at start of training)
# Cannot use copy.deepcopy() due to unpicklable objects (file handles, etc.)
# Instead, create a new model instance and load CLONED state dict
print("[TD3B] Creating reference model for KL regularization...")
# Import Diffusion class
from diffusion import Diffusion
# Create new instance with same config
reference_model = Diffusion(
config=policy_model.config,
tokenizer=policy_model.tokenizer,
mode="eval",
device=policy_model.device if hasattr(policy_model, 'device') else args.device
)
# Get the device from policy model
device = policy_model.device if hasattr(policy_model, 'device') else args.device
if device is None:
device = next(policy_model.parameters()).device
# IMPORTANT: Clone the state dict to create independent tensors
# This ensures no memory sharing between policy and reference model
state_dict_copy = {
key: value.clone().detach()
for key, value in policy_model.state_dict().items()
}
reference_model.load_state_dict(state_dict_copy)
# Move reference model to same device as policy model
reference_model = reference_model.to(device)
# Freeze and set to eval mode
reference_model.eval()
for param in reference_model.parameters():
param.requires_grad = False
print(f"[TD3B] Reference model frozen with {sum(p.numel() for p in reference_model.parameters())} parameters")
print(f"[TD3B] Reference model on device: {device}")
# Verify no parameter sharing
policy_params = {id(p) for p in policy_model.parameters()}
ref_params = {id(p) for p in reference_model.parameters()}
assert len(policy_params.intersection(ref_params)) == 0, \
"ERROR: Reference model shares parameters with policy model!"
print("[TD3B] ✓ Verified: No parameter sharing between policy and reference model")
# Initialize TD3B total loss
td3b_loss_fn = TD3BTotalLoss(
contrastive_weight=contrastive_weight,
contrastive_margin=contrastive_margin,
contrastive_type=contrastive_type,
kl_beta=kl_beta,
reference_model=reference_model
)
# Set model to train mode
policy_model.train()
torch.set_grad_enabled(True)
optim = torch.optim.AdamW(policy_model.parameters(), lr=args.learning_rate)
# Record metrics
batch_losses = []
batch_wdce_losses = []
batch_contrastive_losses = []
batch_kl_losses = []
# Initialize saved trajectories
x_saved, log_rnd_saved, final_rewards_saved = None, None, None
directional_labels_saved, confidences_saved = None, None
# Logs
valid_fraction_log = []
affinity_log = []
gated_reward_log = []
confidence_log = []
direction_prediction_log = [] # Oracle predictions f_φ ∈ [0, 1]
consistency_reward_log = [] # d* × (f_φ - 0.5)
### Fine-Tuning Loop ###
pbar = tqdm(range(args.num_epochs))
for epoch in pbar:
rewards = []
losses = []
policy_model.train()
with torch.no_grad():
if x_saved is None or epoch % args.resample_every_n_step == 0:
# Generate trajectories
if args.no_mcts:
# Direct sampling (not typical for TD3B, but keep for compatibility)
x_final, log_rnd, final_rewards = policy_model.sample_finetuned_with_rnd(
args, reward_model, pretrained
)
directional_labels = torch.zeros(x_final.size(0), dtype=torch.float32)
confidences = torch.ones(x_final.size(0), dtype=torch.float32)
else:
# TD3B MCTS forward pass
# For dual-direction mode, sample BOTH directions in the same batch
if hasattr(args, 'target_direction') and args.target_direction == 'both':
print(f"[Dual-direction] Epoch {epoch}: Sampling BOTH agonist and antagonist binders")
# Sample agonist binders (d* = +1)
reward_model.target_direction = 1.0
if epoch % args.reset_every_n_step == 0:
results_agonist = mcts.forward(resetTree=True)
else:
results_agonist = mcts.forward(resetTree=False)
# Sample antagonist binders (d* = -1)
reward_model.target_direction = -1.0
# Don't reset tree for antagonist to save computation
results_antagonist = mcts.forward(resetTree=False)
# Unpack both results
if len(results_agonist) == 7 and len(results_antagonist) == 7:
x_agonist, log_rnd_agonist, rewards_agonist, _, _, labels_agonist, conf_agonist = results_agonist
x_antagonist, log_rnd_antagonist, rewards_antagonist, _, _, labels_antagonist, conf_antagonist = results_antagonist
# Force labels to be correct (in case oracle is wrong)
labels_agonist = torch.ones(x_agonist.size(0), dtype=torch.float32) * 1.0 # +1 for agonist
labels_antagonist = torch.ones(x_antagonist.size(0), dtype=torch.float32) * -1.0 # -1 for antagonist
# Combine both directions into single batch
x_final = torch.cat([x_agonist, x_antagonist], dim=0)
log_rnd = torch.cat([log_rnd_agonist, log_rnd_antagonist], dim=0)
final_rewards = np.concatenate([rewards_agonist, rewards_antagonist], axis=0)
directional_labels = torch.cat([labels_agonist, labels_antagonist], dim=0)
confidences = torch.cat([
conf_agonist if isinstance(conf_agonist, torch.Tensor) else torch.tensor(conf_agonist),
conf_antagonist if isinstance(conf_antagonist, torch.Tensor) else torch.tensor(conf_antagonist)
], dim=0)
print(f" → Combined batch: {x_agonist.size(0)} agonists + {x_antagonist.size(0)} antagonists = {x_final.size(0)} total")
print(f" → Directional labels: {torch.unique(directional_labels).tolist()} (DIVERSITY CONFIRMED!)")
else:
raise ValueError("Dual-direction mode requires 7-value return from MCTS")
else:
# Single-direction mode
if epoch % args.reset_every_n_step == 0:
results = mcts.forward(resetTree=True)
else:
results = mcts.forward(resetTree=False)
# Unpack results (TD3B version includes directional labels and confidences)
if len(results) == 7:
x_final, log_rnd, final_rewards, score_vectors, sequences, directional_labels, confidences = results
# Convert numpy arrays to tensors immediately for consistency
if not isinstance(directional_labels, torch.Tensor):
directional_labels = torch.tensor(directional_labels, dtype=torch.float32)
if not isinstance(confidences, torch.Tensor):
confidences = torch.tensor(confidences, dtype=torch.float32)
else:
# Fallback for compatibility with base MCTS
x_final, log_rnd, final_rewards, score_vectors, sequences = results
directional_labels = torch.zeros(x_final.size(0), dtype=torch.float32)
confidences = torch.ones(x_final.size(0), dtype=torch.float32)
# Save for next iteration
x_saved = x_final
log_rnd_saved = log_rnd
final_rewards_saved = final_rewards
directional_labels_saved = directional_labels
confidences_saved = confidences
else:
# Reuse cached trajectories
x_final = x_saved
log_rnd = log_rnd_saved
final_rewards = final_rewards_saved
directional_labels = directional_labels_saved
confidences = confidences_saved
# Compute WDCE loss
wdce_loss = loss_wdce(
policy_model,
log_rnd,
x_final,
num_replicates=args.wdce_num_replicates,
centering=args.centering
)
# Compute KL divergence loss
# Use a random masking and forward pass for KL computation
mask_index = policy_model.mask_index
device = x_final.device
# Sample random noise level
lamda = torch.rand(x_final.shape[0], device=device) # (B,)
sigma_kl = -torch.log1p(-(1 - eps) * lamda)
# Apply random masking
masked_index = torch.rand(*x_final.shape, device=device) < lamda[..., None] # (B, L)
perturbed_batch = torch.where(masked_index, mask_index, x_final)
attn_mask_kl = torch.ones_like(perturbed_batch).to(device)
# Compute KL loss
kl_loss = td3b_loss_fn.compute_kl_loss(
policy_model,
perturbed_batch,
attn_mask_kl,
sigma_kl
)
# Extract embeddings for contrastive loss
# Only compute if we have directional labels
if directional_labels is not None and len(torch.unique(directional_labels)) > 1:
# Get device from backbone
device = policy_model.backbone.device if hasattr(policy_model.backbone, 'device') else x_final.device
embeddings = extract_embeddings_from_mdlm(
policy_model,
x_final.to(device),
pool_method=embedding_pool_method
)
# Move directional labels to same device
directional_labels = directional_labels.to(embeddings.device)
# Enable debug mode for first 3 epochs or if loss was zero last epoch
debug_mode = (epoch < 3) or (epoch > 0 and batch_contrastive_losses and batch_contrastive_losses[-1] < 1e-6)
# Compute total TD3B loss
total_loss, loss_dict = td3b_loss_fn.compute_loss(
wdce_loss,
embeddings,
directional_labels,
kl_loss=kl_loss, # Pass KL loss
debug=debug_mode # Enable debugging when needed
)
else:
# If no directional diversity, skip contrastive loss
print(f"[WARNING] Epoch {epoch}: No directional diversity! Skipping contrastive loss.")
print(f" Labels: {directional_labels.cpu().tolist() if directional_labels is not None else 'None'}")
total_loss = wdce_loss + td3b_loss_fn.kl_beta * kl_loss
loss_dict = {
'total_loss': total_loss.item(),
'wdce_loss': wdce_loss.item(),
'contrastive_loss': 0.0,
'kl_loss': kl_loss.item()
}
# Gradient descent
total_loss.backward()
# Gradient clipping
if args.grad_clip:
torch.nn.utils.clip_grad_norm_(policy_model.parameters(), args.gradnorm_clip)
optim.step()
optim.zero_grad()
pbar.set_postfix(
total_loss=loss_dict['total_loss'],
wdce=loss_dict['wdce_loss'],
ctr=loss_dict['contrastive_loss']
)
# Evaluation sampling
x_eval, eval_metrics = policy_model.sample_finetuned_td3b(
args,
reward_model,
batch_size=50,
dataframe=False
)
# Extract metrics (TD3B-specific)
affinity = eval_metrics.get('affinity', [0])
gated_reward = eval_metrics.get('gated_reward', [0])
confidence = eval_metrics.get('confidence', [1])
valid_fraction = eval_metrics.get('valid_fraction', 0)
# Extract direction predictions (f_φ ∈ [0, 1])
direction_predictions = eval_metrics.get('direction_predictions', [0.5])
# Compute consistency reward: d* × (f_φ - 0.5)
# Get target direction d* from reward_model
d_star = reward_model.target_direction # +1 or -1
consistency_rewards = [d_star * (f_phi - 0.5) for f_phi in direction_predictions]
# Append to logs
affinity_log.append(affinity)
gated_reward_log.append(gated_reward)
confidence_log.append(confidence)
valid_fraction_log.append(valid_fraction)
direction_prediction_log.append(direction_predictions)
consistency_reward_log.append(consistency_rewards)
batch_losses.append(loss_dict['total_loss'])
batch_wdce_losses.append(loss_dict['wdce_loss'])
batch_contrastive_losses.append(loss_dict['contrastive_loss'])
batch_kl_losses.append(loss_dict.get('kl_loss', 0.0))
# Compute search statistics
if args.no_mcts:
mean_reward_search = final_rewards.mean().item()
min_reward_search = final_rewards.min().item()
max_reward_search = final_rewards.max().item()
median_reward_search = final_rewards.median().item()
else:
mean_reward_search = np.mean(final_rewards)
min_reward_search = np.min(final_rewards)
max_reward_search = np.max(final_rewards)
median_reward_search = np.median(final_rewards)
# Compute direction oracle and consistency reward statistics
mean_direction = np.mean(direction_predictions) if len(direction_predictions) > 0 else 0.5
std_direction = np.std(direction_predictions) if len(direction_predictions) > 0 else 0.0
mean_consistency = np.mean(consistency_rewards) if len(consistency_rewards) > 0 else 0.0
std_consistency = np.std(consistency_rewards) if len(consistency_rewards) > 0 else 0.0
print(
f"epoch {epoch} | "
f"affinity {np.mean(affinity):.4f} | "
f"gated_reward {np.mean(gated_reward):.4f} | "
f"confidence {np.mean(confidence):.4f} | "
f"valid_frac {valid_fraction:.4f} | "
f"direction_oracle {mean_direction:.4f}±{std_direction:.4f} | "
f"consistency_reward {mean_consistency:.4f}±{std_consistency:.4f} | "
f"total_loss {loss_dict['total_loss']:.4f} | "
f"wdce_loss {loss_dict['wdce_loss']:.4f} | "
f"contrastive_loss {loss_dict['contrastive_loss']:.4f} | "
f"kl_loss {loss_dict.get('kl_loss', 0.0):.4f}"
)
# W&B logging
wandb.log({
"epoch": epoch,
"affinity": np.mean(affinity),
"gated_reward": np.mean(gated_reward),
"confidence": np.mean(confidence),
"valid_fraction": valid_fraction,
"direction_oracle/mean": mean_direction,
"direction_oracle/std": std_direction,
"consistency_reward/mean": mean_consistency,
"consistency_reward/std": std_consistency,
"total_loss": loss_dict['total_loss'],
"wdce_loss": loss_dict['wdce_loss'],
"contrastive_loss": loss_dict['contrastive_loss'],
"kl_loss": loss_dict.get('kl_loss', 0.0),
"mean_reward_search": mean_reward_search,
"min_reward_search": min_reward_search,
"max_reward_search": max_reward_search,
"median_reward_search": median_reward_search
})
# Save checkpoint
if (epoch + 1) % args.save_every_n_epochs == 0:
model_path = os.path.join(args.save_path, f'model_{epoch}.ckpt')
torch.save(policy_model.state_dict(), model_path)
print(f"model saved at epoch {epoch}")
### End of Fine-Tuning Loop ###
wandb.finish()
# Save logs and plots
plot_path = f'{base_path}/TR2-D2/tr2d2-pep/results/{args.run_name}'
os.makedirs(plot_path, exist_ok=True)
output_log_path = f'{base_path}/TR2-D2/tr2d2-pep/results/{args.run_name}/log_{filename}.csv'
save_td3b_logs_to_file(
valid_fraction_log,
affinity_log,
gated_reward_log,
confidence_log,
direction_prediction_log,
consistency_reward_log,
output_log_path
)
plot_data(valid_fraction_log,
save_path=f'{base_path}/TR2-D2/tr2d2-pep/results/{args.run_name}/valid_{filename}.png')
plot_data_with_distribution_seaborn(
log1=affinity_log,
save_path=f'{base_path}/TR2-D2/tr2d2-pep/results/{args.run_name}/affinity_{filename}.png',
label1=f"Average Affinity to {prot_name}",
title=f"Average Affinity to {prot_name} Over Iterations"
)
plot_data_with_distribution_seaborn(
log1=gated_reward_log,
save_path=f'{base_path}/TR2-D2/tr2d2-pep/results/{args.run_name}/gated_reward_{filename}.png',
label1="Average Gated Reward",
title="Average Gated Reward Over Iterations"
)
plot_data_with_distribution_seaborn(
log1=confidence_log,
save_path=f'{base_path}/TR2-D2/tr2d2-pep/results/{args.run_name}/confidence_{filename}.png',
label1="Average Confidence",
title="Average Confidence Over Iterations"
)
# Final evaluation
x_eval, eval_metrics, df = policy_model.sample_finetuned_td3b(
args,
reward_model,
batch_size=200,
dataframe=True
)
df.to_csv(f'{base_path}/TR2-D2/tr2d2-pep/results/{args.run_name}/{prot_name}_generation_results.csv', index=False)
return batch_losses
def save_td3b_logs_to_file(valid_fraction_log, affinity_log, gated_reward_log, confidence_log,
direction_prediction_log, consistency_reward_log, output_path):
"""
Saves TD3B-specific logs to a CSV file.
Parameters:
valid_fraction_log (list): Log of valid fractions over iterations.
affinity_log (list): Log of binding affinity over iterations.
gated_reward_log (list): Log of gated rewards over iterations.
confidence_log (list): Log of confidence scores over iterations.
direction_prediction_log (list): Log of direction oracle predictions over iterations.
consistency_reward_log (list): Log of consistency rewards over iterations.
output_path (str): Path to save the log CSV file.
"""
os.makedirs(os.path.dirname(output_path), exist_ok=True)
# Combine logs into a DataFrame
log_data = {
"Iteration": list(range(1, len(valid_fraction_log) + 1)),
"Valid Fraction": valid_fraction_log,
"Binding Affinity": affinity_log,
"Gated Reward": gated_reward_log,
"Confidence": confidence_log,
"Direction Oracle": direction_prediction_log,
"Consistency Reward": consistency_reward_log
}
df = pd.DataFrame(log_data)
# Save to CSV
df.to_csv(output_path, index=False)
print(f"Logs saved to {output_path}")
# Add sampling method to diffusion model (monkey patch or extend)
def add_td3b_sampling_to_model(model):
"""
Adds TD3B-specific sampling method to the model.
This is a helper function to extend the existing model.
"""
def sample_finetuned_td3b(self, args, reward_model, batch_size=50, dataframe=False):
"""
TD3B-specific sampling that returns directional metrics.
"""
self.backbone.eval()
self.noise.eval()
if batch_size is None:
batch_size = args.batch_size
eps = getattr(args, "sampling_eps", 1e-5)
num_steps = args.total_num_steps
x_rollout = self.sample_prior(
batch_size,
args.seq_length).to(self.device, dtype=torch.long)
timesteps = torch.linspace(1, eps, num_steps + 1, device=self.device)
dt = torch.tensor((1 - eps) / num_steps, device=self.device)
for i in range(num_steps):
t = timesteps[i] * torch.ones(x_rollout.shape[0], 1, device=self.device)
log_p, x_next = self.single_reverse_step(x_rollout, t=t, dt=dt)
x_rollout = x_next.to(self.device)
mask_positions = (x_rollout == self.mask_index)
if mask_positions.any().item():
log_p, x_next = self.single_noise_removal(x_rollout, t=t, dt=dt)
x_rollout = x_next.to(self.device)
# Convert x to sequences to get valid ones
from utils.app import PeptideAnalyzer
analyzer = PeptideAnalyzer()
sequences = self.tokenizer.batch_decode(x_rollout)
valid_mask = torch.tensor([analyzer.is_peptide(seq) for seq in sequences], device=self.device)
valid_sequences = [seq for seq, keep in zip(sequences, valid_mask.tolist()) if keep]
valid_x_final = x_rollout[valid_mask] if valid_mask.any().item() else torch.empty(0, device=self.device)
valid_fraction = len(valid_sequences) / batch_size
if len(valid_sequences) > 0:
result = reward_model(valid_sequences)
if isinstance(result, tuple):
total_rewards, info = result
affinity = np.asarray(info.get('affinities', total_rewards))
confidence = np.asarray(info.get('confidences', np.ones_like(affinity)))
direction_predictions = np.asarray(info.get('directions', np.zeros_like(affinity)))
else:
total_rewards = np.asarray(result)
if total_rewards.ndim > 1:
affinity = total_rewards[:, 0]
else:
affinity = total_rewards
confidence = np.ones_like(affinity)
direction_predictions = np.zeros_like(affinity)
rewards_t = torch.as_tensor(total_rewards, dtype=torch.float32, device=self.device)
alpha = max(float(getattr(args, "alpha", 0.1)), 1e-6)
weights = torch.softmax(rewards_t / alpha, dim=0)
idx = torch.multinomial(weights, num_samples=batch_size, replacement=True)
idx_np = idx.detach().cpu().numpy()
x_resampled = valid_x_final[idx]
sequences = [valid_sequences[i] for i in idx_np]
total_rewards = total_rewards[idx_np]
affinity = affinity[idx_np]
confidence = confidence[idx_np]
direction_predictions = direction_predictions[idx_np]
else:
x_resampled = x_rollout
total_rewards = np.array([])
affinity = np.array([])
confidence = np.array([])
direction_predictions = np.array([])
eval_metrics = {
'affinity': affinity,
'gated_reward': total_rewards,
'confidence': confidence,
'direction_predictions': direction_predictions,
'valid_fraction': valid_fraction
}
if dataframe:
df = pd.DataFrame({
'sequence': sequences if len(total_rewards) else [],
'affinity': affinity,
'gated_reward': total_rewards,
'confidence': confidence
})
return x_resampled, eval_metrics, df
else:
return x_resampled, eval_metrics
# Attach method to model
model.sample_finetuned_td3b = sample_finetuned_td3b.__get__(model, type(model))
return model
|