File size: 22,505 Bytes
388fd6e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 | """
Example demonstrating Gamma SSM's ability to learn copy and reverse tasks.
Copy Task: Given [x₁, x₂, ..., xₗ, ∅, ∅, ..., ∅], predict [∅, ∅, ..., ∅, x₁, x₂, ..., xₗ]
This tests the model's ability to hold information in memory and recall it after a delay.
Reverse Task: Given [x₁, x₂, ..., xₗ], predict [xₗ, xₗ₋₁, ..., x₁]
This tests the model's ability to process sequences bidirectionally.
These are classical synthetic benchmarks for evaluating sequence models.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
import numpy as np
from pathlib import Path
import argparse
from typing import Tuple, Dict, List
import sys
# Add parent directory to path for imports
sys.path.insert(0, str(Path(__file__).parent.parent))
from gamma_space_model.modules.block import GammaSingleBlock
# ============================================================================
# PHASE 1: TASK DATA GENERATION
# ============================================================================
def generate_copy_task(
seq_len: int,
vocab_size: int,
batch_size: int,
device: torch.device,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Generate copy task data.
Input: [token₁, token₂, ..., tokenₗ, 0, 0, ..., 0] (length: 2*seq_len)
Target: [0, 0, ..., 0, token₁, token₂, ..., tokenₗ] (length: 2*seq_len)
Args:
seq_len: Length of the token sequence
vocab_size: Number of possible token values (excluding 0 for padding)
batch_size: Batch size
device: Device to create tensors on
Returns:
inputs: (batch_size, 2*seq_len) tensor of input token indices
targets: (batch_size, 2*seq_len) tensor of target token indices
"""
# Generate random tokens (1 to vocab_size-1; 0 is reserved for padding)
tokens = torch.randint(1, vocab_size, (batch_size, seq_len), device=device)
# Create inputs: [tokens, pad]
padding = torch.zeros((batch_size, seq_len), dtype=torch.long, device=device)
inputs = torch.cat([tokens, padding], dim=1)
# Create targets: [pad, tokens]
targets = torch.cat([padding, tokens], dim=1)
return inputs, targets
def generate_reverse_task(
seq_len: int,
vocab_size: int,
batch_size: int,
device: torch.device,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Generate reverse task data.
Input: [token₁, token₂, ..., tokenₗ]
Target: [tokenₗ, tokenₗ₋₁, ..., token₁]
Args:
seq_len: Length of the token sequence
vocab_size: Number of possible token values (1 to vocab_size-1)
batch_size: Batch size
device: Device to create tensors on
Returns:
inputs: (batch_size, seq_len) tensor of input token indices
targets: (batch_size, seq_len) tensor of target token indices
"""
# Generate random tokens (1 to vocab_size-1)
inputs = torch.randint(1, vocab_size, (batch_size, seq_len), device=device)
# Create targets by reversing along sequence dimension
targets = torch.flip(inputs, dims=[1])
return inputs, targets
class CopyReverseDataset(Dataset):
"""Dataset for copy/reverse tasks."""
def __init__(
self,
task_type: str,
seq_len: int,
vocab_size: int,
num_samples: int,
device: torch.device,
):
"""
Args:
task_type: 'copy' or 'reverse'
seq_len: Sequence length
vocab_size: Number of vocabulary tokens
num_samples: Number of samples to generate
device: Device to generate data on
"""
self.task_type = task_type
self.seq_len = seq_len
self.vocab_size = vocab_size
self.device = device
# Pre-generate all samples
self.inputs = []
self.targets = []
if task_type == 'copy':
gen_fn = generate_copy_task
elif task_type == 'reverse':
gen_fn = generate_reverse_task
else:
raise ValueError(f"Unknown task type: {task_type}")
# Generate samples in batches
batch_size = min(num_samples, 256)
for i in range(0, num_samples, batch_size):
current_batch_size = min(batch_size, num_samples - i)
inputs, targets = gen_fn(seq_len, vocab_size, current_batch_size, device)
self.inputs.append(inputs)
self.targets.append(targets)
self.inputs = torch.cat(self.inputs, dim=0)
self.targets = torch.cat(self.targets, dim=0)
def __len__(self):
return len(self.inputs)
def __getitem__(self, idx):
return self.inputs[idx], self.targets[idx]
# ============================================================================
# PHASE 2: MODEL ARCHITECTURE
# ============================================================================
class CopyReverseModel(nn.Module):
"""
Model for copy/reverse tasks.
Architecture:
input (token indices)
→ embedding (one-hot)
→ GammaSingleBlock (SSM)
→ MLP projection
→ output logits (vocab_size)
"""
def __init__(
self,
vocab_size: int,
d_model: int = 32,
hidden_dim: int = 128,
prenorm: bool = True,
dropout: float = 0.0,
):
"""
Args:
vocab_size: Number of vocabulary tokens
d_model: Model dimension (embedding + SSM input dim)
hidden_dim: SSM hidden state dimension
prenorm: Whether to use prenorm in SSM block
dropout: Dropout rate
"""
super().__init__()
self.vocab_size = vocab_size
self.d_model = d_model
self.hidden_dim = hidden_dim
# Embedding layer converts token indices to d_model dimensional vectors
self.embedding = nn.Embedding(vocab_size, d_model)
# SSM block for sequence processing
self.ssm_block = GammaSingleBlock(
d_model=d_model,
hidden_dim=hidden_dim,
delta_t=0.005,
prenorm=prenorm,
dropout=dropout,
)
# Output MLP projection to vocab logits
self.output_proj = nn.Sequential(
nn.Linear(d_model, d_model),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(d_model, vocab_size),
)
def forward(
self,
token_indices: torch.Tensor,
state: torch.Tensor = None,
mask: torch.Tensor = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Forward pass.
Args:
token_indices: (batch, seq_len) tensor of token indices
state: Optional initial state for SSM
mask: Optional mask for padding
Returns:
logits: (batch, seq_len, vocab_size) output logits
final_state: (batch, hidden_dim) final SSM state
"""
# Embed tokens
embedded = self.embedding(token_indices) # (batch, seq_len, d_model)
# Process through SSM block
ssm_output, final_state = self.ssm_block(embedded, state=state, mask=mask)
# Project to vocabulary
logits = self.output_proj(ssm_output) # (batch, seq_len, vocab_size)
return logits, final_state
# ============================================================================
# PHASE 3: TRAINING & EVALUATION
# ============================================================================
def train_on_task(
model: nn.Module,
train_loader: DataLoader,
val_loader: DataLoader,
task_type: str,
num_epochs: int = 50,
learning_rate: float = 0.001,
device: torch.device = torch.device('cpu'),
verbose: bool = True,
) -> Dict[str, List[float]]:
"""
Train model on copy or reverse task.
Args:
model: CopyReverseModel instance
train_loader: Training data loader
val_loader: Validation data loader
task_type: 'copy' or 'reverse' (for logging)
num_epochs: Number of training epochs
learning_rate: Learning rate for optimizer
device: Device to train on
verbose: Whether to print metrics
Returns:
Dictionary with 'train_loss', 'val_loss', 'train_acc', 'val_acc' lists
"""
model = model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
criterion = nn.CrossEntropyLoss()
history = {
'train_loss': [],
'val_loss': [],
'train_acc': [],
'val_acc': [],
}
for epoch in range(num_epochs):
# Training phase
model.train()
train_loss = 0.0
train_correct = 0
train_total = 0
for batch_inputs, batch_targets in train_loader:
batch_inputs = batch_inputs.to(device)
batch_targets = batch_targets.to(device)
# Forward pass
logits, _ = model(batch_inputs)
# Compute loss (flatten for CrossEntropyLoss)
batch_size, seq_len, vocab_size = logits.shape
loss = criterion(
logits.reshape(-1, vocab_size),
batch_targets.reshape(-1),
)
# Backward pass
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
optimizer.step()
# Metrics
train_loss += loss.item()
predictions = logits.argmax(dim=-1)
train_correct += (predictions == batch_targets).sum().item()
train_total += batch_targets.numel()
train_loss /= len(train_loader)
train_acc = 100 * train_correct / train_total
# Validation phase
model.eval()
val_loss = 0.0
val_correct = 0
val_total = 0
with torch.no_grad():
for batch_inputs, batch_targets in val_loader:
batch_inputs = batch_inputs.to(device)
batch_targets = batch_targets.to(device)
logits, _ = model(batch_inputs)
batch_size, seq_len, vocab_size = logits.shape
loss = criterion(
logits.reshape(-1, vocab_size),
batch_targets.reshape(-1),
)
val_loss += loss.item()
predictions = logits.argmax(dim=-1)
val_correct += (predictions == batch_targets).sum().item()
val_total += batch_targets.numel()
val_loss /= len(val_loader)
val_acc = 100 * val_correct / val_total
history['train_loss'].append(train_loss)
history['val_loss'].append(val_loss)
history['train_acc'].append(train_acc)
history['val_acc'].append(val_acc)
if verbose and (epoch + 1) % 10 == 0:
print(
f"[{task_type.upper()}] Epoch {epoch+1:3d}/{num_epochs} | "
f"Train Loss: {train_loss:.4f} | Train Acc: {train_acc:6.2f}% | "
f"Val Loss: {val_loss:.4f} | Val Acc: {val_acc:6.2f}%"
)
return history
def evaluate_on_task(
model: nn.Module,
test_loader: DataLoader,
task_type: str,
device: torch.device = torch.device('cpu'),
) -> Dict[str, float]:
"""
Evaluate model on test set.
Args:
model: CopyReverseModel instance
test_loader: Test data loader
task_type: 'copy' or 'reverse' (for logging)
device: Device to evaluate on
Returns:
Dictionary with 'loss', 'accuracy', 'per_position_acc' metrics
"""
model = model.to(device)
model.eval()
criterion = nn.CrossEntropyLoss()
total_loss = 0.0
total_correct = 0
total_tokens = 0
# Track per-position accuracy
all_predictions = []
all_targets = []
with torch.no_grad():
for batch_inputs, batch_targets in test_loader:
batch_inputs = batch_inputs.to(device)
batch_targets = batch_targets.to(device)
logits, _ = model(batch_inputs)
batch_size, seq_len, vocab_size = logits.shape
loss = criterion(
logits.reshape(-1, vocab_size),
batch_targets.reshape(-1),
)
total_loss += loss.item()
predictions = logits.argmax(dim=-1)
total_correct += (predictions == batch_targets).sum().item()
total_tokens += batch_targets.numel()
all_predictions.append(predictions.cpu())
all_targets.append(batch_targets.cpu())
avg_loss = total_loss / len(test_loader)
accuracy = 100 * total_correct / total_tokens
# Compute per-position accuracy
all_predictions = torch.cat(all_predictions, dim=0)
all_targets = torch.cat(all_targets, dim=0)
per_pos_correct = (all_predictions == all_targets).float().mean(dim=0)
return {
'loss': avg_loss,
'accuracy': accuracy,
'per_position_acc': per_pos_correct.numpy(),
}
def visualize_predictions(
model: nn.Module,
task_type: str,
seq_len: int,
vocab_size: int,
num_examples: int = 3,
device: torch.device = torch.device('cpu'),
):
"""
Visualize model predictions on sample data.
Args:
model: CopyReverseModel instance
task_type: 'copy' or 'reverse'
seq_len: Sequence length
vocab_size: Vocabulary size
num_examples: Number of examples to show
device: Device to use
"""
model.eval()
if task_type == 'copy':
inputs, targets = generate_copy_task(seq_len, vocab_size, num_examples, device)
else:
inputs, targets = generate_reverse_task(seq_len, vocab_size, num_examples, device)
with torch.no_grad():
logits, _ = model(inputs)
predictions = logits.argmax(dim=-1)
print(f"\n{'='*80}")
print(f"Sample Predictions for {task_type.upper()} Task (seq_len={seq_len}, vocab_size={vocab_size})")
print(f"{'='*80}")
for idx in range(num_examples):
print(f"\nExample {idx + 1}:")
print(f" Input: {inputs[idx].cpu().tolist()}")
print(f" Target: {targets[idx].cpu().tolist()}")
print(f" Predicted: {predictions[idx].cpu().tolist()}")
# Compute accuracy for this example
correct = (predictions[idx] == targets[idx]).sum().item()
acc = 100 * correct / len(targets[idx])
print(f" Accuracy: {acc:.2f}%")
# ============================================================================
# PHASE 4: MAIN ENTRY POINT
# ============================================================================
def main(args):
"""Main training and evaluation script."""
# Setup device
device = torch.device(args.device if torch.cuda.is_available() or args.device == 'cpu' else 'cpu')
print(f"Using device: {device}")
# Create datasets
print(f"\nCreating datasets...")
print(f" Sequence length: {args.seq_len}")
print(f" Vocabulary size: {args.vocab_size}")
print(f" Batch size: {args.batch_size}")
# Copy task
print(f"\n{'='*80}")
print(f"COPY TASK")
print(f"{'='*80}")
copy_train_ds = CopyReverseDataset(
'copy', args.seq_len, args.vocab_size, args.train_samples, device
)
copy_val_ds = CopyReverseDataset(
'copy', args.seq_len, args.vocab_size, args.val_samples, device
)
copy_test_ds = CopyReverseDataset(
'copy', args.seq_len, args.vocab_size, args.test_samples, device
)
copy_train_loader = DataLoader(copy_train_ds, batch_size=args.batch_size, shuffle=True)
copy_val_loader = DataLoader(copy_val_ds, batch_size=args.batch_size, shuffle=False)
copy_test_loader = DataLoader(copy_test_ds, batch_size=args.batch_size, shuffle=False)
# Reverse task
print(f"\n{'='*80}")
print(f"REVERSE TASK")
print(f"{'='*80}")
rev_train_ds = CopyReverseDataset(
'reverse', args.seq_len, args.vocab_size, args.train_samples, device
)
rev_val_ds = CopyReverseDataset(
'reverse', args.seq_len, args.vocab_size, args.val_samples, device
)
rev_test_ds = CopyReverseDataset(
'reverse', args.seq_len, args.vocab_size, args.test_samples, device
)
rev_train_loader = DataLoader(rev_train_ds, batch_size=args.batch_size, shuffle=True)
rev_val_loader = DataLoader(rev_val_ds, batch_size=args.batch_size, shuffle=False)
rev_test_loader = DataLoader(rev_test_ds, batch_size=args.batch_size, shuffle=False)
# Train copy model
print(f"\nTraining COPY task model...")
copy_model = CopyReverseModel(
vocab_size=args.vocab_size,
d_model=args.d_model,
hidden_dim=args.hidden_dim,
prenorm=args.prenorm,
dropout=args.dropout,
)
copy_history = train_on_task(
copy_model,
copy_train_loader,
copy_val_loader,
'copy',
num_epochs=args.num_epochs,
learning_rate=args.lr,
device=device,
verbose=True,
)
# Train reverse model
print(f"\n\nTraining REVERSE task model...")
reverse_model = CopyReverseModel(
vocab_size=args.vocab_size,
d_model=args.d_model,
hidden_dim=args.hidden_dim,
prenorm=args.prenorm,
dropout=args.dropout,
)
rev_history = train_on_task(
reverse_model,
rev_train_loader,
rev_val_loader,
'reverse',
num_epochs=args.num_epochs,
learning_rate=args.lr,
device=device,
verbose=True,
)
# Evaluate on test sets
print(f"\n\n{'='*80}")
print(f"EVALUATION ON TEST SET")
print(f"{'='*80}")
copy_eval = evaluate_on_task(copy_model, copy_test_loader, 'copy', device)
print(f"\nCOPY Task Test Results:")
print(f" Loss: {copy_eval['loss']:.4f}")
print(f" Accuracy: {copy_eval['accuracy']:.2f}%")
rev_eval = evaluate_on_task(reverse_model, rev_test_loader, 'reverse', device)
print(f"\nREVERSE Task Test Results:")
print(f" Loss: {rev_eval['loss']:.4f}")
print(f" Accuracy: {rev_eval['accuracy']:.2f}%")
# Print comparison
print(f"\n{'='*80}")
print(f"COMPARISON")
print(f"{'='*80}")
print(f"Copy accuracy is {'higher' if copy_eval['accuracy'] > rev_eval['accuracy'] else 'lower'} than reverse")
print(f"Difference: {abs(copy_eval['accuracy'] - rev_eval['accuracy']):.2f} percentage points")
# Visualize predictions
if args.visualize:
visualize_predictions(copy_model, 'copy', args.seq_len, args.vocab_size, device=device)
visualize_predictions(reverse_model, 'reverse', args.seq_len, args.vocab_size, device=device)
print(f"\n{'='*80}")
print(f"Training complete!")
print(f"{'='*80}\n")
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='Test Gamma SSM on copy and reverse synthetic tasks.'
)
# Task parameters
parser.add_argument(
'--seq-len',
type=int,
default=20,
help='Sequence length for tasks (default: 20)',
)
parser.add_argument(
'--vocab-size',
type=int,
default=8,
help='Vocabulary size (including 0 for padding, default: 8)',
)
# Dataset parameters
parser.add_argument(
'--train-samples',
type=int,
default=500,
help='Number of training samples (default: 500)',
)
parser.add_argument(
'--val-samples',
type=int,
default=100,
help='Number of validation samples (default: 100)',
)
parser.add_argument(
'--test-samples',
type=int,
default=100,
help='Number of test samples (default: 100)',
)
parser.add_argument(
'--batch-size',
type=int,
default=32,
help='Batch size (default: 32)',
)
# Model parameters
parser.add_argument(
'--d-model',
type=int,
default=32,
help='Model dimension (default: 32)',
)
parser.add_argument(
'--hidden-dim',
type=int,
default=128,
help='SSM hidden dimension (default: 128)',
)
parser.add_argument(
'--prenorm',
type=bool,
default=True,
help='Use prenorm in SSM block (default: True)',
)
parser.add_argument(
'--dropout',
type=float,
default=0.0,
help='Dropout rate (default: 0.0)',
)
# Training parameters
parser.add_argument(
'--num-epochs',
type=int,
default=500,
help='Number of training epochs (default: 500)',
)
parser.add_argument(
'--lr',
type=float,
default=0.001,
help='Learning rate (default: 0.001)',
)
# Other parameters
parser.add_argument(
'--device',
type=str,
default='cuda',
choices=['cpu', 'cuda'],
help='Device to use (default: cuda)',
)
parser.add_argument(
'--visualize',
action='store_true',
help='Visualize predictions on sample data',
)
args = parser.parse_args()
main(args)
|