Spaces:
Sleeping
Sleeping
File size: 50,738 Bytes
af4939b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 | # ---
# jupyter:
# jupytext:
# text_representation:
# extension: .py
# format_name: percent
# format_version: '1.3'
# jupytext_version: 1.17.1
# kernelspec:
# display_name: Python 3 (ipykernel)
# language: python
# name: python3
# ---
# %% [markdown]
"""
# Module 08: Training - Complete Learning Loops
Welcome to Module 08! You're about to build the complete training infrastructure that brings neural networks to life through end-to-end learning.
## π Prerequisites & Progress
**You've Built**: Tensors, activations, layers, losses, DataLoader, gradients, and optimizers
**You'll Build**: Complete training loops with checkpointing, scheduling, and gradient management
**You'll Enable**: Full model training pipeline for the MLP milestone
**Connection Map**:
```
DataLoader β Autograd β Optimizers β Training β Convolutions
(Module 05) (Module 06) (Module 07) (Module 08) (Module 09)
```
## π― Learning Objectives
By the end of this module, you will:
1. Implement a complete Trainer class with train/eval modes
2. Build learning rate scheduling and gradient clipping
3. Create checkpointing for model persistence
4. Test training loops with immediate validation
5. Understand gradient accumulation patterns
Let's get started!
## π¦ Where This Code Lives in the Final Package
**Learning Side:** You work in `modules/08_training/training_dev.py`
**Building Side:** Code exports to `tinytorch.core.training`
```python
# How to use this module:
from tinytorch.core.training import Trainer, CosineSchedule, clip_grad_norm
```
**Why this matters:**
- **Learning:** Complete training system in one focused module for deep understanding
- **Production:** Proper organization like PyTorch's training infrastructure with all training components together
- **Consistency:** All training operations and scheduling functionality in core.training
- **Integration:** Works seamlessly with optimizers and losses for complete learning pipelines
"""
# %% nbgrader={"grade": false, "grade_id": "imports", "locked": false, "solution": false}
#| default_exp core.training
#| export
import numpy as np
import pickle
import time
from typing import Dict, List, Optional, Tuple, Any, Callable
from pathlib import Path
import sys
import os
# Import dependencies from other modules
from tinytorch.core.tensor import Tensor
from tinytorch.core.layers import Linear
from tinytorch.core.losses import MSELoss, CrossEntropyLoss
from tinytorch.core.optimizers import SGD, AdamW
# Constants for learning rate scheduling defaults
DEFAULT_MAX_LR = 0.1 # Default maximum learning rate for cosine schedule
DEFAULT_MIN_LR = 0.01 # Default minimum learning rate for cosine schedule
DEFAULT_TOTAL_EPOCHS = 100 # Default total epochs for learning rate schedule
# %% [markdown]
"""
## π‘ Introduction - What is Training?
Training is where the magic happens - it's the process that transforms a randomly initialized neural network into an intelligent system that can solve problems. Think of training as teaching: you show the model examples, it makes predictions, you measure how wrong it is, and then you adjust its parameters to do better next time.
The training process follows a consistent pattern across all machine learning:
1. **Forward Pass**: Input flows through the model to produce predictions
2. **Loss Calculation**: Compare predictions to true answers
3. **Backward Pass**: Compute gradients showing how to improve
4. **Parameter Update**: Adjust model weights using an optimizer
5. **Repeat**: Continue until the model learns the pattern
But production training systems need much more than this basic loop. They need learning rate scheduling (starting fast, slowing down), gradient clipping (preventing exploding gradients), checkpointing (saving progress), and evaluation modes (testing without learning).
**What we're building today:**
- A complete `Trainer` class that orchestrates the entire learning process
- Learning rate scheduling that adapts during training
- Gradient clipping that prevents training instability
- Checkpointing system for saving and resuming training
- Train/eval modes for proper model behavior
"""
# %% [markdown]
"""
## π Foundations - Mathematical Background
### Training Loop Mathematics
The core training loop implements gradient descent with sophisticated improvements:
**Basic Update Rule:**
```
ΞΈ(t+1) = ΞΈ(t) - Ξ· βL(ΞΈ(t))
```
Where ΞΈ are parameters, Ξ· is learning rate, and βL is the loss gradient.
**Learning Rate Scheduling:**
For cosine annealing over T epochs:
```
Ξ·(t) = Ξ·_min + (Ξ·_max - Ξ·_min) * (1 + cos(Οt/T)) / 2
```
**Gradient Clipping:**
When ||βL|| > max_norm, rescale:
```
βL β βL * max_norm / ||βL||
```
**Gradient Accumulation:**
For effective batch size B_eff = accumulation_steps * B_actual:
```
βL_accumulated = (1/accumulation_steps) * Ξ£ βL_batch_i
```
### Train vs Eval Modes
Many layers behave differently during training vs inference:
- **Dropout**: Active during training, disabled during evaluation
- **BatchNorm**: Updates statistics during training, uses fixed statistics during evaluation
- **Gradient computation**: Enabled during training, disabled during evaluation for efficiency
This mode switching is crucial for proper model behavior and performance.
"""
# %% [markdown]
"""
## ποΈ Implementation - Building Training Infrastructure
Now let's implement the complete training system. We'll build each component step by step: learning rate scheduling, gradient utilities, and finally the complete Trainer class.
Each component will follow the pattern: **Explanation β Implementation β Test** so you understand what you're building before you build it.
"""
# %% [markdown]
"""
### Learning Rate Scheduling - Adaptive Training Speed
Learning rate scheduling is like adjusting your driving speed based on road conditions. You start fast on the highway (high learning rate for quick progress), then slow down in neighborhoods (low learning rate for fine-tuning).
#### Why Cosine Scheduling Works
Cosine annealing follows a smooth curve that provides:
- **Aggressive learning initially** - Fast convergence when far from optimum
- **Gradual slowdown** - Stable convergence as you approach the solution
- **Smooth transitions** - No sudden learning rate drops that shock the model
#### The Mathematics
Cosine annealing uses the cosine function to smoothly transition from max_lr to min_lr:
```
Learning Rate Schedule:
max_lr ββ\
β \
β \
β \
β \
min_lr ββββββββββββ\ββββββββ
0 25 50 75 100 epochs
Formula: lr = min_lr + (max_lr - min_lr) * (1 + cos(Ο * epoch / total_epochs)) / 2
```
This creates a natural learning curve that adapts training speed to the optimization landscape.
"""
# %% nbgrader={"grade": false, "grade_id": "scheduler", "locked": false, "solution": true}
#| export
class CosineSchedule:
"""
Cosine annealing learning rate schedule.
Starts at max_lr, decreases following a cosine curve to min_lr over T epochs.
This provides aggressive learning initially, then fine-tuning at the end.
TODO: Implement cosine annealing schedule
APPROACH:
1. Store max_lr, min_lr, and total_epochs
2. In get_lr(), compute cosine factor: (1 + cos(Ο * epoch / total_epochs)) / 2
3. Interpolate: min_lr + (max_lr - min_lr) * cosine_factor
EXAMPLE:
>>> schedule = CosineSchedule(max_lr=0.1, min_lr=0.01, total_epochs=100)
>>> print(schedule.get_lr(0)) # Start: 0.1
>>> print(schedule.get_lr(50)) # Middle: ~0.055
>>> print(schedule.get_lr(100)) # End: 0.01
HINT: Use np.cos() and np.pi for the cosine calculation
"""
### BEGIN SOLUTION
def __init__(self, max_lr: float = DEFAULT_MAX_LR, min_lr: float = DEFAULT_MIN_LR, total_epochs: int = DEFAULT_TOTAL_EPOCHS):
self.max_lr = max_lr
self.min_lr = min_lr
self.total_epochs = total_epochs
def get_lr(self, epoch: int) -> float:
"""Get learning rate for current epoch."""
if epoch >= self.total_epochs:
return self.min_lr
# Cosine annealing formula
cosine_factor = (1 + np.cos(np.pi * epoch / self.total_epochs)) / 2
return self.min_lr + (self.max_lr - self.min_lr) * cosine_factor
### END SOLUTION
# %% [markdown]
"""
### π§ͺ Unit Test: CosineSchedule
This test validates our learning rate scheduling implementation.
**What we're testing**: Cosine annealing produces correct learning rates
**Why it matters**: Proper scheduling often makes the difference between convergence and failure
**Expected**: Smooth decrease from max_lr to min_lr following cosine curve
"""
# %% nbgrader={"grade": true, "grade_id": "test_scheduler", "locked": true, "points": 10}
def test_unit_cosine_schedule():
"""π¬ Test CosineSchedule implementation."""
print("π¬ Unit Test: CosineSchedule...")
# Test basic schedule
schedule = CosineSchedule(max_lr=0.1, min_lr=0.01, total_epochs=100)
# Test start, middle, and end
lr_start = schedule.get_lr(0)
lr_middle = schedule.get_lr(50)
lr_end = schedule.get_lr(100)
print(f"Learning rate at epoch 0: {lr_start:.4f}")
print(f"Learning rate at epoch 50: {lr_middle:.4f}")
print(f"Learning rate at epoch 100: {lr_end:.4f}")
# Validate behavior
assert abs(lr_start - 0.1) < 1e-6, f"Expected 0.1 at start, got {lr_start}"
assert abs(lr_end - 0.01) < 1e-6, f"Expected 0.01 at end, got {lr_end}"
assert 0.01 < lr_middle < 0.1, f"Middle LR should be between min and max, got {lr_middle}"
# Test monotonic decrease in first half
lr_quarter = schedule.get_lr(25)
assert lr_quarter > lr_middle, "LR should decrease monotonically in first half"
print("β
CosineSchedule works correctly!")
if __name__ == "__main__":
test_unit_cosine_schedule()
# %% [markdown]
"""
### Gradient Clipping - Preventing Training Explosions
Gradient clipping is like having a speed governor on your car - it prevents dangerous situations where gradients become so large they destroy training progress.
#### The Problem: Exploding Gradients
During training, gradients can sometimes become extremely large, causing:
- **Parameter updates that are too big** - Model jumps far from the optimal solution
- **Numerical instability** - Values become NaN or infinite
- **Training collapse** - Model performance suddenly degrades
#### The Solution: Global Norm Clipping
Instead of clipping each gradient individually, we compute the global norm across all parameters and scale uniformly:
```
Gradient Clipping Process:
1. Compute Global Norm:
total_norm = β(sum of all gradient squares)
2. Check if Clipping Needed:
if total_norm > max_norm:
clip_coefficient = max_norm / total_norm
3. Scale All Gradients:
for each gradient:
gradient *= clip_coefficient
Visualization:
Original Gradients: [100, 200, 50] β norm = 230
With max_norm=1.0: [0.43, 0.87, 0.22] β norm = 1.0
```
This preserves the relative magnitudes while preventing explosion.
"""
# %% nbgrader={"grade": false, "grade_id": "gradient_clipping", "locked": false, "solution": true}
#| export
def clip_grad_norm(parameters: List, max_norm: float = 1.0) -> float:
"""
Clip gradients by global norm to prevent exploding gradients.
This is crucial for training stability, especially with RNNs and deep networks.
Instead of clipping each gradient individually, we compute the global norm
across all parameters and scale uniformly if needed.
TODO: Implement gradient clipping by global norm
APPROACH:
1. Compute total norm: sqrt(sum of squared gradients across all parameters)
2. If total_norm > max_norm, compute clip_coef = max_norm / total_norm
3. Scale all gradients by clip_coef: grad *= clip_coef
4. Return the original norm for monitoring
EXAMPLE:
>>> params = [Tensor([1, 2, 3], requires_grad=True)]
>>> params[0].grad = Tensor([10, 20, 30]) # Large gradients
>>> original_norm = clip_grad_norm(params, max_norm=1.0)
>>> print(f"Clipped norm: {np.linalg.norm(params[0].grad.data):.2f}") # Should be β€ 1.0
HINTS:
- Use np.linalg.norm() to compute norms
- Only clip if total_norm > max_norm
- Modify gradients in-place for efficiency
"""
### BEGIN SOLUTION
if not parameters:
return 0.0
# Collect all gradients and compute global norm
total_norm = 0.0
for param in parameters:
if param.grad is not None:
# Handle both Tensor gradients and numpy array gradients
if isinstance(param.grad, np.ndarray):
grad_data = param.grad
else:
# Trust that Tensor has .data attribute
grad_data = param.grad.data
total_norm += np.sum(grad_data ** 2)
total_norm = np.sqrt(total_norm)
# Clip if necessary
if total_norm > max_norm:
clip_coef = max_norm / total_norm
for param in parameters:
if param.grad is not None:
# Handle both Tensor gradients and numpy array gradients
if isinstance(param.grad, np.ndarray):
param.grad = param.grad * clip_coef
else:
# Trust that Tensor has .data attribute
param.grad.data = param.grad.data * clip_coef
return float(total_norm)
### END SOLUTION
# %% [markdown]
"""
### π§ͺ Unit Test: Gradient Clipping
This test validates our gradient clipping implementation.
**What we're testing**: Global norm clipping properly rescales large gradients
**Why it matters**: Prevents exploding gradients that can destroy training
**Expected**: Gradients scaled down when norm exceeds threshold
"""
# %% nbgrader={"grade": true, "grade_id": "test_clipping", "locked": true, "points": 10}
def test_unit_clip_grad_norm():
"""π¬ Test clip_grad_norm implementation."""
print("π¬ Unit Test: Gradient Clipping...")
# Use real Tensor from Module 01
import sys
# Tensor already imported at module level
# Test case 1: Large gradients that need clipping
param1 = Tensor([1.0, 2.0], requires_grad=True)
param1.grad = np.array([3.0, 4.0]) # norm = 5.0
param2 = Tensor([3.0, 4.0], requires_grad=True)
param2.grad = np.array([6.0, 8.0]) # norm = 10.0
params = [param1, param2]
# Total norm = sqrt(5Β² + 10Β²) = sqrt(125) β 11.18
original_norm = clip_grad_norm(params, max_norm=1.0)
# Check original norm was large
assert original_norm > 1.0, f"Original norm should be > 1.0, got {original_norm}"
# Check gradients were clipped
new_norm = 0.0
for param in params:
if isinstance(param.grad, np.ndarray):
grad_data = param.grad
else:
# Trust that Tensor has .data attribute
grad_data = param.grad.data
new_norm += np.sum(grad_data ** 2)
new_norm = np.sqrt(new_norm)
print(f"Original norm: {original_norm:.2f}")
print(f"Clipped norm: {new_norm:.2f}")
assert abs(new_norm - 1.0) < 1e-6, f"Clipped norm should be 1.0, got {new_norm}"
# Test case 2: Small gradients that don't need clipping
small_param = Tensor([1.0, 2.0], requires_grad=True)
small_param.grad = np.array([0.1, 0.2])
small_params = [small_param]
original_small = clip_grad_norm(small_params, max_norm=1.0)
assert original_small < 1.0, "Small gradients shouldn't be clipped"
print("β
Gradient clipping works correctly!")
if __name__ == "__main__":
test_unit_clip_grad_norm()
# %% [markdown]
"""
### The Trainer Class - Orchestrating Complete Training
The Trainer class is like a conductor orchestrating a symphony - it coordinates all the components (model, optimizer, loss function, scheduler) to create beautiful music (successful training).
#### Training Loop Architecture
The training loop follows a consistent pattern across all machine learning:
```
Training Loop Structure:
for epoch in range(num_epochs):
ββββββββββββββββββββ TRAINING PHASE ββββββββββββββββββββ
β β
β for batch in dataloader: β
β ββββ Forward Pass ββββββββ β
β β 1. input β model β β
β β 2. predictions β β
β ββββββββββββββββββββββββββ β
β β β
β ββββ Loss Computation ββββ β
β β 3. loss = loss_fn() β β
β ββββββββββββββββββββββββββ β
β β β
β ββββ Backward Pass βββββββ β
β β 4. loss.backward() β β
β β 5. gradients β β
β ββββββββββββββββββββββββββ β
β β β
β ββββ Parameter Update ββββ β
β β 6. optimizer.step() β β
β β 7. zero gradients β β
β ββββββββββββββββββββββββββ β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
ββββ Learning Rate Update ββββ
β 8. scheduler.step() β
ββββββββββββββββββββββββββββββ
```
#### Key Features
- **Train/Eval Modes**: Different behavior during training vs evaluation
- **Gradient Accumulation**: Effective larger batch sizes with limited memory
- **Checkpointing**: Save/resume training state for long experiments
- **Progress Tracking**: Monitor loss, learning rate, and other metrics
"""
# %% nbgrader={"grade": false, "grade_id": "trainer_class", "locked": false, "solution": true}
#| export
class Trainer:
"""
Complete training orchestrator for neural networks.
Handles the full training lifecycle: forward pass, loss computation,
backward pass, optimization, scheduling, checkpointing, and evaluation.
This is the central class that brings together all the components
you've built in previous modules.
TODO: Implement complete Trainer class
APPROACH:
1. __init__(): Store model, optimizer, loss_fn, scheduler, and grad_clip_norm
2. train_epoch(): Loop through dataloader, forward β loss β backward β step
3. evaluate(): Similar loop but set model.training=False, no grad updates
4. save/load_checkpoint(): Use pickle to persist/restore all training state
EXAMPLE:
>>> model = SimpleModel()
>>> optimizer = SGD(model.parameters(), lr=0.01)
>>> trainer = Trainer(model, optimizer, MSELoss())
>>> # Training data: list of (input, target) tuples
>>> data = [(Tensor([[1.0]]), Tensor([[2.0]]))]
>>> loss = trainer.train_epoch(data)
>>> eval_loss, accuracy = trainer.evaluate(data)
>>> trainer.save_checkpoint('/tmp/checkpoint.pkl')
HINTS:
- In train_epoch(), set model.training = True at start
- For gradient accumulation, scale loss by 1/accumulation_steps
- Use clip_grad_norm() before optimizer.step() if grad_clip_norm is set
- Update scheduler after each epoch: optimizer.lr = scheduler.get_lr(epoch)
- In evaluate(), set model.training = False and don't update gradients
- Checkpoints should include: epoch, step, model state, optimizer state, scheduler state, history
"""
### BEGIN SOLUTION
def __init__(self, model, optimizer, loss_fn, scheduler=None, grad_clip_norm=None):
"""
Initialize trainer with model and training components.
Args:
model: Neural network to train
optimizer: Parameter update strategy (SGD, Adam, etc.)
loss_fn: Loss function (CrossEntropy, MSE, etc.)
scheduler: Optional learning rate scheduler
grad_clip_norm: Optional gradient clipping threshold
"""
self.model = model
self.optimizer = optimizer
self.loss_fn = loss_fn
self.scheduler = scheduler
self.grad_clip_norm = grad_clip_norm
# Training state
self.epoch = 0
self.step = 0
self.training_mode = True
# History tracking
self.history = {
'train_loss': [],
'eval_loss': [],
'learning_rates': []
}
def train_epoch(self, dataloader, accumulation_steps=1):
"""
Train for one epoch through the dataset.
Args:
dataloader: Iterable yielding (inputs, targets) batches
accumulation_steps: Number of batches to accumulate before update
Returns:
Average loss for the epoch
"""
self.model.training = True
self.training_mode = True
total_loss = 0.0
num_batches = 0
accumulated_loss = 0.0
for batch_idx, (inputs, targets) in enumerate(dataloader):
# Forward pass
outputs = self.model.forward(inputs)
loss = self.loss_fn.forward(outputs, targets)
# Scale loss for accumulation
scaled_loss = loss.data / accumulation_steps
accumulated_loss += scaled_loss
# Backward pass
loss.backward()
# Update parameters every accumulation_steps
if (batch_idx + 1) % accumulation_steps == 0:
# Gradient clipping
if self.grad_clip_norm is not None:
params = self.model.parameters()
clip_grad_norm(params, self.grad_clip_norm)
# Optimizer step
self.optimizer.step()
self.optimizer.zero_grad()
total_loss += accumulated_loss
accumulated_loss = 0.0
num_batches += 1
self.step += 1
# Handle remaining accumulated gradients
if accumulated_loss > 0:
if self.grad_clip_norm is not None:
params = self.model.parameters()
clip_grad_norm(params, self.grad_clip_norm)
self.optimizer.step()
self.optimizer.zero_grad()
total_loss += accumulated_loss
num_batches += 1
avg_loss = total_loss / max(num_batches, 1)
self.history['train_loss'].append(avg_loss)
# Update scheduler
if self.scheduler is not None:
current_lr = self.scheduler.get_lr(self.epoch)
# Update optimizer learning rate (trust it has lr attribute)
self.optimizer.lr = current_lr
self.history['learning_rates'].append(current_lr)
self.epoch += 1
return avg_loss
def evaluate(self, dataloader):
"""
Evaluate model on dataset without updating parameters.
Args:
dataloader: Iterable yielding (inputs, targets) batches
Returns:
Average loss and accuracy
"""
self.model.training = False
self.training_mode = False
total_loss = 0.0
correct = 0
total = 0
for inputs, targets in dataloader:
# Forward pass only
outputs = self.model.forward(inputs)
loss = self.loss_fn.forward(outputs, targets)
total_loss += loss.data
# Calculate accuracy (for classification)
# Trust that Tensors have .data attribute
if len(outputs.data.shape) > 1: # Multi-class
predictions = np.argmax(outputs.data, axis=1)
if len(targets.data.shape) == 1: # Integer targets
correct += np.sum(predictions == targets.data)
else: # One-hot targets
correct += np.sum(predictions == np.argmax(targets.data, axis=1))
total += len(predictions)
avg_loss = total_loss / len(dataloader) if len(dataloader) > 0 else 0.0
accuracy = correct / total if total > 0 else 0.0
self.history['eval_loss'].append(avg_loss)
return avg_loss, accuracy
def save_checkpoint(self, path: str):
"""
Save complete training state for resumption.
Args:
path: File path to save checkpoint
"""
checkpoint = {
'epoch': self.epoch,
'step': self.step,
'model_state': self._get_model_state(),
'optimizer_state': self._get_optimizer_state(),
'scheduler_state': self._get_scheduler_state(),
'history': self.history,
'training_mode': self.training_mode
}
Path(path).parent.mkdir(parents=True, exist_ok=True)
with open(path, 'wb') as f:
pickle.dump(checkpoint, f)
def load_checkpoint(self, path: str):
"""
Load training state from checkpoint.
Args:
path: File path to load checkpoint from
"""
with open(path, 'rb') as f:
checkpoint = pickle.load(f)
self.epoch = checkpoint['epoch']
self.step = checkpoint['step']
self.history = checkpoint['history']
self.training_mode = checkpoint['training_mode']
# Restore states (simplified for educational purposes)
if 'model_state' in checkpoint:
self._set_model_state(checkpoint['model_state'])
if 'optimizer_state' in checkpoint:
self._set_optimizer_state(checkpoint['optimizer_state'])
if 'scheduler_state' in checkpoint:
self._set_scheduler_state(checkpoint['scheduler_state'])
def _get_model_state(self):
"""Extract model parameters for checkpointing."""
# Trust model has parameters() method
return {i: param.data.copy() for i, param in enumerate(self.model.parameters())}
def _set_model_state(self, state):
"""Restore model parameters from checkpoint."""
# Trust model has parameters() method
for i, param in enumerate(self.model.parameters()):
if i in state:
param.data = state[i].copy()
def _get_optimizer_state(self):
"""Extract optimizer state for checkpointing."""
state = {}
# Trust optimizer has lr attribute (from Modules 06)
state['lr'] = self.optimizer.lr
# Use explicit API for momentum state (Module 07)
# All optimizers with momentum support have get_momentum_state() method
if hasattr(self.optimizer, 'has_momentum') and self.optimizer.has_momentum():
momentum_state = self.optimizer.get_momentum_state()
if momentum_state is not None:
state['momentum_buffers'] = momentum_state
return state
def _set_optimizer_state(self, state):
"""Restore optimizer state from checkpoint."""
if 'lr' in state:
# Trust optimizer has lr attribute (from Modules 06)
self.optimizer.lr = state['lr']
# Use explicit API for momentum state (Module 07)
# All optimizers with momentum support have set_momentum_state() method
if 'momentum_buffers' in state:
if hasattr(self.optimizer, 'has_momentum') and self.optimizer.has_momentum():
self.optimizer.set_momentum_state(state['momentum_buffers'])
def _get_scheduler_state(self):
"""Extract scheduler state for checkpointing."""
if self.scheduler is None:
return None
return {
'max_lr': getattr(self.scheduler, 'max_lr', None),
'min_lr': getattr(self.scheduler, 'min_lr', None),
'total_epochs': getattr(self.scheduler, 'total_epochs', None)
}
def _set_scheduler_state(self, state):
"""Restore scheduler state from checkpoint."""
if state is None or self.scheduler is None:
return
# Educational Note: hasattr() is legitimate here because:
# 1. Schedulers are user-extensible with custom attributes
# 2. State dict may have keys from different scheduler types
# 3. We safely skip attributes that don't exist on current scheduler
# This is duck-typing for polymorphic checkpoint restoration
for key, value in state.items():
if hasattr(self.scheduler, key):
setattr(self.scheduler, key, value)
### END SOLUTION
# %% [markdown]
"""
### π§ͺ Unit Test: Trainer Class
This test validates our complete training system.
**What we're testing**: Trainer orchestrates training loop correctly
**Why it matters**: This is the backbone that enables all neural network training
**Expected**: Training reduces loss, evaluation works, checkpointing preserves state
"""
# %% nbgrader={"grade": true, "grade_id": "test_trainer", "locked": true, "points": 15}
def test_unit_trainer():
"""π¬ Test Trainer implementation."""
print("π¬ Unit Test: Trainer...")
# Use REAL components from previous modules (already imported at module level)
# Create a simple model using REAL Linear layer
class SimpleModel:
def __init__(self):
self.layer = Linear(2, 1) # Real Linear from Module 03
self.training = True
def forward(self, x):
return self.layer.forward(x)
def parameters(self):
return self.layer.parameters()
# Create trainer with REAL components
model = SimpleModel()
optimizer = SGD(model.parameters(), lr=0.01) # Real SGD from Module 07
loss_fn = MSELoss() # Real MSELoss from Module 04
scheduler = CosineSchedule(max_lr=0.1, min_lr=0.01, total_epochs=10)
trainer = Trainer(model, optimizer, loss_fn, scheduler, grad_clip_norm=1.0)
# Test training
print("Testing training epoch...")
# Use real Tensors for data
dataloader = [
(Tensor([[1.0, 0.5]]), Tensor([[2.0]])),
(Tensor([[0.5, 1.0]]), Tensor([[1.5]]))
]
loss = trainer.train_epoch(dataloader)
assert isinstance(loss, (float, np.floating)), f"Expected float loss, got {type(loss)}"
assert trainer.epoch == 1, f"Expected epoch 1, got {trainer.epoch}"
# Test evaluation
print("Testing evaluation...")
eval_loss, accuracy = trainer.evaluate(dataloader)
assert isinstance(eval_loss, (float, np.floating)), f"Expected float eval_loss, got {type(eval_loss)}"
assert isinstance(accuracy, (float, np.floating)), f"Expected float accuracy, got {type(accuracy)}"
# Test checkpointing
print("Testing checkpointing...")
checkpoint_path = "/tmp/test_checkpoint.pkl"
trainer.save_checkpoint(checkpoint_path)
# Modify trainer state
original_epoch = trainer.epoch
trainer.epoch = 999
# Load checkpoint
trainer.load_checkpoint(checkpoint_path)
assert trainer.epoch == original_epoch, f"Checkpoint didn't restore epoch correctly"
# Clean up
import os
if os.path.exists(checkpoint_path):
os.remove(checkpoint_path)
print(f"β
Trainer works correctly! Final loss: {loss:.4f}")
if __name__ == "__main__":
test_unit_trainer()
# %% [markdown]
"""
## π§ Integration - Complete Training Example
Now let's create a complete training example that demonstrates how all the components work together. This integration shows the full power of our training infrastructure.
### Building a Complete Training Pipeline
```
Training Pipeline Architecture:
Model Creation
β
Optimizer Setup (with parameters)
β
Loss Function Selection
β
Learning Rate Scheduler
β
Trainer Initialization
β
Training Loop (multiple epochs)
β
Evaluation & Checkpointing
```
This example brings together everything you've built in Modules 01-07.
"""
# %% nbgrader={"grade": false, "grade_id": "integration_example", "solution": true}
def demonstrate_complete_training_pipeline():
"""
Complete end-to-end training example using all components.
This demonstrates how Trainer, scheduler, gradient clipping,
and checkpointing work together in a real training scenario.
"""
print("ποΈ Building Complete Training Pipeline...")
print("=" * 60)
# Step 1: Create model using REAL Linear layer
class SimpleNN:
def __init__(self):
self.layer1 = Linear(3, 5)
self.layer2 = Linear(5, 2)
self.training = True
def forward(self, x):
x = self.layer1.forward(x)
# Simple ReLU-like activation (max with 0)
x = Tensor(np.maximum(0, x.data))
x = self.layer2.forward(x)
return x
def parameters(self):
return self.layer1.parameters() + self.layer2.parameters()
print("β Model created: 3 β 5 β 2 network")
# Step 2: Create optimizer
model = SimpleNN()
optimizer = SGD(model.parameters(), lr=0.1, momentum=0.9)
print("β Optimizer: SGD with momentum")
# Step 3: Create loss function
loss_fn = MSELoss()
print("β Loss function: MSE")
# Step 4: Create scheduler
scheduler = CosineSchedule(max_lr=0.1, min_lr=0.001, total_epochs=5)
print("β Scheduler: Cosine annealing (0.1 β 0.001)")
# Step 5: Create trainer with gradient clipping
trainer = Trainer(
model=model,
optimizer=optimizer,
loss_fn=loss_fn,
scheduler=scheduler,
grad_clip_norm=1.0
)
print("β Trainer initialized with gradient clipping")
# Step 6: Create synthetic training data
train_data = [
(Tensor(np.random.randn(4, 3)), Tensor(np.random.randn(4, 2))),
(Tensor(np.random.randn(4, 3)), Tensor(np.random.randn(4, 2))),
(Tensor(np.random.randn(4, 3)), Tensor(np.random.randn(4, 2)))
]
print("β Training data: 3 batches of 4 samples")
# Step 7: Train for multiple epochs
print("\nπ Starting Training...")
print("-" * 60)
print(f"{'Epoch':<8} {'Train Loss':<12} {'Learning Rate':<15}")
print("-" * 60)
for epoch in range(3):
loss = trainer.train_epoch(train_data)
lr = scheduler.get_lr(epoch)
print(f"{epoch:<8} {loss:<12.6f} {lr:<15.6f}")
# Step 8: Save checkpoint
checkpoint_path = "/tmp/training_example_checkpoint.pkl"
trainer.save_checkpoint(checkpoint_path)
print(f"\nβ Checkpoint saved: {checkpoint_path}")
# Step 9: Evaluate
eval_loss, accuracy = trainer.evaluate(train_data)
print(f"β Evaluation - Loss: {eval_loss:.6f}, Accuracy: {accuracy:.6f}")
# Clean up
import os
if os.path.exists(checkpoint_path):
os.remove(checkpoint_path)
print("\n" + "=" * 60)
print("β
Complete training pipeline executed successfully!")
print("\nπ‘ This pipeline demonstrates:")
print(" β’ Model β Optimizer β Loss β Scheduler β Trainer integration")
print(" β’ Training loop with scheduling and gradient clipping")
print(" β’ Checkpointing for training persistence")
print(" β’ Evaluation mode for model assessment")
# %% [markdown]
"""
## π Systems Analysis - Training Performance and Memory
Training systems have significant resource requirements. Understanding memory usage, checkpoint sizes, and training overhead helps optimize production ML pipelines.
### Training Memory Breakdown
```
Training Memory Requirements:
Forward Pass Memory:
βββββββββββββββββββ
β Activations β β Stored for backward pass
βββββββββββββββββββ€
β Model Params β β Network weights
βββββββββββββββββββ
Backward Pass Memory:
βββββββββββββββββββ
β Gradients β β Same size as params
βββββββββββββββββββ€
β Optimizer State β β 2-3Γ params (momentum, Adam buffers)
βββββββββββββββββββ
Checkpoint Memory:
βββββββββββββββββββ
β Model State β β Full parameter snapshot
βββββββββββββββββββ€
β Optimizer State β β All momentum/Adam buffers
βββββββββββββββββββ€
β Training Meta β β Epoch, history, scheduler
βββββββββββββββββββ
Total Training Memory β 5-6Γ Model Parameters
```
### Key Systems Insights
**Gradient Accumulation Trade-off:**
- Effective batch size = accumulation_steps Γ actual_batch_size
- Memory: Fixed (only 1 batch in memory at a time)
- Time: Increases linearly with accumulation steps
- Use case: Large models that don't fit with desired batch size
**Checkpoint Size:**
- Base model: 1Γ parameters
- With optimizer (Adam): ~3Γ parameters
- With full history: Additional metadata
- Compression: Pickle overhead ~10-20%
"""
# %% nbgrader={"grade": false, "grade_id": "analyze_training_memory", "solution": true}
def analyze_training_memory():
"""π Analyze memory overhead of training components."""
print("π Analyzing Training Memory Overhead...")
# Create models of different sizes
model_sizes = [
("Small", 100), # 100 parameters
("Medium", 1000), # 1K parameters
("Large", 10000) # 10K parameters
]
print("\nTraining Memory Analysis:")
print("=" * 90)
print(f"{'Model':<10} {'Params':<10} {'Gradients':<12} {'SGD State':<12} {'Adam State':<12} {'Total':<10}")
print("-" * 90)
for name, param_count in model_sizes:
# Base memory: parameters
param_memory = param_count * 4 # 4 bytes per float32
# Gradients: same as parameters
grad_memory = param_count * 4
# SGD optimizer state: momentum buffer
sgd_memory = param_count * 4
# Adam optimizer state: 2 buffers (m and v)
adam_memory = param_count * 2 * 4
# Total with Adam (worst case)
total_memory = param_memory + grad_memory + adam_memory
# Convert to human-readable
def format_memory(bytes):
if bytes < 1024:
return f"{bytes}B"
elif bytes < 1024 * 1024:
return f"{bytes/1024:.1f}KB"
else:
return f"{bytes/(1024*1024):.1f}MB"
print(f"{name:<10} {format_memory(param_memory):<10} "
f"{format_memory(grad_memory):<12} {format_memory(sgd_memory):<12} "
f"{format_memory(adam_memory):<12} {format_memory(total_memory):<10}")
print("\nπ‘ Key Insights:")
print("- Training memory = Parameters + Gradients + Optimizer State")
print("- SGD: 3Γ parameter memory (params + grads + momentum)")
print("- Adam: 4Γ parameter memory (params + grads + 2 moment buffers)")
print("- Gradient accumulation reduces memory but increases training time")
# %% nbgrader={"grade": false, "grade_id": "analyze_checkpoint_overhead", "solution": true}
def analyze_checkpoint_overhead():
"""π Analyze checkpoint size and overhead."""
print("\nπ Analyzing Checkpoint Overhead...")
# Create a simple model
class TinyModel:
def __init__(self, size):
self.layer = Linear(size, size)
self.training = True
def forward(self, x):
return self.layer.forward(x)
def parameters(self):
return self.layer.parameters()
sizes = [10, 50, 100]
print("\nCheckpoint Size Analysis:")
print("=" * 70)
print(f"{'Model Size':<12} {'Raw Params':<15} {'Checkpoint':<15} {'Overhead':<10}")
print("-" * 70)
import pickle
import sys
for size in sizes:
# Create model and trainer
model = TinyModel(size)
optimizer = SGD(model.parameters(), lr=0.01)
trainer = Trainer(model, optimizer, MSELoss())
# Estimate raw parameter size
param_count = size * size + size # W + b
raw_size = param_count * 4 # 4 bytes per float32
# Create checkpoint and measure size
checkpoint_path = f"/tmp/checkpoint_test_{size}.pkl"
trainer.save_checkpoint(checkpoint_path)
import os
checkpoint_size = os.path.getsize(checkpoint_path)
overhead = (checkpoint_size / raw_size - 1) * 100
# Clean up
os.remove(checkpoint_path)
def format_size(bytes):
if bytes < 1024:
return f"{bytes}B"
return f"{bytes/1024:.1f}KB"
print(f"{size}Γ{size:<8} {format_size(raw_size):<15} "
f"{format_size(checkpoint_size):<15} {overhead:.1f}%")
print("\nπ‘ Key Insights:")
print("- Checkpoints include model state + optimizer state + training metadata")
print("- Pickle serialization adds 10-30% overhead")
print("- Adam optimizer doubles checkpoint size vs SGD")
print("- Use checkpoint frequency wisely in production (memory vs fault tolerance)")
# %% [markdown]
"""
## π§ͺ Module Integration Test
Final validation that everything works together correctly.
"""
# %% nbgrader={"grade": true, "grade_id": "test_module", "locked": true, "points": 20}
def test_module():
"""π§ͺ Module Test: Complete Integration
Comprehensive test of entire module functionality.
This final test runs before module summary to ensure:
- All unit tests pass
- Functions work together correctly
- Module is ready for integration with TinyTorch
"""
print("π§ͺ RUNNING MODULE INTEGRATION TEST")
print("=" * 50)
# Run all unit tests
print("Running unit tests...")
test_unit_cosine_schedule()
test_unit_clip_grad_norm()
test_unit_trainer()
print("\nRunning integration scenarios...")
# Test complete training pipeline integration with REAL components
print("π¬ Integration Test: Complete Training Pipeline...")
# Use REAL components from previous modules (already imported at module level)
# Create a simple model using REAL Linear layer
class SimpleModel:
def __init__(self):
self.layer = Linear(2, 1) # Real Linear from Module 03
self.training = True
def forward(self, x):
return self.layer.forward(x)
def parameters(self):
return self.layer.parameters()
# Create integrated system with REAL components
model = SimpleModel()
optimizer = SGD(model.parameters(), lr=0.01) # Real SGD from Module 07
loss_fn = MSELoss() # Real MSELoss from Module 04
scheduler = CosineSchedule(max_lr=0.1, min_lr=0.001, total_epochs=3)
trainer = Trainer(
model=model,
optimizer=optimizer,
loss_fn=loss_fn,
scheduler=scheduler,
grad_clip_norm=0.5
)
# Test data using REAL Tensors
data = [
(Tensor([[1.0, 0.5]]), Tensor([[0.8]])),
(Tensor([[0.5, 1.0]]), Tensor([[0.2]]))
]
# Test training
initial_loss = trainer.train_epoch(data)
assert isinstance(initial_loss, (float, np.floating)), "Training should return float loss"
assert trainer.epoch == 1, "Epoch should increment"
# Test evaluation
eval_loss, accuracy = trainer.evaluate(data)
assert isinstance(eval_loss, (float, np.floating)), "Evaluation should return float loss"
assert isinstance(accuracy, (float, np.floating)), "Evaluation should return float accuracy"
# Test scheduling
lr_epoch_0 = scheduler.get_lr(0)
lr_epoch_1 = scheduler.get_lr(1)
assert lr_epoch_0 > lr_epoch_1, "Learning rate should decrease"
# Test gradient clipping with large gradients using real Tensor
large_param = Tensor([1.0, 2.0], requires_grad=True)
large_param.grad = np.array([100.0, 200.0])
large_params = [large_param]
original_norm = clip_grad_norm(large_params, max_norm=1.0)
assert original_norm > 1.0, "Original norm should be large"
if isinstance(large_params[0].grad, np.ndarray):
grad_data = large_params[0].grad
else:
# Trust that Tensor has .data attribute
grad_data = large_params[0].grad.data
new_norm = np.linalg.norm(grad_data)
assert abs(new_norm - 1.0) < 1e-6, "Clipped norm should equal max_norm"
# Test checkpointing
checkpoint_path = "/tmp/integration_test_checkpoint.pkl"
trainer.save_checkpoint(checkpoint_path)
original_epoch = trainer.epoch
trainer.epoch = 999
trainer.load_checkpoint(checkpoint_path)
assert trainer.epoch == original_epoch, "Checkpoint should restore state"
# Clean up
import os
if os.path.exists(checkpoint_path):
os.remove(checkpoint_path)
print("β
End-to-end training pipeline works!")
print("\n" + "=" * 50)
print("π ALL TESTS PASSED! Module ready for export.")
print("Run: tito module complete 07")
if __name__ == "__main__":
test_module()
# %% [markdown]
"""
## π€ ML Systems Reflection Questions
Before we complete this module, let's reflect on the systems aspects of training infrastructure.
Use only knowledge from Modules 01-07 to answer these questions:
**1. Memory Trade-offs:**
- If you have a model with 1 million parameters and use Adam optimizer, estimate the total training memory required (parameters + gradients + optimizer state).
- How does gradient accumulation help when you want batch_size=128 but can only fit batch_size=32 in memory?
**2. Gradient Clipping:**
- Why do we clip gradients by *global norm* rather than clipping each gradient independently?
- What happens to training if gradients consistently exceed max_norm? What does this signal?
**3. Learning Rate Scheduling:**
- Why does cosine annealing start with high learning rate and end with low learning rate?
- Compare training with fixed lr=0.1 vs cosine schedule (0.1 β 0.01). When might fixed LR be better?
**4. Checkpointing Strategy:**
- You're training for 100 epochs. Checkpoints are large (1GB each). How often should you save checkpoints?
- What information MUST be in a checkpoint to resume training exactly where you left off?
**5. Train vs Eval Modes:**
- Why is it crucial to set model.training = False during evaluation?
- What would happen if you forgot to zero gradients between training steps?
**Think about these questions. The answers reveal deep understanding of training systems!**
"""
# %% [markdown]
"""
## β Aha Moment: Training Just Works
**What you built:** A complete training infrastructure with Trainer, schedulers, and checkpoints.
**Why it matters:** You've assembled all the pieces: tensors β layers β losses β autograd β
optimizers β training loop. This is the complete ML training pipeline! The Trainer orchestrates
forward pass, loss computation, backward pass, and weight updatesβjust like PyTorch Lightning.
In the milestones, you'll use this training infrastructure to train real models on real data!
"""
# %%
def demo_training():
"""π― See the training loop in action."""
print("π― AHA MOMENT: Training Just Works")
print("=" * 45)
# Simple linear regression: learn y = 2x + 1
np.random.seed(42)
X = Tensor(np.random.randn(20, 1))
y = Tensor(X.data * 2 + 1) # True relationship
# Simple model: one weight, one bias
w = Tensor(np.array([[0.0]]), requires_grad=True)
b = Tensor(np.array([0.0]), requires_grad=True)
optimizer = SGD([w, b], lr=0.1)
loss_fn = MSELoss()
print("Learning y = 2x + 1:")
for epoch in range(5):
# Forward
pred = X.matmul(w) + b
loss = loss_fn(pred, y)
# Backward
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(f" Epoch {epoch+1}: w={w.data[0,0]:.2f}, b={b.data[0]:.2f}, loss={float(loss.data):.4f}")
print(f"\nLearned: y = {w.data[0,0]:.1f}x + {b.data[0]:.1f}")
print("Target: y = 2.0x + 1.0")
print("\n⨠Your training loop learned the pattern!")
# %%
if __name__ == "__main__":
test_module()
print("\n")
demo_training()
# %% [markdown]
"""
## π MODULE SUMMARY: Training
Congratulations! You've built a complete training infrastructure that orchestrates the entire machine learning training process!
### Key Accomplishments
- Built Trainer class with complete training/evaluation loops
- Implemented CosineSchedule for adaptive learning rate management
- Created clip_grad_norm for training stability and gradient management
- Added comprehensive checkpointing for training persistence
- Analyzed training memory overhead and checkpoint costs
- All tests pass β
(validated by `test_module()`)
### Systems Insights Gained
Through hands-on implementation and analysis, you discovered:
- **Training memory is 4-6Γ model size**: Parameters + gradients + optimizer buffers
- **Gradient accumulation trades time for memory**: Enables larger effective batch sizes
- **Checkpoints include full training state**: Model + optimizer + scheduler + metadata
- **Learning rate scheduling improves convergence**: Cosine annealing balances speed and stability
- **Gradient clipping prevents instability**: Global norm preserves gradient direction
### Real-World Context
Your training infrastructure mirrors production ML systems:
- **PyTorch Lightning Trainer**: Similar architecture with training/eval loops
- **Hugging Face Transformers**: Uses same checkpoint patterns
- **Production Training**: All major ML frameworks use gradient clipping and scheduling
### Ready for Next Steps
Your training implementation enables sophisticated model training with proper scheduling, stability controls, and state management.
**Export with:** `tito module complete 07`
**Next**: Module 09 (Convolutions) will add spatial neural network operations, enabling CNN architectures for computer vision!
**π You now understand the complete training infrastructure that powers modern ML systems!**
""" |