File size: 24,173 Bytes
fc0f7bd | 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 | {
"cells": [
{
"cell_type": "markdown",
"metadata": {
"id": "qEIr4kn4W6Zs"
},
"source": [
"# Sharpness-Aware Minimization (SAM)\n",
"\n",
"[](https://colab.sandbox.google.com/github/google-deepmind/optax/blob/main/examples/sam.ipynb)\n",
"\n",
"\n",
"This serves a testing ground for a simple SAM type optimizer implementation in JAX with existing apis."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "AxR7ryYMXHcr"
},
"outputs": [],
"source": [
"import jax\n",
"import jax.numpy as np\n",
"import matplotlib.pyplot as plt\n",
"import optax\n",
"import chex\n",
"from optax import contrib"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "ImgJJV9_iq-v"
},
"source": [
"## Transparent Mode\n",
"\n",
"This implementation of SAM can be used in two different modes: transparent and opaque.\n",
" - Transparent mode exposes all gradient updates (described below) to the training loop, but it is easier to set up.\n",
" - Opaque mode hides the adversarial updates from the training loop, which is necessary when other state depends on the updates, such as BatchNorm parameters.\n",
"\n",
"Opaque mode is slightly more work to set up, so we will start with transparent mode."
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "TocZvhcDJoyY"
},
"source": [
"One way to describe what SAM does is that it does some number of steps (usually 1) of adversarial updates, followed by an outer gradient update.\n",
"\n",
"What this means is that we have to do a bunch of steps:\n",
"\n",
"\n",
" # adversarial step\n",
" params = params + sam_rho * normalize(gradient)\n",
"\n",
" # outer update step\n",
" params = cache - learning_rate * gradient\n",
" cache = params\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "7-p_W8vkhnO1"
},
"source": [
"To actually use SAM then, you create your adversarial optimizer, here SGD with normalized gradients, an outer optimzer, and then wrap them with SAM.\n",
"\n",
"The drop-in SAM optimizer described in the paper uses SGD for both optimizers."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "ueMKkNw7jLNJ"
},
"outputs": [],
"source": [
"lr = 0.01\n",
"rho = 0.1\n",
"opt = optax.sgd(lr)\n",
"adv_opt = optax.chain(contrib.normalize(), optax.sgd(rho))\n",
"sam_opt = contrib.sam(opt, adv_opt, sync_period=2) # This is the drop-in SAM optimizer.\n",
"\n",
"sgd_opt = optax.sgd(lr) # Baseline SGD optimizer"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "jgFuHGHPAIfU"
},
"source": [
"However, it is possible to use SGD for the adversarial optimizer, and, for example, SGD with momentum for the outer optimizer."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "TsLVwjHywg55"
},
"outputs": [],
"source": [
"def sam_mom(lr=1e-3, momentum=0.1, rho=0.1, sync_period=2):\n",
" opt = optax.sgd(lr, momentum=momentum)\n",
" adv_opt = optax.chain(contrib.normalize(), optax.sgd(rho))\n",
" return contrib.sam(opt, adv_opt, sync_period=sync_period)\n",
"\n",
"mom = 0.9\n",
"sam_mom_opt = sam_mom(lr, momentum=mom)\n",
"mom_opt = optax.sgd(lr, momentum=mom)"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "K6CCwKIIASsN"
},
"source": [
"It's even possible to use Adam for **both** optimizers. In this case, we'll need to increase the number of adversarial steps between syncs, but the resulting optimization will still be much faster than using SGD by itself with SAM."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "0FesjRbUsT80"
},
"outputs": [],
"source": [
"def sam_adam(lr=1e-3, b1=0.9, b2=0.999, rho=0.03, sync_period=5):\n",
" \"\"\"A SAM optimizer using Adam for the outer optimizer.\"\"\"\n",
" opt = optax.adam(lr, b1=b1, b2=b2)\n",
" adv_opt = optax.chain(contrib.normalize(), optax.adam(rho))\n",
" return contrib.sam(opt, adv_opt, sync_period=sync_period)\n",
"\n",
"sam_adam_opt = sam_adam(lr)\n",
"adam_opt = optax.adam(lr)"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "DyTChHZr2Yw6"
},
"source": [
"We'll set up a simple test problem below, we're going to try to optimize a sum of two exponentials that has two minima, one at (0,0) and another at (2,0) and compare the performance of both SAM and ordinary SGD."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "PSE3mM2FZGio"
},
"outputs": [],
"source": [
"# An example 2D loss function. It has two minima at (0,0) and (2,0).\n",
"# Both points attain almost zero loss value, but the first one is much sharper.\n",
"\n",
"def loss(params):\n",
" x, y = params\n",
" return -np.exp(-(x - 2)**2 - y**2) - 1.0*np.exp(-((x)**2 + (y)**2*100))\n",
"\n",
"x, y = np.meshgrid(np.linspace(0, 2, 100), np.linspace(0, 2, 100))\n",
"l = loss((x, y))\n",
"plt.matshow(l)\n",
"plt.xticks([0, 50, 100], [0, 1, 2])\n",
"plt.yticks([0, 50, 100], [0, 1, 2])\n",
"plt.title('Loss Surface')\n",
"plt.show();"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "Zi3tzM1AZbN_"
},
"outputs": [],
"source": [
"params = np.array([-0.4, -0.4])\n",
"\n",
"@chex.dataclass\n",
"class Store:\n",
" params: chex.Array\n",
" state: optax.OptState\n",
" step: int = 0\n",
"\n",
"sam_store = Store(params=params, state=sam_opt.init(params))\n",
"sgd_store = Store(params=params, state=sgd_opt.init(params))\n",
"sam_mom_store = Store(params=params, state=sam_mom_opt.init(params))\n",
"mom_store = Store(params=params, state=mom_opt.init(params))\n",
"sam_adam_store = Store(params=params, state=sam_adam_opt.init(params))\n",
"adam_store = Store(params=params, state=adam_opt.init(params))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "UhFr0AwqZjRk"
},
"outputs": [],
"source": [
"def make_step(opt):\n",
" @jax.jit\n",
" def step(store):\n",
" value, grads = jax.value_and_grad(loss)(store.params)\n",
" updates, state = opt.update(grads, store.state, store.params)\n",
" params = optax.apply_updates(store.params, updates)\n",
" return store.replace(\n",
" params=params,\n",
" state=state,\n",
" step=store.step+1), value\n",
" return step"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "bTkjju6IinJx"
},
"outputs": [],
"source": [
"sam_step = make_step(sam_opt)\n",
"sgd_step = make_step(sgd_opt)\n",
"\n",
"sam_mom_step = make_step(sam_mom_opt)\n",
"mom_step = make_step(mom_opt)\n",
"\n",
"sam_adam_step = make_step(sam_adam_opt)\n",
"adam_step = make_step(adam_opt)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "MEF-PriWcLSa"
},
"outputs": [],
"source": [
"sam_vals = []\n",
"sam_params = []\n",
"sgd_vals = []\n",
"sgd_params = []\n",
"\n",
"sam_mom_vals = []\n",
"sam_mom_params = []\n",
"mom_vals = []\n",
"mom_params = []\n",
"\n",
"sam_adam_vals = []\n",
"sam_adam_params = []\n",
"adam_vals = []\n",
"adam_params = []"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "8Em3xy9PaEbH"
},
"outputs": [],
"source": [
"T = 8000\n",
"for i in range(T):\n",
" sam_store, sam_val = sam_step(sam_store)\n",
" sgd_store, sgd_val = sgd_step(sgd_store)\n",
" sam_mom_store, sam_mom_val = sam_mom_step(sam_mom_store)\n",
" mom_store, mom_val = mom_step(mom_store)\n",
" sam_adam_store, sam_adam_val = sam_adam_step(sam_adam_store)\n",
" adam_store, adam_val = adam_step(adam_store)\n",
"\n",
" sam_vals.append(sam_val)\n",
" sgd_vals.append(sgd_val)\n",
" sam_mom_vals.append(sam_mom_val)\n",
" mom_vals.append(mom_val)\n",
" sam_adam_vals.append(sam_adam_val)\n",
" adam_vals.append(adam_val)\n",
"\n",
" sam_params.append(sam_store.params)\n",
" sgd_params.append(sgd_store.params)\n",
" sam_mom_params.append(sam_mom_store.params)\n",
" mom_params.append(mom_store.params)\n",
" sam_adam_params.append(sam_adam_store.params)\n",
" adam_params.append(adam_store.params)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "sCrf_qJzdDmk"
},
"outputs": [],
"source": [
"ts = np.arange(T)\n",
"fig, axs = plt.subplots(6, figsize=(10, 12))\n",
"axs[0].plot(ts, sgd_vals, label='SGD')\n",
"axs[0].plot(ts[::2], sam_vals[0::2], label='SAM Outer Loss', lw=3, zorder=100)\n",
"axs[0].plot(ts[1::2], sam_vals[1::2], label='SAM Adv Loss', alpha=0.5)\n",
"axs[0].legend();\n",
"\n",
"axs[1].plot(ts, sgd_vals, label='SGD')\n",
"axs[1].plot(ts[::2] / 2, sam_vals[::2], label='1/2 SAM', lw=3)\n",
"axs[1].legend();\n",
"\n",
"axs[2].plot(ts, mom_vals, label='Mom')\n",
"axs[2].plot(ts[::2], sam_mom_vals[::2], label='SAM Mom Outer Loss', lw=3, zorder=100)\n",
"axs[2].plot(ts[1::2], sam_mom_vals[1::2], label='SAM Mom Adv Loss', alpha=0.5)\n",
"axs[2].legend();\n",
"\n",
"axs[3].plot(ts, mom_vals, label='Mom')\n",
"axs[3].plot(ts[::2] / 2, sam_mom_vals[::2], label='1/2 SAM Mom', lw=3)\n",
"axs[3].legend();\n",
"\n",
"axs[4].plot(ts, adam_vals, label='Adam')\n",
"axs[4].plot(ts[::5], sam_adam_vals[::5], label='SAM Adam Outer Loss', lw=3, zorder=100)\n",
"axs[4].plot(ts[4::5], sam_adam_vals[4::5], label='SAM Adam Adv Loss', alpha=0.5)\n",
"axs[4].legend();\n",
"\n",
"axs[5].plot(ts, adam_vals, label='Adam')\n",
"axs[5].plot(ts[::5] / 5, sam_adam_vals[::5], label='1/5 SAM Adam', lw=3)\n",
"axs[5].legend();"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "kEmX6vdp_n50"
},
"source": [
"On this problem, SAM Mom is the fastest of the three SAM optimizers in terms of real steps, but in terms of outer gradient steps, SAM Adam is slightly faster, since it has 4 inner gradient steps for every outer gradient step, compared with 1 inner per outer for SAM and SAM Mom."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "o1kIeonUeA0x"
},
"outputs": [],
"source": [
"fig, axs = plt.subplots(ncols=3, figsize=(8 * 3, 6))\n",
"axs[0].plot(*np.array(sgd_params).T, label='SGD')\n",
"axs[0].plot(*np.array(sam_params)[1::2].T, label='SAM Outer Steps', zorder=100)\n",
"axs[0].plot(*np.array(sam_params)[::2].T, label='SAM Adv Steps', alpha=0.5)\n",
"axs[0].legend(loc=4);\n",
"\n",
"axs[1].plot(*np.array(mom_params).T, label='Mom')\n",
"axs[1].plot(*np.array(sam_mom_params)[1::2].T, label='SAM Mom Outer Steps', zorder=100)\n",
"axs[1].plot(*np.array(sam_mom_params)[::2].T, label='SAM Mom Adv Steps', alpha=0.5)\n",
"axs[1].legend(loc=4);\n",
"\n",
"axs[2].plot(*np.array(adam_params).T, label='Adam')\n",
"axs[2].plot(*np.array(sam_adam_params)[4::5].T, label='SAM Adam Outer Steps', zorder=100)\n",
"axs[2].plot(*np.array(sam_adam_params)[3::5].T, label='SAM Adam Adv Steps', alpha=0.5)\n",
"axs[2].legend(loc=4);"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "8nVNiUsweApJ"
},
"source": [
"As you can see, all three SAM optimizers find the smooth optimum, while all three standard optimizers get stuck in the sharp optimum.\n",
"\n",
"SAM and SAM Mom follow fairly similar paths (although SAM Mom is much faster), but Sam Adam actually passes through the sharp optimum on the way to the smooth optimum.\n",
"\n",
"The adversarial steps are quite different between the three SAM optimizers, demonstrating that the choice of both outer and inner optimizer have strong impacts on how the loss landscape gets explored."
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "0s8UWJB-joV-"
},
"source": [
"## Opaque Mode\n",
"\n",
"Here, we'll demonstrate how to use opaque mode on the same setting.\n",
"\n",
"The main difference is that we need to pass a gradient function to the update call. The gradient function needs to take as arguments params and an integer (indicating the current adversarial step). It's generally safe to ignore the second argument:\n",
"```python\n",
"grad_fn = jax.grad(\n",
" lambda params, _: loss(params, batch, and_other_args, to_loss))\n",
"updates, sam_state = sam_opt.update(updates, sam_state, params, grad_fn=grad_fn)\n",
"params = optax.apply_updates(params, updates)\n",
"```"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "EwXxXOFVkjou"
},
"source": [
"Here's the opaque drop-in SAM optimizer again.\n",
"\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "UMvJqqxRjmIF"
},
"outputs": [],
"source": [
"lr = 0.01\n",
"rho = 0.1\n",
"opt = optax.sgd(lr)\n",
"adv_opt = optax.chain(contrib.normalize(), optax.sgd(rho))\n",
"sam_opt = contrib.sam(opt, adv_opt, sync_period=2, opaque_mode=True)\n",
"\n",
"sgd_opt = optax.sgd(lr) # Baseline SGD optimizer"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "woThldyjjmIH"
},
"source": [
"Here's an opaque momentum SAM optimizer."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "YcaxX4-4jmIH"
},
"outputs": [],
"source": [
"def sam_mom(lr=1e-3, momentum=0.1, rho=0.1, sync_period=2):\n",
" opt = optax.sgd(lr, momentum=momentum)\n",
" adv_opt = optax.chain(contrib.normalize(), optax.sgd(rho))\n",
" return contrib.sam(opt, adv_opt, sync_period=sync_period, opaque_mode=True)\n",
"\n",
"mom = 0.9\n",
"sam_mom_opt = sam_mom(lr, momentum=mom)\n",
"mom_opt = optax.sgd(lr, momentum=mom)"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "0MXyU4WrjmIH"
},
"source": [
"Here's an opaque Adam-based SAM optimizer."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "Fb8Cy5xgjmIJ"
},
"outputs": [],
"source": [
"def sam_adam(lr=1e-3, b1=0.9, b2=0.999, rho=0.03, sync_period=5):\n",
" \"\"\"A SAM optimizer using Adam for the outer optimizer.\"\"\"\n",
" opt = optax.adam(lr, b1=b1, b2=b2)\n",
" adv_opt = optax.chain(contrib.normalize(), optax.adam(rho))\n",
" return contrib.sam(opt, adv_opt, sync_period=sync_period, opaque_mode=True)\n",
"\n",
"sam_adam_opt = sam_adam(lr)\n",
"adam_opt = optax.adam(lr)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "-z4NxcZujmIJ"
},
"outputs": [],
"source": [
"params = np.array([-0.4, -0.4])\n",
"\n",
"@chex.dataclass\n",
"class Store:\n",
" params: chex.Array\n",
" state: optax.OptState\n",
" step: int = 0\n",
"\n",
"sam_store = Store(params=params, state=sam_opt.init(params))\n",
"sgd_store = Store(params=params, state=sgd_opt.init(params))\n",
"sam_mom_store = Store(params=params, state=sam_mom_opt.init(params))\n",
"mom_store = Store(params=params, state=mom_opt.init(params))\n",
"sam_adam_store = Store(params=params, state=sam_adam_opt.init(params))\n",
"adam_store = Store(params=params, state=adam_opt.init(params))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "bwLv3UTCjmIK"
},
"outputs": [],
"source": [
"def make_step(opt):\n",
" @jax.jit\n",
" def step(store):\n",
" value, grads = jax.value_and_grad(loss)(store.params)\n",
" if isinstance(store.state, contrib.SAMState):\n",
" updates, state = opt.update(\n",
" grads, store.state, store.params,\n",
" grad_fn=jax.grad(lambda p, _: loss(p))) # NOTICE THE ADDITIONAL grad_fn ARGUMENT!\n",
" else:\n",
" updates, state = opt.update(grads, store.state, store.params)\n",
" params = optax.apply_updates(store.params, updates)\n",
" return store.replace(\n",
" params=params,\n",
" state=state,\n",
" step=store.step+1), value\n",
" return step"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "VRJiF7TWjmIK"
},
"outputs": [],
"source": [
"sam_step = make_step(sam_opt)\n",
"sgd_step = make_step(sgd_opt)\n",
"\n",
"sam_mom_step = make_step(sam_mom_opt)\n",
"mom_step = make_step(mom_opt)\n",
"\n",
"sam_adam_step = make_step(sam_adam_opt)\n",
"adam_step = make_step(adam_opt)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "cwSnRxfQjmIK"
},
"outputs": [],
"source": [
"sam_vals = []\n",
"sam_params = []\n",
"sgd_vals = []\n",
"sgd_params = []\n",
"\n",
"sam_mom_vals = []\n",
"sam_mom_params = []\n",
"mom_vals = []\n",
"mom_params = []\n",
"\n",
"sam_adam_vals = []\n",
"sam_adam_params = []\n",
"adam_vals = []\n",
"adam_params = []"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "m20mYyBHjmIK"
},
"outputs": [],
"source": [
"T = 4000\n",
"for i in range(T):\n",
" sam_store, sam_val = sam_step(sam_store)\n",
" sgd_store, sgd_val = sgd_step(sgd_store)\n",
" sam_mom_store, sam_mom_val = sam_mom_step(sam_mom_store)\n",
" mom_store, mom_val = mom_step(mom_store)\n",
" sam_adam_store, sam_adam_val = sam_adam_step(sam_adam_store)\n",
" adam_store, adam_val = adam_step(adam_store)\n",
"\n",
" sam_vals.append(sam_val)\n",
" sgd_vals.append(sgd_val)\n",
" sam_mom_vals.append(sam_mom_val)\n",
" mom_vals.append(mom_val)\n",
" sam_adam_vals.append(sam_adam_val)\n",
" adam_vals.append(adam_val)\n",
"\n",
" sam_params.append(sam_store.params)\n",
" sgd_params.append(sgd_store.params)\n",
" sam_mom_params.append(sam_mom_store.params)\n",
" mom_params.append(mom_store.params)\n",
" sam_adam_params.append(sam_adam_store.params)\n",
" adam_params.append(adam_store.params)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "zlDP-42ajmIK"
},
"outputs": [],
"source": [
"ts = np.arange(T)\n",
"fig, axs = plt.subplots(6, figsize=(10, 12))\n",
"axs[0].plot(ts, sgd_vals, label='SGD')\n",
"axs[0].plot(ts, sam_vals, label='SAM', lw=3, zorder=100)\n",
"axs[0].legend();\n",
"\n",
"axs[1].plot(ts, sgd_vals, label='SGD')\n",
"axs[1].plot(ts * 2, sam_vals, label='2 * SAM', lw=3)\n",
"axs[1].legend();\n",
"\n",
"axs[2].plot(ts, mom_vals, label='Mom')\n",
"axs[2].plot(ts, sam_mom_vals, label='SAM Mom', lw=3, zorder=100)\n",
"axs[2].legend();\n",
"\n",
"axs[3].plot(ts, mom_vals, label='Mom')\n",
"axs[3].plot(ts * 2, sam_mom_vals, label='2 * SAM Mom', lw=3)\n",
"axs[3].legend();\n",
"\n",
"axs[4].plot(ts, adam_vals, label='Adam')\n",
"axs[4].plot(ts, sam_adam_vals, label='SAM Adam', lw=3, zorder=100)\n",
"axs[4].legend();\n",
"\n",
"axs[5].plot(ts, adam_vals, label='Adam')\n",
"axs[5].plot(ts * 5, sam_adam_vals, label='5 * SAM Adam', lw=3)\n",
"axs[5].legend();"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "NLAJFJ-SjmIK"
},
"source": [
"The behavior is identical to transparent mode, but the percieved number of gradient steps is half as many as in transparent mode (or 1/5 as many for SAM Adam)."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "mY5FzqMLjmIK"
},
"outputs": [],
"source": [
"fig, axs = plt.subplots(ncols=3, figsize=(8 * 3, 6))\n",
"axs[0].plot(*np.array(sgd_params).T, label='SGD')\n",
"axs[0].plot(*np.array(sam_params).T, label='SAM', zorder=100)\n",
"axs[0].legend(loc=4);\n",
"\n",
"axs[1].plot(*np.array(mom_params).T, label='Mom')\n",
"axs[1].plot(*np.array(sam_mom_params).T, label='SAM Mom')\n",
"axs[1].legend(loc=4);\n",
"\n",
"axs[2].plot(*np.array(adam_params).T, label='Adam')\n",
"axs[2].plot(*np.array(sam_adam_params).T, label='SAM Adam')\n",
"axs[2].legend(loc=4);"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "P_MjRgofjmIK"
},
"source": [
"The behavior is identical to transparent mode here as well, but we don't get to see the adversarial updates."
]
}
],
"metadata": {
"colab": {
"last_runtime": {
"build_target": "//learning/grp/tools/ml_python:ml_notebook",
"kind": "private"
},
"private_outputs": true,
"provenance": [
{
"file_id": "1ap7lwdf3vgoQdyNogmSVfncYZcH1ijch",
"timestamp": 1697129954516
}
],
"toc_visible": true
},
"kernelspec": {
"display_name": "Python 3",
"name": "python3"
},
"language_info": {
"name": "python"
}
},
"nbformat": 4,
"nbformat_minor": 0
}
|