File size: 40,430 Bytes
31126ff | 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 | {
"cells": [
{
"cell_type": "code",
"execution_count": null,
"id": "9c91ef58",
"metadata": {},
"outputs": [],
"source": [
"import pandas as pd\n",
"import tensorflow as tf\n",
"import tensorflow.keras.backend as K\n",
"from tensorflow.keras import layers\n",
"\n",
"import collections\n",
"import logging\n",
"import numpy as np\n",
"import tensorflow as tf\n",
"from tensorflow.keras.callbacks import EarlyStopping, LearningRateScheduler, ModelCheckpoint, ReduceLROnPlateau\n",
"import pandas as pd\n",
"from sklearn.preprocessing import StandardScaler\n",
"import os\n",
"import math\n",
"import gc"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "9a965750",
"metadata": {},
"outputs": [],
"source": [
"name = \"ETTm1\" # dataset name\n",
"seq_len = 512\n",
"\n",
"batch_size = 32\n",
"pred_len = 192\n",
"\n",
"feature_type = \"M\"\n",
"target = \"\"\n",
"learning_rate = 0.0001\n",
"patience = 5\n",
"\n",
"num_heads=1\n",
"d_model=16\n",
"rho = 0.1"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "d17cf421",
"metadata": {},
"outputs": [],
"source": [
"# please provide your own absolute links here\n",
"LOCAL_CACHE_DIR = '../Data/Benchmark/' # please include your dataset directory here\n",
"\n",
"checkpoint_path = \"../Train_SAM/models/checkpoint\" + name + str(pred_len) + \"_SAM\" + \".model\" # where the model checkpoint should be saved + the .model file type\n",
"name_df = \"../Results_SAM/\" + name + \"_\" + str(pred_len) + \"_SAM\" + \".csv\" # a filepath .csv file type to save the mse and mae results"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "5a6e0049",
"metadata": {},
"outputs": [],
"source": [
"class CaptureWeightsCallback(tf.keras.callbacks.Callback):\n",
" \"\"\"\n",
" Custom TensorFlow callback for capturing and logging model weights during training, with a focus on attention weights.\n",
"\n",
" This callback is designed to monitor the evolution of model weights, particularly attention weights, across training epochs.\n",
" It facilitates the analysis of training dynamics and model behavior by storing weight snapshots at specified intervals.\n",
"\n",
" Attributes:\n",
" model (tf.keras.Model): Instance of the TensorFlow model being trained. The model should have a method\n",
" `get_last_attention_weights()` that this callback can invoke to obtain attention weights.\n",
" attention_weights_history (list): Accumulates the attention weights captured at the end of specified epochs. \n",
" This history facilitates post-training analysis of weight adjustments.\n",
"\n",
" Methods:\n",
" on_epoch_end(epoch, logs=None): Overrides the base class method to capture attention weights at the end of each epoch.\n",
" Weights are captured based on specified criteria, e.g., every 5 epochs.\n",
" get_attention_weights_history(): Provides access to the accumulated history of attention weights captured during training.\n",
" \"\"\"\n",
" \n",
" def __init__(self, model):\n",
" \"\"\"\n",
" Initializes the callback with a specific model to monitor its attention weights during training.\n",
"\n",
" Parameters:\n",
" model (tf.keras.Model): The model whose attention weights are to be monitored and captured.\n",
" \"\"\"\n",
" super().__init__()\n",
" self.model = model\n",
" self.penultimate_weights = None\n",
" self.attention_weights_history = []\n",
"\n",
" def on_epoch_end(self, epoch, logs=None):\n",
" \"\"\"\n",
" Called at the end of an epoch during training to capture and store attention weights if the current\n",
" epoch satisfies the capture criteria (e.g., every 5 epochs).\n",
"\n",
" Parameters:\n",
" epoch (int): The current epoch number.\n",
" logs (dict): Currently unused. Contains logs from the training epoch.\n",
" \"\"\"\n",
" if epoch % 5 == 0: # Perform analysis every 5 epochs\n",
" # Retrieve attention weights from the model\n",
" last_attention_weights = self.model.get_last_attention_weights()\n",
" if last_attention_weights is not None:\n",
" self.attention_weights_history.append(last_attention_weights)\n",
" \n",
" def get_attention_weights_history(self):\n",
" \"\"\"\n",
" Returns the history of attention weights captured during training.\n",
"\n",
" Returns:\n",
" A list of attention weights captured at specified intervals during training.\n",
" \"\"\"\n",
" return self.attention_weights_history"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "22e2e8be",
"metadata": {},
"outputs": [],
"source": [
"def setup_callbacks(learning_rate, patience, checkpoint_path, model):\n",
" \"\"\"\n",
" Sets up and returns TensorFlow callbacks for use during model training. These callbacks include early stopping,\n",
" learning rate scheduling, model checkpointing, and a custom callback for capturing model weights.\n",
"\n",
" This function is tailored to support flexible training configurations, allowing for dynamic adjustment of training\n",
" behavior based on model performance and training progress.\n",
"\n",
" Parameters:\n",
" args (argparse.Namespace): Parsed command-line arguments containing training configurations such as patience for early stopping,\n",
" total training epochs, and initial learning rate.\n",
" checkpoint_path (str): File path where model checkpoints will be saved. The best model according to validation loss is checkpointed.\n",
" model (tf.keras.Model): The TensorFlow model being trained. Required for initializing the `CaptureWeightsCallback`.\n",
" model_name (str): Name of the model being trained. This can be used to adjust callback behavior for different models.\n",
"\n",
" Returns:\n",
" tuple: A tuple containing:\n",
" - A list of TensorFlow callbacks configured for the training session.\n",
" - An instance of `CaptureWeightsCallback`, which can be used post-training to access captured weights.\n",
"\n",
" Raises:\n",
" Exception: If an error occurs in the setup of callbacks, an exception is logged and raised to prevent silent training failures.\n",
"\n",
" Example:\n",
" >>> callbacks, capture_weights_callback = setup_callbacks(args, './model_checkpoints', model, 'my_model')\n",
" This example demonstrates how to invoke `setup_callbacks` to obtain configured callbacks for training, including a custom\n",
" weight capture callback for post-training analysis.\n",
" \"\"\"\n",
" try:\n",
" checkpoint_callback = ModelCheckpoint(\n",
" filepath=checkpoint_path,\n",
" monitor='val_loss', \n",
" verbose=1,\n",
" save_best_only=True,\n",
" save_weights_only=True,\n",
" )\n",
"\n",
" early_stop_callback = EarlyStopping(\n",
" monitor='val_loss',\n",
" patience=patience,\n",
" verbose=1,\n",
" )\n",
"\n",
" lr_schedule_callback = LearningRateScheduler(\n",
" lambda epoch: cosine_annealing(epoch, 5, learning_rate, 1e-6),\n",
" verbose=1,\n",
" )\n",
"\n",
" lrdecay = ReduceLROnPlateau(monitor='val_loss', factor=0.90, patience=3, min_lr=learning_rate * 0.001, verbose=1)\n",
"\n",
" capture_weights_callback = CaptureWeightsCallback(model)\n",
"\n",
" callbacks = [checkpoint_callback, lr_schedule_callback, capture_weights_callback, early_stop_callback]\n",
" \n",
" return callbacks, capture_weights_callback\n",
" except Exception as e:\n",
" logging.error(f\"Error setting up callbacks: {e}\")\n",
" raise\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "5604ad15",
"metadata": {},
"outputs": [],
"source": [
"def cosine_annealing(epoch, max_epochs, initial_lr, min_lr):\n",
" \"\"\"\n",
" Applies cosine annealing to the learning rate.\n",
" \n",
" Parameters:\n",
" epoch (int): Current epoch.\n",
" max_epochs (int): Maximum number of epochs.\n",
" initial_lr (float): Initial learning rate.\n",
" min_lr (float): Minimum learning rate.\n",
" \n",
" Returns:\n",
" float: Adjusted learning rate.\n",
" \"\"\"\n",
" cos_inner = (math.pi * (epoch % max_epochs)) / max_epochs\n",
" return min_lr + (initial_lr - min_lr) * (math.cos(cos_inner) + 1) / 2\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "430e2b9e",
"metadata": {},
"outputs": [],
"source": [
"class RevNorm(layers.Layer):\n",
" \"\"\"\n",
" Implements Reversible Instance Normalization (RevNorm).\n",
"\n",
" This layer normalizes input features per instance and can reverse the normalization process. It is designed\n",
" to maintain the statistical properties of the input data, making it particularly useful in generative models\n",
" where the exact inverse operation is necessary.\n",
"\n",
" Attributes:\n",
" axis (int): The axis along which to compute the mean and standard deviation for normalization.\n",
" eps (float): A small constant added to the standard deviation to prevent division by zero.\n",
" affine (bool): Whether to apply a learnable affine transformation after normalization.\n",
"\n",
" The original implementation can be found in the Google Research repository:\n",
" https://github.com/google-research/google-research/blob/master/tsmixer/tsmixer_basic/models/rev_in.py\n",
"\n",
" Example usage:\n",
" rev_norm = RevNorm(axis=-1, eps=1e-5, affine=True)\n",
" normalized_output = rev_norm(input_tensor, mode='norm')\n",
" denormalized_output = rev_norm(normalized_output, mode='denorm', target_slice=slice_indices)\n",
" \"\"\"\n",
"\n",
" def __init__(self, axis, eps=1e-5, affine=True):\n",
" super().__init__()\n",
" self.axis = axis # Defines the dimension along which normalization is performed.\n",
" self.eps = eps # Small epsilon value to ensure numerical stability.\n",
" self.affine = affine # Determines if learnable affine parameters should be used.\n",
"\n",
" def build(self, input_shape):\n",
" \"\"\"\n",
" Initializes the layer's weights.\n",
"\n",
" This method creates affine transformation weights if the `affine` attribute is set to True.\n",
" It defines two trainable weights, `affine_weight` and `affine_bias`, which are used to scale and shift\n",
" the normalized data respectively.\n",
"\n",
" Args:\n",
" input_shape (TensorShape): The shape of the input tensor to the layer. The last dimension is used\n",
" to determine the shape of the affine weights.\n",
"\n",
" Note: This method is automatically called during the first use of the layer.\n",
" \"\"\"\n",
" if self.affine:\n",
" self.affine_weight = self.add_weight(\n",
" 'affine_weight', shape=input_shape[-1], initializer='ones'\n",
" )\n",
" self.affine_bias = self.add_weight(\n",
" 'affine_bias', shape=input_shape[-1], initializer='zeros'\n",
" )\n",
"\n",
" def call(self, x, mode, target_slice=None):\n",
" \"\"\"\n",
" Performs normalization or denormalization on the input tensor.\n",
"\n",
" Args:\n",
" x (Tensor): Input tensor to be normalized or denormalized.\n",
" mode (str): 'norm' for normalization and 'denorm' for denormalization.\n",
" target_slice (slice, optional): Target slice for denormalization.\n",
"\n",
" Returns:\n",
" Tensor: The normalized or denormalized output.\n",
" \"\"\"\n",
" if mode == 'norm':\n",
" self._get_statistics(x)\n",
" x = self._normalize(x)\n",
" elif mode == 'denorm':\n",
" x = self._denormalize(x, target_slice)\n",
" else:\n",
" raise NotImplementedError\n",
" return x\n",
"\n",
" def _get_statistics(self, x):\n",
" \"\"\"\n",
" Computes the mean and standard deviation of the input tensor along the specified axis.\n",
"\n",
" The calculated mean and standard deviation are used for normalizing the input data. They are computed\n",
" using `tf.reduce_mean` and `tf.sqrt(tf.reduce_variance(...) + self.eps)` to ensure numerical stability.\n",
"\n",
" Args:\n",
" x (Tensor): Input tensor from which the statistics are computed.\n",
"\n",
" Updates:\n",
" self.mean (Tensor): The mean of the input tensor, calculated along the specified axis.\n",
" self.stdev (Tensor): The standard deviation of the input tensor, ensuring numerical stability by adding `self.eps`.\n",
" \"\"\"\n",
" self.mean = tf.stop_gradient(\n",
" tf.reduce_mean(x, axis=self.axis, keepdims=True)\n",
" )\n",
" self.stdev = tf.stop_gradient(\n",
" tf.sqrt(\n",
" tf.math.reduce_variance(x, axis=self.axis, keepdims=True) + self.eps\n",
" )\n",
" )\n",
"\n",
" def _normalize(self, x):\n",
" \"\"\"\n",
" Normalizes the input tensor using the computed mean and standard deviation.\n",
"\n",
" This method subtracts the mean from the input tensor and divides it by the standard deviation, effectively\n",
" standardizing the input to have a mean of 0 and a standard deviation of 1. If affine transformation is enabled,\n",
" it further applies scaling and shifting to the standardized input.\n",
"\n",
" Args:\n",
" x (Tensor): Input tensor to be normalized.\n",
"\n",
" Returns:\n",
" Tensor: The normalized tensor.\n",
" \"\"\"\n",
" x = x - self.mean\n",
" x = x / self.stdev\n",
" if self.affine:\n",
" x = x * self.affine_weight\n",
" x = x + self.affine_bias\n",
" return x\n",
"\n",
" def _denormalize(self, x, target_slice=None):\n",
" \"\"\"\n",
" Reverses the normalization process for the given slice of the input tensor.\n",
"\n",
" This method applies the inverse of the normalization operation. If affine transformation was applied during\n",
" normalization, it reverses this process first. Then, it multiplies the tensor by the standard deviation and adds\n",
" the mean to denormalize the data.\n",
"\n",
" Args:\n",
" x (Tensor): Normalized tensor that needs to be denormalized.\n",
" target_slice (slice, optional): Specific slice of the tensor to denormalize. Useful when different parts\n",
" of the tensor require different reverse operations.\n",
"\n",
" Returns:\n",
" Tensor: The denormalized tensor.\n",
" \"\"\"\n",
" if self.affine:\n",
" x = x - self.affine_bias[target_slice]\n",
" x = x / self.affine_weight[target_slice]\n",
" x = x * self.stdev[:, :, target_slice]\n",
" x = x + self.mean[:, :, target_slice]\n",
" return x\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "323930c3",
"metadata": {},
"outputs": [],
"source": [
"class SAM:\n",
" \"\"\"\n",
" Sharpness-Aware Minimization (SAM) for Enhanced Training Stability.\n",
" \n",
" SAM optimizes a model's parameters in the direction that enhances model\n",
" performance while simultaneously minimizing loss sharpness, aiming to improve\n",
" generalization. This implementation wraps around a base TensorFlow optimizer\n",
" to apply the SAM methodology.\n",
" \n",
" Reference:\n",
" \"Sharpness-Aware Minimization for Efficiently Improving Generalization\"\n",
" by Foret, Kleiner, Mobahi, and Neyshabur. https://openreview.net/pdf?id=6Tm1mposlrM\n",
"\n",
" The original implementation can be found at:\n",
" - https://github.com/davda54/sam\n",
" \n",
" Attributes:\n",
" base_optimizer (tf.keras.optimizers.Optimizer): The TensorFlow optimizer to wrap.\n",
" rho (float): The neighborhood size for sharpness-aware optimization.\n",
" eps (float): A small epsilon value to prevent division by zero.\n",
" \"\"\"\n",
"\n",
" def __init__(self, base_optimizer, rho=0.05, eps=1e-12):\n",
" \"\"\"\n",
" Initializes the SAM optimizer wrapper.\n",
" \n",
" Parameters:\n",
" base_optimizer (tf.keras.optimizers.Optimizer): The base optimizer.\n",
" rho (float): The neighborhood size for sharpness-aware optimization.\n",
" eps (float): A small epsilon value to prevent division by zero.\n",
" \"\"\"\n",
" assert rho >= 0.0, f\"Invalid rho, should be non-negative: {rho}\"\n",
" self.rho = rho\n",
" self.eps = eps\n",
" self.base_optimizer = base_optimizer\n",
"\n",
" def first_step(self, gradients, trainable_vars):\n",
" \"\"\"\n",
" Performs the first optimization step, moving weights in the direction\n",
" that increases loss sharpness.\n",
" \n",
" Parameters:\n",
" gradients (List[tf.Tensor]): Gradients of the loss with respect to the model parameters.\n",
" trainable_vars (List[tf.Variable]): The model's trainable variables.\n",
" \"\"\"\n",
" self.e_ws = []\n",
" grad_norm = tf.linalg.global_norm(gradients)\n",
" ew_multiplier = self.rho / (grad_norm + self.eps)\n",
"\n",
" for i in range(len(trainable_vars)):\n",
" e_w = tf.math.multiply(gradients[i], ew_multiplier)\n",
" trainable_vars[i].assign_add(e_w)\n",
" self.e_ws.append(e_w)\n",
"\n",
" def second_step(self, gradients, trainable_variables):\n",
" \"\"\"\n",
" Performs the second optimization step, applying the base optimizer\n",
" update after reverting the first step's perturbation.\n",
" \n",
" Parameters:\n",
" gradients (List[tf.Tensor]): Gradients of the loss with respect to the model parameters after the first step.\n",
" trainable_variables (List[tf.Variable]): The model's trainable variables.\n",
" \"\"\"\n",
" for i in range(len(trainable_variables)):\n",
" trainable_variables[i].assign_add(-self.e_ws[i]) # Revert first step\n",
" self.base_optimizer.apply_gradients(zip(gradients, trainable_variables))\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "3adf7700",
"metadata": {},
"outputs": [],
"source": [
"class SpectralNormalizedAttention(layers.MultiHeadAttention):\n",
" \"\"\"\n",
" Spectral Normalized Multi-Head Attention Layer.\n",
" \n",
" This layer extends the MultiHeadAttention layer with spectral normalization on the\n",
" query, key, and value weights, implementing the sigma-reparam method described in\n",
" \"Stabilizing Transformer Training by Preventing Attention Entropy Collapse\".\n",
" \n",
" Paper URL: https://openreview.net/forum?id=LL8gz8FHxH\n",
" GitHub Code: https://github.com/apple/ml-sigma-reparam\n",
" \n",
" Attributes:\n",
" gamma (tf.Variable): Scaling factor for the normalized weights, trainable.\n",
" \"\"\"\n",
"\n",
" def __init__(self, *args, **kwargs):\n",
" \"\"\"Initializes the SpectralNormalizedAttention layer with standard arguments for MultiHeadAttention.\"\"\"\n",
" super(SpectralNormalizedAttention, self).__init__(*args, **kwargs)\n",
" self.gamma = self.add_weight(name='gamma', shape=[], initializer='ones', trainable=True)\n",
"\n",
" def build(self, input_shape):\n",
" \"\"\"Builds the layer, initializing weights.\"\"\"\n",
" super(SpectralNormalizedAttention, self).build(input_shape)\n",
" # Additional initializations can be added here if necessary.\n",
"\n",
" def _normalize_weights(self, W):\n",
" \"\"\"\n",
" Normalizes the weights matrix W using its spectral norm.\n",
" \n",
" Parameters:\n",
" W (tf.Tensor): The weight matrix to normalize.\n",
" \n",
" Returns:\n",
" tf.Tensor: Spectrally normalized weights.\n",
" \"\"\"\n",
" singular_values = tf.linalg.svd(W, compute_uv=False)\n",
" spectral_norm = tf.reduce_max(singular_values)\n",
" return W / spectral_norm\n",
"\n",
" def call(self, query, value, key=None, attention_mask=None, return_attention_scores=False):\n",
" \"\"\"\n",
" Calls the SpectralNormalizedAttention layer. Normalizes the query, key, and value weights\n",
" before calling the parent MultiHeadAttention layer.\n",
" \n",
" Parameters:\n",
" query (tf.Tensor): Query tensor.\n",
" value (tf.Tensor): Value tensor.\n",
" key (tf.Tensor): Key tensor. Defaults to None, in which case the query is used as the key.\n",
" attention_mask (tf.Tensor): Optional tensor to mask out certain positions from attending to others.\n",
" return_attention_scores (bool): Flag to return attention scores along with output.\n",
" \n",
" Returns:\n",
" A tuple of (output tensor, attention scores) if return_attention_scores is True, otherwise just the output tensor.\n",
" \"\"\"\n",
" key = query if key is None else key\n",
" query = self._normalize_weights(query) * self.gamma\n",
" key = self._normalize_weights(key) * self.gamma\n",
" value = self._normalize_weights(value) * self.gamma\n",
"\n",
" return super(SpectralNormalizedAttention, self).call(query, value, key, attention_mask, return_attention_scores)\n",
"\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "663d4c7e",
"metadata": {},
"outputs": [],
"source": [
"class BaseModel(tf.keras.Model):\n",
" \"\"\"\n",
" A base model class that integrates various enhancements including \n",
" Reversible Instance Normalization and Channel-Wise Attention, and optionally,\n",
" spectral normalization and SAM optimization. To use SAMformer, enable use_sam,\n",
" use_attention, use_revin and trainable.\n",
" \n",
" Attributes:\n",
" pred_len (int): The length of the output predictions.\n",
" num_heads (int): The number of heads in the multi-head attention mechanism. \n",
" d_model (int): The dimensionality of the embedding vectors.\n",
" use_sam (bool): If True, applies Sharpness-Aware Minimization (SAM) optimization technique during training, \n",
" aiming to improve model generalization by considering the loss landscape's sharpness.\n",
" use_attention (bool): If True, enables the multi-head attention mechanism in the model. If False, the model is\n",
" equivalent to a simple linear layer.\n",
" use_revin (bool): If True, applies Reversible Instance Normalization (RevIN) to the model.\n",
" trainable (bool): Specifies if the model's weights should be updated or frozen during training. Useful to \n",
" highlight some attention layer issues in Time Series Forecasting.\n",
" rho (float): The neighborhood size parameter for SAM optimization. It determines the radius within which SAM \n",
" seeks to minimize the sharpness of the loss landscape.\n",
" spec (bool): If True, applies spectral normalization (sigma-reparam) to the attention mechanism, aiming to\n",
" stabilize the training by constraining the spectral norm of the weight matrices.\n",
" \n",
" Methods:\n",
" call(inputs, training=False): Defines the computation from inputs to outputs, optionally applying SAM, \n",
" spectral normalization, and reversible instance normalization based on the \n",
" configuration.\n",
" _apply_attention(x): Applies the attention mechanism to the input tensor, capturing the inter-dependencies \n",
" within the data thanks to the Channel-Wise Attention mechanism.\n",
" get_last_attention_weights(): Retrieves the attention weights from the last but one batch, useful for \n",
" analysis and debugging purposes.\n",
" train_step(data): Custom training logic, including the application of SAM's two-step optimization process, \n",
" to improve model generalization and performance stability.\n",
"\n",
" \"\"\"\n",
"\n",
" def __init__(self, pred_len, num_heads=1, d_model=16, use_sam=None, \n",
" use_attention=None, use_revin=None, \n",
" trainable=None, rho=None, spec=None):\n",
" super(BaseModel, self).__init__()\n",
" self.pred_len = pred_len\n",
" self.num_heads = num_heads\n",
" self.d_model = d_model\n",
" self.use_sam = use_sam\n",
" self.use_attention = use_attention\n",
" self.use_revin = use_revin\n",
" self.rho = rho if use_sam and trainable else 0.0\n",
" self.spec = spec\n",
"\n",
" # Define model layers\n",
" self.rev_norm = RevNorm(axis=-2)\n",
" self.attention_layer = layers.MultiHeadAttention(num_heads=num_heads, key_dim=d_model)\n",
" self.dense = layers.Dense(pred_len)\n",
" self.all_attention_weights = collections.deque(maxlen=2)\n",
" self.all_dense_weights = collections.deque(maxlen=2)\n",
"\n",
" if self.spec:\n",
" self.spec_layer = SpectralNormalizedAttention(num_heads=num_heads, key_dim=d_model)\n",
"\n",
" #Define trainability of attention layer\n",
" self.attention_layer.trainable = trainable\n",
"\n",
" def call(self, inputs, training=False):\n",
" \"\"\"\n",
" The forward pass for the model.\n",
" \n",
" Parameters:\n",
" inputs (Tensor): Input tensor.\n",
" training (bool): Whether the call is for training.\n",
" \n",
" Returns:\n",
" Tensor: The output of the model.\n",
" \"\"\"\n",
"\n",
" x = inputs\n",
" if self.use_revin:\n",
" x = self.rev_norm(x, mode='norm')\n",
" x = tf.transpose(x, perm=[0, 2, 1])\n",
"\n",
" if self.use_attention:\n",
" attention_output = self._apply_attention(x)\n",
" x = layers.Add()([x, attention_output])\n",
"\n",
" x = self.dense(x)\n",
" outputs = tf.transpose(x, perm=[0, 2, 1])\n",
"\n",
" if self.use_revin:\n",
" outputs = self.rev_norm(outputs, mode='denorm')\n",
"\n",
" return outputs\n",
"\n",
" def _apply_attention(self, x):\n",
" \"\"\"\n",
" Applies the attention mechanism to the input tensor.\n",
" \n",
" Parameters:\n",
" x (Tensor): The input tensor.\n",
" training (bool): Whether the call is for training.\n",
" \n",
" Returns:\n",
" Tensor: The output tensor after applying attention.\n",
" \"\"\"\n",
" if self.spec:\n",
" attention_output, weights = self.spec_layer(x, x, return_attention_scores=True)\n",
" else:\n",
" attention_output, weights = self.attention_layer(x, x, return_attention_scores=True)\n",
" \n",
" self.all_attention_weights.append(weights.numpy())\n",
" return attention_output\n",
"\n",
" def get_last_attention_weights(self):\n",
" \"\"\"Returns the attention weights from the last but one batch.\"\"\"\n",
" if len(self.all_attention_weights) > 1:\n",
" return self.all_attention_weights[-2]\n",
" return None\n",
"\n",
" def train_step(self, data):\n",
" sam_optimizer = SAM(self.optimizer, rho=self.rho, eps=1e-12) \n",
"\n",
" # Unpack the data.\n",
" x, y = data\n",
"\n",
" with tf.GradientTape() as tape:\n",
" y_pred = self(x, training=True) # Forward pass\n",
" loss = self.compiled_loss(y, y_pred, regularization_losses=self.losses)\n",
"\n",
" # Compute gradients\n",
" gradients = tape.gradient(loss, self.trainable_variables)\n",
"\n",
" # Apply SAM's first step\n",
" sam_optimizer.first_step(gradients, self.trainable_variables)\n",
"\n",
" with tf.GradientTape() as tape:\n",
" y_pred = self(x, training=True) # Forward pass again\n",
" loss = self.compiled_loss(y, y_pred, regularization_losses=self.losses)\n",
"\n",
" # Compute gradients again\n",
" gradients = tape.gradient(loss, self.trainable_variables)\n",
"\n",
" # Apply SAM's second step\n",
" sam_optimizer.second_step(gradients, self.trainable_variables)\n",
"\n",
" # Update metrics\n",
" self.compiled_metrics.update_state(y, y_pred)\n",
"\n",
" # Return a dict mapping metric names to current value\n",
" return {m.name: m.result() for m in self.metrics}\n",
" "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cdc39fe9",
"metadata": {},
"outputs": [],
"source": [
"class TSFDataLoader:\n",
" \"\"\"Generate data loader from raw data.\"\"\"\n",
"\n",
" def __init__(\n",
" self, data, batch_size, seq_len, pred_len, feature_type, target='OT'\n",
" ):\n",
" self.data = data\n",
" self.batch_size = batch_size\n",
" self.seq_len = seq_len\n",
" self.pred_len = pred_len\n",
" self.feature_type = feature_type\n",
" self.target = target\n",
" self.target_slice = slice(0, None)\n",
"\n",
" self._read_data()\n",
"\n",
" def _read_data(self):\n",
" \"\"\"Load raw data and split datasets.\"\"\"\n",
"\n",
" # copy data from cloud storage if not exists\n",
" if not os.path.isdir(LOCAL_CACHE_DIR):\n",
" os.mkdir(LOCAL_CACHE_DIR)\n",
"\n",
" file_name = self.data + '.csv'\n",
" cache_filepath = os.path.join(LOCAL_CACHE_DIR, file_name)\n",
" if not os.path.isfile(cache_filepath):\n",
" tf.io.gfile.copy(\n",
" os.path.join(DATA_DIR, file_name), cache_filepath, overwrite=True\n",
" )\n",
" df_raw = pd.read_csv(cache_filepath)\n",
"\n",
" \n",
" \n",
" # S: univariate-univariate, M: multivariate-multivariate, MS:\n",
" # multivariate-univariate\n",
" df = df_raw.set_index('date')\n",
" if self.feature_type == 'S':\n",
" df = df[[self.target]]\n",
" elif self.feature_type == 'MS':\n",
" target_idx = df.columns.get_loc(self.target)\n",
" self.target_slice = slice(target_idx, target_idx + 1)\n",
"\n",
" # split train/valid/test\n",
" n = len(df)\n",
" if self.data.startswith('ETTm'):\n",
" train_end = 12 * 30 * 24 * 4\n",
" val_end = train_end + 4 * 30 * 24 * 4\n",
" test_end = val_end + 4 * 30 * 24 * 4\n",
" elif self.data.startswith('ETTh'):\n",
" train_end = 12 * 30 * 24\n",
" val_end = train_end + 4 * 30 * 24\n",
" test_end = val_end + 4 * 30 * 24\n",
" else:\n",
" train_end = int(n * 0.7)\n",
" val_end = n - int(n * 0.2)\n",
" test_end = n\n",
" \n",
" train_df = df[:train_end]\n",
" val_df = df[train_end - self.seq_len : val_end]\n",
" test_df = df[val_end - self.seq_len : test_end]\n",
"\n",
" # standardize by training set\n",
" self.scaler = StandardScaler()\n",
" self.scaler1 = StandardScaler()\n",
" \n",
" self.scaler.fit(train_df.values)\n",
" # self.scaler1.fit(df.iloc[:train_end, :-5].values)\n",
"\n",
" def scale_df(df, scaler):\n",
" data = scaler.transform(df.values)\n",
" # return pd.DataFrame(df, index=df.index, columns=df.columns)\n",
" return pd.DataFrame(data, index=df.index, columns=df.columns)\n",
"\n",
" self.train_df = scale_df(train_df, self.scaler)\n",
" self.val_df = scale_df(val_df, self.scaler)\n",
" self.test_df = scale_df(test_df, self.scaler)\n",
" self.n_feature = self.train_df.shape[-1]\n",
"\n",
" def _split_window(self, data):\n",
" inputs = data[:, : self.seq_len, :]\n",
" labels = data[:, self.seq_len :, self.target_slice]\n",
" # Slicing doesn't preserve static shape information, so set the shapes\n",
" # manually. This way the `tf.data.Datasets` are easier to inspect.\n",
" inputs.set_shape([None, self.seq_len, None])\n",
" labels.set_shape([None, self.pred_len, self.n_feature])\n",
" return inputs, labels\n",
"\n",
" def _make_dataset(self, data, shuffle=True):\n",
" data = np.array(data, dtype=np.float32)\n",
" ds = tf.keras.utils.timeseries_dataset_from_array(\n",
" data=data,\n",
" targets=None,\n",
" sequence_length=(self.seq_len + self.pred_len),\n",
" sequence_stride=1,\n",
" shuffle=False,\n",
" batch_size=self.batch_size,\n",
" )\n",
" ds = ds.map(self._split_window)\n",
" return ds\n",
"\n",
" def _make_dataset_test(self, data, shuffle=True):\n",
" data = np.array(data, dtype=np.float32)\n",
" ds = tf.keras.utils.timeseries_dataset_from_array(\n",
" data=data,\n",
" targets=None,\n",
" sequence_length=(self.seq_len + self.pred_len),\n",
" sequence_stride=1,\n",
" shuffle=False,\n",
" batch_size=1)\n",
" ds = ds.map(self._split_window)\n",
" return ds\n",
"\n",
"\n",
" def inverse_transform(self, data):\n",
" return self.scaler.inverse_transform(data)\n",
"\n",
" def get_train(self, shuffle=True):\n",
" return self._make_dataset(self.train_df, shuffle=shuffle)\n",
"\n",
" def get_val(self):\n",
" return self._make_dataset(self.val_df, shuffle=False)\n",
"\n",
" def get_test(self):\n",
" return self._make_dataset(self.test_df, shuffle=False)\n",
"\n",
"\n",
"def load_data(name, batch_size, seq_len, pred_len, feature_type, target):\n",
" \"\"\"\n",
" Loads or generates training, validation, and testing datasets based on the specified configurations.\n",
"\n",
" Parameters:\n",
" args (argparse.Namespace): Command line arguments specifying dataset configurations.\n",
"\n",
" Returns:\n",
" tuple: Training, validation, and test datasets as tf.data.Dataset objects.\n",
" \"\"\"\n",
" data_loader = TSFDataLoader(name, batch_size, seq_len, pred_len, feature_type, target)\n",
" train_data, val_data, test_data = data_loader.get_train(), data_loader.get_val(), data_loader.get_test()\n",
"\n",
" return train_data, val_data, test_data, data_loader.n_feature"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "fe7abdc1",
"metadata": {},
"outputs": [],
"source": [
"train_data, val_data, test_data, n_features = load_data(name, batch_size, seq_len, pred_len, feature_type, target)\n",
"new_cols = ['mse', 'mae', \"seq_len\", \"learning_rate\", \"batch_size\", \"num_heads\", \n",
" \"d_model\", \"minimization_neighbourhood\"]\n",
"\n",
"df = pd.DataFrame(columns=new_cols)\n",
"\n",
"numEvalSamples = 1\n",
"seeds = [2022]"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "29470575",
"metadata": {},
"outputs": [],
"source": [
"for sd in range(numEvalSamples):\n",
" tf.keras.backend.clear_session() \n",
"\n",
" tf.keras.utils.set_random_seed(seeds[sd])\n",
" \n",
" print(f\"..............iteration{sd}.................\")\n",
" print()\n",
" model = BaseModel(pred_len, num_heads=num_heads, d_model=d_model, use_sam=True, use_attention=True, use_revin=True, trainable=True, rho=rho, \n",
" spec=False)\n",
" optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)\n",
" model.compile(optimizer=optimizer, loss='mae', metrics=['mse', 'mae'], run_eagerly=True)\n",
" \n",
" callbacks, capture_weights_callback = setup_callbacks(learning_rate, patience, checkpoint_path, model)\n",
" \n",
" history = model.fit(\n",
" train_data,\n",
" epochs=300,\n",
" validation_data=val_data,\n",
" callbacks=callbacks)\n",
" \n",
" model.load_weights(checkpoint_path)\n",
" df.loc[sd, :] = model.evaluate(test_data)[-2:] + [ seq_len, learning_rate, batch_size, num_heads, \n",
" d_model, rho]\n",
" df.to_csv(name_df, header=True, index=False)\n",
"\n",
" del model\n",
" tf.keras.backend.clear_session()\n",
" # garbage collector\n",
" gc.collect()\n",
"\n",
" # remove_dir(checkpoint_path)\n",
"\n",
"print(\"................completed Successfully......................\")\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "852cd02e",
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.11.5"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
|