fmda/fmda_rnn_spatial.ipynb

   1 {
   2  "cells": [
   3   {
   4    "cell_type": "markdown",
   5    "id": "83b774b3-ef55-480a-b999-506676e49145",
   6    "metadata": {},
   7    "source": [
   8     "# v2.1 run RNN with Spatial Training\n",
   9     "\n",
  10     "This notebook is intended to set up a test where the RNN is run serial by location and compared to the spatial training scheme. Additionally, the ODE model with the augmented KF will be run as a comparison, but note that the RNN models will be predicting entirely without knowledge of the heldout locations, while the augmented KF will be run directly on the test locations.\n"
  11    ]
  12   },
  13   {
  14    "cell_type": "markdown",
  15    "id": "bbd84d61-a9cd-47b4-b538-4986fb10b98d",
  16    "metadata": {},
  17    "source": [
  18     "## Environment Setup"
  19    ]
  20   },
  21   {
  22    "cell_type": "code",
  23    "execution_count": null,
  24    "id": "83cc1dc4-3dcb-4325-9263-58101a3dc378",
  25    "metadata": {},
  26    "outputs": [],
  27    "source": [
  28     "import numpy as np\n",
  29     "from utils import print_dict_summary, print_first, str2time, logging_setup\n",
  30     "import pickle\n",
  31     "import logging\n",
  32     "import os.path as osp\n",
  33     "from moisture_rnn_pkl import pkl2train\n",
  34     "from moisture_rnn import RNNParams, RNNData, RNN \n",
  35     "from utils import hash2, read_yml, read_pkl, retrieve_url, Dict\n",
  36     "from moisture_rnn import RNN\n",
  37     "import reproducibility\n",
  38     "from data_funcs import rmse, to_json, combine_nested, process_train_dict\n",
  39     "from moisture_models import run_augmented_kf\n",
  40     "import copy\n",
  41     "import pandas as pd\n",
  42     "import matplotlib.pyplot as plt\n",
  43     "import yaml\n",
  44     "import time"
  45    ]
  46   },
  47   {
  48    "cell_type": "code",
  49    "execution_count": null,
  50    "id": "17db9b90-a931-4674-a447-5b8ffbcdc86a",
  51    "metadata": {},
  52    "outputs": [],
  53    "source": [
  54     "logging_setup()"
  55    ]
  56   },
  57   {
  58    "cell_type": "code",
  59    "execution_count": null,
  60    "id": "35319c1c-7849-4b8c-8262-f5aa6656e0c7",
  61    "metadata": {},
  62    "outputs": [],
  63    "source": [
  64     "retrieve_url(\n",
  65     "    url = \"https://demo.openwfm.org/web/data/fmda/dicts/fmda_nw_202401-05_f05.pkl\", \n",
  66     "    dest_path = \"data/fmda_nw_202401-05_f05.pkl\")"
  67    ]
  68   },
  69   {
  70    "cell_type": "code",
  71    "execution_count": null,
  72    "id": "eabdbd9c-07d9-4bae-9851-cca79f321895",
  73    "metadata": {},
  74    "outputs": [],
  75    "source": [
  76     "file_paths = ['data/fmda_nw_202401-05_f05.pkl']"
  77    ]
  78   },
  79   {
  80    "cell_type": "code",
  81    "execution_count": null,
  82    "id": "dcca6185-e799-4dd1-8acb-87ad33c411d7",
  83    "metadata": {},
  84    "outputs": [],
  85    "source": [
  86     "# read/write control\n",
  87     "train_file='data/train.pkl'\n",
  88     "train_create=True   # if false, read\n",
  89     "train_write=True\n",
  90     "train_read=True"
  91    ]
  92   },
  93   {
  94    "cell_type": "code",
  95    "execution_count": null,
  96    "id": "604388de-11ab-45c3-9f0d-80bdff0cca60",
  97    "metadata": {},
  98    "outputs": [],
  99    "source": [
 100     "# Params used for data filtering\n",
 101     "params_data = read_yml(\"params_data.yaml\") \n",
 102     "params_data"
 103    ]
 104   },
 105   {
 106    "cell_type": "code",
 107    "execution_count": null,
 108    "id": "211a1c2f-ba8d-40b8-b29c-daa38af97a26",
 109    "metadata": {},
 110    "outputs": [],
 111    "source": [
 112     "# Params used for setting up RNN\n",
 113     "params = read_yml(\"params.yaml\", subkey='rnn') \n",
 114     "params"
 115    ]
 116   },
 117   {
 118    "cell_type": "code",
 119    "execution_count": null,
 120    "id": "bc0a775b-b587-42ef-8576-e36dc0be3a75",
 121    "metadata": {
 122     "scrolled": true
 123    },
 124    "outputs": [],
 125    "source": [
 126     "if train_create:\n",
 127     "    logging.info('creating the training cases from files %s',file_paths)\n",
 128     "    # osp.join works on windows too, joins paths using \\ or /\n",
 129     "    train = process_train_dict(file_paths, params_data = params_data, verbose=True)\n",
 130     "if train_write:\n",
 131     "    with open(train_file, 'wb') as file:\n",
 132     "        logging.info('Writing the rain cases into file %s',train_file)\n",
 133     "        pickle.dump(train, file)\n",
 134     "if train_read:\n",
 135     "    logging.info('Reading the train cases from file %s',train_file)\n",
 136     "    train = read_pkl(train_file)"
 137    ]
 138   },
 139   {
 140    "cell_type": "code",
 141    "execution_count": null,
 142    "id": "23cd60c0-9865-4314-9a96-948c3d400c08",
 143    "metadata": {},
 144    "outputs": [],
 145    "source": [
 146     "from itertools import islice\n",
 147     "train = {k: train[k] for k in islice(train, 200)}"
 148    ]
 149   },
 150   {
 151    "cell_type": "markdown",
 152    "id": "efc10cdc-f18b-4781-84da-b8e2eef39981",
 153    "metadata": {},
 154    "source": [
 155     "## Setup Validation Runs"
 156    ]
 157   },
 158   {
 159    "cell_type": "markdown",
 160    "id": "2d9cd5c5-87ed-41f9-b36c-e0c18d58c841",
 161    "metadata": {},
 162    "source": [
 163     "The following parameters will be used for both serial and spatial models."
 164    ]
 165   },
 166   {
 167    "cell_type": "code",
 168    "execution_count": null,
 169    "id": "66f40c9f-c1c2-4b12-bf14-2ada8c26113d",
 170    "metadata": {},
 171    "outputs": [],
 172    "source": [
 173     "params = RNNParams(params)\n",
 174     "params.update({'epochs': 200, \n",
 175     "               'learning_rate': 0.001,\n",
 176     "               'activation': ['tanh', 'tanh'], # Activation for RNN Layers, Dense layers respectively.\n",
 177     "               'recurrent_layers': 2, 'recurrent_units': 30, \n",
 178     "               'dense_layers': 2, 'dense_units': 30,\n",
 179     "               'early_stopping_patience': 30, # how many epochs of no validation accuracy gain to wait before stopping\n",
 180     "               'batch_schedule_type': 'exp', # Hidden state batch reset schedule\n",
 181     "               'bmin': 20, # Lower bound of hidden state batch reset, \n",
 182     "               'bmax': params_data['hours'], # Upper bound of hidden state batch reset, using max hours\n",
 183     "               'features_list': ['Ed', 'Ew', 'rain', 'elev', 'lon', 'lat', 'solar', 'wind']\n",
 184     "              })"
 185    ]
 186   },
 187   {
 188    "cell_type": "code",
 189    "execution_count": null,
 190    "id": "36823193-b93c-421e-b699-8c1ae5719309",
 191    "metadata": {},
 192    "outputs": [],
 193    "source": [
 194     "reproducibility.set_seed(123)"
 195    ]
 196   },
 197   {
 198    "cell_type": "markdown",
 199    "id": "a24d76fc-6c25-43e7-99df-3cd5dbf84fc3",
 200    "metadata": {},
 201    "source": [
 202     "## Spatial Data Training\n",
 203     "\n",
 204     "This method combines the training timeseries data into a single 3-d array, with timeseries at the same location arranged appropriately in the right order for a given `batch_size` hyperparameter. The hidden state of the recurrent layers are set up reset when the location changes. "
 205    ]
 206   },
 207   {
 208    "cell_type": "code",
 209    "execution_count": null,
 210    "id": "3b5371a9-c1e8-4df5-b360-210746f7cd52",
 211    "metadata": {},
 212    "outputs": [],
 213    "source": [
 214     "# Start timer for code \n",
 215     "start_time = time.time()"
 216    ]
 217   },
 218   {
 219    "cell_type": "code",
 220    "execution_count": null,
 221    "id": "faf93470-b55f-4770-9fa9-3288a2f13fcc",
 222    "metadata": {},
 223    "outputs": [],
 224    "source": [
 225     "# Combine Nested Dictionary into Spatial Data\n",
 226     "train_sp = Dict(combine_nested(train))"
 227    ]
 228   },
 229   {
 230    "cell_type": "code",
 231    "execution_count": null,
 232    "id": "c0c7f5fb-4c33-45f8-9a2e-38c9ab1cd4e3",
 233    "metadata": {},
 234    "outputs": [],
 235    "source": [
 236     "rnn_dat_sp = RNNData(\n",
 237     "    train_sp, # input dictionary\n",
 238     "    scaler=\"standard\",  # data scaling type\n",
 239     "    features_list = params['features_list'] # features for predicting outcome\n",
 240     ")\n",
 241     "\n",
 242     "\n",
 243     "rnn_dat_sp.train_test_split(   \n",
 244     "    time_fracs = [.8, .1, .1], # Percent of total time steps used for train/val/test\n",
 245     "    space_fracs = [.8, .1, .1] # Percent of total timeseries used for train/val/test\n",
 246     ")\n",
 247     "rnn_dat_sp.scale_data()\n",
 248     "\n",
 249     "rnn_dat_sp.batch_reshape(\n",
 250     "    timesteps = params['timesteps'], # Timesteps aka sequence length for RNN input data. \n",
 251     "    batch_size = params['batch_size'] # Number of samples of length timesteps for a single round of grad. descent\n",
 252     ")"
 253    ]
 254   },
 255   {
 256    "cell_type": "code",
 257    "execution_count": null,
 258    "id": "7431bc95-d384-40fd-a622-bbc0ee68e5cd",
 259    "metadata": {},
 260    "outputs": [],
 261    "source": [
 262     "# Update Params specific to spatial training\n",
 263     "params.update({\n",
 264     "    'loc_batch_reset': rnn_dat_sp.n_seqs # Used to reset hidden state when location changes for a given batch\n",
 265     "})"
 266    ]
 267   },
 268   {
 269    "cell_type": "code",
 270    "execution_count": null,
 271    "id": "4bc11474-fed8-47f2-b9cf-dfdda0d3d3b2",
 272    "metadata": {},
 273    "outputs": [],
 274    "source": [
 275     "rnn_sp = RNN(params)\n",
 276     "m_sp, errs = rnn_sp.run_model(rnn_dat_sp)"
 277    ]
 278   },
 279   {
 280    "cell_type": "code",
 281    "execution_count": null,
 282    "id": "704ad662-d81a-488d-be3d-e90bf775a5b8",
 283    "metadata": {},
 284    "outputs": [],
 285    "source": [
 286     "errs.mean()"
 287    ]
 288   },
 289   {
 290    "cell_type": "code",
 291    "execution_count": null,
 292    "id": "d53571e3-b6cf-49aa-9848-e3c77053283d",
 293    "metadata": {},
 294    "outputs": [],
 295    "source": [
 296     "# End Timer\n",
 297     "end_time = time.time()\n",
 298     "\n",
 299     "# Calculate Code Runtime\n",
 300     "elapsed_time_sp = end_time - start_time\n",
 301     "print(f\"Spatial Training Elapsed time: {elapsed_time_sp:.4f} seconds\")"
 302    ]
 303   },
 304   {
 305    "cell_type": "markdown",
 306    "id": "7d8292a2-418c-48ed-aff7-ccbe98b046d3",
 307    "metadata": {},
 308    "source": [
 309     "## Run ODE + KF "
 310    ]
 311   },
 312   {
 313    "cell_type": "code",
 314    "execution_count": null,
 315    "id": "cca12d8c-c0e1-4df4-b2ca-20440485f2f3",
 316    "metadata": {},
 317    "outputs": [],
 318    "source": [
 319     "# Get timeseries IDs from previous RNNData object\n",
 320     "test_cases = rnn_dat_sp.loc['test_locs']\n",
 321     "print(len(test_cases))"
 322    ]
 323   },
 324   {
 325    "cell_type": "code",
 326    "execution_count": null,
 327    "id": "997f2534-7e77-45b3-93bf-d988837dfc0b",
 328    "metadata": {},
 329    "outputs": [],
 330    "source": [
 331     "test_ind = rnn_dat_sp.test_ind # Time index for test period start\n",
 332     "print(test_ind)"
 333    ]
 334   },
 335   {
 336    "cell_type": "code",
 337    "execution_count": null,
 338    "id": "1e4ffc68-c775-41c6-ac42-f49c76824b43",
 339    "metadata": {
 340     "scrolled": true
 341    },
 342    "outputs": [],
 343    "source": [
 344     "outputs_kf = {}\n",
 345     "for case in test_cases:\n",
 346     "    print(\"~\"*50)\n",
 347     "    print(case)\n",
 348     "    # Run Augmented KF\n",
 349     "    print('Running Augmented KF')\n",
 350     "    train[case]['h2'] = test_ind\n",
 351     "    train[case]['scale_fm'] = 1\n",
 352     "    m, Ec = run_augmented_kf(train[case])\n",
 353     "    y = train[case]['y']        \n",
 354     "    train[case]['m_kf'] = m\n",
 355     "    print(f\"KF RMSE: {rmse(m[test_ind:],y[test_ind:])}\")\n",
 356     "    outputs_kf[case] = {'case':case, 'errs': rmse(m[test_ind:],y[test_ind:])}"
 357    ]
 358   },
 359   {
 360    "cell_type": "code",
 361    "execution_count": null,
 362    "id": "57b19ec5-23f6-44ec-9f71-16d4d69aec68",
 363    "metadata": {},
 364    "outputs": [],
 365    "source": [
 366     "df_kf = pd.DataFrame.from_dict(outputs_kf).transpose()\n",
 367     "df_kf.head()"
 368    ]
 369   },
 370   {
 371    "cell_type": "code",
 372    "execution_count": null,
 373    "id": "25a9d2fe-83f7-4ef3-a04b-14c970b6e2ba",
 374    "metadata": {},
 375    "outputs": [],
 376    "source": [
 377     "df_kf.errs.mean()"
 378    ]
 379   },
 380   {
 381    "cell_type": "markdown",
 382    "id": "f616bbf8-d89e-4c5b-9e47-59f02246b6f2",
 383    "metadata": {},
 384    "source": [
 385     "## Serial Training\n",
 386     "\n",
 387     "This method initializes a RNN and uses successive `.fit` calls to train the model one location at a time. This is the naive approach to training a RNN on multiple timeseries, and is used as a baseline to see whether the spatial training scheme improves things."
 388    ]
 389   },
 390   {
 391    "cell_type": "code",
 392    "execution_count": null,
 393    "id": "6fa20e9f-604a-4938-ab68-b71fbb7326df",
 394    "metadata": {},
 395    "outputs": [],
 396    "source": [
 397     "# Start timer for code \n",
 398     "start_time = time.time()"
 399    ]
 400   },
 401   {
 402    "cell_type": "code",
 403    "execution_count": null,
 404    "id": "f033e78c-a506-4508-a23c-8e6574014872",
 405    "metadata": {},
 406    "outputs": [],
 407    "source": [
 408     "# Update Params specific to Serial training\n",
 409     "params.update({\n",
 410     "    'loc_batch_reset': None, # Used to reset hidden state when location changes for a given batch\n",
 411     "    'epochs': 1 # less epochs since fit will be run multiple times over locations\n",
 412     "})"
 413    ]
 414   },
 415   {
 416    "cell_type": "code",
 417    "execution_count": null,
 418    "id": "ff1788ec-081b-403f-bcfa-b625f0e3dbe1",
 419    "metadata": {},
 420    "outputs": [],
 421    "source": [
 422     "train_cases = rnn_dat_sp.loc['train_locs']\n",
 423     "test_cases = rnn_dat_sp.loc['test_locs']"
 424    ]
 425   },
 426   {
 427    "cell_type": "code",
 428    "execution_count": null,
 429    "id": "8a2af45e-e81b-421f-b940-e8779177dd5d",
 430    "metadata": {},
 431    "outputs": [],
 432    "source": [
 433     "# Initialize Model with first train case\n",
 434     "rnn_dat = RNNData(train[train_cases[0]], params['scaler'], params['features_list'])\n",
 435     "rnn_dat.train_test_split(\n",
 436     "    time_fracs = [.8, .1, .1]\n",
 437     ")\n",
 438     "rnn_dat.scale_data()\n",
 439     "rnn_dat.batch_reshape(timesteps = params['timesteps'], batch_size = params['batch_size'])"
 440    ]
 441   },
 442   {
 443    "cell_type": "code",
 444    "execution_count": null,
 445    "id": "ac6fecc2-f614-4506-b5f9-05a6eca3b62e",
 446    "metadata": {},
 447    "outputs": [],
 448    "source": [
 449     "reproducibility.set_seed()\n",
 450     "rnn = RNN(params)"
 451    ]
 452   },
 453   {
 454    "cell_type": "code",
 455    "execution_count": null,
 456    "id": "79b5af30-7d52-410c-9595-e89e9756fd38",
 457    "metadata": {
 458     "scrolled": true
 459    },
 460    "outputs": [],
 461    "source": [
 462     "# Train\n",
 463     "for case in train_cases:\n",
 464     "    print(\"~\"*50)\n",
 465     "    print(f\"Training with Case {case}\")\n",
 466     "    rnn_dat_temp = RNNData(train[case], params['scaler'], params['features_list'])\n",
 467     "    rnn_dat_temp.train_test_split(\n",
 468     "        time_fracs = [.8, .1, .1]\n",
 469     "    )\n",
 470     "    rnn_dat_temp.scale_data()\n",
 471     "    rnn_dat_temp.batch_reshape(timesteps = params['timesteps'], batch_size = params['batch_size'])\n",
 472     "    rnn.fit(rnn_dat_temp['X_train'], rnn_dat_temp['y_train'],\n",
 473     "           validation_data=(rnn_dat_temp['X_val'], rnn_dat_temp['y_val']))    "
 474    ]
 475   },
 476   {
 477    "cell_type": "code",
 478    "execution_count": null,
 479    "id": "03d716b4-0ff5-4b80-a241-440543ba9b46",
 480    "metadata": {
 481     "scrolled": true
 482    },
 483    "outputs": [],
 484    "source": [
 485     "# Predict\n",
 486     "outputs_rnn_serial = {}\n",
 487     "test_ind = rnn_dat.test_ind\n",
 488     "for i, case in enumerate(test_cases):\n",
 489     "    print(\"~\"*50)\n",
 490     "    rnn_dat_temp = RNNData(train[case], params['scaler'], params['features_list'])\n",
 491     "    rnn_dat_temp.train_test_split(\n",
 492     "        time_fracs = [.8, .1, .1]\n",
 493     "    )\n",
 494     "    rnn_dat_temp.scale_data()\n",
 495     "    rnn_dat_temp.batch_reshape(timesteps = params['timesteps'], batch_size = params['batch_size'])    \n",
 496     "    X_temp = rnn_dat_temp.scale_all_X()\n",
 497     "    m = rnn.predict(X_temp)\n",
 498     "    outputs_rnn_serial[case] = {'case':case, 'errs': rmse(m[test_ind:], rnn_dat.y_test)}"
 499    ]
 500   },
 501   {
 502    "cell_type": "code",
 503    "execution_count": null,
 504    "id": "e5a80bae-fe1a-4ec9-b9ac-31d540eaba40",
 505    "metadata": {},
 506    "outputs": [],
 507    "source": [
 508     "df_rnn_serial = pd.DataFrame.from_dict(outputs_rnn_serial).transpose()\n",
 509     "df_rnn_serial.head()"
 510    ]
 511   },
 512   {
 513    "cell_type": "code",
 514    "execution_count": null,
 515    "id": "0c5b866e-c2bf-4bc1-8f6f-3ba8a9448d07",
 516    "metadata": {},
 517    "outputs": [],
 518    "source": [
 519     "df_rnn_serial.errs.mean()"
 520    ]
 521   },
 522   {
 523    "cell_type": "code",
 524    "execution_count": null,
 525    "id": "f5a364cb-01bf-49ad-a704-5aa3c9564967",
 526    "metadata": {},
 527    "outputs": [],
 528    "source": [
 529     "# End Timer\n",
 530     "end_time = time.time()\n",
 531     "\n",
 532     "# Calculate Code Runtime\n",
 533     "elapsed_time_ser = end_time - start_time\n",
 534     "print(f\"Serial Training Elapsed time: {elapsed_time_ser:.4f} seconds\")"
 535    ]
 536   },
 537   {
 538    "cell_type": "markdown",
 539    "id": "86795281-f8ea-4141-81ea-c53fae830e80",
 540    "metadata": {},
 541    "source": [
 542     "## Compare"
 543    ]
 544   },
 545   {
 546    "cell_type": "code",
 547    "execution_count": null,
 548    "id": "508a6392-49bc-4471-ad8e-814f60119283",
 549    "metadata": {},
 550    "outputs": [],
 551    "source": [
 552     "print(f\"Total Test Cases: {len(test_cases)}\")\n",
 553     "print(f\"Total Test Hours: {rnn_dat_temp.y_test.shape[0]}\")"
 554    ]
 555   },
 556   {
 557    "cell_type": "code",
 558    "execution_count": null,
 559    "id": "73e8ca05-d17b-4e72-8def-fa77664e7bb0",
 560    "metadata": {},
 561    "outputs": [],
 562    "source": [
 563     "print(f\"Spatial Training RMSE: {errs.mean()}\")\n",
 564     "print(f\"Serial Training RMSE: {df_rnn_serial.errs.mean()}\")\n",
 565     "print(f\"Augmented KF RMSE: {df_kf.errs.mean()}\")"
 566    ]
 567   },
 568   {
 569    "cell_type": "code",
 570    "execution_count": null,
 571    "id": "a73d22ee-707b-44a3-80ab-ad6e671731cf",
 572    "metadata": {},
 573    "outputs": [],
 574    "source": []
 575   },
 576   {
 577    "cell_type": "code",
 578    "execution_count": null,
 579    "id": "272bfb32-e8e2-49dd-8f90-4b5b09c3a2a2",
 580    "metadata": {},
 581    "outputs": [],
 582    "source": [
 583     "print(f\"Spatial Training Elapsed time: {elapsed_time_sp:.4f} seconds\")\n",
 584     "print(f\"Serial Training Elapsed time: {elapsed_time_ser:.4f} seconds\")"
 585    ]
 586   },
 587   {
 588    "cell_type": "code",
 589    "execution_count": null,
 590    "id": "38ab08fb-ac97-45be-8907-6f9cd124243b",
 591    "metadata": {},
 592    "outputs": [],
 593    "source": []
 594   }
 595  ],
 596  "metadata": {
 597   "kernelspec": {
 598    "display_name": "Python 3 (ipykernel)",
 599    "language": "python",
 600    "name": "python3"
 601   },
 602   "language_info": {
 603    "codemirror_mode": {
 604     "name": "ipython",
 605     "version": 3
 606    },
 607    "file_extension": ".py",
 608    "mimetype": "text/x-python",
 609    "name": "python",
 610    "nbconvert_exporter": "python",
 611    "pygments_lexer": "ipython3",
 612    "version": "3.12.5"
 613   }
 614  },
 615  "nbformat": 4,
 616  "nbformat_minor": 5
 617 }