fmda/fmda_rnn_train_and_save.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     "import tensorflow as tf\n",
  34     "from moisture_rnn_pkl import pkl2train\n",
  35     "from moisture_rnn import RNNParams, RNNData, RNN \n",
  36     "from utils import hash2, read_yml, read_pkl, retrieve_url, Dict\n",
  37     "from moisture_rnn import RNN\n",
  38     "import reproducibility\n",
  39     "from data_funcs import rmse, to_json, combine_nested, subset_by_features, build_train_dict\n",
  40     "from moisture_models import run_augmented_kf\n",
  41     "import copy\n",
  42     "import pandas as pd\n",
  43     "import matplotlib.pyplot as plt\n",
  44     "import yaml\n",
  45     "import time"
  46    ]
  47   },
  48   {
  49    "cell_type": "code",
  50    "execution_count": null,
  51    "id": "17db9b90-a931-4674-a447-5b8ffbcdc86a",
  52    "metadata": {},
  53    "outputs": [],
  54    "source": [
  55     "logging_setup()"
  56    ]
  57   },
  58   {
  59    "cell_type": "code",
  60    "execution_count": null,
  61    "id": "35319c1c-7849-4b8c-8262-f5aa6656e0c7",
  62    "metadata": {},
  63    "outputs": [],
  64    "source": [
  65     "retrieve_url(\n",
  66     "    url = \"https://demo.openwfm.org/web/data/fmda/dicts/fmda_rocky_202403-05_f05.pkl\", \n",
  67     "    dest_path = \"data/fmda_rocky_202403-05_f05.pkl\")"
  68    ]
  69   },
  70   {
  71    "cell_type": "code",
  72    "execution_count": null,
  73    "id": "eabdbd9c-07d9-4bae-9851-cca79f321895",
  74    "metadata": {},
  75    "outputs": [],
  76    "source": [
  77     "file_paths = ['data/fmda_rocky_202403-05_f05.pkl']"
  78    ]
  79   },
  80   {
  81    "cell_type": "code",
  82    "execution_count": null,
  83    "id": "dcca6185-e799-4dd1-8acb-87ad33c411d7",
  84    "metadata": {},
  85    "outputs": [],
  86    "source": [
  87     "# read/write control\n",
  88     "train_file='data/train.pkl'\n",
  89     "train_create=True   # if false, read\n",
  90     "train_write=False\n",
  91     "train_read=False"
  92    ]
  93   },
  94   {
  95    "cell_type": "code",
  96    "execution_count": null,
  97    "id": "604388de-11ab-45c3-9f0d-80bdff0cca60",
  98    "metadata": {},
  99    "outputs": [],
 100    "source": [
 101     "# Params used for data filtering\n",
 102     "params_data = read_yml(\"params_data.yaml\") \n",
 103     "params_data"
 104    ]
 105   },
 106   {
 107    "cell_type": "code",
 108    "execution_count": null,
 109    "id": "211a1c2f-ba8d-40b8-b29c-daa38af97a26",
 110    "metadata": {},
 111    "outputs": [],
 112    "source": [
 113     "# Params used for setting up RNN\n",
 114     "params = read_yml(\"params.yaml\", subkey='rnn') \n",
 115     "params"
 116    ]
 117   },
 118   {
 119    "cell_type": "code",
 120    "execution_count": null,
 121    "id": "38e6bc61-e123-4cc9-bdee-54b051bbb352",
 122    "metadata": {},
 123    "outputs": [],
 124    "source": [
 125     "feats = ['Ed', 'Ew', 'solar', 'wind', 'elev', 'lon', 'lat', 'rain']\n",
 126     "params.update({'features_list': feats})"
 127    ]
 128   },
 129   {
 130    "cell_type": "code",
 131    "execution_count": null,
 132    "id": "ef84104f-9898-4cd9-be54-7c480536ee0e",
 133    "metadata": {
 134     "scrolled": true
 135    },
 136    "outputs": [],
 137    "source": [
 138     "train = build_train_dict(file_paths, atm_source=\"RAWS\", params_data = params_data,\n",
 139     "                         features_subset = feats, spatial=True, verbose=True)"
 140    ]
 141   },
 142   {
 143    "cell_type": "code",
 144    "execution_count": null,
 145    "id": "bc0a775b-b587-42ef-8576-e36dc0be3a75",
 146    "metadata": {
 147     "scrolled": true
 148    },
 149    "outputs": [],
 150    "source": [
 151     "# if train_create:\n",
 152     "#     params_data.update({'hours': 1440})\n",
 153     "#     logging.info('creating the training cases from files %s',file_paths)\n",
 154     "#     # osp.join works on windows too, joins paths using \\ or /\n",
 155     "#     train = process_train_dict(file_paths, atm_dict = \"RAWS\", params_data = params_data, verbose=True)\n",
 156     "#     train = subset_by_features(train, feats)\n",
 157     "#     train = combine_nested(train)\n",
 158     "# if train_write:\n",
 159     "#     with open(train_file, 'wb') as file:\n",
 160     "#         logging.info('Writing the rain cases into file %s',train_file)\n",
 161     "#         pickle.dump(train, file)\n",
 162     "# if train_read:\n",
 163     "#     logging.info('Reading the train cases from file %s',train_file)\n",
 164     "#     train = read_pkl(train_file)"
 165    ]
 166   },
 167   {
 168    "cell_type": "markdown",
 169    "id": "a24d76fc-6c25-43e7-99df-3cd5dbf84fc3",
 170    "metadata": {},
 171    "source": [
 172     "## Spatial Data Training\n",
 173     "\n",
 174     "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. "
 175    ]
 176   },
 177   {
 178    "cell_type": "code",
 179    "execution_count": null,
 180    "id": "36823193-b93c-421e-b699-8c1ae5719309",
 181    "metadata": {},
 182    "outputs": [],
 183    "source": [
 184     "reproducibility.set_seed(123)"
 185    ]
 186   },
 187   {
 188    "cell_type": "code",
 189    "execution_count": null,
 190    "id": "66f40c9f-c1c2-4b12-bf14-2ada8c26113d",
 191    "metadata": {},
 192    "outputs": [],
 193    "source": [
 194     "params = RNNParams(params)\n",
 195     "params.update({'epochs': 200, \n",
 196     "               'learning_rate': 0.001,\n",
 197     "               'activation': ['relu', 'relu'], # Activation for RNN Layers, Dense layers respectively.\n",
 198     "               'recurrent_layers': 1, 'recurrent_units': 30, \n",
 199     "               'dense_layers': 1, 'dense_units': 30,\n",
 200     "               'early_stopping_patience': 30, # how many epochs of no validation accuracy gain to wait before stopping\n",
 201     "               'batch_schedule_type': 'exp', # Hidden state batch reset schedule\n",
 202     "               'bmin': 20, # Lower bound of hidden state batch reset, \n",
 203     "               'bmax': params_data['hours'], # Upper bound of hidden state batch reset, using max hours\n",
 204     "               'batch_size': 60\n",
 205     "              })"
 206    ]
 207   },
 208   {
 209    "cell_type": "code",
 210    "execution_count": null,
 211    "id": "82bc407d-9d26-41e3-8b58-ab3f7238e105",
 212    "metadata": {},
 213    "outputs": [],
 214    "source": [
 215     "import importlib\n",
 216     "import moisture_rnn\n",
 217     "importlib.reload(moisture_rnn)\n",
 218     "from moisture_rnn import RNNData"
 219    ]
 220   },
 221   {
 222    "cell_type": "code",
 223    "execution_count": null,
 224    "id": "c0c7f5fb-4c33-45f8-9a2e-38c9ab1cd4e3",
 225    "metadata": {},
 226    "outputs": [],
 227    "source": [
 228     "rnn_dat_sp = RNNData(\n",
 229     "    train, # input dictionary\n",
 230     "    scaler=\"standard\",  # data scaling type\n",
 231     "    features_list = params['features_list'] # features for predicting outcome\n",
 232     ")\n",
 233     "\n",
 234     "\n",
 235     "rnn_dat_sp.train_test_split(   \n",
 236     "    time_fracs = [.8, .1, .1], # Percent of total time steps used for train/val/test\n",
 237     "    space_fracs = [.8, .1, .1] # Percent of total timeseries used for train/val/test\n",
 238     ")\n",
 239     "rnn_dat_sp.scale_data()\n",
 240     "\n",
 241     "rnn_dat_sp.batch_reshape(\n",
 242     "    timesteps = params['timesteps'], # Timesteps aka sequence length for RNN input data. \n",
 243     "    batch_size = params['batch_size'] # Number of samples of length timesteps for a single round of grad. descent\n",
 244     ")\n",
 245     "# Update Params specific to spatial training\n",
 246     "params.update({\n",
 247     "    'loc_batch_reset': rnn_dat_sp.n_seqs # Used to reset hidden state when location changes for a given batch\n",
 248     "})"
 249    ]
 250   },
 251   {
 252    "cell_type": "code",
 253    "execution_count": null,
 254    "id": "4bc11474-fed8-47f2-b9cf-dfdda0d3d3b2",
 255    "metadata": {},
 256    "outputs": [],
 257    "source": [
 258     "rnn_sp = RNN(params)\n",
 259     "m_sp, errs = rnn_sp.run_model(rnn_dat_sp)"
 260    ]
 261   },
 262   {
 263    "cell_type": "code",
 264    "execution_count": null,
 265    "id": "704ad662-d81a-488d-be3d-e90bf775a5b8",
 266    "metadata": {},
 267    "outputs": [],
 268    "source": [
 269     "errs.mean()"
 270    ]
 271   },
 272   {
 273    "cell_type": "markdown",
 274    "id": "62c1b049-304e-4c90-b1d2-b9b96b9a202f",
 275    "metadata": {},
 276    "source": [
 277     "## Save Model"
 278    ]
 279   },
 280   {
 281    "cell_type": "code",
 282    "execution_count": null,
 283    "id": "f333521f-c724-40bf-8c1c-32735aea52cc",
 284    "metadata": {},
 285    "outputs": [],
 286    "source": [
 287     "outpath = \"data/outputs/models\"\n",
 288     "filename = osp.join(outpath, f\"model_predict_raws_rocky.keras\")\n",
 289     "rnn_sp.model_predict.save(filename)"
 290    ]
 291   },
 292   {
 293    "cell_type": "markdown",
 294    "id": "bc1c601f-23a9-41b0-b921-47f1340f2a47",
 295    "metadata": {},
 296    "source": [
 297     "## Load and Check"
 298    ]
 299   },
 300   {
 301    "cell_type": "code",
 302    "execution_count": null,
 303    "id": "3c27b3c1-6f60-450e-82ea-18eaf012fece",
 304    "metadata": {},
 305    "outputs": [],
 306    "source": [
 307     "mod = tf.keras.models.load_model(filename)"
 308    ]
 309   },
 310   {
 311    "cell_type": "code",
 312    "execution_count": null,
 313    "id": "25bf5420-d681-40ec-9eb8-aed784ca4e5a",
 314    "metadata": {},
 315    "outputs": [],
 316    "source": [
 317     "from utils import hash_weights\n",
 318     "\n",
 319     "hash_weights(mod)"
 320    ]
 321   },
 322   {
 323    "cell_type": "code",
 324    "execution_count": null,
 325    "id": "d773b2ab-18de-4b13-a243-b6353c57f192",
 326    "metadata": {},
 327    "outputs": [],
 328    "source": [
 329     "type(rnn_dat_sp.X_test)"
 330    ]
 331   },
 332   {
 333    "cell_type": "code",
 334    "execution_count": null,
 335    "id": "253ba437-c3a2-452b-b8e6-078aa17c8408",
 336    "metadata": {},
 337    "outputs": [],
 338    "source": [
 339     "X_test = np.stack(rnn_dat_sp.X_test, axis=0)\n",
 340     "y_array = np.stack(rnn_dat_sp.y_test, axis=0)"
 341    ]
 342   },
 343   {
 344    "cell_type": "code",
 345    "execution_count": null,
 346    "id": "f4332dd8-57cd-4f5b-a864-dc72f96d72b2",
 347    "metadata": {},
 348    "outputs": [],
 349    "source": [
 350     "preds = mod.predict(X_test)\n",
 351     "preds.shape"
 352    ]
 353   },
 354   {
 355    "cell_type": "code",
 356    "execution_count": null,
 357    "id": "4e4cd809-6701-4bd7-b4fe-37c5e35d8999",
 358    "metadata": {},
 359    "outputs": [],
 360    "source": [
 361     "np.mean(np.sqrt(np.mean(np.square(preds - y_array), axis=(1,2))))"
 362    ]
 363   },
 364   {
 365    "cell_type": "code",
 366    "execution_count": null,
 367    "id": "4f4d80cb-edef-4720-b335-4af5a04992c3",
 368    "metadata": {},
 369    "outputs": [],
 370    "source": []
 371   },
 372   {
 373    "cell_type": "code",
 374    "execution_count": null,
 375    "id": "e9d7f913-b391-4e14-9b64-46a0a9786f4a",
 376    "metadata": {},
 377    "outputs": [],
 378    "source": []
 379   }
 380  ],
 381  "metadata": {
 382   "kernelspec": {
 383    "display_name": "Python 3 (ipykernel)",
 384    "language": "python",
 385    "name": "python3"
 386   },
 387   "language_info": {
 388    "codemirror_mode": {
 389     "name": "ipython",
 390     "version": 3
 391    },
 392    "file_extension": ".py",
 393    "mimetype": "text/x-python",
 394    "name": "python",
 395    "nbconvert_exporter": "python",
 396    "pygments_lexer": "ipython3",
 397    "version": "3.12.5"
 398   }
 399  },
 400  "nbformat": 4,
 401  "nbformat_minor": 5
 402 }