├── .gitignore
├── AWS-S3-Dask-XGBoost.ipynb
├── LICENSE
├── README.md
├── brenowitz-data-loading.ipynb
├── djgagne_partial_dependence_plot.ipynb
└── rasp-data-loading.ipynb


/.gitignore:
--------------------------------------------------------------------------------
  1 | # Byte-compiled / optimized / DLL files
  2 | __pycache__/
  3 | *.py[cod]
  4 | *$py.class
  5 | 
  6 | # C extensions
  7 | *.so
  8 | 
  9 | # Distribution / packaging
 10 | .Python
 11 | build/
 12 | develop-eggs/
 13 | dist/
 14 | downloads/
 15 | eggs/
 16 | .eggs/
 17 | lib/
 18 | lib64/
 19 | parts/
 20 | sdist/
 21 | var/
 22 | wheels/
 23 | *.egg-info/
 24 | .installed.cfg
 25 | *.egg
 26 | MANIFEST
 27 | 
 28 | # PyInstaller
 29 | #  Usually these files are written by a python script from a template
 30 | #  before PyInstaller builds the exe, so as to inject date/other infos into it.
 31 | *.manifest
 32 | *.spec
 33 | 
 34 | # Installer logs
 35 | pip-log.txt
 36 | pip-delete-this-directory.txt
 37 | 
 38 | # Unit test / coverage reports
 39 | htmlcov/
 40 | .tox/
 41 | .coverage
 42 | .coverage.*
 43 | .cache
 44 | nosetests.xml
 45 | coverage.xml
 46 | *.cover
 47 | .hypothesis/
 48 | .pytest_cache/
 49 | 
 50 | # Translations
 51 | *.mo
 52 | *.pot
 53 | 
 54 | # Django stuff:
 55 | *.log
 56 | local_settings.py
 57 | db.sqlite3
 58 | 
 59 | # Flask stuff:
 60 | instance/
 61 | .webassets-cache
 62 | 
 63 | # Scrapy stuff:
 64 | .scrapy
 65 | 
 66 | # Sphinx documentation
 67 | docs/_build/
 68 | 
 69 | # PyBuilder
 70 | target/
 71 | 
 72 | # Jupyter Notebook
 73 | .ipynb_checkpoints
 74 | 
 75 | # pyenv
 76 | .python-version
 77 | 
 78 | # celery beat schedule file
 79 | celerybeat-schedule
 80 | 
 81 | # SageMath parsed files
 82 | *.sage.py
 83 | 
 84 | # Environments
 85 | .env
 86 | .venv
 87 | env/
 88 | venv/
 89 | ENV/
 90 | env.bak/
 91 | venv.bak/
 92 | 
 93 | # Spyder project settings
 94 | .spyderproject
 95 | .spyproject
 96 | 
 97 | # Rope project settings
 98 | .ropeproject
 99 | 
100 | # mkdocs documentation
101 | /site
102 | 
103 | # mypy
104 | .mypy_cache/
105 | 


--------------------------------------------------------------------------------
/AWS-S3-Dask-XGBoost.ipynb:
--------------------------------------------------------------------------------
  1 | {
  2 |  "cells": [
  3 |   {
  4 |    "cell_type": "markdown",
  5 |    "metadata": {
  6 |     "toc": true
  7 |    },
  8 |    "source": [
  9 |     "<h1>Table of Contents<span class=\"tocSkip\"></span></h1>\n",
 10 |     "<div class=\"toc\"><ul class=\"toc-item\"><li><span><a href=\"#Overview\" data-toc-modified-id=\"Overview-1\">Overview</a></span></li><li><span><a href=\"#Load-libraries\" data-toc-modified-id=\"Load-libraries-2\">Load libraries</a></span></li><li><span><a href=\"#Load-Data\" data-toc-modified-id=\"Load-Data-3\">Load Data</a></span></li><li><span><a href=\"#Train-the-XGB-model\" data-toc-modified-id=\"Train-the-XGB-model-4\">Train the XGB model</a></span><ul class=\"toc-item\"><li><span><a href=\"#Determine-feature-importance\" data-toc-modified-id=\"Determine-feature-importance-4.1\">Determine feature importance</a></span></li></ul></li><li><span><a href=\"#Test-model-performance\" data-toc-modified-id=\"Test-model-performance-5\">Test model performance</a></span></li></ul></div>"
 11 |    ]
 12 |   },
 13 |   {
 14 |    "cell_type": "markdown",
 15 |    "metadata": {},
 16 |    "source": [
 17 |     "# Overview"
 18 |    ]
 19 |   },
 20 |   {
 21 |    "cell_type": "markdown",
 22 |    "metadata": {},
 23 |    "source": [
 24 |     "**Author**:\n",
 25 |     "**Zhonghua Zheng** (zzheng25@illinois.edu)   \n",
 26 |     "GitHub: **zzheng93**   \n",
 27 |     "\n",
 28 |     "**Motivation**:\n",
 29 |     "This notebook shows the workflow of \n",
 30 |     "* Loading Climate Data ([CESM-LE from AWS S3](https://medium.com/pangeo/cesm-lens-on-aws-4e2a996397a1))\n",
 31 |     "* Scalable [Dask-XGBoost](https://examples.dask.org/machine-learning/xgboost.html) *regression* implementation (the official documentation of [Dask-XGBoost](https://examples.dask.org/machine-learning/xgboost.html) only provides the *classfication* implementation.\n",
 32 |     "\n",
 33 |     "**Task**:\n",
 34 |     "How to predict **maximum reference height temperature (TREFHTMX)** based on related features?\n",
 35 |     "\n",
 36 |     "* Response (Y):   \n",
 37 |     "\"TREFHTMX\": Maximum reference height temperature over output period  \n",
 38 |     "\n",
 39 |     "* Features (X):   \n",
 40 |     "\"PRECT\": Total (convective and large-scale) precipitation rate (liq + ice)  \n",
 41 |     "\"WSPDSRFAV\": Horizontal total wind speed average at the surface  \n",
 42 |     "\"TS\": Surface temperature (radiative)  \n",
 43 |     "\"TREFHT\": Reference height temperature    \n",
 44 |     "\n",
 45 |     "**Prerequisite**  \n",
 46 |     "How to create environment for this workflow?  \n",
 47 |     "Recommendation: \n",
 48 |     "```bash\n",
 49 |     "git clone http://github.com/dask/dask-tutorial\n",
 50 |     "cd dask-tutorial\n",
 51 |     "conda env create -f binder/environment.yml \n",
 52 |     "conda activate dask-tutorial\n",
 53 |     "conda install xarray cftime=1.0.3.4\n",
 54 |     "conda install -c conda-forge dask-ml dask-xgboost\n",
 55 |     "```\n"
 56 |    ]
 57 |   },
 58 |   {
 59 |    "cell_type": "markdown",
 60 |    "metadata": {},
 61 |    "source": [
 62 |     "# Load libraries"
 63 |    ]
 64 |   },
 65 |   {
 66 |    "cell_type": "markdown",
 67 |    "metadata": {},
 68 |    "source": [
 69 |     "Here we load the necessary libraries, connect to cluster, and set up anonymous access to s3"
 70 |    ]
 71 |   },
 72 |   {
 73 |    "cell_type": "code",
 74 |    "execution_count": 1,
 75 |    "metadata": {},
 76 |    "outputs": [],
 77 |    "source": [
 78 |     "import numpy as np\n",
 79 |     "import s3fs\n",
 80 |     "import xarray as xr\n",
 81 |     "import matplotlib.pyplot as plt\n",
 82 |     "import dask\n",
 83 |     "import dask.array as da\n",
 84 |     "import dask.dataframe as dd\n",
 85 |     "import dask_xgboost\n",
 86 |     "import matplotlib.pyplot as plt\n",
 87 |     "from dask.distributed import Client\n",
 88 |     "from dask_ml.model_selection import train_test_split\n",
 89 |     "from sklearn import metrics\n",
 90 |     "import xgboost\n",
 91 |     "\n",
 92 |     "client = Client() # connect to cluster\n",
 93 |     "s3 = s3fs.S3FileSystem(anon=True) # anonymous access to s3"
 94 |    ]
 95 |   },
 96 |   {
 97 |    "cell_type": "code",
 98 |    "execution_count": 2,
 99 |    "metadata": {},
100 |    "outputs": [
101 |     {
102 |      "data": {
103 |       "text/html": [
104 |        "<table style=\"border: 2px solid white;\">\n",
105 |        "<tr>\n",
106 |        "<td style=\"vertical-align: top; border: 0px solid white\">\n",
107 |        "<h3 style=\"text-align: left;\">Client</h3>\n",
108 |        "<ul style=\"text-align: left; list-style: none; margin: 0; padding: 0;\">\n",
109 |        "  <li><b>Scheduler: </b>tcp://127.0.0.1:34195</li>\n",
110 |        "  <li><b>Dashboard: </b><a href='http://127.0.0.1:8787/status' target='_blank'>http://127.0.0.1:8787/status</a>\n",
111 |        "</ul>\n",
112 |        "</td>\n",
113 |        "<td style=\"vertical-align: top; border: 0px solid white\">\n",
114 |        "<h3 style=\"text-align: left;\">Cluster</h3>\n",
115 |        "<ul style=\"text-align: left; list-style:none; margin: 0; padding: 0;\">\n",
116 |        "  <li><b>Workers: </b>4</li>\n",
117 |        "  <li><b>Cores: </b>16</li>\n",
118 |        "  <li><b>Memory: </b>32.66 GB</li>\n",
119 |        "</ul>\n",
120 |        "</td>\n",
121 |        "</tr>\n",
122 |        "</table>"
123 |       ],
124 |       "text/plain": [
125 |        "<Client: 'tcp://127.0.0.1:34195' processes=4 threads=16, memory=32.66 GB>"
126 |       ]
127 |      },
128 |      "execution_count": 2,
129 |      "metadata": {},
130 |      "output_type": "execute_result"
131 |     }
132 |    ],
133 |    "source": [
134 |     "client"
135 |    ]
136 |   },
137 |   {
138 |    "cell_type": "markdown",
139 |    "metadata": {},
140 |    "source": [
141 |     "# Load Data"
142 |    ]
143 |   },
144 |   {
145 |    "cell_type": "markdown",
146 |    "metadata": {},
147 |    "source": [
148 |     "Here we load the CESM-LE RCP8.5 data from [AWS S3](https://medium.com/pangeo/cesm-lens-on-aws-4e2a996397a1).   \n",
149 |     "The original format of the data is [Zarr](https://zarr.readthedocs.io/en/stable/), but Dask-XGBoost only accepts Dask.array or Dask.dataframe collections.\n",
150 |     "\n",
151 |     "So we use [xarray](http://xarray.pydata.org/en/stable/) to load the data, and convert to [Dask Dataframe](https://docs.dask.org/en/latest/dataframe.html).  \n",
152 |     "\n",
153 |     "As an illustration, here I only use \"member_id=1\" and \"time=\"2010-01-01\"\" for this tutorial.\n"
154 |    ]
155 |   },
156 |   {
157 |    "cell_type": "code",
158 |    "execution_count": 3,
159 |    "metadata": {},
160 |    "outputs": [
161 |     {
162 |      "name": "stdout",
163 |      "output_type": "stream",
164 |      "text": [
165 |       "CPU times: user 6.55 s, sys: 308 ms, total: 6.85 s\n",
166 |       "Wall time: 1min 34s\n"
167 |      ]
168 |     }
169 |    ],
170 |    "source": [
171 |     "%%time\n",
172 |     "zarr_ls = [xr.open_zarr(s3fs.S3Map(root=\"ncar-cesm-lens/atm/daily/cesmLE-RCP85-\"+var+\".zarr\", s3=s3))[var]\n",
173 |     "           .sel(member_id=1,time=\"2010-01-01\") \n",
174 |     "           for var in [\"TREFHT\",\"TREFHTMX\",\"PRECT\",\"TS\",\"WSPDSRFAV\"]]  # read data from s3"
175 |    ]
176 |   },
177 |   {
178 |    "cell_type": "code",
179 |    "execution_count": 4,
180 |    "metadata": {},
181 |    "outputs": [],
182 |    "source": [
183 |     "dddf=xr.merge(zarr_ls).to_dask_dataframe()  # merge arrays and convert to dask dataframe"
184 |    ]
185 |   },
186 |   {
187 |    "cell_type": "code",
188 |    "execution_count": 5,
189 |    "metadata": {},
190 |    "outputs": [],
191 |    "source": [
192 |     "X=dddf[[\"TREFHT\",\"PRECT\",\"TS\",\"WSPDSRFAV\"]]  # features for Dask-XGBoost\n",
193 |     "Y=dddf[\"TREFHTMX\"]  # response for Dask-XGBoost"
194 |    ]
195 |   },
196 |   {
197 |    "cell_type": "code",
198 |    "execution_count": 6,
199 |    "metadata": {},
200 |    "outputs": [],
201 |    "source": [
202 |     "X_train, X_test, y_train, y_test = train_test_split(X, Y, test_size=0.15)  # split into training and testing"
203 |    ]
204 |   },
205 |   {
206 |    "cell_type": "markdown",
207 |    "metadata": {},
208 |    "source": [
209 |     "# Train the XGB model"
210 |    ]
211 |   },
212 |   {
213 |    "cell_type": "markdown",
214 |    "metadata": {},
215 |    "source": [
216 |     "This parameters are available at [Learning Task Parameters](https://xgboost.readthedocs.io/en/latest/parameter.html).\n",
217 |     "* **reg:squarederror**: regression with squared loss.\n",
218 |     "* **eta (learning_rate)**: Step size shrinkage used in update to prevents overfitting. After each boosting step, we can directly get the weights of new features, and eta shrinks the feature weights to make the boosting process more conservative.  \n",
219 |     "* **max_depth**: Maximum depth of a tree. Increasing this value will make the model more complex and more likely to overfit. 0 is only accepted in lossguided growing policy when tree_method is set as hist and it indicates no limit on depth. Beware that XGBoost aggressively consumes memory when training a deep tree. \n",
220 |     "* **num_boost_round**:  [Number of boosting](iterationshttps://xgboost.readthedocs.io/en/latest/python/python_api.html)."
221 |    ]
222 |   },
223 |   {
224 |    "cell_type": "code",
225 |    "execution_count": 7,
226 |    "metadata": {},
227 |    "outputs": [
228 |     {
229 |      "name": "stdout",
230 |      "output_type": "stream",
231 |      "text": [
232 |       "CPU times: user 306 ms, sys: 57.8 ms, total: 364 ms\n",
233 |       "Wall time: 13 s\n"
234 |      ]
235 |     }
236 |    ],
237 |    "source": [
238 |     "%%time\n",
239 |     "params = {'objective': 'reg:squarederror',\n",
240 |     "          'max_depth': 6, 'eta': 0.01}\n",
241 |     "\n",
242 |     "bst = dask_xgboost.train(client, params, X_train, y_train, num_boost_round=500)"
243 |    ]
244 |   },
245 |   {
246 |    "cell_type": "markdown",
247 |    "metadata": {},
248 |    "source": [
249 |     "## Determine feature importance"
250 |    ]
251 |   },
252 |   {
253 |    "cell_type": "markdown",
254 |    "metadata": {},
255 |    "source": [
256 |     "Generally, importance provides a score that indicates how useful or valuable each feature was in the construction of the boosted decision trees within the model. The more an attribute is used to make key decisions with decision trees, the higher its relative importance (reference: [Feature Importance and Feature Selection With XGBoost in Python](https://machinelearningmastery.com/feature-importance-and-feature-selection-with-xgboost-in-python/))."
257 |    ]
258 |   },
259 |   {
260 |    "cell_type": "code",
261 |    "execution_count": 8,
262 |    "metadata": {},
263 |    "outputs": [
264 |     {
265 |      "data": {
266 |       "image/png": "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\n",
267 |       "text/plain": [
268 |        "<Figure size 432x288 with 1 Axes>"
269 |       ]
270 |      },
271 |      "metadata": {
272 |       "needs_background": "light"
273 |      },
274 |      "output_type": "display_data"
275 |     }
276 |    ],
277 |    "source": [
278 |     "%matplotlib inline\n",
279 |     "ax = xgboost.plot_importance(bst)\n",
280 |     "ax.grid(False, axis=\"y\")\n",
281 |     "ax.set_title('Estimated feature importance')\n",
282 |     "plt.show()"
283 |    ]
284 |   },
285 |   {
286 |    "cell_type": "markdown",
287 |    "metadata": {},
288 |    "source": [
289 |     "# Test model performance"
290 |    ]
291 |   },
292 |   {
293 |    "cell_type": "markdown",
294 |    "metadata": {},
295 |    "source": [
296 |     "Here we test our XGBoost emulator performance using the testing data"
297 |    ]
298 |   },
299 |   {
300 |    "cell_type": "code",
301 |    "execution_count": 9,
302 |    "metadata": {},
303 |    "outputs": [],
304 |    "source": [
305 |     "y_hat = dask_xgboost.predict(client, bst, X_test).persist()\n",
306 |     "y_test, y_hat = dask.compute(y_test, y_hat) # get the predicted data and true data"
307 |    ]
308 |   },
309 |   {
310 |    "cell_type": "code",
311 |    "execution_count": 10,
312 |    "metadata": {},
313 |    "outputs": [
314 |     {
315 |      "data": {
316 |       "image/png": "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\n",
317 |       "text/plain": [
318 |        "<Figure size 360x360 with 1 Axes>"
319 |       ]
320 |      },
321 |      "metadata": {
322 |       "needs_background": "light"
323 |      },
324 |      "output_type": "display_data"
325 |     },
326 |     {
327 |      "name": "stdout",
328 |      "output_type": "stream",
329 |      "text": [
330 |       "rmse: 2.5904884\n",
331 |       "r2_score: 0.98544579149015\n"
332 |      ]
333 |     }
334 |    ],
335 |    "source": [
336 |     "fig, ax = plt.subplots(figsize=(5, 5))\n",
337 |     "ax.scatter(y_hat,y_test,s=0.1)\n",
338 |     "ax.plot([y_test.min(), y_test.min()], [y_test.max(), y_test.max()],c=\"black\")\n",
339 |     "ax.set(\n",
340 |     "    xlabel=\"Predicted\",\n",
341 |     "    ylabel=\"True\",\n",
342 |     ")\n",
343 |     "plt.show()\n",
344 |     "print(\"rmse:\",np.sqrt(metrics.mean_squared_error(y_test, y_hat)))\n",
345 |     "print(\"r2_score:\",metrics.r2_score(y_test,y_hat))"
346 |    ]
347 |   }
348 |  ],
349 |  "metadata": {
350 |   "kernelspec": {
351 |    "display_name": "Python 3",
352 |    "language": "python",
353 |    "name": "python3"
354 |   },
355 |   "language_info": {
356 |    "codemirror_mode": {
357 |     "name": "ipython",
358 |     "version": 3
359 |    },
360 |    "file_extension": ".py",
361 |    "mimetype": "text/x-python",
362 |    "name": "python",
363 |    "nbconvert_exporter": "python",
364 |    "pygments_lexer": "ipython3",
365 |    "version": "3.7.4"
366 |   },
367 |   "toc": {
368 |    "base_numbering": 1,
369 |    "nav_menu": {},
370 |    "number_sections": false,
371 |    "sideBar": true,
372 |    "skip_h1_title": false,
373 |    "title_cell": "Table of Contents",
374 |    "title_sidebar": "Contents",
375 |    "toc_cell": true,
376 |    "toc_position": {},
377 |    "toc_section_display": true,
378 |    "toc_window_display": false
379 |   }
380 |  },
381 |  "nbformat": 4,
382 |  "nbformat_minor": 2
383 | }
384 | 


--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
  1 |                                  Apache License
  2 |                            Version 2.0, January 2004
  3 |                         http://www.apache.org/licenses/
  4 | 
  5 |    TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
  6 | 
  7 |    1. Definitions.
  8 | 
  9 |       "License" shall mean the terms and conditions for use, reproduction,
 10 |       and distribution as defined by Sections 1 through 9 of this document.
 11 | 
 12 |       "Licensor" shall mean the copyright owner or entity authorized by
 13 |       the copyright owner that is granting the License.
 14 | 
 15 |       "Legal Entity" shall mean the union of the acting entity and all
 16 |       other entities that control, are controlled by, or are under common
 17 |       control with that entity. For the purposes of this definition,
 18 |       "control" means (i) the power, direct or indirect, to cause the
 19 |       direction or management of such entity, whether by contract or
 20 |       otherwise, or (ii) ownership of fifty percent (50%) or more of the
 21 |       outstanding shares, or (iii) beneficial ownership of such entity.
 22 | 
 23 |       "You" (or "Your") shall mean an individual or Legal Entity
 24 |       exercising permissions granted by this License.
 25 | 
 26 |       "Source" form shall mean the preferred form for making modifications,
 27 |       including but not limited to software source code, documentation
 28 |       source, and configuration files.
 29 | 
 30 |       "Object" form shall mean any form resulting from mechanical
 31 |       transformation or translation of a Source form, including but
 32 |       not limited to compiled object code, generated documentation,
 33 |       and conversions to other media types.
 34 | 
 35 |       "Work" shall mean the work of authorship, whether in Source or
 36 |       Object form, made available under the License, as indicated by a
 37 |       copyright notice that is included in or attached to the work
 38 |       (an example is provided in the Appendix below).
 39 | 
 40 |       "Derivative Works" shall mean any work, whether in Source or Object
 41 |       form, that is based on (or derived from) the Work and for which the
 42 |       editorial revisions, annotations, elaborations, or other modifications
 43 |       represent, as a whole, an original work of authorship. For the purposes
 44 |       of this License, Derivative Works shall not include works that remain
 45 |       separable from, or merely link (or bind by name) to the interfaces of,
 46 |       the Work and Derivative Works thereof.
 47 | 
 48 |       "Contribution" shall mean any work of authorship, including
 49 |       the original version of the Work and any modifications or additions
 50 |       to that Work or Derivative Works thereof, that is intentionally
 51 |       submitted to Licensor for inclusion in the Work by the copyright owner
 52 |       or by an individual or Legal Entity authorized to submit on behalf of
 53 |       the copyright owner. For the purposes of this definition, "submitted"
 54 |       means any form of electronic, verbal, or written communication sent
 55 |       to the Licensor or its representatives, including but not limited to
 56 |       communication on electronic mailing lists, source code control systems,
 57 |       and issue tracking systems that are managed by, or on behalf of, the
 58 |       Licensor for the purpose of discussing and improving the Work, but
 59 |       excluding communication that is conspicuously marked or otherwise
 60 |       designated in writing by the copyright owner as "Not a Contribution."
 61 | 
 62 |       "Contributor" shall mean Licensor and any individual or Legal Entity
 63 |       on behalf of whom a Contribution has been received by Licensor and
 64 |       subsequently incorporated within the Work.
 65 | 
 66 |    2. Grant of Copyright License. Subject to the terms and conditions of
 67 |       this License, each Contributor hereby grants to You a perpetual,
 68 |       worldwide, non-exclusive, no-charge, royalty-free, irrevocable
 69 |       copyright license to reproduce, prepare Derivative Works of,
 70 |       publicly display, publicly perform, sublicense, and distribute the
 71 |       Work and such Derivative Works in Source or Object form.
 72 | 
 73 |    3. Grant of Patent License. Subject to the terms and conditions of
 74 |       this License, each Contributor hereby grants to You a perpetual,
 75 |       worldwide, non-exclusive, no-charge, royalty-free, irrevocable
 76 |       (except as stated in this section) patent license to make, have made,
 77 |       use, offer to sell, sell, import, and otherwise transfer the Work,
 78 |       where such license applies only to those patent claims licensable
 79 |       by such Contributor that are necessarily infringed by their
 80 |       Contribution(s) alone or by combination of their Contribution(s)
 81 |       with the Work to which such Contribution(s) was submitted. If You
 82 |       institute patent litigation against any entity (including a
 83 |       cross-claim or counterclaim in a lawsuit) alleging that the Work
 84 |       or a Contribution incorporated within the Work constitutes direct
 85 |       or contributory patent infringement, then any patent licenses
 86 |       granted to You under this License for that Work shall terminate
 87 |       as of the date such litigation is filed.
 88 | 
 89 |    4. Redistribution. You may reproduce and distribute copies of the
 90 |       Work or Derivative Works thereof in any medium, with or without
 91 |       modifications, and in Source or Object form, provided that You
 92 |       meet the following conditions:
 93 | 
 94 |       (a) You must give any other recipients of the Work or
 95 |           Derivative Works a copy of this License; and
 96 | 
 97 |       (b) You must cause any modified files to carry prominent notices
 98 |           stating that You changed the files; and
 99 | 
100 |       (c) You must retain, in the Source form of any Derivative Works
101 |           that You distribute, all copyright, patent, trademark, and
102 |           attribution notices from the Source form of the Work,
103 |           excluding those notices that do not pertain to any part of
104 |           the Derivative Works; and
105 | 
106 |       (d) If the Work includes a "NOTICE" text file as part of its
107 |           distribution, then any Derivative Works that You distribute must
108 |           include a readable copy of the attribution notices contained
109 |           within such NOTICE file, excluding those notices that do not
110 |           pertain to any part of the Derivative Works, in at least one
111 |           of the following places: within a NOTICE text file distributed
112 |           as part of the Derivative Works; within the Source form or
113 |           documentation, if provided along with the Derivative Works; or,
114 |           within a display generated by the Derivative Works, if and
115 |           wherever such third-party notices normally appear. The contents
116 |           of the NOTICE file are for informational purposes only and
117 |           do not modify the License. You may add Your own attribution
118 |           notices within Derivative Works that You distribute, alongside
119 |           or as an addendum to the NOTICE text from the Work, provided
120 |           that such additional attribution notices cannot be construed
121 |           as modifying the License.
122 | 
123 |       You may add Your own copyright statement to Your modifications and
124 |       may provide additional or different license terms and conditions
125 |       for use, reproduction, or distribution of Your modifications, or
126 |       for any such Derivative Works as a whole, provided Your use,
127 |       reproduction, and distribution of the Work otherwise complies with
128 |       the conditions stated in this License.
129 | 
130 |    5. Submission of Contributions. Unless You explicitly state otherwise,
131 |       any Contribution intentionally submitted for inclusion in the Work
132 |       by You to the Licensor shall be under the terms and conditions of
133 |       this License, without any additional terms or conditions.
134 |       Notwithstanding the above, nothing herein shall supersede or modify
135 |       the terms of any separate license agreement you may have executed
136 |       with Licensor regarding such Contributions.
137 | 
138 |    6. Trademarks. This License does not grant permission to use the trade
139 |       names, trademarks, service marks, or product names of the Licensor,
140 |       except as required for reasonable and customary use in describing the
141 |       origin of the Work and reproducing the content of the NOTICE file.
142 | 
143 |    7. Disclaimer of Warranty. Unless required by applicable law or
144 |       agreed to in writing, Licensor provides the Work (and each
145 |       Contributor provides its Contributions) on an "AS IS" BASIS,
146 |       WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147 |       implied, including, without limitation, any warranties or conditions
148 |       of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149 |       PARTICULAR PURPOSE. You are solely responsible for determining the
150 |       appropriateness of using or redistributing the Work and assume any
151 |       risks associated with Your exercise of permissions under this License.
152 | 
153 |    8. Limitation of Liability. In no event and under no legal theory,
154 |       whether in tort (including negligence), contract, or otherwise,
155 |       unless required by applicable law (such as deliberate and grossly
156 |       negligent acts) or agreed to in writing, shall any Contributor be
157 |       liable to You for damages, including any direct, indirect, special,
158 |       incidental, or consequential damages of any character arising as a
159 |       result of this License or out of the use or inability to use the
160 |       Work (including but not limited to damages for loss of goodwill,
161 |       work stoppage, computer failure or malfunction, or any and all
162 |       other commercial damages or losses), even if such Contributor
163 |       has been advised of the possibility of such damages.
164 | 
165 |    9. Accepting Warranty or Additional Liability. While redistributing
166 |       the Work or Derivative Works thereof, You may choose to offer,
167 |       and charge a fee for, acceptance of support, warranty, indemnity,
168 |       or other liability obligations and/or rights consistent with this
169 |       License. However, in accepting such obligations, You may act only
170 |       on Your own behalf and on Your sole responsibility, not on behalf
171 |       of any other Contributor, and only if You agree to indemnify,
172 |       defend, and hold each Contributor harmless for any liability
173 |       incurred by, or claims asserted against, such Contributor by reason
174 |       of your accepting any such warranty or additional liability.
175 | 
176 |    END OF TERMS AND CONDITIONS
177 | 
178 |    APPENDIX: How to apply the Apache License to your work.
179 | 
180 |       To apply the Apache License to your work, attach the following
181 |       boilerplate notice, with the fields enclosed by brackets "[]"
182 |       replaced with your own identifying information. (Don't include
183 |       the brackets!)  The text should be enclosed in the appropriate
184 |       comment syntax for the file format. We also recommend that a
185 |       file or class name and description of purpose be included on the
186 |       same "printed page" as the copyright notice for easier
187 |       identification within third-party archives.
188 | 
189 |    Copyright [yyyy] [name of copyright owner]
190 | 
191 |    Licensed under the Apache License, Version 2.0 (the "License");
192 |    you may not use this file except in compliance with the License.
193 |    You may obtain a copy of the License at
194 | 
195 |        http://www.apache.org/licenses/LICENSE-2.0
196 | 
197 |    Unless required by applicable law or agreed to in writing, software
198 |    distributed under the License is distributed on an "AS IS" BASIS,
199 |    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200 |    See the License for the specific language governing permissions and
201 |    limitations under the License.
202 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # ml-workflow-examples
2 | Simple examples of data pipelines from xarray to ML training
3 | 


--------------------------------------------------------------------------------
/brenowitz-data-loading.ipynb:
--------------------------------------------------------------------------------
  1 | {
  2 |  "cells": [
  3 |   {
  4 |    "cell_type": "markdown",
  5 |    "metadata": {},
  6 |    "source": [
  7 |     "Here is a toy xarray dataset. It has a few 3D and 2D variables."
  8 |    ]
  9 |   },
 10 |   {
 11 |    "cell_type": "code",
 12 |    "execution_count": 1,
 13 |    "metadata": {},
 14 |    "outputs": [
 15 |     {
 16 |      "data": {
 17 |       "text/plain": [
 18 |        "<xarray.Dataset>\n",
 19 |        "Dimensions:  (time: 4, x: 2, y: 5, z: 4)\n",
 20 |        "Dimensions without coordinates: time, x, y, z\n",
 21 |        "Data variables:\n",
 22 |        "    a        (time, z, y, x) float64 1.0 1.0 1.0 1.0 1.0 ... 1.0 1.0 1.0 1.0 1.0\n",
 23 |        "    b        (time, y, x) float64 1.0 1.0 1.0 1.0 1.0 ... 1.0 1.0 1.0 1.0 1.0"
 24 |       ]
 25 |      },
 26 |      "execution_count": 1,
 27 |      "metadata": {},
 28 |      "output_type": "execute_result"
 29 |     }
 30 |    ],
 31 |    "source": [
 32 |     "import numpy as np\n",
 33 |     "import xarray as xr\n",
 34 |     "from itertools import product\n",
 35 |     "\n",
 36 |     "from torch.utils.data import Dataset, DataLoader\n",
 37 |     "\n",
 38 |     "class XRTimeSeries(Dataset):\n",
 39 |     "    \"\"\"A pytorch Dataset class for time series data in xarray format\n",
 40 |     "\n",
 41 |     "    This function assumes the data has dimensions ['time', 'z', 'y', 'x'], and\n",
 42 |     "    that the axes of the data arrays are all stored in that order.\n",
 43 |     "\n",
 44 |     "    An individual \"sample\" is the full time time series from a single\n",
 45 |     "    horizontal location. The time-varying variables in this sample will have\n",
 46 |     "    shape (time, z, 1, 1).\n",
 47 |     "\n",
 48 |     "    Examples\n",
 49 |     "    --------\n",
 50 |     "    >>> ds = xr.open_dataset(\"in.nc\")\n",
 51 |     "    >>> dataset = XRTimeSeries(ds)\n",
 52 |     "    >>> dataset[0]\n",
 53 |     "\n",
 54 |     "    \"\"\"\n",
 55 |     "    dims = ['time', 'z', 'x', 'y']\n",
 56 |     "\n",
 57 |     "    def __init__(self, data, time_length=None):\n",
 58 |     "        \"\"\"\n",
 59 |     "        Parameters\n",
 60 |     "        ----------\n",
 61 |     "        data : xr.DataArray\n",
 62 |     "            An input dataset. This dataset must contain at least some variables\n",
 63 |     "            with all of the dimensions ['time' , 'z', 'x', 'y'].\n",
 64 |     "        time_length : int, optional\n",
 65 |     "            The length of the time sequences to use, must evenly divide the\n",
 66 |     "            total number of time points.\n",
 67 |     "        \"\"\"\n",
 68 |     "        self.time_length = time_length or len(data.time)\n",
 69 |     "        self.data = data\n",
 70 |     "        self.numpy_data = {key: data[key].values for key in data.data_vars}\n",
 71 |     "        self.data_vars = set(data.data_vars)\n",
 72 |     "        self.dims = {key: data[key].dims for key in data.data_vars}\n",
 73 |     "        self.constants = {\n",
 74 |     "            key\n",
 75 |     "            for key in data.data_vars\n",
 76 |     "            if len({'x', 'y', 'time'} & set(data[key].dims)) == 0\n",
 77 |     "        }\n",
 78 |     "        self.setup_indices()\n",
 79 |     "\n",
 80 |     "    def setup_indices(self):\n",
 81 |     "        len_x = len(self.data['x'].values)\n",
 82 |     "        len_y = len(self.data['y'].values)\n",
 83 |     "        len_t = len(self.data['time'].values)\n",
 84 |     "\n",
 85 |     "        x_iter = range(0, len_x, 1)\n",
 86 |     "        y_iter = range(0, len_y, 1)\n",
 87 |     "        t_iter = range(0, len_t, self.time_length)\n",
 88 |     "        assert len_t % self.time_length == 0\n",
 89 |     "        self.indices = list(product(t_iter, y_iter, x_iter))\n",
 90 |     "\n",
 91 |     "    def __len__(self):\n",
 92 |     "        return len(self.indices)\n",
 93 |     "\n",
 94 |     "    def __getitem__(self, i):\n",
 95 |     "        t, y, x = self.indices[i]\n",
 96 |     "        output_tensors = {}\n",
 97 |     "        for key in self.data_vars:\n",
 98 |     "            if key in self.constants:\n",
 99 |     "                continue\n",
100 |     "\n",
101 |     "            data_array = self.numpy_data[key]\n",
102 |     "            if 'z' in self.dims[key]:\n",
103 |     "                this_array_index = (slice(t, t + self.time_length),\n",
104 |     "                                    slice(None), y, x)\n",
105 |     "            else:\n",
106 |     "                this_array_index = (slice(t, t + self.time_length), None, y, x)\n",
107 |     "\n",
108 |     "            sample = data_array[this_array_index][:, :, np.newaxis, np.newaxis]\n",
109 |     "            output_tensors[key] = sample.astype(np.float32)\n",
110 |     "        return output_tensors\n",
111 |     "\n",
112 |     "    @property\n",
113 |     "    def time_dim(self):\n",
114 |     "        return self.dims[0][0]\n",
115 |     "\n",
116 |     "    def torch_constants(self):\n",
117 |     "        return {\n",
118 |     "            key: torch.tensor(self.data[key].values, requires_grad=False)\n",
119 |     "            .float()\n",
120 |     "            for key in self.constants\n",
121 |     "        }\n",
122 |     "\n",
123 |     "    @property\n",
124 |     "    def scale(self):\n",
125 |     "        std = self.std\n",
126 |     "        return valmap(lambda x: x.max(), std)\n",
127 |     "    \n",
128 |     "\n",
129 |     "def get_xarray_dataset():\n",
130 |     "\n",
131 |     "    dims_3d = ['time', 'z', 'y', 'x']\n",
132 |     "    dims_2d = ['time', 'y', 'x']\n",
133 |     "\n",
134 |     "    data_3d = np.ones((4, 4, 5, 2))\n",
135 |     "    data_2d = np.ones((4, 5, 2))\n",
136 |     "\n",
137 |     "    return xr.Dataset({\n",
138 |     "        'a': (dims_3d, data_3d),\n",
139 |     "        'b': (dims_2d, data_2d)\n",
140 |     "    })\n",
141 |     "\n",
142 |     "ds = get_xarray_dataset()\n",
143 |     "ds"
144 |    ]
145 |   },
146 |   {
147 |    "cell_type": "code",
148 |    "execution_count": 2,
149 |    "metadata": {},
150 |    "outputs": [],
151 |    "source": [
152 |     "torch_dataset = XRTimeSeries(ds, time_length=4)"
153 |    ]
154 |   },
155 |   {
156 |    "cell_type": "code",
157 |    "execution_count": 3,
158 |    "metadata": {},
159 |    "outputs": [
160 |     {
161 |      "data": {
162 |       "text/plain": [
163 |        "10"
164 |       ]
165 |      },
166 |      "execution_count": 3,
167 |      "metadata": {},
168 |      "output_type": "execute_result"
169 |     }
170 |    ],
171 |    "source": [
172 |     "len(torch_dataset)"
173 |    ]
174 |   },
175 |   {
176 |    "cell_type": "markdown",
177 |    "metadata": {},
178 |    "source": [
179 |     "The length of the dataset is $x\\dot y$"
180 |    ]
181 |   },
182 |   {
183 |    "cell_type": "code",
184 |    "execution_count": 4,
185 |    "metadata": {},
186 |    "outputs": [
187 |     {
188 |      "data": {
189 |       "text/plain": [
190 |        "{'a': array([[[[1.]],\n",
191 |        " \n",
192 |        "         [[1.]],\n",
193 |        " \n",
194 |        "         [[1.]],\n",
195 |        " \n",
196 |        "         [[1.]]],\n",
197 |        " \n",
198 |        " \n",
199 |        "        [[[1.]],\n",
200 |        " \n",
201 |        "         [[1.]],\n",
202 |        " \n",
203 |        "         [[1.]],\n",
204 |        " \n",
205 |        "         [[1.]]],\n",
206 |        " \n",
207 |        " \n",
208 |        "        [[[1.]],\n",
209 |        " \n",
210 |        "         [[1.]],\n",
211 |        " \n",
212 |        "         [[1.]],\n",
213 |        " \n",
214 |        "         [[1.]]],\n",
215 |        " \n",
216 |        " \n",
217 |        "        [[[1.]],\n",
218 |        " \n",
219 |        "         [[1.]],\n",
220 |        " \n",
221 |        "         [[1.]],\n",
222 |        " \n",
223 |        "         [[1.]]]], dtype=float32), 'b': array([[[[1.]]],\n",
224 |        " \n",
225 |        " \n",
226 |        "        [[[1.]]],\n",
227 |        " \n",
228 |        " \n",
229 |        "        [[[1.]]],\n",
230 |        " \n",
231 |        " \n",
232 |        "        [[[1.]]]], dtype=float32)}"
233 |       ]
234 |      },
235 |      "execution_count": 4,
236 |      "metadata": {},
237 |      "output_type": "execute_result"
238 |     }
239 |    ],
240 |    "source": [
241 |     "sample = torch_dataset[0]\n",
242 |     "sample"
243 |    ]
244 |   },
245 |   {
246 |    "cell_type": "markdown",
247 |    "metadata": {},
248 |    "source": [
249 |     "Two dimensional and three dimensional variables have broadcastable shapes"
250 |    ]
251 |   },
252 |   {
253 |    "cell_type": "code",
254 |    "execution_count": 5,
255 |    "metadata": {},
256 |    "outputs": [
257 |     {
258 |      "data": {
259 |       "text/plain": [
260 |        "(4, 1, 1, 1)"
261 |       ]
262 |      },
263 |      "execution_count": 5,
264 |      "metadata": {},
265 |      "output_type": "execute_result"
266 |     }
267 |    ],
268 |    "source": [
269 |     "sample['b'].shape"
270 |    ]
271 |   },
272 |   {
273 |    "cell_type": "code",
274 |    "execution_count": 6,
275 |    "metadata": {},
276 |    "outputs": [
277 |     {
278 |      "data": {
279 |       "text/plain": [
280 |        "(4, 4, 1, 1)"
281 |       ]
282 |      },
283 |      "execution_count": 6,
284 |      "metadata": {},
285 |      "output_type": "execute_result"
286 |     }
287 |    ],
288 |    "source": [
289 |     "sample['a'].shape"
290 |    ]
291 |   },
292 |   {
293 |    "cell_type": "markdown",
294 |    "metadata": {},
295 |    "source": [
296 |     "Now that we have made the torch dataset object, we can pass it to pytorch's DataLoader class."
297 |    ]
298 |   },
299 |   {
300 |    "cell_type": "code",
301 |    "execution_count": 7,
302 |    "metadata": {},
303 |    "outputs": [],
304 |    "source": [
305 |     "train_loader = DataLoader(torch_dataset, batch_size=4)"
306 |    ]
307 |   },
308 |   {
309 |    "cell_type": "code",
310 |    "execution_count": 8,
311 |    "metadata": {},
312 |    "outputs": [
313 |     {
314 |      "name": "stdout",
315 |      "output_type": "stream",
316 |      "text": [
317 |       "shape of b torch.Size([4, 4, 1, 1, 1])\n",
318 |       "shape of b torch.Size([4, 4, 1, 1, 1])\n",
319 |       "shape of b torch.Size([2, 4, 1, 1, 1])\n"
320 |      ]
321 |     }
322 |    ],
323 |    "source": [
324 |     "for batch in train_loader:\n",
325 |     "    print(\"shape of b\", batch['b'].shape)"
326 |    ]
327 |   },
328 |   {
329 |    "cell_type": "markdown",
330 |    "metadata": {},
331 |    "source": [
332 |     "The first dimension becomes the \"batch\" dimension. The other dimensions are the physical dimensions (time, z, y, x). My [model classes](https://github.com/nbren12/uwnet/blob/047a63b70985b12e17013355ecd25c908681ab76/uwnet/modules.py) accept data in this format."
333 |    ]
334 |   }
335 |  ],
336 |  "metadata": {
337 |   "kernelspec": {
338 |    "display_name": "Python 3",
339 |    "language": "python",
340 |    "name": "python3"
341 |   },
342 |   "language_info": {
343 |    "codemirror_mode": {
344 |     "name": "ipython",
345 |     "version": 3
346 |    },
347 |    "file_extension": ".py",
348 |    "mimetype": "text/x-python",
349 |    "name": "python",
350 |    "nbconvert_exporter": "python",
351 |    "pygments_lexer": "ipython3",
352 |    "version": "3.6.6"
353 |   }
354 |  },
355 |  "nbformat": 4,
356 |  "nbformat_minor": 2
357 | }
358 | 


--------------------------------------------------------------------------------
/djgagne_partial_dependence_plot.ipynb:
--------------------------------------------------------------------------------
  1 | {
  2 |  "cells": [
  3 |   {
  4 |    "cell_type": "markdown",
  5 |    "metadata": {},
  6 |    "source": [
  7 |     "## Partial Dependence Plot Example\n",
  8 |     "David John Gagne\n",
  9 |     "\n",
 10 |     "The goal of this notebook is to show an example of a serial and potentially parallel partial dependence plot in order to figure out ways to scale this better with Pangeo tools."
 11 |    ]
 12 |   },
 13 |   {
 14 |    "cell_type": "code",
 15 |    "execution_count": 4,
 16 |    "metadata": {},
 17 |    "outputs": [],
 18 |    "source": [
 19 |     "%matplotlib inline\n",
 20 |     "import numpy as np\n",
 21 |     "import pandas as pd\n",
 22 |     "from sklearn.ensemble import RandomForestRegressor\n",
 23 |     "from sklearn.linear_model import Ridge\n",
 24 |     "import matplotlib.pyplot as plt\n",
 25 |     "from dask.distributed import LocalCluster, Client\n",
 26 |     "import os\n",
 27 |     "from os.path import exists, join\n",
 28 |     "from urllib.request import urlretrieve\n",
 29 |     "import tarfile\n",
 30 |     "import glob"
 31 |    ]
 32 |   },
 33 |   {
 34 |    "cell_type": "markdown",
 35 |    "metadata": {},
 36 |    "source": [
 37 |     "Download CSV Data."
 38 |    ]
 39 |   },
 40 |   {
 41 |    "cell_type": "code",
 42 |    "execution_count": 3,
 43 |    "metadata": {},
 44 |    "outputs": [
 45 |     {
 46 |      "name": "stdout",
 47 |      "output_type": "stream",
 48 |      "text": [
 49 |       "Get csv files\n",
 50 |       "Extract csv tar file\n"
 51 |      ]
 52 |     }
 53 |    ],
 54 |    "source": [
 55 |     "if not exists(\"tornado_data\"):\n",
 56 |     "    os.mkdir(\"tornado_data\")\n",
 57 |     "csv_tar_file = \"https://storage.googleapis.com/track_data_ncar_ams_3km_csv_small/track_data_ncar_ams_3km_csv_small.tar.gz\"\n",
 58 |     "print(\"Get csv files\")\n",
 59 |     "urlretrieve(csv_tar_file, join(\"tornado_data\", csv_tar_file.split(\"/\")[-1]))\n",
 60 |     "print(\"Extract csv tar file\")\n",
 61 |     "csv_tar = tarfile.open(join(\"tornado_data\", csv_tar_file.split(\"/\")[-1]))\n",
 62 |     "csv_tar.extractall(\"tornado_data/\")\n",
 63 |     "csv_tar.close()"
 64 |    ]
 65 |   },
 66 |   {
 67 |    "cell_type": "markdown",
 68 |    "metadata": {},
 69 |    "source": [
 70 |     "Load CSV data"
 71 |    ]
 72 |   },
 73 |   {
 74 |    "cell_type": "code",
 75 |    "execution_count": 11,
 76 |    "metadata": {},
 77 |    "outputs": [
 78 |     {
 79 |      "name": "stdout",
 80 |      "output_type": "stream",
 81 |      "text": [
 82 |       "['tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20101024-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20101122-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110201-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110308-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110326-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110404-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110414-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110420-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110425-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110509-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110522-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110527-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110605-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110610-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110615-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110620-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110625-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110704-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20110712-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20111116-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120218-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120315-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120323-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120401-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120409-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120426-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120503-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120510-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120529-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120606-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120622-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120701-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120706-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20120715-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20121225-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130318-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130331-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130411-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130429-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130513-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130519-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130527-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130602-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130613-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130619-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130625-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130701-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130708-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20130715-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140220-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140328-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140407-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140425-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140508-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140514-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140526-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140604-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140609-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140617-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140622-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140628-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140705-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20140710-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20141123-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150331-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150416-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150422-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150505-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150510-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150523-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150528-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150605-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150612-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150620-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150625-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150630-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150706-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20150712-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20151031-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20151227-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160224-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160323-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160401-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160415-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160429-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160505-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160511-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160522-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160528-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160608-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160616-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160622-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160628-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160707-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20160712-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20161129-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20170121-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20170228-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20170323-0000.csv', 'tornado_data/track_data_ncar_ams_3km_csv_small/track_step_NCARSTORM_d01_20170329-0000.csv']\n"
 83 |      ]
 84 |     }
 85 |    ],
 86 |    "source": [
 87 |     "path = \"tornado_data/track_data_ncar_ams_3km_csv_small/\"\n",
 88 |     "files = sorted(glob.glob(path+\"/*.csv\"))\n",
 89 |     "print(files)\n",
 90 |     "df = pd.concat([pd.read_csv(f, parse_dates=[\"Run_Date\", \"Valid_Date\"]) for f in files], ignore_index=True)\n"
 91 |    ]
 92 |   },
 93 |   {
 94 |    "cell_type": "code",
 95 |    "execution_count": 9,
 96 |    "metadata": {},
 97 |    "outputs": [
 98 |     {
 99 |      "name": "stdout",
100 |      "output_type": "stream",
101 |      "text": [
102 |       "Step_ID\n",
103 |       "Track_ID\n",
104 |       "Ensemble_Name\n",
105 |       "Ensemble_Member\n",
106 |       "Run_Date\n",
107 |       "Valid_Date\n",
108 |       "Forecast_Hour\n",
109 |       "Valid_Hour_UTC\n",
110 |       "Duration\n",
111 |       "Centroid_Lon\n",
112 |       "Centroid_Lat\n",
113 |       "Centroid_X\n",
114 |       "Centroid_Y\n",
115 |       "Storm_Motion_U\n",
116 |       "Storm_Motion_V\n",
117 |       "REFL_COM_mean\n",
118 |       "REFL_COM_max\n",
119 |       "REFL_COM_min\n",
120 |       "REFL_COM_std\n",
121 |       "REFL_COM_percentile_10\n",
122 |       "REFL_COM_percentile_25\n",
123 |       "REFL_COM_percentile_50\n",
124 |       "REFL_COM_percentile_75\n",
125 |       "REFL_COM_percentile_90\n",
126 |       "U10_mean\n",
127 |       "U10_max\n",
128 |       "U10_min\n",
129 |       "U10_std\n",
130 |       "U10_percentile_10\n",
131 |       "U10_percentile_25\n",
132 |       "U10_percentile_50\n",
133 |       "U10_percentile_75\n",
134 |       "U10_percentile_90\n",
135 |       "V10_mean\n",
136 |       "V10_max\n",
137 |       "V10_min\n",
138 |       "V10_std\n",
139 |       "V10_percentile_10\n",
140 |       "V10_percentile_25\n",
141 |       "V10_percentile_50\n",
142 |       "V10_percentile_75\n",
143 |       "V10_percentile_90\n",
144 |       "T2_mean\n",
145 |       "T2_max\n",
146 |       "T2_min\n",
147 |       "T2_std\n",
148 |       "T2_percentile_10\n",
149 |       "T2_percentile_25\n",
150 |       "T2_percentile_50\n",
151 |       "T2_percentile_75\n",
152 |       "T2_percentile_90\n",
153 |       "RVORT1_MAX-future_mean\n",
154 |       "RVORT1_MAX-future_max\n",
155 |       "RVORT1_MAX-future_min\n",
156 |       "RVORT1_MAX-future_std\n",
157 |       "RVORT1_MAX-future_percentile_10\n",
158 |       "RVORT1_MAX-future_percentile_25\n",
159 |       "RVORT1_MAX-future_percentile_50\n",
160 |       "RVORT1_MAX-future_percentile_75\n",
161 |       "RVORT1_MAX-future_percentile_90\n",
162 |       "HAIL_MAXK1-future_mean\n",
163 |       "HAIL_MAXK1-future_max\n",
164 |       "HAIL_MAXK1-future_min\n",
165 |       "HAIL_MAXK1-future_std\n",
166 |       "HAIL_MAXK1-future_percentile_10\n",
167 |       "HAIL_MAXK1-future_percentile_25\n",
168 |       "HAIL_MAXK1-future_percentile_50\n",
169 |       "HAIL_MAXK1-future_percentile_75\n",
170 |       "HAIL_MAXK1-future_percentile_90\n",
171 |       "area\n",
172 |       "eccentricity\n",
173 |       "major_axis_length\n",
174 |       "minor_axis_length\n",
175 |       "orientation\n",
176 |       "Matched\n",
177 |       "Max_Hail_Size\n",
178 |       "Num_Matches\n",
179 |       "Shape\n",
180 |       "Location\n",
181 |       "Scale\n"
182 |      ]
183 |     }
184 |    ],
185 |    "source": [
186 |     "for col in df.columns:\n",
187 |     "    print(col)"
188 |    ]
189 |   },
190 |   {
191 |    "cell_type": "markdown",
192 |    "metadata": {},
193 |    "source": [
194 |     "We are using reflectivity, u-wind, v-wind, and 2 m temperature to predict vorticity for a given storm in a dataset of storms."
195 |    ]
196 |   },
197 |   {
198 |    "cell_type": "code",
199 |    "execution_count": 12,
200 |    "metadata": {},
201 |    "outputs": [],
202 |    "source": [
203 |     "input_cols = [\"REFL_COM_mean\", \"U10_mean\", \"V10_mean\", \"T2_mean\"]\n",
204 |     "output_col = [\"RVORT1_MAX-future_max\"]\n",
205 |     "split_date = pd.Timestamp(\"2015-01-01\")\n",
206 |     "train_in = df.loc[df[\"Run_Date\"] < split_date, input_cols]\n",
207 |     "train_out = df.loc[df[\"Run_Date\"] < split_date, output_col]\n",
208 |     "test_in = df.loc[df[\"Run_Date\"] >= split_date, input_cols]\n",
209 |     "test_out = df.loc[df[\"Run_Date\"]>= split_date, output_col]\n"
210 |    ]
211 |   },
212 |   {
213 |    "cell_type": "code",
214 |    "execution_count": 16,
215 |    "metadata": {},
216 |    "outputs": [
217 |     {
218 |      "data": {
219 |       "text/plain": [
220 |        "(76377, 1)"
221 |       ]
222 |      },
223 |      "execution_count": 16,
224 |      "metadata": {},
225 |      "output_type": "execute_result"
226 |     }
227 |    ],
228 |    "source": [
229 |     "train_out.shape"
230 |    ]
231 |   },
232 |   {
233 |    "cell_type": "markdown",
234 |    "metadata": {},
235 |    "source": [
236 |     "Train the random forest"
237 |    ]
238 |   },
239 |   {
240 |    "cell_type": "code",
241 |    "execution_count": 17,
242 |    "metadata": {},
243 |    "outputs": [
244 |     {
245 |      "data": {
246 |       "text/plain": [
247 |        "RandomForestRegressor(bootstrap=True, criterion='mse', max_depth=8,\n",
248 |        "           max_features='auto', max_leaf_nodes=None,\n",
249 |        "           min_impurity_decrease=0.0, min_impurity_split=None,\n",
250 |        "           min_samples_leaf=1, min_samples_split=2,\n",
251 |        "           min_weight_fraction_leaf=0.0, n_estimators=50, n_jobs=None,\n",
252 |        "           oob_score=False, random_state=None, verbose=0, warm_start=False)"
253 |       ]
254 |      },
255 |      "execution_count": 17,
256 |      "metadata": {},
257 |      "output_type": "execute_result"
258 |     }
259 |    ],
260 |    "source": [
261 |     "rf = RandomForestRegressor(n_estimators=50, max_depth=8)\n",
262 |     "rf.fit(train_in, train_out.values.ravel())"
263 |    ]
264 |   },
265 |   {
266 |    "cell_type": "markdown",
267 |    "metadata": {},
268 |    "source": [
269 |     "We want to interpret the input sensitivities for the random forest. We will use a technique called partial dependence plots, which changes the inputs in a way that can affect the mean output and reveal sensitivities of the model to certain ranges of inputs. "
270 |    ]
271 |   },
272 |   {
273 |    "cell_type": "code",
274 |    "execution_count": 18,
275 |    "metadata": {},
276 |    "outputs": [],
277 |    "source": [
278 |     "def partial_dependence_1d(x, model, var_index, var_vals):\n",
279 |     "    \"\"\"\n",
280 |     "    Calculate how the mean prediction of an ML model varies if one variable's value is fixed across all input\n",
281 |     "    examples.\n",
282 |     "\n",
283 |     "    Args:\n",
284 |     "        x: array of input variables\n",
285 |     "        model: scikit-learn style model object\n",
286 |     "        var_index: column index of the variable being investigated\n",
287 |     "        var_vals: values of the input variable that are fixed.\n",
288 |     "\n",
289 |     "    Returns:\n",
290 |     "        Array of partial dependence values.\n",
291 |     "    \"\"\"\n",
292 |     "    partial_dependence = np.zeros(var_vals.shape)\n",
293 |     "    x_copy = np.copy(x)\n",
294 |     "    for v, var_val in enumerate(var_vals):\n",
295 |     "        x_copy[:, var_index] = var_val\n",
296 |     "        partial_dependence[v] = model.predict(x_copy).mean()\n",
297 |     "    return partial_dependence"
298 |    ]
299 |   },
300 |   {
301 |    "cell_type": "code",
302 |    "execution_count": 19,
303 |    "metadata": {},
304 |    "outputs": [
305 |     {
306 |      "data": {
307 |       "text/html": [
308 |        "<div>\n",
309 |        "<style scoped>\n",
310 |        "    .dataframe tbody tr th:only-of-type {\n",
311 |        "        vertical-align: middle;\n",
312 |        "    }\n",
313 |        "\n",
314 |        "    .dataframe tbody tr th {\n",
315 |        "        vertical-align: top;\n",
316 |        "    }\n",
317 |        "\n",
318 |        "    .dataframe thead th {\n",
319 |        "        text-align: right;\n",
320 |        "    }\n",
321 |        "</style>\n",
322 |        "<table border=\"1\" class=\"dataframe\">\n",
323 |        "  <thead>\n",
324 |        "    <tr style=\"text-align: right;\">\n",
325 |        "      <th></th>\n",
326 |        "      <th>REFL_COM_mean</th>\n",
327 |        "      <th>U10_mean</th>\n",
328 |        "      <th>V10_mean</th>\n",
329 |        "      <th>T2_mean</th>\n",
330 |        "    </tr>\n",
331 |        "  </thead>\n",
332 |        "  <tbody>\n",
333 |        "    <tr>\n",
334 |        "      <th>count</th>\n",
335 |        "      <td>76377.000000</td>\n",
336 |        "      <td>76377.000000</td>\n",
337 |        "      <td>76377.000000</td>\n",
338 |        "      <td>76377.000000</td>\n",
339 |        "    </tr>\n",
340 |        "    <tr>\n",
341 |        "      <th>mean</th>\n",
342 |        "      <td>46.846515</td>\n",
343 |        "      <td>0.449749</td>\n",
344 |        "      <td>0.542538</td>\n",
345 |        "      <td>289.408528</td>\n",
346 |        "    </tr>\n",
347 |        "    <tr>\n",
348 |        "      <th>std</th>\n",
349 |        "      <td>3.984091</td>\n",
350 |        "      <td>4.347625</td>\n",
351 |        "      <td>4.461936</td>\n",
352 |        "      <td>6.931555</td>\n",
353 |        "    </tr>\n",
354 |        "    <tr>\n",
355 |        "      <th>min</th>\n",
356 |        "      <td>40.082820</td>\n",
357 |        "      <td>-18.653700</td>\n",
358 |        "      <td>-20.561280</td>\n",
359 |        "      <td>262.921600</td>\n",
360 |        "    </tr>\n",
361 |        "    <tr>\n",
362 |        "      <th>25%</th>\n",
363 |        "      <td>43.755930</td>\n",
364 |        "      <td>-2.366580</td>\n",
365 |        "      <td>-2.301950</td>\n",
366 |        "      <td>285.114260</td>\n",
367 |        "    </tr>\n",
368 |        "    <tr>\n",
369 |        "      <th>50%</th>\n",
370 |        "      <td>46.170650</td>\n",
371 |        "      <td>0.396810</td>\n",
372 |        "      <td>0.659590</td>\n",
373 |        "      <td>290.625850</td>\n",
374 |        "    </tr>\n",
375 |        "    <tr>\n",
376 |        "      <th>75%</th>\n",
377 |        "      <td>49.128190</td>\n",
378 |        "      <td>3.271980</td>\n",
379 |        "      <td>3.496990</td>\n",
380 |        "      <td>294.398190</td>\n",
381 |        "    </tr>\n",
382 |        "    <tr>\n",
383 |        "      <th>max</th>\n",
384 |        "      <td>68.743160</td>\n",
385 |        "      <td>19.686640</td>\n",
386 |        "      <td>18.616930</td>\n",
387 |        "      <td>312.149230</td>\n",
388 |        "    </tr>\n",
389 |        "  </tbody>\n",
390 |        "</table>\n",
391 |        "</div>"
392 |       ],
393 |       "text/plain": [
394 |        "       REFL_COM_mean      U10_mean      V10_mean       T2_mean\n",
395 |        "count   76377.000000  76377.000000  76377.000000  76377.000000\n",
396 |        "mean       46.846515      0.449749      0.542538    289.408528\n",
397 |        "std         3.984091      4.347625      4.461936      6.931555\n",
398 |        "min        40.082820    -18.653700    -20.561280    262.921600\n",
399 |        "25%        43.755930     -2.366580     -2.301950    285.114260\n",
400 |        "50%        46.170650      0.396810      0.659590    290.625850\n",
401 |        "75%        49.128190      3.271980      3.496990    294.398190\n",
402 |        "max        68.743160     19.686640     18.616930    312.149230"
403 |       ]
404 |      },
405 |      "execution_count": 19,
406 |      "metadata": {},
407 |      "output_type": "execute_result"
408 |     }
409 |    ],
410 |    "source": [
411 |     "train_in.describe()"
412 |    ]
413 |   },
414 |   {
415 |    "cell_type": "markdown",
416 |    "metadata": {},
417 |    "source": [
418 |     "Here is an example for a single variable."
419 |    ]
420 |   },
421 |   {
422 |    "cell_type": "code",
423 |    "execution_count": 41,
424 |    "metadata": {},
425 |    "outputs": [],
426 |    "source": [
427 |     "pd_count = 50\n",
428 |     "index = 0\n",
429 |     "var_vals = np.linspace(train_in.iloc[:, 0].min(), train_in.iloc[:, 0].max(), pd_count)\n",
430 |     "pd_vals = partial_dependence_1d(train_in, rf, index, var_vals)"
431 |    ]
432 |   },
433 |   {
434 |    "cell_type": "code",
435 |    "execution_count": 43,
436 |    "metadata": {},
437 |    "outputs": [
438 |     {
439 |      "name": "stdout",
440 |      "output_type": "stream",
441 |      "text": [
442 |       "4.99 s ± 236 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
443 |      ]
444 |     }
445 |    ],
446 |    "source": [
447 |     "%timeit pd_vals = partial_dependence_1d(train_in, rf, index, var_vals)"
448 |    ]
449 |   },
450 |   {
451 |    "cell_type": "code",
452 |    "execution_count": 45,
453 |    "metadata": {},
454 |    "outputs": [
455 |     {
456 |      "data": {
457 |       "text/plain": [
458 |        "Text(0.5, 0, 'REFL_COM_mean')"
459 |       ]
460 |      },
461 |      "execution_count": 45,
462 |      "metadata": {},
463 |      "output_type": "execute_result"
464 |     },
465 |     {
466 |      "data": {
467 |       "image/png": "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\n",
468 |       "text/plain": [
469 |        "<Figure size 432x288 with 1 Axes>"
470 |       ]
471 |      },
472 |      "metadata": {
473 |       "needs_background": "light"
474 |      },
475 |      "output_type": "display_data"
476 |     }
477 |    ],
478 |    "source": [
479 |     "plt.plot(var_vals, pd_vals, 'ko-')\n",
480 |     "plt.xlabel(input_cols[index])"
481 |    ]
482 |   },
483 |   {
484 |    "cell_type": "markdown",
485 |    "metadata": {},
486 |    "source": [
487 |     "How well can Dask parallelize the task across all the input variables?"
488 |    ]
489 |   },
490 |   {
491 |    "cell_type": "code",
492 |    "execution_count": 25,
493 |    "metadata": {},
494 |    "outputs": [],
495 |    "source": [
496 |     "cluster = LocalCluster(n_workers=4)\n",
497 |     "client = Client(cluster)"
498 |    ]
499 |   },
500 |   {
501 |    "cell_type": "code",
502 |    "execution_count": 26,
503 |    "metadata": {},
504 |    "outputs": [
505 |     {
506 |      "data": {
507 |       "text/html": [
508 |        "<table style=\"border: 2px solid white;\">\n",
509 |        "<tr>\n",
510 |        "<td style=\"vertical-align: top; border: 0px solid white\">\n",
511 |        "<h3>Client</h3>\n",
512 |        "<ul>\n",
513 |        "  <li><b>Scheduler: </b>tcp://127.0.0.1:51668\n",
514 |        "  <li><b>Dashboard: </b><a href='http://127.0.0.1:8787/status' target='_blank'>http://127.0.0.1:8787/status</a>\n",
515 |        "</ul>\n",
516 |        "</td>\n",
517 |        "<td style=\"vertical-align: top; border: 0px solid white\">\n",
518 |        "<h3>Cluster</h3>\n",
519 |        "<ul>\n",
520 |        "  <li><b>Workers: </b>4</li>\n",
521 |        "  <li><b>Cores: </b>8</li>\n",
522 |        "  <li><b>Memory: </b>17.18 GB</li>\n",
523 |        "</ul>\n",
524 |        "</td>\n",
525 |        "</tr>\n",
526 |        "</table>"
527 |       ],
528 |       "text/plain": [
529 |        "<Client: scheduler='tcp://127.0.0.1:51668' processes=4 cores=8>"
530 |       ]
531 |      },
532 |      "execution_count": 26,
533 |      "metadata": {},
534 |      "output_type": "execute_result"
535 |     }
536 |    ],
537 |    "source": [
538 |     "client"
539 |    ]
540 |   },
541 |   {
542 |    "cell_type": "code",
543 |    "execution_count": 38,
544 |    "metadata": {},
545 |    "outputs": [],
546 |    "source": [
547 |     "pd_count = 50\n",
548 |     "var_vals = np.zeros((len(input_cols), pd_count))\n",
549 |     "futures = []\n",
550 |     "train_future = client.scatter(train_in.values)\n",
551 |     "for i, input_col in enumerate(input_cols):\n",
552 |     "    var_vals[i] = np.linspace(train_in[input_col].min(), train_in[input_col].max(), pd_count)\n",
553 |     "    futures.append(client.submit(partial_dependence_1d, train_future, rf, i, var_vals[i]))"
554 |    ]
555 |   },
556 |   {
557 |    "cell_type": "code",
558 |    "execution_count": 39,
559 |    "metadata": {},
560 |    "outputs": [],
561 |    "source": [
562 |     "results = client.gather(futures)"
563 |    ]
564 |   },
565 |   {
566 |    "cell_type": "code",
567 |    "execution_count": 40,
568 |    "metadata": {},
569 |    "outputs": [
570 |     {
571 |      "data": {
572 |       "image/png": "iVBORw0KGgoAAAANSUhEUgAAAmkAAAHkCAYAAABlt7hJAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDMuMC4yLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvOIA7rQAAIABJREFUeJzs3Xt8XGW99/3PL2l6CC2UpqXQQ5KiRSnKQUJBEQUqu4UNFBW2YNQ+wGN2D2oV3duycz+4hZ19w6OAsKVoldZuOghYTkUrFQvK9hYKKWCh1GKhSRusUAq2pSk95Xf/MWvKZLImmWRmMqfv+/WaV2euda0115pm1vzWdTR3R0RERETyS1muCyAiIiIiXSlIExEREclDCtJERERE8pCCNBEREZE8pCBNREREJA8pSBMRERHJQwrSRERERPKQgjQRKQpmNs3M1pvZBjObF7J9kJndE2xfZWa1cduuDtLXm9nUIG28mT1uZuvMbK2ZzY3Lf52ZrTGz583sN2Y2Jkg3M7s1ONYaM/tI9s9cRIqVaTJbESl0ZlYOvAycA7QBzwCXuftLcXlmA8e7+0wzuxT4tLt/zswmAT8HJgNjgN8CxwBHAEe5+7NmNgxYDVzk7i+Z2aHuviM47teAScFxzwO+CpwHnArc4u6n9suHICJFZ0CuC5AJI0eO9Nra2lwXQ0T60erVq99091HBy8nABnd/FcDM7gamAy/F7TId+Pfg+VLgh2ZmQfrd7r4H2GhmG4DJ7v4ksAXA3Xea2TpgLPBSLEALHAJ43Hv8t0fvfp8ys+FmdpS7b0l2Hrp+iZSehOtXUkURpNXW1tLc3JzrYohIPzKz1riXY4HNca/biNZkEZbH3feb2XagKkh/KmHfsQnvVQucBKyKS2sCvgRsB87qphxjCYK9uH0bgAaA6upqXb9ESkzC9Ssp9UkTkWJgIWmJfTmS5el2XzMbCtwHfD2+Bs3dG919PBABvtKLcuDuC9y9zt3rRo3q8WZaREqUgjQRKQZtwPi41+OAvybLY2YDgMOAt7rb18wqiAZoEXe/P8l73wV8thflEBFJiYI0ESkGzwATzWyCmQ0ELgWWJeRZBswInl8MPBb0HVsGXBqM/pwATASeDvqr3QGsc/eb4g9kZhPjXl4I/DnuPb4UjPI8DdjeXX80EZHuFEWfNBEpbUEfs68AK4ByYKG7rzWza4Fmd19GNOC6MxgY8BbRQI4g371EBxnsB+a4+wEz+zjwReAFM3s+eKt/c/flwPVm9gGgA2gFZgbblxMd2bkBaAcuz/rJi0jRUk2aiORUJBKhtraWsrIyamtriUQifTqOuy9392Pc/X3u3hSkXRMEaLj7u+5+ibu/390nx0aCBtuagv0+4O6/DtL+4O7m7se7+4nBY3mw7bPu/qFg2wXu/lqQ7u4+JzjWh909YyMCwj6nxLTZs2f3mKevn6+I9L+imCetrq7ONTpKpPBEIhEaGhpob28/mFZZWcmCBQuor6/vdl8zW+3uddkuY7alcv0K+5wqKiowM/bu3Zt0v7A8lZWVzJgxg+XLl7Np0yaqq6s577zzOr1uamrq8fMXkb5L9fqlIE1Ecqa2tpbW1q4j0Wtqamhpael231IK0pJ9Tn1lZnR37U81UBaRvkn1+qXmThHJmU2bNvUqvVRl+vPo6ea8vb2dxsbGjL6niPSegjQRyZnq6upepZeqXHweCpRFci+tIC3TCxoH6cPNbKmZ/TlY2Pij6ZRRRPLXd77znS5plZWVNDU15aA0+aupqYnKyspOaRUVFQwcOLDb/cLyRGcW6dmYMWN6zKNBCSLZ1ecgLVjQ+DbgXGAScFmwUHG8K4G33f39wM3ADcG+k4gOfz8OmAbMD44HcAvwiLt/EDgBWNfXMopIftu9ezcAo0ePxsyoqalRX6gQ9fX1LFiwgJqamoOf06JFi1i4cGGntFmzZvWYZ+bMmV0CvjB///vfGTt2bNKRoiNHjuSKK66gtbUVd6e1tZWGhgYFaiIZ1OeBA0EN17+7+9Tg9dUA7v6/4/KsCPI8Gczw/TdgFDAvPm8sH7AW+BNwtPeiYBo4IFJ49u/fz8SJEznyyCP54x//mHINT0wpDRzItEgkQmNjY9LRnaeccgpLly7ttE8qo0khtUEfIqUu1etXOpPZZmNB493AVmCRmZ0ArAbmuvuuxDdPXKBYRArLvffeS0tLCz/4wQ96HaBJeurr67utraytre2Stm/fvpSOrb5sIpmTTp+0bCxoPAD4CHC7u58E7CKodeuSWQsUixQsd+eGG27g2GOP5YILLsh1cSRBOoGWbppFMiedIC0bCxq3AW3uvipIX0o0aBORIhDr01ReXs6aNWs444wzKCvTIPN809dAa9CgQRr0IZJB6VwdM76gsbv/DdgcrIkHMIXoenoiUuBis+bHOpoDLFmyRB3N81Cqo0krKiqoqqrCzCgrK+PYY4/VoA+RDOpzkObu+4HYgsbrgHtjCxqb2YVBtjuAqmBB46t4b8DAWiC2oPEjBAsaB/t8FYiY2RrgROA/+1pGEckfjY2NnZY1Ak2amq9SHU26aNEi3nzzTTo6Orj66qtZs2YNmzdv7vkNRCQlWhZKRPpFWVlZ6Ez3ZkZHR0evj6fRnfll48aNvO997+M73/lO6Px3IvIeLQslInlFqwsUtwkTJnDOOedwxx13cODAgZ53EJEeKUgTkX4xZ86cLmmZXF0g0yugmNl4M3s8WPlkrZnNjcv/vWBVlDVm9oCZDQ/Sa81st5k9Hzx+lJGTKxANDQ1s3ryZFStW5LooIkVBQZqIZF1HRwfLli2jsrKScePGZXx1gSytgLIf+Ka7HwucBsyJO+ajwIfc/XjgZeDquPd5xd1PDB4z0z65AnLhhRcyevRoFixYkOuiiBQFBWkikjXxU2784Q9/4Atf+AKbN2+mo6ODlpaWTI4EnAxscPdX3X0vcDcwPSHPdGBx8HwpMMWis+hOB+529z3uvhHYAEx29y3u/iyAu+8kOkBqbPD6N8HgKYhOzD0uUydSyCoqKjjllFN46KGHtJ6nSAYoSBMpYdlcIDt+yo2YLE65EbYCythkeYIAK34FlG73DZpGTwJW0dUVwK/jXk8ws+fM7PdmdkZYYc2swcyazax569at3Z9ZAYlEIvz2t78F0HqeIhmgIE2kRCXOW5bpH9R+nnIjGyugRHcyGwrcB3zd3Xd0OqBZI9Fm0diHtgWoDlZMuQq4y8wO7XLwIl0xpbGxkXfffbdTmqZZEek7BWkiJSrbQVSypYWytLZjNlZAwcwqiAZoEXe/P/5gZjYDOB+oDybpJmgy3RY8Xw28AhyTgfMrCP38fy5S9BSkiZSobP+gJqshytKUGxlfASXor3YHsM7db4o/kJlNA74NXOju7XHpo4JBB5jZ0cGxXs3wueYtTbMiklkK0kRKVDZ/UDdu3MiuXbuIxjnvyeSUG/GytALK6cAXgbPjptQ4LzjWD4FhwKMJU218AlhjZn8iOjhhpru/lfETzlNhy0ll6/9cpCS4e8E/Tj75ZBeR3lmyZImbmRPtf+WADx482JcsWdLn49XU1LiZeUVFhQ8ZMsRvvPHGg2k1NTV9PnYYoNnz4PqT7qPYrl+xvwPAy8rK/M4778x1kUTyTqrXL9WkiZSoo446CndnxIgRmBlmxqRJk/o0LUbiIIR9+/bR0dHB6NGjaWlpycaUG5Kn6uvraWlp4c4776Sjo4NJkxKnqxORVClIEylR3/ve9xg9ejSvvfYaHR0dXHvttTz77LM888wzvT5W2CCEPXv2aFRfCTv77LMBWLlyZY5LIlK4FKSJlKA1a9bwyCOP8LWvfY3BgwcDMHfuXKqqqrjmmmt6fTyN6pNEY8aM4dhjj+Wxxx7LdVFECpaCNJES9P3vf59DDjmEWbNmHUwbNmwY3/72t3nkkUc48sgjU57gtqOjg6FDh4Zu06i+0jZlyhSeeOIJ9u7dm+uiiBQkBWkiJSJ+dYE777yTM844g8MPP7xTnpEjRwLw+uuvd5rgdvbs2Z1WJoh/PWzYMHbu3MmAAQM6HUuj+mTKlCm0t7ezalXYQg0i0hMFaSJFKj4oGzlyJFdcccXBjv0Av//977vUkn33u9/tcpz29nZ+9KMfdVqZ4Pbbbz/4ur29nYqKCr785S9TU1OT8cXTpXCdeeaZlJWVqV+aSB8pSBMpAolrcM6ePbvTaMtt27Z1aXLavXt3l479yfqQxQK7ZPbt28fy5cs1klM6GT58OB/5yEcUpIn0kYI0kQLTU0AWq+lKHG0ZJjEoS6cPmQYJSJgpU6bw1FNP8c477+S6KCIFR0GaSAEJWxQ91YAsTGJQFjZjfOKqAakeSwSiQdr+/fv5n//5n1wXRaTgKEgTKSBh85H1VVjH/vr6ehYsWNCpb9nMmTO7BG6pHEsE4PTTT2fgwIFq8hTpAwVpInkuvnmztbU15f0Sa8AqKiqoqqrqsWN/bMb4WN+y+fPndwncZs2apUECkpLKyko+9rGPab40kT5QkCaSI4l9yyKRSI/9zZIJW8h85syZnQKpRYsW8eabb/apY39Y4KZBApKqkSNH8txzz6U8956IRA3oOYuIZFqsb1ms6bK1tZXLL78cMzs4CjPW36wnlZWVzJgxg+XLl7Np0yaqq6tpampS4CR5IRKJ8PDDDwN0mnsP0N+oSA9UkyaSA2F9y/bt29ermdnjmxpVswVmNs3M1pvZBjObF7J9kJndE2xfZWa1cduuDtLXm9nUIG28mT1uZuvMbK2ZzY3L/z0z+7OZrTGzB8xseHfHKmWNjY3s2bOnU1p7e7vWdRVJgWrSRHIg3ekqampqaGlpyUxhioCZlQO3AecAbcAzZrbM3V+Ky3Yl8La7v9/MLgVuAD5nZpOAS4HjgDHAb83sGGA/8E13f9bMhgGrzezR4JiPAle7+34zuwG4Gvh2smO5+4F++BjyktZ1Fem7tGrSMn3nGqS3mNkLZva8mTWnUz6RfDVu3LiU84b1N9NIyi4mAxvc/VV33wvcDUxPyDMdWBw8XwpMseiHOx242933uPtGYAMw2d23uPuzAO6+E1gHjA1e/8bd9wfHegoYF/ceXY6VhfMtGMmmZtGULSI963OQFnfnei4wCbgsuIuMd/DOFbiZ6J0rCXeb04D5wfFiznL3E929rq/lE8lnJ554Ype0iooKBg4c2CktbACARlKGGgtsjnvdFqSF5gkCrO1AVSr7BjeYJwFhi1BeAfy6F+XAzBrMrNnMmrdu3drNaRW+sLn3dKMhkpp0atIyfueaRllECkZzczPLly/nk5/8ZJfRlwsXLuwSkKm/WUrCZtxNHA6bLE+3+5rZUOA+4OvuvqPTAc0aiTaLxoYrplIO3H2Bu9e5e92oUaNCdikesbn3Dj/8cCBai6wbDZHUpNMnLeyO8dRkeYK+G/F3rk8l7Bu723TgN2bmwI/dfUHYm5tZA9AAqjaX/BeJRGhsbGTTpk0MGDCAYcOG8eCDDzJ8+PAuefXj1SdtwPi41+OAvybJ02ZmA4DDgLe629fMKogGaBF3vz/+YGY2AzgfmOLvzY+SSjlKTn19PUOHDuWiiy7igQceoK5OjSQiqUinJi1bd66nu/tHiDajzjGzT4S9eSndiUphS1zKad++fezevZtf/epXuS5aMXkGmGhmE8xsINHuFMsS8iwDZgTPLwYeC4KrZcClQR/aCcBE4Omg1v8OYJ273xR/IDObBnwbuNDd44fphh4ro2daoGI30xowIJK6dIK03ty5kuqdq7vH/n0DeAA1g0qBC5tuY8+ePZqCIIOCPmZfAVYQ7eB/r7uvNbNrzezCINsdQJWZbQCuAuYF+64F7gVeAh4B5gSjMU8HvgicHQxket7MzguO9UNgGPBokP6jHo5V8hSkifReOs2dB+9cgdeI3rl+PiFP7M71SeLuXM1sGXCXmd1EdJh67M71EKDM3XcGz/8BuDaNMorknKYg6B/uvhxYnpB2Tdzzd4FLkuzbBDQlpP2B8Fp/gsFQycrR5VgCI0aMoLKyUn/3Ir3Q5yAt6GMWu3MtBxbG7lyBZndfRvTO9c7gzvUtooEcQb7Y3eZ+grtNMxsNPBBMOTAAuMvdH0nj/ERyrrq6OnTNTfWllFJiZlRXVytIE+mFtCazzcKd66vACemUSSTfXHLJJXz/+9/vlKYpCKQUKUgT6R0tCyXSC2GLondn69at3HnnnYwbN47x48drrjMpaQrSRHpHQZpIihJHacYWig4L1CKRCDU1NRxxxBG8/vrrzJ49m02bNmmuMylp1dXVvP7667z77ru5LopIQVCQJpKisFGa7e3tzJ07t1Pt2uzZs2loaOhUY/Af//EfPda6iRS7WD/Mtra2HJdEpDAoSBNJUbJmmm3btnWqXbv99ttDgzlNuSGlTtNwiPSOgjSRFKU7GlM/TFLqFKSJ9I6CNJEUzZw5M639NeWGlLpx48YBCtJEUqUgTSQFu3fvZsmSJRx66KGMGzfu4CjNqqqq0PzBXH8HacoNERg0aBBHHnmkgjSRFClIE+lGbMqNyspK1q5dy8yZM9m8efPBUZq33HILlZWVnfaprKxk5syZ1NTUaMoNkQSahkMkdWlNZitSzGJTbsQPAvjhD3/I8ccffzDgiv3b2NjIpk2bqK6upqmpSQGZSBLV1dW88MILuS6GSEFQTZpIEsmm3EgcpVlfX09LS4vmQBNJQawmzd1zXRSRvKcgTSQJLYwuknnV1dXs3r2bbdu25booInlPQZpIErGRaIk0SlOk72Lfn9bW1hyXRCT/KUgTSeLUU0/tkqZRmvnLzKaZ2Xoz22Bm80K2DzKze4Ltq8ysNm7b1UH6ejObGqSNN7PHzWydma01s7lx+S8J0jrMrC4uvdbMdpvZ88HjR9k968KjudJEUqcgTSTEzp07WblyJSeeeKJGaRYAMysHbgPOBSYBl5nZpIRsVwJvu/v7gZuBG4J9JwGXAscB04D5wfH2A99092OB04A5ccd8EfgM8ERIcV5x9xODR3qT6xUhBWkiqVOQJiUhNpVGbH3NntbRvO2223j77bdZsGCBBgUUhsnABnd/1d33AncD0xPyTAcWB8+XAlMsOqHddOBud9/j7huBDcBkd9/i7s8CuPtOYB0wNni9zt3XZ/2sitDIkSMZMmSIgjSRFChIk6IXm0ojfn3NhoaGpIHaO++8w4033si5557LKaec0s+llT4aC2yOe90WpIXmcff9wHagKpV9g6bRk4BVKZRlgpk9Z2a/N7MzwjKYWYOZNZtZ89atW1M4ZPEwM82VJpIiBWlS9JJNpTFjxoxONWux2rZhw4bx5ptvMnny5ByVWPrAQtIS53hIlqfbfc1sKHAf8HV339FDObYA1e5+EnAVcJeZHdrl4O4L3L3O3etGjRrVwyGLj4I0kdRoMlspesl+DA4cOABER5ldfvnlmBl79+49uP173/seEydOVBNnYWgDxse9Hgf8NUmeNjMbABwGvNXdvmZWQTRAi7j7/T0Vwt33AHuC56vN7BXgGKC5D+dUtKqrq/nVr36V62KI5D3VpEnRS2XKjH379nUK0CB84lrJW88AE81sgpkNJDoQYFlCnmXAjOD5xcBjHp1RdRlwaTD6cwIwEXg66K92B7DO3W9KpRBmNioYdICZHR0c69U0z63oVFdX87e//Y09e/bkuigieU1BmhS96667rs/7qkmmMAR9zL4CrCDawf9ed19rZtea2YVBtjuAKjPbQLQpcl6w71rgXuAl4BFgjrsfAE4HvgicHTelxnkAZvZpM2sDPgr8ysxWBO/xCWCNmf2J6OCEme7+VtY/gAITu3Fqa2vLcUlE8puaO6XoxX4QRo4cybZt2ygrKzvY1JnqvpL/3H05sDwh7Zq45+8ClyTZtwloSkj7A+H91XD3B4AHQtLvI9o8Kt2In4bjfe97X45LI5K/VJMmRW/p0qUMGTKEjRs30tHRweLFi6msrOyUp6KigoEDB3ZK08S1ItmhudJEUqMgTYpaR0cH9913H+eeey5Dhw4FoguiL1iwoNMktYsWLWLhwoWauFakH8SWXFOQJtI9NXdKUXvyySfZsmULn/3sZzul19fXhwZgCspEsm/w4MGMHj1aQZpID1STJkXtvvvuY+DAgZx//vm5LoqIxNFcaSI9SytIy/SCxnHbyoMZu3+ZTvmktLk7S5cuZerUqRx6aJf5REUkhxSkifSsz0FalhY0jplLdBi9SJ8988wzbN68mYsvvjjXRRGRBLEgLTpVnYiESacmLeMLGgOY2TjgH4GfplE2EZYuXcqAAQO44IILcl0UEUnwxhtv0N7eTnl5+cGl2USks3SCtGwtaPwD4F+Bju7evJQXKJaexZo6P/WpT3H44YfnujgiEicSibB06VIg+l1tbW2loaFBgZpIgnSCtIwvaGxm5wNvuPvqnt681BcoluQikQhjx45l48aNPP3007rwi+SZxsbGLktCtbe3M2PGDMrKylSzJhJIJ0jrzYLGpLig8enAhWbWQrT59GwzW5JGGaXERCIRGhoa2LJlCwBvvfWW7tBF8kyyAQMHDhxQzZpInHSCtIwvaOzuV7v7OHevDY73mLt/IY0ySolpbGykvb29U5oWShfJL6kst6bvrUgaQVqWFjQWSUuyO3QN9RfJH01NTV2WZguj762UurRWHMj0gsYJ238H/C6d8knpGT9+fOiFXQuli+SP2MoejY2NbNq0ibKyMg4c6Hqfru+tlDqtOCBF5YwzzuiSpoXSS0OmJ9c2s/Fm9riZrTOztWY2Ny7/JUFah5nVJbxP0om65T319fW0tLTQ0dHB4sWLu9Ss6XsroiBNisiOHTtYsWIFkyZN0kLpJSZLk2vvB77p7scCpwFz4o75IvAZ4ImEcvQ0UbeEqK+vZ8GCBdTU1BxMa2xs1PdWSp6CNCkaN954I2+++SaLFy8+eIfe0tKiC31pyPjk2u6+xd2fBXD3nUT73o4NXq9z9/Uh5Ug6Ubd0L1az9tprrwFQXq7YVkRBmhSF119/nRtvvJFLLrmEurq6nneQYpOtybUBCJpGTwJWZaAcmoy7G2PGjOHDH/4wK1asyHVRRHJOQZoUtEgkQm1tLUceeSS7du3i1FNPzXWRJDcyPrn2wZ3MhgL3AV939x0ZKIcm4+7B1KlT+cMf/sA777yT66KI5JSCNClYsYlrW1tbD6Zdc801mgCzNGVjcm3MrIJogBZx9/szVA7pwdSpU9m3bx+/+93vcl0UkZxSkCYFSxPXSpyMT64d9Fe7A1jn7jelWI7QY6V1ZiXo4x//OEOGDFGTp5Q8BWlSMGJNm7G1/eJr0OJpAszSk6XJtU8Hvkh0ebrng8d5AGb2aTNrAz4K/MrMVvRwLOmFwYMHc9ZZZylIk5Jn0RvJwlZXV+fNzc25LoZkUaxpM7HmLExNTQ0tLS3ZL5TklJmtdveCHyWi61e4W2+9lblz5/Lqq68yYcKEXBdHJKNSvX6pJk0KQljTJkC0Reo9mgBTpDhMnRqdB1i1aVLKFKRJQUjWhOnumrhWpAgdc8wx1NTUKEiTkqYgTQrC+PHjQ9NjTZuauFakuJgZRx99NA899NDBfqgauV3aEvsll8Lfg4I0KQhak1OktEQiEf74xz/i7rg7ra2tNDQ0lMQPs3QVP+VSd38PRRfIxb4Ahfw4+eSTXQrbkiVLvKamxs3Ma2pqfMmSJZ3SAJ8wYYJXV1d3yiOlC2j2PLj+pPvQ9StcTU2NE50IuNOjpqYm10WTDAu7/ieqrq4O/XsoLy8/uN+sWbO8srKy0/bKykqfNWtWt78vsX17ypPqfqlI9fql0Z2Sc2EjNysqKjAz9u7dezBtyJAh/OQnP1GTpgAa3VnsysrKCPt9MjM6OjpyUCLJlEgkQmNjI5s2bWLEiBHs3Lmz07W+srKSGTNmsHz5cjZt2sT48ePTmlrJzDr9LYX9viQKy5PKfpWVlSn1jU71+qUgTXKuuznPEml6DYlRkFbckl0XdA0obKlOp5QYWBWSVP5GNQWHFIze3CFpolqR0tDU1ERlZWWnNPVDLXzJplNKlKwWtRBk8ndKQZrkXHV1dVbyikjhqq+vZ8GCBYwYMQKAMWPGaIqdIpBOAONxUy6Vl5eH5kkM5HIR2GXyd0pBmuTc3Llzu6RVVFQwcODATmm6ixYpLfX19dxzzz0A3HXXXQrQikCy6ZTiJQus4qdcWrx4cWhN68yZMzvNnTlz5swu+cJ+XxKF5Ullv0z/TilIk5xbs2YNAwYMYOzYsQe/WIsWLWLhwoWaqFakxMVqJVLttyr57cwzz+ySVlFRQVVVVbeBVWLwE6tpTfyNmD9/fqe5M+fPn98lX9jvy6xZs3rMk8p+Gf+dSmUIaL4/NIS9cL388steVlbmV111Va6LIgUGTcFREtrb2x3w6667LtdFkTS98cYbPnz4cD/uuON6nE4plWk5Clmq168BmQv3RHrvu9/9LoMHD+Zf//Vfc10UEclDQ4YM4YgjjlBNWhFobGzknXfe4Re/+AXHHntst3nr6+vVcoKaOyUH4meEjkQinH322YwePTrXxRKRPFVTU6OR3QUq/nr/k5/8hHPOOafHAE3eoyBN+lXi0h4AK1euLPylOyTnzGyama03sw1mNi9k+yAzuyfYvsrMauO2XR2krzezqUHaeDN73MzWmdlaM5sbl3+EmT1qZn8J/j08SD/TzLab2fPB45rsn3nxq66uVk1aAQq73v/ud7/T9b4XFKRJvwqbI2f37t00NjbmqERSDMysHLgNOBeYBFxmZpMSsl0JvO3u7wduBm4I9p0EXAocB0wD5gfH2w98092PBU4D5sQdcx6w0t0nAiuD1zH/4+4nBo9rs3C6JSdWkxb7oZfCoOt9+tIK0rJw5zrYzJ42sz8Fd67fTad8kn+SNVmoKUPSNBnY4O6vuvte4G5gekKe6cDi4PlSYIpFx/pPB+529z3uvhHYAEx29y3u/iyAu+8E1gFjQ461GLgoS+clRGvSdu/ezZtvvpnrokgv6Hqfvj4HaVm6c90DnO3uJwAnAtPM7LS+llHyT7I5cjRJraRpLLA57nUb7wVUXfK4+35gO1CVyr7BDeZJwKogabS7bwmOtQU4Ii7KmscYAAAgAElEQVT7R4MbzV+b2XFhhTWzBjNrNrPmrVu3pnqOJaumpgbQj3uhGTVqVGi6rvepS6cmLRt3ru7u7wT5K4KH6reLyCc/+ckuaZqkVjIgbPbLxGtHsjzd7mtmQ4H7gK+7+44eyvEsUBPcaP4X8GBYJndf4O517l6X7IdM3qO50gpD/CCBI444gq1bt3aZmFbX+95JJ0jLyp2rmZWb2fPAG8Cj7r6KELoTLRzxX9w777yTmpoaqqurNUmtZFIbEF9NOw74a7I8ZjYAOAx4q7t9zayCaIAWcff74/K8bmZHBXmOInq9wt13xG403X05UGFmIzNxgqVMNWn5L3GQQOx3+fLLL9ek5GlIZ560rNy5uvsB4EQzGw48YGYfcvcXu2R2XwAsAKirq1NtW56KfXHjO4++8cYb/OQnP9EXVTLpGWCimU0AXiPaneLzCXmWATOAJ4GLgcfc3c1sGXCXmd0EjAEmAk8Htf53AOvc/aYkx7o++PchADM7Eng9OO5kojfC2zJ+tiVmxIgRVFZWqiYtj4UNEnB3Vq5cSUtLS24KVQTSqUnLyp1rjLv/Hfgd0T5rUiDia81qa2uZO3euRvdI1gU19V8BVhDt4H+vu681s2vN7MIg2x1AlZltAK4iGJHp7muBe4GXgEeAOcHN4unAF4Gz46bUOC841vXAOWb2F+Cc4DVEg78XzexPwK3Apa4hiWmL1cKoJi1/aZBAdqRTk5aNO9dRwD53/7uZDQE+RTDYQPJfYq1Zd3e9+uJKpgXNi8sT0q6Je/4ucEmSfZuApoS0PxBe64+7bwOmhKT/EPhhb8suPdNcafkt2f+PBgmkp881aVm6cz0KeNzM1hANAh9191/2tYzSv8Kqu5PRF1dEekM1afmtqamJgQMHdkrTIIH0pbV2ZxbuXNcQHeYuBSjVC6i+uCLSW9XV1WzdupX29nYqKytzXRxJUF9fz6233kpzczPuTnV1NU1NTep7nCatOCAZM2bMmND0qqoqje4RkbTERnhu3ry5h5ySC7t372bt2rV8+ctfpqOjg5aWFl3nM0BBmmTMxIkTu6RVVlZyyy230NLSoi+uiPRZrIuEmjzz06OPPsquXbv4zGc+k+uiFBUFaZIRL7zwAk888QTTpk1TrZmIZFysJk2DB/LT/fffz/DhwznzzDNzXZSiklafNJGYb33rWxx66KFEIhFGjBiR6+KISJEZM2YMZWVlqknLQ/v27ePhhx/mggsu6DJ4QNKjmjTps/g50X7zm99w3nnnKUATkayoqKhgzJgxqknLQ0888QRvvfUWn/70p3NdlKKjIE36JHEJEIAHH3yQSCSS45KJSLHSNBz56f7772fIkCFMnTo110UpOgrSpE/C5kRrb2/XSgIikjWa0Db/dHR08MADD3DuuedqapQsUJAmfaIlQESkv9XU1NDW1saBAwdyXRQJrFq1ii1btmhUZ5YoSJM+SbZigFYSEJFsqa6uZt++ffztb3/LdVGEaLeXadOiy2vPmzdP3V2yQEGa9Mk//dM/dUnTSgIikk2xaThUY597sX7JO3bsAKCtrY2GhgYFahmmIE16bdeuXfziF7/gqKOOorq6WnOiiUi/iNXUq19a7qlfcv9QkCYpiZ9uY8yYMbS0tHDPPffQ2tqqlQQkL5jZNDNbb2YbzGxeyPZBZnZPsH2VmdXGbbs6SF9vZlODtPFm9riZrTOztWY2Ny7/CDN71Mz+Evx7eJBuZnZrcKw1ZvaR7J956dCqA/lD/ZL7h4I06VHidBs7duxgwIAB+jJK3jCzcuA24FxgEnCZmU1KyHYl8La7vx+4Gbgh2HcScClwHDANmB8cbz/wTXc/FjgNmBN3zHnASnefCKwMXhO8/8Tg0QDcnoXTLVmHHnoow4cPV01aHhg1alRouvolZ5aCNOlRWLX2/v37Va0t+WQysMHdX3X3vcDdwPSEPNOBxcHzpcAUM7Mg/W533+PuG4ENwGR33+LuzwK4+05gHTA25FiLgYvi0v/bo54ChpvZUZk+2VKmudJyz9055JBDiH593qN+yZmnIE16pGptKQBjgc1xr9t4L6Dqksfd9wPbgapU9g2aRk8CVgVJo919S3CsLcARvSgHZtZgZs1m1rx169aUTlCiNFda7t13331s3LiRL3/5y1qrOcsUpEmo+D5oZWXhfyaq1pY8YiFpnmKebvc1s6HAfcDX3X1HBsqBuy9w9zp3r0vWbCThVJOWW7FWlOOOO4758+fT0tKifslZpAXWpYtYH7RYE2fYxJGq1pY80waMj3s9DvhrkjxtZjYAOAx4q7t9zayCaIAWcff74/K8bmZHufuWoDnzjV6UQ9JQXV3N9u3b2b59O4cddliui1NyFi1axMsvv8xDDz1EeXl5rotT9FSTJl2E9UEDKC8vV7W25KtngIlmNsHMBhIdCLAsIc8yYEbw/GLgMY8uPLsMuDQY/TmBaKf/p4P+ancA69z9pm6ONQN4KC79S8Eoz9OA7bFmUcmMWC3a4YcfTm1trebl6ieRSITq6moaGhoYNGgQO3fuzHWRSoJq0qSLZE0JHR0ddHR09HNpRHrm7vvN7CvACqAcWOjua83sWqDZ3ZcRDbjuNLMNRGvQLg32XWtm9wIvER3ROcfdD5jZx4EvAi+Y2fPBW/2buy8HrgfuNbMrgU3AJcH25cB5RAcftAOXZ/3kS0gkEuGnP/0pEO283traSkNDA4BuGrMosXVlz549+tz7iUVvJAtbXV2dNzc357oYBSsSidDY2MimTZsYM2YMr7/+Ovv37++Sr6amhpaWlv4voEgIM1vt7nW5Lke6dP1KXW1tbeigAV2bskufe+alev1STVqJS7xDeu211wAYMGBAp0BNfdBEJNc00jw39Lnnjvqklbhk/c8OO+wwDa0WkbySbES5Rppnlz733FGQVuKS3Qm99dZbGlotInmlqamJysrKTmlDhgxRLX+WzZo1q0uaWlf6h4K0EpO4BmfijNExukMSkXxTX1/PggULDtbyA5x//vm6icyy1157jfLycsaNG6fWlX6mPmklJLH/2ZYt0ZkB1P9MRApFfX39weDgoosu4pFHHmHr1q1J15KU9LS3t/Pf//3ffO5zn9N0JzmQVk2amU0zs/VmtsHM5oVsH2Rm9wTbVwVLq8S2XR2krzezqUHaeDN73MzWmdlaM5ubTvmkM/U/E5Ficv3119Pe3s51112X66IUrXvvvZft27fzz//8z7kuSknqc5BmZuXAbcC5wCTgMjOblJDtSuBtd38/cDNwQ7DvJKJzFB0HTAPmB8fbD3zT3Y8FTgPmhBxT+kj9z0SkmHzwgx/kk5/8JP/1X/9FWVmZJrfNgh//+Mcce+yxnHHGGbkuSklKpyZtMrDB3V91973A3cD0hDzTgcXB86XAlGAW7+nA3e6+x903Ep34cbK7b3H3ZwHcfSewjpDFiSV18X3Qks2Jp/5nIlKIIpEITz75JNB5clsFapmxZs0annrqKRoaGpL2X5bsSidIGwtsjnvdRteA6mAed98PbAeqUtk3aBo9CViVRhlLWqwPWmtra9IATf3PRKRQNTY2snv37k5p7e3tNDY25qhExSF2c3/CCScAdBlRK/0nnSAtLKxOjASS5el2XzMbSnRR46+7+47QNzdrMLNmM2veunVrikUuLVqDU0SKmSZZzbz4m/uYb3zjG6qdzJF0grQ2YHzc63HAX5PlMbMBwGFE18xLuq+ZVRAN0CLufn+yN3f3Be5e5+51GtUTrqc1ONX/TEQKWbKuGuPHjw9Nl56F3dyrdjJ30gnSngEmmtkEMxtIdCDAsoQ8y4AZwfOLgcc82u62DLg0GP05AZgIPB30V7sDWOfuN6VRtpK3c+dOKioqQrepD5qIFIOwyW0BzjrrrByUpjiodjK/9DlIC/qYfQVYQbSD/73uvtbMrjWzC4NsdwBVZrYBuAqYF+y7FrgXeAl4BJjj7geA04EvAmeb2fPB47y+lrHUxA8SGDVqFHv37mXQoEGd8qgPmogUi8TJbaurqzn++ONZsmQJY8aM6TTiM/76qFGgyWkJqDzj7gX/OPnkk73ULVmyxCsrK51o3z4HfODAgT5r1iyvqalxM/OamhpfsmRJrosqkhFAs+fB9Sfdh65fmTV//vxO10HAKyoqfODAgZ3SKisrdT0MsWTJEh8wYIA+qyxL9fqlZaGKRFg/gr1797J8+XLNgSYlIdOTawfpC83sDTN7MeFYJ5jZk2b2gpk9bGaHBum1ZrY7riXgR9k7Ywlzww03dEnbt28fe/fu7ZSmflbhLrzwQsrLyznkkEM0wCwPKEgrUIlV9/EjceKpH4GUgixNrg3wsyAt0U+Bee7+YeAB4F/itr3i7icGj5mZOD9JXW+uebo+dnXXXXexZ88efvvb3+rmPg8oSCsQ8UHZyJEjueKKKw7Of5YsQAP1I5CSkfHJtQHc/QmiI9ITfQB4Inj+KPDZTJ6M9F1vrnkjRoxQP7U47s6Pf/xjjj/+eE499dRcF0dQkFYQEiel3bZtW5eqe6DLjNAaJCAlJKuTa4d4EYgNkLqEzlMKTTCz58zs92YWupaO5nnMnrARnxUVFQwcOLBL3rfffrvTzW6pr1bQ3NzMc889x8yZM7XCQJ5QkFYAkk1Km8jdtVC6lKqsTa6dxBVE1xZeDQwDYndNW4Bqdz+J6Ij2u2L91TodXPM8Zk3iiM+amhoWLVrEwoULO40CHTx4MB0dHZ32LfV+aj/+8Y855JBD9LuRRwbkugDSs1T7TdTU1NDS0pLdwojkp95Mrt2W6uTaybj7n4F/ADCzY4B/DNL3AHuC56vN7BXgGKC5T2clfVJfXx8aaMSnlZWF11GUaj+17du38/Of/5zPf/7zHHpol/sKyRHVpBWA4cOH95hHTZtS4jI+uXZ3b2ZmRwT/lgH/C/hR8HpUbNCBmR0dHOvVDJyfZJjmA4uK9XcePnw47e3tJXf++U5BWp578skn+fvf/055eXmn9IqKCqqqqtS0KULWJtfGzH4OPAl8wMzazOzK4FiXmdnLwJ+J1rotCtI/Aawxsz8RHZww093DBh5IjoX1XSu1m92wdTqvv/76ku6Xl3dSmUwt3x/FNhnkkiVLDk5AW15e7qNGjfIFCxZoUlqROGgyW0lT7FpLMGnrzTffnOsi9av4c49/1NTU5LpoRS/V65f6pOWZ2J1NbKDAgQMH2LlzJ5WVlepvJiKSQbG+a21tbUycOJFnn30210XqV1qnM/+puTPHEielnTt3bpeRnO+++25JjzgSEcmmcePG8dWvfpUlS5awZs2aXBen36hfXv5TkJZDifOftba2sm3bttC8urMREcmeefPmMXjwYE477bSSmdz2G9/4Rpe0UuuXl+8UpOVQqvOfge5sRESy6de//jX79+9n9+7dJTG5rbuzcuVKKioqGDt2rAah5SkFaf0o1fU2E+nORkQkuxobG9m3b1+ntGKc3Db2O1ReXs7DDz/MxRdfTFtbm9bpzFMK0vpJWNNmMlVVVVo5QESkH5VCJ/rE3yGAhx56qGhrC4uBgrR+kqxpM2y9zVtuuYWWlhbd2YiI9JNkXUoOOeSQolmEPex3qBhrC4uJgrQsim/eTFZz5lpvU0Qk58ImtwV45513imYR9lKoLSw2CtKyJKxaOUxsvU3VmomI5E7YwuxVVVVd8hVyzdNRRx0Vmq6BaflLQVqGpDLfWSINCBARyR/19fWdbprfeit8Ra9CrHnatWtX6KLy+h3KbwrSMqA3850BatoUESkAhT7Za3zlwRFHHEFbWxvf/va31cWmgGhZqD6IRCI0NjayadMmqqureeedd1Ke7yzWvCkiIvmtqamp0zJ9MVdddVWOSpS6xCUG29vbqaio4MMf/jDXX399jksnqVJNWi/1ttYsnqqVRUQKR2I/tTFjxjBo0CBuueUWampq8nrEZ9hIzn379hVsf7pSpSCtl3qzSoDmOxPpP2Y2zczWm9kGM5sXsn2Qmd0TbF9lZrVx264O0teb2dS49IVm9oaZvZhwrBPM7Ekze8HMHjazQ3s6lhSm+H5qr732GjNnzuTVV19l06ZNeT3iUyM5i4OCtF5K9Q9c852J9B8zKwduA84FJgGXmdmkhGxXAm+7+/uBm4Ebgn0nAZcCxwHTgPnB8QB+FqQl+ikwz90/DDwA/EsKx5Ii8OCDD3ZJy4cRn/H9z2pqahg8eHBovkLpTydRCtJ6KdkfuGrNRHJqMrDB3V91973A3cD0hDzTgcXB86XAFIvOJj0duNvd97j7RmBDcDzc/QkgbIjfB4AnguePAp+Ne4/QY0lxyMcaqsRuOJs2bWL37t1dRnOqy03hUZDWSxdccEGXNNWaieTcWGBz3Ou2IC00j7vvB7YDVSnum+hF4MLg+SXA+F6UAzNrMLNmM2veunVrD28l+STVEZ+J0zJlszk0WTec4cOHq/KgwKUVpPVnH5Bciv+y3XbbbRxxxBFUV1frD18kf1hIWuIs0snypLJvoiuAOWa2GhgG7O1FOXD3Be5e5+51o0aN6uGtJJ8kW5ngzTffPBiQzZ49u8sAs2z2W0tWi/f222+r8qDA9TlIy0EfkJxIrEZ2d3bu3Ml//ud/6g9fJH+08V5tFsA44K/J8pjZAOAwok2Zqezbibv/2d3/wd1PBn4OvNKLckgBSxzxOWLECMyMXbt2HQzIbr/99qyvkRlfeZC4BnSM+p8VvnRq0vq7D0hOhFUj7969O+edREWkk2eAiWY2wcwGEr0JXJaQZxkwI3h+MfCYR9dsWwZcGtT8TwAmAk9392ZmdkTwbxnwv4Afxb1Hr44lhSd+xOewYcO6XfovXmtra49NoKk0kyZWHnR0dHTJo/5nxSGdIK2/+4B0kok+HWFfhsQRMskWRtcwZpH8EVxfvgKsANYB97r7WjO71sxifcfuAKrMbANwFTAv2HctcC/wEvAIMMfdDwCY2c+BJ4EPmFmbmV0ZHOsyM3sZ+DPRmrJFPR1LilNvfwu6awINm4ezoaGB2bNnd/qt+trXvhbaB628vFzdcIqMpXoH0GVHs0uAqe7+/wavvwhMdvevxuVZG+RpC16/QrTG7FrgSXdfEqTfASx39/uC17XAL939Q6mUpa6uzpubm7vNk7hKwHnnncfixYs7/aFXVFRgZuzdu7ebI0Vp5QCR3DKz1e5el+typCuV65fkr9ra2tCbeTNLqYatvLycjo6Og6vXpDo5ehgzC61Vk/yT6vUrnZq0fu0Dko6wu5OwPgP79u0LDdAS2/tVjSwiIhA+kKCyspKZM2d2GlmZzIEDB3q9ek0y6oNWfNIJ0vq1D0g6erNKQBh31zBmERHpInEgQew3Yv78+Z1GVnYXqGWCKg+KU58XWHf3/WYW6wNSDiyM9QEBmt19GdE+IHcGfUDeIhrIEeSL9dvYT9c+IGcCI82sDfiOu9/R5zMk/f5jatoUEZFk6uvre7xxT7ZYe0+SNZtWVVUxdOjQg114mpqaVHlQhPocpAG4+3JgeULaNXHP3yU60WPYvk1Al7Df3S9Lp0xhqqurU+ozENYnTXcnIiKSrlgAFesbXVZWxoEDXceUJAZfYf2nYxOoKygrfiWx4kCqfQYWLVrEwoUL1bQpIiIZFz91x+LFi0N/lxJXr5k/f35oc6p+l0pDWjVphSLxDqanqmH98YuISDb15ncpleZUKU59noIjn2gIu0jp0RQcIlKo+mMKDhERERHJEgVpIiIiInlIQZqIiIhIHlKQJiIiIpKHimLggJltBcJXQu9qJPBmFouTLpUvPflePsj/MhZK+WrcfVSuC5OukOtXvn/+mabzLW4633ApXb+KIkjrDTNrzucRYSpfevK9fJD/ZVT5cqvYzy+Rzre46XzTo+ZOERERkTykIE1EREQkD5VikLYg1wXogcqXnnwvH+R/GVW+3Cr280uk8y1uOt80lFyfNBEREZFCUIo1aSIiIiJ5r+iDNDMrN7PnzOyXwesJZrbKzP5iZveY2cA8K9/PzGyjmT0fPE7McflazOyFoCzNQdoIM3s0+AwfNbPD86x8/25mr8V9huflsHzDzWypmf3ZzNaZ2Ufz7PMLK18+fX4fiCvH82a2w8y+nk+fYSaY2feC/4M1ZvaAmQ2P23a1mW0ws/VmNjWX5cwUM7vEzNaaWYeZ1SVsK7rzjTGzacF5bTCzebkuT6aZ2UIze8PMXoxLK6rvajwzG29mjwfXzrVmNjdIz9g5F32QBswF1sW9vgG42d0nAm8DV+akVO9JLB/Av7j7icHj+VwUKsFZQVliF9N5wMrgM1wZvM6lxPJB9P849hkuz1nJ4BbgEXf/IHAC0f/rfPr8wsoHefL5ufv6WDmAk4F24AHy6zPMhEeBD7n78cDLwNUAZjYJuBQ4DpgGzDez8pyVMnNeBD4DPBGfWMTnS3AetwHnApOAy4LzLSY/I/r/Fq/Yvqvx9gPfdPdjgdOAOcH/acbOuaiDNDMbB/wj8NPgtQFnA0uDLIuBi3JTuq7lKyDTiX52kOPPMJ+Z2aHAJ4A7ANx9r7v/nTz5/LopX76aArzi7q3kyWeYKe7+G3ffH7x8ChgXPJ8O3O3ue9x9I7ABmJyLMmaSu69z9/Uhm4ryfAOTgQ3u/qq77wXuJnq+RcPdnwDeSkguqu9qPHff4u7PBs93Er3JHUsGz7mogzTgB8C/Ah3B6yrg73EXwzaiH2iuJJYvpilo9rjZzAbloFzxHPiNma02s4YgbbS7b4HoHylwRM5KF14+gK8En+HCHFavHw1sBRYFTdo/NbNDyJ/PL1n5ID8+v0SXAj8PnufLZ5gNVwC/Dp6PBTbHbcv1NSvbivl8i/nculPM39WDzKwWOAlYRQbPuWiDNDM7H3jD3VfHJ4dkzcnw1iTlg2gzxweBU4ARwLf7u2wJTnf3jxCtop9jZp/IcXkShZXvduB9wInAFuDGHJVtAPAR4HZ3PwnYRX5V9ScrX758fgdZtO/ohcAvcl2WvjKz35rZiyGP6XF5Gok2oURiSSGHKogh+amcb9huIWkFcb4pKOZzK2lmNhS4D/i6u+/I5LEHZPJgeeZ04MKg0/Ng4FCiNVfDzWxAUJs2DvhrvpTPzJa4+xeC7XvMbBHwrRyVDwB3/2vw7xtm9gDRKvvXzewod99iZkcBb+RT+YIqdwDM7CfAL3NUvDagzd1XBa+XEg2C8uXzCy2fu78ey5Djzy/eucCzcWXLl88wZe7+qe62m9kM4Hxgir83N1IbMD4uWy6vWb3S0/kmUbDnm4JiPrfuFNx3tTfMrIJogBZx9/uD5Iydc9HWpLn71e4+zt1riTaTPObu9cDjwMVBthnAQ3lUvi8E/6Gx/nMXEe1gmxNmdoiZDYs9B/4hKM8yop8d5PAzTFa+2GcY+DQ5+gzd/W/AZjP7QJA0BXiJPPn8kpUvXz6/BJfxXlMn5MlnmClmNo1orfmF7t4et2kZcKmZDTKzCcBE4OlclLGfFPP5PgNMtOgMAwOJXveX5bhM/aGovqvxgt/pO4B17n5T3KbMnbO7F/0DOBP4ZfD8aKJf+g1Em04G5Vn5HgNeIPrDuAQYmsNyHQ38KXisBRqD9CqiI1b+Evw7Is/Kd2fwGa4JvixH5fAzPBFoDsryIHB4vnx+3ZQvbz6/oIyVwDbgsLi0vPkMM3SOG4j2V3o+ePwoblsj8AqwHjg312XN0Pl+mmjN0h7gdWBFMZ9v3LmdR3T07iux61UxPYjeSG0B9gX/v1cW23c14Xw/TrTJek3cd/e8TJ6zVhwQERERyUNF29wpIiIiUsgUpImIiIjkIQVpIiIiInlIQZqIiIhIHlKQJiIiIpKHFKSJiIiI5CEFaRLKzA6Y2fPBMi4Pm9nwIL3WzHYH22KPLwXbWszshbj0jwX5U54M1cy+ZWZ/Dt73T3HHHmhmPzCzV8zsL2b2ULBAfWw/N7M7414PMLOtZpYPs+WLSAEJu26Z2b8H16dLzGytmXWYWV1CnqvNbIOZrTezqf1bailGxbwslKRnt7ufCGBmi4E5QFOw7ZXYthBnufubsRfBorMpMbOZwDlEl3baYWaHEV11AeA/gWHAMe5+wMwuB+43s1M9OtnfLuBDZjbE3XcHx3kt1fcWEUnRi8BngB/HJ5rZJKKrCBwHjAF+a2bHuPuB/i+iFAvVpEkqngTG9sP7/Bsw24MFat19u7svNrNK4HLgG7ELnrsvIjpb+dlx+/8a+MfgeeIyQl0Ed8aLzew3QS3gZ8zs/w9qAx8J1mTDzE42s9+b2WozWxG3dNeXzeyZoMbvvqCcmNnPzOxWM/ujmb1qZhd3Vw4RKRzuvs7d14dsmg7c7e573H0j0VUkJocdI6ip+7OZ/TRoNYiY2afM7P8ELQWTg3yHmNnC4DrznAWL0wf7/4+ZPRs8Phakn2lmvzOzpcHxI8HSRVKgFKRJt8ysnOiajvFrzL0vobnzjLhtjwdpq+gFi67BOczdXwnZ/H5gUyx4i9NM9K415m6i6/4NBo4HUinD+4gGdtOJLsP1uLt/GNgN/GMQqP0XcLG7nwws5L0axfvd/RR3PwFYR3QJlJijiC4Zcj5wfQrlEJHCNpbo0l4xbXR/c/t+4Bai16oPAp8nes34FtEbVogukfWYu58CnAV8z6LrFL8BnOPuHwE+B9wad9yTgK8Dk4gunXd6eqcluaTmTklmiJk9D9QCq4FH47al3NzZC0Z0DbTebOuU7u5rgubVy4DlKb7vr919n5m9AJQDjwTpLxA99w8AHwIeDW5Iy4muTQfR5tX/AIYDQ4EVccd90N07iC5YPjrFsohIfkh2LepuHcWwGqvu8m909xcAzGwtsNLdPbgW1QZ5/gG40My+FbweDFQDfwV+aGYnAgeAY+KO+7S7twXHjV3D/9BNOSSPKUiTZHa7+4lBv7BfEu2TdmsP+/RZ0Adtl5kd7e6vJmzeANSY2TB33xmX/hHg4UiBRUkAACAASURBVIS8y4DvE120viqFt94TvH+Hme3z9xaz7SD6/TBgrbt/NGTfnwEXufufzOz/Cd6z03EDam4QKSzbgMMT0kYAG7vZpw0YH/d6HNFgKpn4a0RH3OvYtQei147PJjavmtm/E12Y/gSiLWLvJjnuAfQ7X9DU3CndcvftwNeAb8X6aGXR/wZuM7NDAczsUDNrcPddwGLgpqD5lWDUZyXwWMIxFgLXxu5QM2A9MMrMPhq8b4WZxZpYhwFbgs+lPkPvJyI55u7vEP1uTwEwsxHANLqvkVpGtLvFIDObAEwEnk6zKCuAr8b6lZnZSUH6YcCWoLb+i0Rr+KUIKUiTHrn7c8CfiI5cgq590r7WwyE+YGZtcY9LkuS7HXgceCYY/v57oD3YdjXRu8WXzewvwCXAp+NqvmJlbXP3W3p/luHcfS9wMXCDmf0JeB74WLD5/yPa7+1R4M+Zek8RyQtfAv5X0GT4GPBdd3/FzD5tZm3AR4FfmdkKAHdfC9wLvES028ScDIzsvA6oANYE18TrgvT5wAwze4poU+euNN9H8pQl/MaJiIiISB5QTZqIiIhIHlKHQul3ZnYbXYeF3xLMfZaN97scmJuQ/H/cfU423k9EBMDMqoCVIZumuPu2/i6PFB41d4qIiIjkITV3ioiIiOQhBWkiIiIieUhBmoiIiEgeUpAmIiIikocUpImIiIjkoaKYgmPkyJFeW1ub62KISD9avXr1m+4+KtflSJeuXyKlJ9XrV1EEabW1tTQ3N+e6GCLSj8ysNddlyARdv0RKT6rXLzV3ioiIiOQhBWkiIiIieUhBmoiIiEgeUpAmIiIikocUpIlIr0UiEWpraykrK6O2tpZIJJLrIomUNH0ni1NKQZqZTTOz9Wa2wczmhWwfZGb3BNtXmVlt3Larg/T1ZjY1SBtvZo+b2TozW2tmc+PyX2dma8zseTP7jZmNSf80RSRTIpEIDQ0NtLa24u60trbS0NDA7Nmz9SMh0k/ig7KRI0dyxRVX6DtZhMzdu89gVg68DJwDtAHPAJe5+0txeWYDx7v7TDO7FPi0u3/OzCYBPwcmA2OA3wLHAEcAR7n7s2Y2DFgNXOTuL5nZoe6+Izju14BJ7j6zuzLW1dW5hrCL9I/a2lpaW3sePV5ZWcmCBQuor6/PSjnMbLW712Xl4P1I1y/prdiNUnt7e6/2y/Z3UlKX6vUrlZq0ycAGd3/V3fcCdwPTE/JMBxYHz5cCU8zMgvS73X2Pu28ENgCT3X2Luz8L4O47gXXA2OD1jrjjHgJ0H0WKSL/atGlTSvna29uZO3dupzv5sDt7NdOIdC/xOzJ37txeB2gQ/U42NjZmoYSSLalMZjsW2Bz3ug04NVked99vZtuBqiD9qYR9x8bvGDSNngSsiktrAr4EbAfOCiuUmTUA/7e9e4+PqrwX/f/55goRJRIRMJCLgNSA1Aqya73wqliNHJV2V1t85XRbtScFcf+wejzFV/Z2a2u6y95Wt2dX0FhAdhMFvLRia71UZaOnlYvWC4hIigkEMEBAoAK5zff3x6zEyWQmGZLJrJm1vu/Xa17MPPOsNd8VZq35rmet53nKAQoKCmLYDGNMPBQUFMTUkgbQ1NREU1MTAPX19SxevLjzvfr6em688UZEhJaWls6y8vJyADvbN55XU1NDRUUFO3bsoKCggMrKSoAuZTNnzmT58uWdSVms+140sZ5kmeQQS0uaRCgLb92KVqfHZUVkCPAMcFtoC5qqVqjqGKAGuDVSUKpapapTVXXq8OEpPzOMMSmjsrKSjIyu53fBhvMT19ra2pmgdbCzfeMHke7tvPHGG7vdW/bII4/0qdUs2j5pjRqpJZYkrQEYE/J6NLA7Wh0RyQCGAgd6WlZEMgkmaDWq+myUz34C+HYMMRpjEuSKK64AYMiQIYgIhYWFzJkzh5ycnLh9hp3tG6+rqKjolnxFOmnp7b7xDpmZmeTl5fW4T+bk5HS21pnUEEuStgEYLyLFIpIFzAZWh9VZDdzgPL8WeE2D36zVwGyn92cxMB5Y79yvtgTYoqoPhK5IRMaHvLwG+OhEN8oYM3AeeeQR2traWL9+PYFAgLq6OhYtWkRVVRWFhYWdPxJ5eXl9/gw72zde198Tkby8vC7727Jly9i/f3/EfRIgLS2NRx55xG4jSDG9Jmmq2kbwkuNLBG/wX6Wqm0XkJyJyjVNtCZAnIrXA7cACZ9nNwCrgQ+BFYJ6qtgMXAt8DLnWG2nhXRGY66/q5iGwSkfeBy4HO4TmMMe5qbm7m4YcfprS0lLPPPrvLe2VlZdTV1XX+SDz00EO9tq5lZmaSlZXVpczO9o1XhXYAOJFbBMLr5uTk8NBDD3XZ3yIlXx375MqVKwkEApx55pn93gaTYKqa8o8pU6aoMWbgLV++XAF96aWXYqpfXV2thYWFKiJaWFioc+fO7fK6urq6W53q6uqY1g1s1CQ4/vT3Yccvf6iurtacnBwleF92xEdmZqZmZWV1KcvJyYm435yIzz77TDMyMvTHP/7xAG2dOVGxHr96HSctFdg4Q8YMPFXlvPPOo7W1lQ8++KDPnQXixcZJM6kk2viC6enpBAKBqL07Kysr43KJcsaMGTQ2NrJp06Z+r8v0X6zHr1iG4DDG+FxNTQ133HEHjY2NDBs2jCeeeMLubTHmBES7By0QCBAIBLqUDcS+ddVVV3H77bfzySefUFxcHPf1m4Fhc3caY3rUMVRAY2MjAAcOHKC8vNwGnTUmRsePH2fQoEER30tUJ5mrrroKgN/97ncJ+TwTH5akGWN6FGmoABvLzJiehXYSyM3N5dixY2RmZnapk8hOMuPHj2fChAmWpKUYS9KMMT2KdpnGxjIzJrLwgWqbm5vJysriBz/4QZdhMxI9j+ZVV13FmjVrOHLkSMI+0/SPJWnGmKgCgQDZ2dkR37OxzIyJLFLrc0tLCy+88EKvw2YMpKuuuoqWlhZeeeWVhH6u6TtL0owxUf30pz/l+PHjrl6mMSbVJGvr84UXXsjgwYP5h3/4h87J2u3e0uRmSZoxpovQe2nuueceLrroIpYtW+bqZRpjklnoPjNs2LCoUzm53fq8atUqWlpa+PzzzzvnBrVOQMnNhuAwxnTquJcm9FLN22+/DUBdXZ1LURmTvML3mYMHDyIiZGZmdpmHMxlanysqKmhvb+9S1tEJyE66kpO1pBljOkW6l+bYsWPWk9OYKCLtM6rKySefnHStz8l6GdZEZy1pxphOdhA35sRE2zcOHDjA/v37ExxNzwoKCiLOeuD2ZVgTnbWkGWM6jRo1KmK5HcSNiSzavpGM+0xlZSU5OTldypLhMqyJzpI0YwwQHG7j5JNP7lae6gdxESkVka0iUisiCyK8ny0iK53314lIUch7dznlW0XkCqdsjIi8LiJbRGSziMwPqT9MRF4RkW3Ov6cmYhuNe26//fZuZcm6z5SVlVFVVdV5GRagtLTU9cuwJjpL0owxAPzqV79i69atrg+4GU8ikg48DFwJlADXi0hJWLWbgYOqOg54EFjoLFsCzAYmAqXAImd9bcAdqno28FVgXsg6FwCvqup44FXntfGwv/zlL6Snp5Ofn58S+0xZWVnnWG3f/e53+d3vfsfHH3/sdlgmCkvSjPGx0KED5syZw9lnn01VVZWrA27G2TSgVlW3q2oLsAKYFVZnFrDcef40MEOCzQyzgBWq2qyqnwC1wDRV3aOq7wCo6hFgC5AfYV3LgW8O0HaZJLBp0yaWL1/ObbfdRkNDQ8rtMw8++CDp6elMnjzZxk1LUpakGeNT4VPXqCqffPIJTzzxhNuhxVM+sDPkdQNfJFTd6qhqG3AIyItlWefS6FeAdU7RCFXd46xrD3B6HLbBJJmOk5tzzjkHCM6LmYpee+012tvbaW5utnHTkpQlacb4VKShA44fP+614TYkQln4SKPR6vS4rIgMAZ4BblPVwycUlEi5iGwUkY379u07kUWNy0JPbiA43Mbtt9+ekolNRUVFl7Hc4Itx00xysCTNGJ/yyXAbDcCYkNejgd3R6ohIBjAUONDTsiKSSTBBq1HVZ0PqNIrIKKfOKGBvpKBUtUpVp6rq1OHDh/dx04wbIp3cpGpi45NjQEqzJM0Ynwi9/6ywsJBBgwZFrJeMQwf0wwZgvIgUi0gWwY4Aq8PqrAZucJ5fC7ymwXl9VgOznd6fxcB4YL1zv9oSYIuqPtDDum4Anov7FhlXeSmxSaXhQ/zKkjRjfCD8/rMdO3Zw7Ngx0tK6HgKSdeiAvnLuMbsVeIngDf6rVHWziPxERK5xqi0B8kSkFrgdp0emqm4GVgEfAi8C81S1HbgQ+B5wqYi86zxmOuv6OfANEdkGfMN5bTzES4lNpHHTABYssE7JycKSNGN8INIlGoDc3FzPDLcRjaq+oKpnqepYVa10yu5W1dXO8+Oqep2qjlPVaaq6PWTZSme5Car6B6fsTVUVVZ2squc6jxec95pUdYaqjnf+PeDGNpuBU1lZSVZWVpeyVD25CR83bdSoUaSlpfHaa6+5HZpxWJJmjA9EuxRz8OBBLw23YcyAKysr49prrwXwxMlN6Lhpu3fv5r777uOpp55i+PDhNixHErC5O43xAZuzz5j4GTFiBCeddBJHjhzpHLnfK0aPHk1aWlrnvKMdw3IAKZuIpjJrSTPGBy699NJuZal6icYYtzU0NDB69GjPJWgA//zP/0wgEOhSlqq9V73AkjRjPCq0N+eyZcsYM2YMBQUFnrhEY4ybOpI0L/JS71UvsMudxnhQR2/O0M4C+/fv57HHHrPEzJh+amhoiNg67QV2a0RysZY0YzwoUm/OY8eO2SULY/qpvb2d3bt3e7YlLdKwHHZrhHssSTPGg+yShTEDo7Gxkfb2ds8maR3DcowZE5xsY+jQoXZrhIssSTPGg7w04KYxyaShoQHAs0kaBBO1HTt2UFRURGlpqSVoLrIkzRgP+sEPftCtzC5ZGNN/fkjSOkyaNInNmze7HYavWZJmjAfV1taSlZXFmDFjrDenMXG0c+dOwD9J2kcffURLS4vbofiW9e40xmP27dvHihUruPnmm1m0aJHb4RjjKQ0NDQwaNIi8vDy3QxlwkyZNoq2tjW3btjFx4kS3w/Ela0kzxmMee+wxmpubufXWW90OxRjP8fJAtuE6ErNNmza5HIl/xZSkiUipiGwVkVoRWRDh/WwRWem8v05EikLeu8sp3yoiVzhlY0TkdRHZIiKbRWR+SP1/F5GPROR9EfmNiOT2fzON8Ye2tjYWL17MjBkzKCkpcTscYzzHywPZhvvSl75EWlqa3Zfmol6TNBFJBx4GrgRKgOtFJPzofzNwUFXHAQ8CC51lS4DZwESgFFjkrK8NuENVzwa+CswLWecrwCRVnQx8DNzVv000xj9++9vf0tDQwD/+4z+6HYoxnuSnJG3QoEGMHz/eWtJcFEtL2jSgVlW3q2oLsAKYFVZnFrDcef40MEOCbcGzgBWq2qyqnwC1wDRV3aOq7wCo6hFgC5DvvH5ZVducdb0F+GNvMKYfOqaAuu6660hPT+fw4cNuh2SM5wQCAXbt2uWbJA2C96VZkuaeWJK0fGBnyOsGpyxiHSfBOgTkxbKsc2n0K8C6CJ99E/CHGGI0xrc6poDqmMqlvb2dOXPmUFNT43JkxnjLvn37aG1t9VWSNnHiRGprazl27JjbofhSLElapLsjNcY6PS4rIkOAZ4DbVLXLqb+IVBC8LBrxl0ZEykVko4hs3LdvXw/hG+NtkaaAOnr0qE0BZUyc+WmMtA6TJk1CVfnoo4/cDsWXYknSGoAxIa9HA7uj1RGRDGAocKCnZUUkk2CCVqOqz4auTERuAK4CylQ1PCEEQFWrVHWqqk4dPnx4DJthjDfZFFDGJIZfkzSwHp5uiSVJ2wCMF5FiEcki2BFgdVid1cANzvNrgdec5Go1MNvp/VkMjAfWO/erLQG2qOoDoSsSkVLgx8A1qtq1ecAY041NAWVMYvgxSRs3bhxZWVmWpLmk1yTNucfsVuAlgjf4r1LVzSLyExG5xqm2BMgTkVrgdmCBs+xmYBXwIfAiME9V24ELge8Bl4rIu85jprOuXwInA6845Y/Ea2ON8aLKykqys7O7lNkUUMbEX0NDA5mZmfjp6k1mZiYTJkywJM0lMc04oKovAC+Eld0d8vw4cF2UZSuByrCyN4l8vxrOMB7GmBiVlZXx7LPP8uyzzyIiFBQUUFlZaVNAGRNnO3fuJD8/n7Q0f40DP2nSJP70pz+5HYYv2bRQxnhAS0sLJSUlNuikMQPIT2OkhZo0aRJPPvkkhw8f5pRTTnE7HF/x1+mAMR6kqqxfv57zzz/f7VCM8TQ/J2kAH374ocuR+I8lacakuJ07d7J3715L0owZQKrq2yTN5vB0jyVpxqS4DRs2AFiSZswAampqorm52ZdJWnFxMYMHD7bbKVxgSZoxKW7Dhg1kZmby5S9/2e1QjPEsPw6/0SEtLY2JEydaS5oLLEkzJsVt2LCByZMndxuGwxgTP35O0sDm8HSLJWnGpLBAIMDGjRvtUqcxA8zvSdrEiRP59NNPaWpqcjsUX7EkzZgUtm3bNg4fPmxJmjEDrKGhgfT0dEaOHOl2KK7Yu3cvAMOHD6eoqIiamojTaps4syTNmBTW0Wlg2rRpLkdijLc1NDRwxhlnkJ6e7nYoCVdTU8Mvf/lLINjLtb6+nvLyckvUEsCSNGNS2IYNGzjppJM4++yz3Q7FGE/z6/AbABUVFRw7dqxL2dGjR6moqHApIv+wJM2YFLZhwwbOO+88X57dG5NIO3fu9G2StmPHjhMqN/FjSZoxKaq1tZW//OUvdj+aMQPMzwPZAhQUFJxQuYkfS9KMSVGbNm3i+PHjlqQZM8A+++wzjh496tskrbKykpycnC5lOTk5VFZWuhSRf1iSZkyKspkGjEkMvw+/UVZWRlVVFcOGDQPgjDPOoKqqirKyMpcj8z5L0oxJURs2bGDYsGGceeaZboeS1ESkVES2ikitiCyI8H62iKx03l8nIkUh793llG8VkStCypeKyF4R2RS2rnNF5C0ReVdENoqIdbv1AL8naRBM1J555hkAHn/8cUvQEsSSNGNS1IYNG5g6dSoi4nYoSUtE0oGHgSuBEuB6ESkJq3YzcFBVxwEPAgudZUuA2cBEoBRY5KwP4HGnLNy/Afeq6rnA3c5rk+IsSQsaO3YsAH/9619djsQ/LEkzJsXU1NRQUFDAe++9x7p162ysop5NA2pVdbuqtgArgFlhdWYBy53nTwMzJJj5zgJWqGqzqn4C1DrrQ1XXAgcifJ4CpzjPhwK747kxJvFqamq48847Abj44ot9vb/l5+eTnZ1tSVoCZbgdgDEmdjU1NZSXl3P06FEADh06RHl5OYBdfogsH9gZ8roB+LtodVS1TUQOAXlO+Vthy+b38nm3AS+JyP0ET4K/FqmSiJQD5WA95JJZ+P62Y8cOX+9vaWlpFBcXW5KWQNaSZkwKqaio6PzB6GCDSvYo0rVgjbFOLMuGmwv8SFXHAD8ClkSqpKpVqjpVVacOHz68l1Uat9j+1t3YsWMtSUsgS9KMSSE2qOQJawDGhLweTfdLkJ11RCSD4GXKAzEuG+4G4Fnn+VM4l0dNarL9rbuOJE21t/MVEw+WpBmTQmxQyRO2ARgvIsUikkWwI8DqsDqrCSZXANcCr2nwF2g1MNvp/VkMjAfW9/J5u4HpzvNLgW1x2AbjEtvfuhs7diyff/5554TrZmBZkmZMCqmsrCQ7O7tLmQ0qGZ2qtgG3Ai8BW4BVqrpZRH4iItc41ZYAeSJSC9wOLHCW3QysAj4EXgTmqWo7gIg8CfwZmCAiDSJys7Ou/wX8QkTeA36Gc9+ZSU133313tzK/72/WwzOxrOOAMSmkrKyMVatWsXr1akSEgoICKisrfXkTc6xU9QXghbCyu0OeHweui7JsJdDtF1lVr49S/01gSn/iNcnj8OHDAIwYMYK9e/fa/kbXJO1rX4vYL8bEkSVpxqSYXbt2cfHFF7N27Vq3QzHGs5qbm7n//vuZPn06a9ascTucpFFcXIyIWEtagtjlTmNSyL59+3jnnXe4/PLL3Q7FGE/79a9/za5du7jrrrvcDiWpZGdnM3r0aEvSEsSSNGNSyKuvvoqqWpJmzABqb29n4cKFnHfeebavRWDDcCSOJWnGpJCXX36ZU089lSlT7LYnY+KtpqaGoqIiMjMzqa2t5aKLLrJp1yKwJC1xLEkzJkWoKq+88gqXXXYZ6enpvS9gjIlZx+wC9fX1nWOA/epXv/L1NFDRjB07lr1793LkyBG3Q/E8S9KMSREfffQRDQ0NfOMb33A7FGM8oaPlLC0tjRtuuMFmF4hRRw/P7du3uxyJ91mSZkyKePnllwEsSTMmDsJbztrb2yPW8/PsAtHYWGmJY0maMSni5Zdf5qyzzqKoqMjtUIxJOaGtZkVFRcyfP79by1kkfp5dIBpL0hLHxkkzJgU0NzezZs0abrrpJrdDMSbldLSadSRl9fX1MS3n99kFosnNzWXYsGGWpCWAtaQZkwL+9Kc/cfToURsOwJg+qKioiKnVDCA9PR0RobCwkKqqKl/PLtAT6+GZGJakGZPkampq+OY3vwnAvHnzrLeZMb0Iv7R5Ii1ny5cvJxAIUFdXZwlaDyxJS4yYkjQRKRWRrSJSKyILIryfLSIrnffXiUhRyHt3OeVbReQKp2yMiLwuIltEZLOIzA+pf51TFhCRqf3fRGNSV8dlmo45BHfu3El5ebklasZEEd4hoKcELS8vj8LCQms564OxY8eyY8cOWltb3Q7F03pN0kQkHXgYuBIoAa4XkZKwajcDB1V1HPAgsNBZtgSYDUwESoFFzvragDtU9Wzgq8C8kHVuAv4esIkJje9FukxjwwIYE120S5vhg9Lm5OTw0EMPUVdXZy1nfTB27Fja29tjbqU0fRNLS9o0oFZVt6tqC7ACmBVWZxaw3Hn+NDBDgnvELGCFqjar6idALTBNVfeo6jsAqnoE2ALkO6+3qOrW/m6YMV4Qrfu/DQtgTGTR9g1VtVazOLIenokRS+/OfGBnyOsG4O+i1VHVNhE5BOQ55W+FLZsfuqBzafQrwLoTiBsRKQfKwbpIG+8qKCiIeKZq33ljIhs2bBhNTU3dygsLC6mrq0t8QB5lSVpixNKSFmniMo2xTo/LisgQ4BngNlU9HEMsX6xEtUpVp6rq1OHDh/daP/xG0pqamj6X3XLLLXFbLlaR1m+8r7KykqysrC5lNiyAMV8IPTbm5eXR1NREWlrXnzbbZ+Jv1KhRDBo0yJK0gaaqPT6AC4CXQl7fBdwVVucl4ALneQawn2CC1qVuWL1M5/XtUT53DTC1t/hUlSlTpmhPqqurNScnRwkmiApoZmamZmVl9aks/NHX5XJycnTu3LlaWFioIqKFhYVaXV0dc/x5eXldlquuru51XSb1fPvb31bA/l/DABs1huNDsj96O36Z6CIdG9PS0rS8vNyOhQlQUlKis2bNcjuMlBTr8SuWJC0D2A4UA1nAe8DEsDrzgEec57OBVc7ziU79bGf57UC6k8D9F/AfPXxu3JK0wsLCHpOlZHpEStzy8vJ6XS5SUpiTk2MHJw+YN2+e5ubmuh1G0rEkzUQ7thcWFrodmi9cffXVOmnSJLfDSEmxHr96vdypqm3ArQRbvbY4CdhmEfmJiFzjVFsC5IlILXA7sMBZdjOwCvgQeBGYp6rtwIXA94BLReRd5zETQES+JSINBFvwfi8iL/UWY29S6Sbro0ePsnjx4i7dxyPdXxGutbWVlpaWbuuyXoCpr7GxkREjRrgdhjFJxzrWuKu1tZXNmzfbbTgDKKZpoVT1BeCFsLK7Q54fB66LsmwlUBlW9iaR71dDVX8D/CaWuGIV7eZrP7CDVeqzJM2YyMaMGRPxGGcdawZeTU0Nr776aseVL+rr6ykvLwewXrNx5IsZByorK8nJyelSlpmZ2e2G7FjLwvV1ufBxewaCHaxSX2NjIyNHjnQ7DGOSzvTp07uVWSeBxKioqOg2kK1dvYk/XyRpZWVlVFVVdRkjZ9myZSxdurRPZXPnzo3LcnPmzOmWPEZL3EJHxs7Ly4spKbSDlTdYS5oxXwjtzfnrX/+aMWPGUFBQYOOfJZhdak6QWG5cS/ZHKt94G94jc+7cud16K0XqABCpJ2d1dbUWFBQooCeddJJ1GvCAY8eOKaD33Xef26EkHazjgO9E6s05ePBgO9a5wDpt9E+sxy/XD1DxeHjtINffoTSuu+46HTNmjAYCgQGK0CRKfX29AvrYY4+5HUrSsSTNfywxSB6REmYbUSB2sR6/fHG5M9WUlZX1az656dOns3PnThtd2wMaGxsB7HKnMdgltmTScRtRx7Fp+PDhdql5AFiS5kEdN9P+93//t8uRmP6yJM34Wej9Z6effnrUetZByh1lZWXs2LGDIUOG8O1vf9sStAFgSZoHlZSUkJeXZ0maB1iSZvyqpqaG8vLyzjEj9+3bh6qSmZnZpZ51kHJXVlYWM2bM4MUXXwzeQ2XiypI0D0pLS+OSSy6xJM0DLEkzflVRUcHRo0e7lZ9yyildesnbJTb3lZaWUldXx8cff+x2KJ4T02C2JvVMnz6d3/zmN+zcuZMxY8a4HY7po08//ZRTTjmFQYMGuR2KMQkV7T6zAwcOsH///gRHY3pyxRVXAPDiiy8yYcIEl6PxFmtJ8yi7L80bbIw041fR7jOz+8+ST3FxMRMmTODFF190OxTPsSTNo8455xxyc3MtSUtxlqQZv6qsrLT7z1JIaWkpa9as4dixY26H4imWpHlUeno6F198sSVpKc6SNONXZWVlTJw4kYyMDLv/LAWUlpZy/Phx1q5d63YonmJJmodNnz6dbdu2sWfPHrdDMX1kSZrxq/b2dj755BNuvPHGPo8ZaRJneEosTAAAIABJREFU+vTpDBo0yC55xpklaR5m96WltpaWFg4ePGhJmvGlTZs2cejQIS655BK3QzExGDx4MNOnT7ckLc4sSfOwc889l5NPPpk1a9a4HYrpg7179wI2/EZ/iUipiGwVkVoRWRDh/WwRWem8v05EikLeu8sp3yoiV4SULxWRvSKyKcL6/tGpv1lE/m2gtsvr3njjDQBL0lJIaWkpH330kc12E0eWpHlYRkYGxcXFLFmyhLS0NIqKirjllls6R/AuKiqipqbG7TBNFDZGWv+JSDrwMHAlUAJcLyIlYdVuBg6q6jjgQWChs2wJMBuYCJQCi5z1ATzulIV/3teBWcBkVZ0I3B/vbfKLtWvXUlBQYL05U0hraysQ7O1pvy/xYUmah9XU1LBlyxba2tpQVerr61m8eHHnCN719fWUl5fbjpSkLEmLi2lArapuV9UWYAXBJCrULGC58/xpYIaIiFO+QlWbVfUToNZZH6q6FjgQ4fPmAj9X1Wan3t54b5AfqCpr1661VrQUUlNTwz333NP52n5f4sOSNA+rqKjoPLOJ5ujRo1RUVCQoInMiOpK0kSNHuhxJSssHdoa8bnDKItZR1TbgEJAX47LhzgIudi6b/reInB+pkoiUi8hGEdm4b9++mDfGL7Zt20ZjY6MlaSkk0gwR9vvSf5akeVi0Ebv7Ws8klrWkxYVEKAufYDBanViWDZcBnAp8FbgTWOW0ynVdiWqVqk5V1anDhw/vZZX+0zGMgyVpqSPa74j9vvSPJWkeFuu9HHbPR3JqbGxkyJAh5OTkuB1KKmsAQudFGw3sjlZHRDKAoQQvZcaybKTPe1aD1gMB4LQ+R+9Tb7zxBqeffjpnnXWW26GYGNkMEQPDkjQPq6ys7PUHPisry0bwTlI2RlpcbADGi0ixiGQR7AiwOqzOauAG5/m1wGuqqk75bKf3ZzEwHljfy+f9FrgUQETOArIAm2jyBK1du5aLL76YCI2QJklF+r2xGSL6z5I0DysrK6OqqorCwsLOEbvnzp3b+TojI4PCwkIbIDJJWZLWf849ZrcCLwFbgFWqullEfiIi1zjVlgB5IlIL3A4scJbdDKwCPgReBOapajuAiDwJ/BmYICINInKzs66lwJnO0BwrgBuchM/EaMeOHdTV1dmlzhQT+nvTYeHChfb70k8ZbgdgBlZZWVnUneSBBx7gjjvu4P3332fy5MkJjsz0prGx0S73xIGqvgC8EFZ2d8jz48B1UZatBLo1Bajq9VHqtwD/sz/x+p2Nj5a6On5vPvjgAyZPnmy3asSBtaT52A033EB2djaPPvqo26GYCKwlzfjR2rVrGTp0KOecc47boZg+mjRpEvn5+Tb7QBxYkuZjeXl5fOc73+HXv/41f/vb39wOx4RobW2lqanJkjTjGzU1NRQVFVFVVUVzczMrVqxwOyTTRyJCaWkpr7zyCm1tbW6Hk9IsSfO5OXPmcOTIETsgJpl9+/ahqpakGV+oqamhvLyc+vp6AI4fP24Doaa40tJSPvvsM9566y23Q0lplqT53AUXXMDo0aOZN2+eTRWVRGyMNOMnNhCq91x22WWkp6fbJc9+siTN55544gkaGxtpaWmxqaKSiCVpxk9sIFTvyc3N5YILLuAPf/iD26GkNEvSfC7S1FF2Bus+S9KMn9hAqN505ZVX8s4773Qez8yJsyTN5+wMNjlZkmb8xAZC9abS0lIAXnrpJZcjSV2WpPmcncEmp8bGRgYPHsyQIUPcDsWYAdcxEGpmZiYAhYWFVFVV2UCoKe7cc89lxIgRdl9aP1iS5nORzmBtqij3dYyRZtPiGL8oKysjNzeX8vJy6urqLEHzgLS0NM466yxWrlxpHdP6yJI0nwufOiorK4vBgwdz9dVXux2ar9lAtsZvmpub2bdvH/n5+W6HYuKkpqaGdevWEQgErGNaH1mSZigrK6Ouro5AIMCbb77J4cOHuffee90Oy9csSTN+s2fPHgBL0jykoqKClpaWLmXWMe3ExJSkiUipiGwVkVoRWRDh/WwRWem8v05EikLeu8sp3yoiVzhlY0TkdRHZIiKbRWR+SP1hIvKKiGxz/j21/5tpYnX++efzgx/8gAcffJAzzjjDmqhd0tjYyMiRI90Ow5iE2bVrFwCjR492ORITL9Yxrf96TdJEJB14GLgSKAGuF5GSsGo3AwdVdRzwILDQWbYEmA1MBEqBRc762oA7VPVs4KvAvJB1LgBeVdXxwKvOa5NA5513HqrKnj17rInaBe3t7ezfv99a0oyvdCRp1pLmHdYxrf9iaUmbBtSq6nZVbQFWALPC6swCljvPnwZmSPCO51nAClVtVtVPgFpgmqruUdV3AFT1CLAFyI+wruXAN/u2aaavfv7zn3crsybqxNm/fz+BQMCSNOMrlqR5jxtDq3TMAeuVq0CxJGn5wM6Q1w18kVB1q6OqbcAhIC+WZZ1Lo18B1jlFI1R1j7OuPcDpkYISkXIR2SgiG/ft2xfDZphYWRO1u2yMNONHDQ0NDB48mNzcXLdDMXHS0TGto+Vs0KBBAzq0SugcsF65ChRLkhZpDACNsU6Py4rIEOAZ4DZVPRxDLF+sRLVKVaeq6tThw4efyKKmF9ZE7S5L0owf7dq1i/z8fBt2xmPKysqor6/njjvuoL29fUBHDvDiHLCxJGkNwJiQ16OB3dHqiEgGMBQ40NOyIpJJMEGrUdVnQ+o0isgop84oYG+sG2PiI1ITdVpaGvfcc487AfmMJWnGjzqSNONNs2bNorW1dUBnH4h2tae+vj5lL4HGkqRtAMaLSLGIZBHsCLA6rM5q4Abn+bXAa6qqTvlsp/dnMTAeWO/cr7YE2KKqD/SwrhuA5050o0z/hI+dNnz4cAKBALfeemtKfslTjSVpxo8sSfO2r33ta5x22mk899zA/aT3dLUn/BLoLbfc0iVxC39dU1MT8f628LJIy8WVqvb6AGYCHwN/BSqcsp8A1zjPBwFPEewYsB44M2TZCme5rcCVTtlFBC97vg+86zxmOu/lEezVuc35d1hv8U2ZMkXNwKmurtbMzEx1/s8U0JycHK2urnY7NE+68847NTs7WwOBgNuhJDVgo8Zw/Er2hx2/VAOBgGZnZ+udd97pdihmAH3/+9/X3NxcbWlpGZD1V1dXq4h0+a3q6yMzM7Pb715GRka3svBHrL+NsR6/YhonTVVfUNWzVHWsqlY6ZXer6mrn+XFVvU5Vx6nqNFXdHrJspbPcBFX9g1P2pqqKqk5W1XOdxwvOe02qOkNVxzv/HoglRjNwKioqaG1t7VKW6tf5k1VNTQ2LFi2iubmZ4uJia7E0vtDU1ERzc7ONkeZxs2bN4rPPPuONN94YkPWPGDECVSUvLw8RobCwsM/ram1t7fa719bW1q0sXLx/G23GAdMr6+2ZGB09kz7//HMAT/RMMiYWNvyGP3zjG99g0KBBA3bJ8z/+4z84/fTT2bVrF4FAgLq6un4lan0Vz99GS9JMr6y3Z2J4sWeSMbFoaGgALEnzupNOOonLLruM1atXd9wOFTfbtm3j97//PXPnziU7O7uzPFJHuIHuQRzP30ZL0kyvovX2HMgBCf3IWiyNX1lLmn/MmjWLuro6Pvjgg7iu9z//8z/JzMxkzpw5XcrDO8IVFhYyZ86cbr9p4TIzM8nKyuq1LFy8B+u1JM30KvxLPmzYMAKBAHl5eW6H5inWYmn8ateuXYiIzVfrA1dffTUiEtdLnocOHWLZsmVcf/31Eb9DZWVl1NXVdV4CXbRoUbfEbe7cuV1eL1u2jKVLl/ZaFr5c3AfrjaV3QbI/rHdUYjU3N2txcbFOmTLFeiDGUaSeSdaLNjqsd6dn3HzzzTpy5Ei3wzAJMm7cOM3KylIR0cLCwj4f46qrq7WwsLDzeHnffffFOdKBE+vxy1rSzAnLysriX/7lX3j77bf57W9/63Y4npGfn9+tZ9JATqFiTLKwMdL8o6amhvr6elpaWlDt+9RNoVNAdfjZz37muY5WlqSZPikrK2PkyJF897vftQFu46Sqqorc3Fx27tzZ2SxvCZrxA0vS/CNeQzr5paOVJWmmT1auXMmBAwdobW3t19mQCdq/fz/PPPMM3/ve9xg8eLDb4RiTULt27bIx0nwiXh2k/NLRypI00ycVFRW0tLR0KfPiWUyiLF++nJaWFn74wx+6HYoxCXXs2DEOHDhgLWk+EWsHqUhTMvVlPanOkjTTJ345i0kEVaWqqooLL7yQiRMnuh2OMQllw2/4S7Rxy0JP8EPvN4t2peaee+7pNt5ZvIe/SAaWpJk+8ctZTCKsWbOGjz/+mPLycrdDMSbhLEnzl/AhnTqmcpo/f35nq9n8+fN7vd/swIEDqCojRozwdEcrS9JMn0Q6G8rMzPTcWcxA6mjOv/TSS0lLS6O9vd3tkIxJOEvS/Cd03LJf/OIXZGZmcuzYsc5Ws6ampojLdVypOXz4MD/72c+4/PLL+fTTTz3d0cqSNNMn4WdDOTk5iAgzZsxwO7SUEN59PBAIcOutt1rHC+M7lqT5W6TentGceuqpFBUVMXToUJqamrj44osHODr3WZJm+iz0bOi9994jEAhw7733uh1WSvBL93FjerNr1y6GDBnCKaec4nYoxgUnch/zwYMHu4yL9q//+q+eP7G1JM3Exbhx45gzZw6PPvoo+fn5NnZaL6zjReKISKmIbBWRWhFZEOH9bBFZ6by/TkSKQt67yynfKiJXhJQvFZG9IrIpymf+bxFRETltILbJS2yMNH+Ldh9zXl5el+mWhgwZ0m1Sdj+c2FqSZuKmpKQEVWX37t1deuTccsstPXal9iPreJEYIpIOPAxcCZQA14tISVi1m4GDqjoOeBBY6CxbAswGJgKlwCJnfQCPO2WRPnMM8A3AMu4Y2Bhp/hbp/uacnBweeuihLvNtfv755xGX9/qJrSVpJm4WLlzYrezo0aM88sgjPXal9qPKykrS0rrufl7sPp4EpgG1qrpdVVuAFcCssDqzgOXO86eBGRLs2z8LWKGqzar6CVDrrA9VXQsciPKZDwL/h+B8gqYXDQ0N1pLmY+H3N0frpenXE1tL0kzcRDuj8WMTdW8uuOACAoEAQ4cO9XT38SSQD+wMed3glEWso6ptwCEgL8ZluxCRa4Bdqvpe/8L2h0AgwJ49eyxJ87nQ+5uj9dKM1uLm9RPbDLcDMN5RUFDQ5abOnni9ibo3ixcvJj09nc2bN9sP1MCSCGXhLVzR6sSy7BcrEckBKoDLew1KpBwoB++3BPRk7969tLW12T5getWRuFVUVLBjxw4KCgqorKz0/ImttaSZuIk2knQkfv5hOnbsGEuXLuVb3/qW/TgNvAZgTMjr0cDuaHVEJAMYSvBSZizLhhoLFAPviUidU/8dERkZXlFVq1R1qqpOHT58+AltkJfY8BvmRMTS4uY1lqSZuIl0b8GcOXO6JW6DBg3yfBN1T1asWMGBAweYN2+e26H4wQZgvIgUi0gWwY4Aq8PqrAZucJ5fC7ymwWv0q4HZTu/PYmA8sD7aB6nqB6p6uqoWqWoRwSTvPFX9NL6b5B2WpBnTM0vSTFyFn+ksWrSoS+KWnp7OsGHDuPbaa90O1RWqysMPP8zEiROZPn262+F4nnOP2a3AS8AWYJWqbhaRnzj3jwEsAfJEpBa4HVjgLLsZWAV8CLwIzFPVdgAReRL4MzBBRBpE5OZEbpdXWJJmTM8sSTMDLjRxe/7559m9ezcjRozw1ZAcHVNApaen8/bbbzN16tSol4JNfKnqC6p6lqqOVdVKp+xuVV3tPD+uqtep6jhVnaaq20OWrXSWm6Cqfwgpv15VR6lqpqqOVtUlET63SFX3J2IbU9WuXbtIT09nxIgRbodiTFKyJM0k1IEDB0hPT+fQoUO+GZIjdAqojp6uq1at8vQ2GxOLhoYGRo0aRXp6eu+VjfEhS9JMQlVUVHSbSPzo0aPMnz/fswPeRpoC6tixY74fhsQYm23AmJ5ZkmYSKtrQG01NTZ4d8NamgDKmu5qaGtasWcO6des8d2JmTLxYkmYSKtahN7w04K1fR8o2JpqOWwDa2toAPHdiZky8WJJmEirSWGrR1NfXe+IS6D/90z916yTgh5GyjYkm0i0AXjoxMyZeLEkzCRVpLLW8vLyo9b1wCXTjxo2oKiNHjrQpoIzBbgEwJlaWpJmECx9L7aGHHoqpdS2VzrQ7htxIS0vj0UcfZebMmezZs8dXI2UbE43dAmBMbCxJM66L1LoWTSqcaUcacuP1119PyVZAYwZCZWUlaWldf37sFgBjurMkzSSF8Na1aIlaKpxp25AbxvTs61//OoFAgNzcXLsFwJgeWJJmklKkDgZZWVkpcaZt99sY07Pf//73AKxdu9ZuATCmBzElaSJSKiJbRaRWRBZEeD9bRFY6768TkaKQ9+5yyreKyBUh5UtFZK+IbApb15dF5M8i8oGIPC8ip/R980yqCr8Emp2djYjwta99ze3QejVmzJiI5anQCmhMIqxevZqioiImTZrkdijGJLVekzQRSQceBq4ESoDrRaQkrNrNwEFVHQc8CCx0li0BZgMTgVJgkbM+gMedsnC/Ahao6jnAb4A7T3CbjEeEXgLdunUrAGeffXbSD8lx5ZVXdiuz+22MCTp69Ch//OMfufrqq23+WmN6EUtL2jSgVlW3q2oLsAKYFVZnFrDcef40MEOCe98sYIWqNqvqJ0Ctsz5UdS1wIMLnTQDWOs9fAb59AttjPOrNN99EVWlubk7qITkOHz7Mc889x9ixYykoKLD7bYwJ88c//pHjx49zzTXXuB2KMUkvI4Y6+cDOkNcNwN9Fq6OqbSJyCMhzyt8KW7a3ido2AdcAzwHXAZGvHRlfqaiooKWlpUtZx5AcyZT8/PSnP+XTTz9l/fr1nH/++W6HY0zSef755znllFO45JJL3A7FmKQXS0tapPZojbFOLMuGuwmYJyJvAycDLZEqiUi5iGwUkY379u3rZZUm1SXzzfihY6Ldf//9TJ8+3RI044rQ72Iy3hIQCAR4/vnnKS0tJSsry+1wjEl6sSRpDXRtzRoN7I5WR0QygKEEL2XGsmwXqvqRql6uqlOAJ4G/RqlXpapTVXXq8OHDY9gMk8qi3XQ/aNAgV3+UIo2Jtn79+qT7cTSpL1ICFlp22mmncdNNN3WbpeOWW25JmsRt48aNNDY2cvXVV7sWgzEpRVV7fBC8JLodKAaygPeAiWF15gGPOM9nA6uc5xOd+tnO8tuB9JDlioBNYes63fk3Dfgv4KbeYpwyZYoab6uurtacnBwl2BKrgIpIl9eA5uTk6Ny5c7WwsFBFRAsLC7W6unrA4iosLOwWA6CFhYUD9pkmCNiovRwbUuERy/Er0vc/MzNTs7KyIn7/etpPcnJyBnSf6ElFRYWmp6drU1OTK59vTLKI9fgV00EEmAl8TLBVq8Ip+wlwjfN8EPAUwY4B64EzQ5atcJbbClwZUv4ksAdoJdjidrNTPt/5rI+BnwPSW3yWpPlDdXV1l+QrLy/P9R+lSIliRwxmYPkpSYt2MtDXR15eXpd9KdKJTfj+Fq3sREyePFkvueSSE1rGGC+Ka5KW7A9L0vwpWoIUS8tWrD9A4WWhP2b5+fmakZFhLWku8VOSdiLf9Xg8IrXSRSqLdAIUbT/Kz89XQHNzc11ryTMmWcR6/JJg3dQ2depU3bhxo9thmAQrKiqivr4+5vqFhYXs2LGDYcOGceTIkS69RTMzMxGRXssiSU9Pp729vfN1Tk6ODbmRACLytqpOdTuO/orl+HWi3/UOIsJAH+PT09MJBAIUFBQwc+ZMli9f3mVatEj7ke0jxu9iPX7ZtFAmZUWaOqqnwTE7bqhuamrqlni1trbGVBZJbm5ul8nh7cfHxFuk73pmZma3HpKZmZnk5eV1fhfnzJnTbbl4a29vRzXYUWHx4sXd5q2NtB91DJ9jjOmZJWkmZYVPHZWoH6VwBw4c6DI5vCVoJt4ifdeXLVvG0qVLu5Xt37+/87u4aNGibsvl5eW5vTlAcgyfY0yys8udxnNqamqoqKhgx44dFBQU9Oky0YkoLCykrq5uQD/DdOeny53x1DFsTHiLV6j+XP6Ple03xs/scqfxrdA5P+vq6igsLOx1mWiXjnobcNPm5DSpJlKr3Ny5c2NqpQstS09Pj7j+8FsOIu1Htt8YE6NYehck+8N6d5qeRBtjKi8vr1+9Owd6DDbTM3zUuzMZRdqvoo1T2N+hO4zxmliPX7HM3WlMSuu4Ryz0EmhlZWXEe8diLTPG705kvwqtb4yJnd2TZoxJSXZPmjEmVdk9acYYY4wxKcySNGOMMcaYJGRJmjHGGGNMErIkzRhjjDEmCXmi44CI7AP6OmLpacD+OIaTzPyyrX7ZTvD3thaq6nC3gomXfh6/3JLK3zuL3R2pGvtAxR3T8csTSVp/iMhGL/QQi4VfttUv2wm2rcYdqfx/YbG7I1Vjdztuu9xpjDHGGJOELEkzxhhjjElClqRBldsBJJBfttUv2wm2rcYdqfx/YbG7I1VjdzVu39+TZowxxhiTjKwlzRhjjDEmCVmSZowxxhiThHyZpInIv4vIRyLyvoj8RkRyQ967S0RqRWSriFzhZpzxICLXichmEQmIyNSw9zy1rQAiUupsT62ILHA7nngSkaUisldENoWUDRORV0Rkm/PvqW7GGC8iMkZEXheRLc73d75T7sntTSY9/O1Xisi7zqNORN4NWSYpjiU9xH6uiLzlxL5RRKY55SIi/9eJ/X0ROS8JY/+yiPxZRD4QkedF5JSQZZLl7z5IRNaLyHtO7Pc65cUiss7ZX1eKSJZTnu28rnXeL0rC2G914lMROS2kfmK/M6rquwdwOZDhPF8ILHSelwDvAdlAMfBXIN3tePu5rWcDE4A1wNSQci9ua7qzHWcCWc72lbgdVxy37xLgPGBTSNm/AQuc5ws6vsup/gBGAec5z08GPna+s57c3mR6RPvbh9X5BXC38zxpjiU9fG9eBq50ymcCa0Ke/wEQ4KvAumT7uwMbgOlO+U3AT5Pw7y7AEOd5JrDO+XuuAmY75Y8Ac53ntwCPOM9nAytd/LtHi/0rQBFQB5wWUj+h3xlftqSp6suq2ua8fAsY7TyfBaxQ1WZV/QSoBaa5EWO8qOoWVd0a4S3PbSvB+GtVdbuqtgArCG6nJ6jqWuBAWPEsYLnzfDnwzYQGNUBUdY+qvuM8PwJsAfLx6PYmkx7+9kCwJQH4DvCkU5Q0x5IeYlegowVqKLDbeT4L+C8NegvIFZFRCQ4b6DH2CcBap9orwLed58n0d1dV/ZvzMtN5KHAp8LRTHrq/hu7HTwMznO9VwkWLXVX/oqp1ERZJ6HfGl0lamJsIZsUQ3CF2hrzXQMjByWO8uK1e3KbejFDVPRA8yAOnuxxP3DmXQr5C8AzX89ubTML+9h0uBhpVdZvzOin3u7DYbwP+XUR2AvcDdznVUiH2TcA1zlvXAWOc50kVu4ikO5fA9xJMJv8KfBbSIBIaX2fszvuHgLzERvyF8NhVdV0P1RP6d/dskiYifxSRTREes0LqVABtQE1HUYRVJf0YJbFsa6TFIpQl/bb2wovb5GsiMgR4BrhNVQ+7HY+f9PC3v54vWtEgCfe7CLHPBX6kqmOAHwFLOqpGWDzZYr8JmCcibxO8DNrSUTXC4q7FrqrtqnouwStT0wjeatOtmvNvUscuIpN6qJ7Q2DMGasVuU9XLenpfRG4ArgJmqHOhmWBGPCak2mi+aBZPWr1taxQpua298OI29aZRREap6h6nyX2v2wHFi4hkEvyxqlHVZ51iz25vMonyt0dEMoC/B6aEVE+q/S5K7DcA853nTwG/cp4nfeyq+hHB+6gRkbOA/+FUT6rYO6jqZyKyhuD9WrkikuG0loXG1xF7g/OdGkr3WzkSLiT2UoItmJEk9O/u2Za0nohIKfBj4BpVPRry1mpgttPzpBgYD6x3I8YE8OK2bgDGOz2KsgjekLra5ZgG2mqCP0A4/z7nYixx49yfsgTYoqoPhLzlye1NJj387QEuAz5S1YaQsqQ5lvQQ+25guvP8UqDjUu1q4B+cHntfBQ51XE5PtGixi8jpzr9pwD8RvAEfkuvvPlycURJEZDDB78kW4HXgWqda6P4auh9fC7wW0liSUFFi/6iHRRL7nRnIXgnJ+iB4g+VO4F3n8UjIexUEr6VvxekNlMoP4FsEM/9moBF4yavb6mzTTIK9ov4KVLgdT5y37UlgD9Dq/J/eTPA+jlcJ/ui8CgxzO844betFBC8hvB+yn8706vYm0yPa395573FgToRlkuJY0sP35iLgbYK9IdcBU5z6AjzsxP4BIT3gkyj2+c4x7WPg5zgzBSXZ330y8Bcn9k180fP3TIKJYy3BFsxsp3yQ87rWef/MJIz9/3OOs20Ek/xfufGdsWmhjDHGGGOSkC8vdxpjjDHGJDtL0owxxhhjkpAlacYYY4wxSciSNGOMMcaYJGRJmjHGGGNMErIkzcSdiKwRkSvCym4TkUUi8qKIfCYivwt7v1hE1onINhFZ6YxzZowxrhKRPBF513l8KiK7nOd/FZHXRWSLiGwWkfm9r82YE2NDcJi4E5EfAl9V1RtDyt4C7gSygBzgh6p6Vcj7q4BnVXWFiDwCvKeqixMcujHGRCUi9wB/U9X7nRkvRqnqOyJyMsFx2L6pqh+6GqTxFGtJMwPhaeAqEcmGzsmCzwDeVNVXgSOhlZ2Rti91lgNYDnwz2spF5HERWeycxW4XkekistQ5o308pN7lIvJnEXlHRJ5y5sRDRO4WkQ3O/KZVzud3tAAuFJH1IvKxiFwcp7+HMcZjVHWPqr7jPD9CcIT9qBNtO8eXB0VkrXOsOl9EnnWuHtwXUu9/Osegd0XkURFJd8oXi8hGp9Xu3pD6dSJyr3Oc+0BEvjRwW20SzZI0E3fX4+y1AAAC1klEQVSq2kRwFOlSp2g2sFKjN9vmAZ9pcH43CI7yHPVg5ziVYGL3I+B54EFgInCOiJwrIqcRnELlMlU9D9gI3O4s+0tVPV9VJwGDCc7h2iFDVacBtwH/EtMGG2N8zTkR/QrB2Qx60qKqlxCc2uk5YB4wCfi+c1n1bOC7wIUanPC7HShzlq1Q1akER8ifLiKTQ9a73znOLQb+d3y2yiQDz06wblz3JMHk7Dnn35t6qCsRynq7Dv+8qqqIfAA0quoHACKyGSgiOOltCfD/nIayLODPzrJfF5H/Q/Cy6zBgM8FED6BjQua3nfUYY0xUTgv9M8Btqnq4l+odcwl/AGxWZ85HEdlOcNLuiwhOXr/BOW4NBvY6y3xHRMoJ/m6PInh8e995L/S49ff93SaTPCxJMwPlt8ADInIeMLjjskAU+4FcEclwWtNGE5wrrSfNzr+BkOcdrzMInoG+oqrXhy4kIoOARQTnW9vp3GMyKMJ627H9wxjTAxHJJJig1ajqs73Vp/fjlgDLVfWusM8pJthCdr6qHnRu67Djlg/Y5U4zIFT1b8AaYCnBVrWe6irwOnCtU3QDwRa4/ngLuFBExgGISI6InMUXB7b9zhnwtdFWYIwx0Tj3si4BtqjqA3Fa7avAtSJyuvMZw0SkEDgF+Bw4JCIjgCvj9HkmyVmSZgbSk8CXgRUdBSLyBvAUMENEGkKG6vgxcLuI1BK8R21Jfz5YVfcB3weeFJH3CSZtX1LVz4DHCF5u+C2woT+fY4zxrQuB7wGXhgzRMbM/K3R6hv4T8LJz3HqFYA/S94C/ELw1Yynw//oXukkVNgSHMcYYY0wSspY0Y4wxxpgkZDcYmqQlIhXAdWHFT6lqpRvxGGNMb0TkYYKXQkM9pKrL3IjHpDa73GmMMcYYk4TscqcxxhhjTBKyJM0YY4wxJglZkmaMMcYYk4QsSTPGGGOMSUKWpBljjDHGJKH/HxVA/rQPlQ1LAAAAAElFTkSuQmCC\n",
573 |       "text/plain": [
574 |        "<Figure size 720x576 with 4 Axes>"
575 |       ]
576 |      },
577 |      "metadata": {
578 |       "needs_background": "light"
579 |      },
580 |      "output_type": "display_data"
581 |     }
582 |    ],
583 |    "source": [
584 |     "fig, axes = plt.subplots(2, 2, figsize=(10, 8))\n",
585 |     "axef = axes.ravel()\n",
586 |     "for r, res in enumerate(results):\n",
587 |     "    axef[r].plot(var_vals[r], res, 'ko-')\n",
588 |     "    axef[r].set_xlabel(input_cols[r])"
589 |    ]
590 |   },
591 |   {
592 |    "cell_type": "markdown",
593 |    "metadata": {},
594 |    "source": [
595 |     "It turns out that in this case, dask already does quite well with paralellizing the task? \n",
596 |     "\n",
597 |     "Is there a way we can make dask struggle or fail? At what dataset size do we see problems?\n",
598 |     "\n",
599 |     "Is there a more concise way to paralellize the task?"
600 |    ]
601 |   },
602 |   {
603 |    "cell_type": "code",
604 |    "execution_count": null,
605 |    "metadata": {},
606 |    "outputs": [],
607 |    "source": []
608 |   }
609 |  ],
610 |  "metadata": {
611 |   "kernelspec": {
612 |    "display_name": "Python 3",
613 |    "language": "python",
614 |    "name": "python3"
615 |   },
616 |   "language_info": {
617 |    "codemirror_mode": {
618 |     "name": "ipython",
619 |     "version": 3
620 |    },
621 |    "file_extension": ".py",
622 |    "mimetype": "text/x-python",
623 |    "name": "python",
624 |    "nbconvert_exporter": "python",
625 |    "pygments_lexer": "ipython3",
626 |    "version": "3.6.7"
627 |   }
628 |  },
629 |  "nbformat": 4,
630 |  "nbformat_minor": 2
631 | }
632 | 


--------------------------------------------------------------------------------
/rasp-data-loading.ipynb:
--------------------------------------------------------------------------------
   1 | {
   2 |  "cells": [
   3 |   {
   4 |    "cell_type": "markdown",
   5 |    "metadata": {},
   6 |    "source": [
   7 |     "# xarray use case: Neural network training\n",
   8 |     "\n",
   9 |     "\n",
  10 |     "**tl;dr**\n",
  11 |     "\n",
  12 |     "1. This notebook is an example of reading from a climate model netCDF file to train a neural network. Neural networks (for use in parameterization research) require random columns of several stacked variables at a time. \n",
  13 |     "\n",
  14 |     "2. Experiments in this notebook show:\n",
  15 |     "    1. Reading from raw climate model output files is super slow (1s per batch... need speeds on the order of ms)\n",
  16 |     "    2. open_mfdataset is half as fast as opening the same dataset with open_dataset\n",
  17 |     "    3. Pure h5py is much faster than reading the same dataset using xarray (even using the h5 backend)\n",
  18 |     "\n",
  19 |     "3. Currently, I revert to preformatting the dataset (flatten time, lat, lon). This gets the reading speed down to milliseconds per batch.\n",
  20 |     "\n",
  21 |     "**Conclusions**\n",
  22 |     "\n",
  23 |     "Reading straight from the raw netCDF files (with all dimensions intact) is handy and might be necessary for later applications (using continuous time slices or lat-lon regions for RNNs or CNNs).\n",
  24 |     "\n",
  25 |     "However, at the moment this is many orders of magnitude too slow. Preprocessing seems required.\n",
  26 |     "\n",
  27 |     "What would be a good way of speeding this up without too extensive post processing?\n"
  28 |    ]
  29 |   },
  30 |   {
  31 |    "cell_type": "code",
  32 |    "execution_count": 1,
  33 |    "metadata": {},
  34 |    "outputs": [],
  35 |    "source": [
  36 |     "import xarray as xr\n",
  37 |     "import numpy as np"
  38 |    ]
  39 |   },
  40 |   {
  41 |    "cell_type": "code",
  42 |    "execution_count": 2,
  43 |    "metadata": {},
  44 |    "outputs": [
  45 |     {
  46 |      "data": {
  47 |       "text/plain": [
  48 |        "'0.11.2'"
  49 |       ]
  50 |      },
  51 |      "execution_count": 2,
  52 |      "metadata": {},
  53 |      "output_type": "execute_result"
  54 |     }
  55 |    ],
  56 |    "source": [
  57 |     "xr.__version__"
  58 |    ]
  59 |   },
  60 |   {
  61 |    "cell_type": "markdown",
  62 |    "metadata": {},
  63 |    "source": [
  64 |     "## Load an example dataset\n",
  65 |     "\n",
  66 |     "I uploaded a sample dataset here: http://doi.org/10.5281/zenodo.2559313\n",
  67 |     "\n",
  68 |     "The files are around 1GB large. Let's download it.\n",
  69 |     "\n",
  70 |     "NOTE: I have all my data on an SSD"
  71 |    ]
  72 |   },
  73 |   {
  74 |    "cell_type": "code",
  75 |    "execution_count": 47,
  76 |    "metadata": {},
  77 |    "outputs": [],
  78 |    "source": [
  79 |     "# Modify this path!\n",
  80 |     "DATADIR = '/local/S.Rasp/tmp/'"
  81 |    ]
  82 |   },
  83 |   {
  84 |    "cell_type": "code",
  85 |    "execution_count": 48,
  86 |    "metadata": {},
  87 |    "outputs": [
  88 |     {
  89 |      "name": "stdout",
  90 |      "output_type": "stream",
  91 |      "text": [
  92 |       "--2019-02-07 13:08:52--  https://zenodo.org/record/2559183/files/sample_SPCAM_1.nc\n",
  93 |       "Resolving zenodo.org (zenodo.org)... 137.138.76.77\n",
  94 |       "Connecting to zenodo.org (zenodo.org)|137.138.76.77|:443... connected.\n",
  95 |       "HTTP request sent, awaiting response... 200 OK\n",
  96 |       "Length: 923498891 (881M) [application/octet-stream]\n",
  97 |       "Saving to: ‘/local/S.Rasp/tmp/sample_SPCAM_1.nc’\n",
  98 |       "\n",
  99 |       "sample_SPCAM_1.nc   100%[===================>] 880.72M  6.59MB/s    in 1m 49s  \n",
 100 |       "\n",
 101 |       "2019-02-07 13:10:42 (8.09 MB/s) - ‘/local/S.Rasp/tmp/sample_SPCAM_1.nc’ saved [923498891/923498891]\n",
 102 |       "\n",
 103 |       "--2019-02-07 13:10:42--  https://zenodo.org/record/2559183/files/sample_SPCAM_2.nc\n",
 104 |       "Resolving zenodo.org (zenodo.org)... 137.138.76.77\n",
 105 |       "Connecting to zenodo.org (zenodo.org)|137.138.76.77|:443... connected.\n",
 106 |       "HTTP request sent, awaiting response... 200 OK\n",
 107 |       "Length: 923498891 (881M) [application/octet-stream]\n",
 108 |       "Saving to: ‘/local/S.Rasp/tmp/sample_SPCAM_2.nc’\n",
 109 |       "\n",
 110 |       "sample_SPCAM_2.nc   100%[===================>] 880.72M  24.5MB/s    in 87s     \n",
 111 |       "\n",
 112 |       "2019-02-07 13:12:09 (10.1 MB/s) - ‘/local/S.Rasp/tmp/sample_SPCAM_2.nc’ saved [923498891/923498891]\n",
 113 |       "\n",
 114 |       "--2019-02-07 13:12:09--  https://zenodo.org/record/2559183/files/sample_SPCAM_concat.nc\n",
 115 |       "Resolving zenodo.org (zenodo.org)... 137.138.76.77\n",
 116 |       "Connecting to zenodo.org (zenodo.org)|137.138.76.77|:443... connected.\n",
 117 |       "HTTP request sent, awaiting response... 200 OK\n",
 118 |       "Length: 1846816429 (1.7G) [application/octet-stream]\n",
 119 |       "Saving to: ‘/local/S.Rasp/tmp/sample_SPCAM_concat.nc’\n",
 120 |       "\n",
 121 |       "sample_SPCAM_concat 100%[===================>]   1.72G  3.80MB/s    in 3m 28s  \n",
 122 |       "\n",
 123 |       "2019-02-07 13:15:38 (8.46 MB/s) - ‘/local/S.Rasp/tmp/sample_SPCAM_concat.nc’ saved [1846816429/1846816429]\n",
 124 |       "\n"
 125 |      ]
 126 |     }
 127 |    ],
 128 |    "source": [
 129 |     "!wget -P $DATADIR https://zenodo.org/record/2559313/files/sample_SPCAM_1.nc\n",
 130 |     "!wget -P $DATADIR https://zenodo.org/record/2559313/files/sample_SPCAM_2.nc\n",
 131 |     "!wget -P $DATADIR https://zenodo.org/record/2559313/files/sample_SPCAM_concat.nc"
 132 |    ]
 133 |   },
 134 |   {
 135 |    "cell_type": "code",
 136 |    "execution_count": 49,
 137 |    "metadata": {},
 138 |    "outputs": [
 139 |     {
 140 |      "name": "stdout",
 141 |      "output_type": "stream",
 142 |      "text": [
 143 |       "-rw-r--r-- 1 S.Rasp ls-craig 881M Feb  7 13:00 /local/S.Rasp/tmp//sample_SPCAM_1.nc\r\n",
 144 |       "-rw-r--r-- 1 S.Rasp ls-craig 881M Feb  7 13:00 /local/S.Rasp/tmp//sample_SPCAM_2.nc\r\n",
 145 |       "-rw-r--r-- 1 S.Rasp ls-craig 1.8G Feb  7 13:00 /local/S.Rasp/tmp//sample_SPCAM_concat.nc\r\n"
 146 |      ]
 147 |     }
 148 |    ],
 149 |    "source": [
 150 |     "!ls -lh $DATADIR/sample_SPCAM*"
 151 |    ]
 152 |   },
 153 |   {
 154 |    "cell_type": "markdown",
 155 |    "metadata": {},
 156 |    "source": [
 157 |     "The files are typical climate model output files. `sample_SPCAM_1.nc` and `sample_SPCAM_2.nc` are two contiguous output files. `sample_SPCAM_concat.nc` is the concatenated version of the two files."
 158 |    ]
 159 |   },
 160 |   {
 161 |    "cell_type": "code",
 162 |    "execution_count": 53,
 163 |    "metadata": {},
 164 |    "outputs": [
 165 |     {
 166 |      "name": "stdout",
 167 |      "output_type": "stream",
 168 |      "text": [
 169 |       "CPU times: user 56 ms, sys: 0 ns, total: 56 ms\n",
 170 |       "Wall time: 54.7 ms\n"
 171 |      ]
 172 |     }
 173 |    ],
 174 |    "source": [
 175 |     "%%time\n",
 176 |     "ds = xr.open_mfdataset(DATADIR + 'sample_SPCAM_1.nc')"
 177 |    ]
 178 |   },
 179 |   {
 180 |    "cell_type": "code",
 181 |    "execution_count": 54,
 182 |    "metadata": {},
 183 |    "outputs": [
 184 |     {
 185 |      "data": {
 186 |       "text/plain": [
 187 |        "<xarray.Dataset>\n",
 188 |        "Dimensions:       (crm_x: 32, crm_y: 1, crm_z: 28, ilev: 31, isccp_prs: 7, isccp_prstau: 49, isccp_tau: 7, lat: 64, lev: 30, lon: 128, tbnd: 2, time: 48)\n",
 189 |        "Coordinates:\n",
 190 |        "  * lat           (lat) float64 -87.86 -85.1 -82.31 -79.53 ... 82.31 85.1 87.86\n",
 191 |        "  * lon           (lon) float64 0.0 2.812 5.625 8.438 ... 351.6 354.4 357.2\n",
 192 |        "  * crm_x         (crm_x) float64 0.0 4.0 8.0 12.0 ... 112.0 116.0 120.0 124.0\n",
 193 |        "  * crm_y         (crm_y) float64 0.0\n",
 194 |        "  * crm_z         (crm_z) float64 992.6 976.3 957.5 936.2 ... 38.27 24.61 14.36\n",
 195 |        "  * lev           (lev) float64 3.643 7.595 14.36 24.61 ... 957.5 976.3 992.6\n",
 196 |        "  * ilev          (ilev) float64 2.255 5.032 10.16 18.56 ... 967.5 985.1 1e+03\n",
 197 |        "  * isccp_prs     (isccp_prs) float64 90.0 245.0 375.0 500.0 620.0 740.0 900.0\n",
 198 |        "  * isccp_tau     (isccp_tau) float64 0.15 0.8 2.45 6.5 16.2 41.5 219.5\n",
 199 |        "  * isccp_prstau  (isccp_prstau) float64 90.0 90.0 90.0 ... 900.0 900.0 900.2\n",
 200 |        "  * time          (time) object 0000-01-01 00:00:00 ... 0000-01-01 23:29:59\n",
 201 |        "Dimensions without coordinates: tbnd\n",
 202 |        "Data variables:\n",
 203 |        "    P0            float64 ...\n",
 204 |        "    time_bnds     (time, tbnd) object dask.array<shape=(48, 2), chunksize=(48, 2)>\n",
 205 |        "    date_written  (time) |S8 dask.array<shape=(48,), chunksize=(48,)>\n",
 206 |        "    time_written  (time) |S8 dask.array<shape=(48,), chunksize=(48,)>\n",
 207 |        "    ntrm          int32 ...\n",
 208 |        "    ntrn          int32 ...\n",
 209 |        "    ntrk          int32 ...\n",
 210 |        "    ndbase        int32 ...\n",
 211 |        "    nsbase        int32 ...\n",
 212 |        "    nbdate        int32 ...\n",
 213 |        "    nbsec         int32 ...\n",
 214 |        "    mdt           int32 ...\n",
 215 |        "    nlon          (lat) int32 dask.array<shape=(64,), chunksize=(64,)>\n",
 216 |        "    wnummax       (lat) int32 dask.array<shape=(64,), chunksize=(64,)>\n",
 217 |        "    hyai          (ilev) float64 dask.array<shape=(31,), chunksize=(31,)>\n",
 218 |        "    hybi          (ilev) float64 dask.array<shape=(31,), chunksize=(31,)>\n",
 219 |        "    hyam          (lev) float64 dask.array<shape=(30,), chunksize=(30,)>\n",
 220 |        "    hybm          (lev) float64 dask.array<shape=(30,), chunksize=(30,)>\n",
 221 |        "    gw            (lat) float64 dask.array<shape=(64,), chunksize=(64,)>\n",
 222 |        "    ndcur         (time) int32 dask.array<shape=(48,), chunksize=(48,)>\n",
 223 |        "    nscur         (time) int32 dask.array<shape=(48,), chunksize=(48,)>\n",
 224 |        "    date          (time) int32 dask.array<shape=(48,), chunksize=(48,)>\n",
 225 |        "    datesec       (time) int32 dask.array<shape=(48,), chunksize=(48,)>\n",
 226 |        "    nsteph        (time) int32 dask.array<shape=(48,), chunksize=(48,)>\n",
 227 |        "    DTV           (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 228 |        "    DTVKE         (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 229 |        "    FLNS          (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 230 |        "    FLNT          (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 231 |        "    FLUT          (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 232 |        "    FSNS          (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 233 |        "    FSNT          (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 234 |        "    LHFLX         (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 235 |        "    PHCLDICE      (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 236 |        "    PHCLDLIQ      (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 237 |        "    PHQ           (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 238 |        "    PRECC         (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 239 |        "    PRECL         (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 240 |        "    PRECSC        (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 241 |        "    PRECSL        (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 242 |        "    PRECSTEN      (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 243 |        "    PRECT         (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 244 |        "    PRECTEND      (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 245 |        "    PS            (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 246 |        "    QAP           (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 247 |        "    QCAP          (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 248 |        "    QIAP          (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 249 |        "    QRL           (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 250 |        "    QRS           (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 251 |        "    SHFLX         (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 252 |        "    SOLIN         (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 253 |        "    SPDQ          (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 254 |        "    SPDT          (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 255 |        "    T             (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 256 |        "    TAP           (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 257 |        "    TPHYSTND      (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 258 |        "    TS            (time, lat, lon) float32 dask.array<shape=(48, 64, 128), chunksize=(48, 64, 128)>\n",
 259 |        "    UAP           (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 260 |        "    VAP           (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 261 |        "    VD01          (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 262 |        "    VPHYSTND      (time, lev, lat, lon) float32 dask.array<shape=(48, 30, 64, 128), chunksize=(48, 30, 64, 128)>\n",
 263 |        "Attributes:\n",
 264 |        "    Conventions:  CF-1.0\n",
 265 |        "    source:       CAM\n",
 266 |        "    case:         AndKua_aqua_SPCAM3.0_sp_fbp32\n",
 267 |        "    title:        \n",
 268 |        "    logname:      tg847872\n",
 269 |        "    host:         \n",
 270 |        "    Version:      $Name:  $\n",
 271 |        "    revision_Id:  $Id: history.F90,v 1.26.2.38 2003/12/15 18:52:35 hender Exp $"
 272 |       ]
 273 |      },
 274 |      "execution_count": 54,
 275 |      "metadata": {},
 276 |      "output_type": "execute_result"
 277 |     }
 278 |    ],
 279 |    "source": [
 280 |     "ds"
 281 |    ]
 282 |   },
 283 |   {
 284 |    "cell_type": "markdown",
 285 |    "metadata": {},
 286 |    "source": [
 287 |     "## Random columns for machine learning parameterizations\n",
 288 |     "\n",
 289 |     "For the work on ML parameterizations that a few of us are doing now, we would like to work one column at a time. One simple example would be predicting the temperature and humidity tendencies (TPHYSTND and PHQ) from the temperature and humidity profiles (TAP and QAP). \n",
 290 |     "\n",
 291 |     "This means we would like to give the neural network a stacked vector containing the inputs (2 x 30 levels) and ask it to predict the outputs (also 2 x 30 levels).\n",
 292 |     "\n",
 293 |     "In NN training, we usually train on a batch of data at a time. Batches typically have a few hundred samples (columns in our case). It is really important that the samples in a batch are not correlated but rather represent a random sample of the entire dataset.\n",
 294 |     "\n",
 295 |     "To achieve this we will write a data generator that loads the batches by randomly selecting along the time, lat and lon dimensions."
 296 |    ]
 297 |   },
 298 |   {
 299 |    "cell_type": "code",
 300 |    "execution_count": 57,
 301 |    "metadata": {},
 302 |    "outputs": [],
 303 |    "source": [
 304 |     "class DataGenerator(object):\n",
 305 |     "    \"\"\"\n",
 306 |     "    Data generator that randomly (if shuffle = True) picks columns from the dataset and returns them in \n",
 307 |     "    batches. For each column the input variables and output variables will be stacked.\n",
 308 |     "    \"\"\"\n",
 309 |     "    def __init__(self, fn_or_ds, batch_size=128, input_vars=['TAP', 'QAP'], output_vars=['TPHYSTND', 'PHQ'], \n",
 310 |     "                 shuffle=True, engine='netcdf4'):\n",
 311 |     "        self.ds = xr.open_mfdataset(fn_or_ds, engine=engine) if type(fn_or_ds) is str else fn_or_ds\n",
 312 |     "        self.batch_size = batch_size\n",
 313 |     "        self.input_vars = input_vars\n",
 314 |     "        self.output_vars = output_vars\n",
 315 |     "        self.ntime, self.nlat, self.nlon = self.ds.time.size, self.ds.lat.size, self.ds.lon.size\n",
 316 |     "        self.ntot = self.ntime * self.nlat * self.ntime\n",
 317 |     "        self.n_batches = self.ntot // batch_size\n",
 318 |     "        self.indices = np.arange(self.ntot)\n",
 319 |     "        if shuffle:\n",
 320 |     "            self.indices = np.random.permutation(self.indices)\n",
 321 |     "    def __getitem__(self, index):\n",
 322 |     "        time_indices, lat_indices, lon_indices = np.unravel_index(\n",
 323 |     "            self.indices[index*self.batch_size:(index+1)*self.batch_size], (self.ntime, self.nlat, self.nlon)\n",
 324 |     "        )\n",
 325 |     "        \n",
 326 |     "        X, Y = [], []\n",
 327 |     "        for itime, ilat, ilon in zip(time_indices, lat_indices, lon_indices):\n",
 328 |     "            X.append(\n",
 329 |     "                np.concatenate(\n",
 330 |     "                    [self.ds[v].isel(time=itime, lat=ilat, lon=ilon).values for v in self.input_vars]\n",
 331 |     "                )\n",
 332 |     "            )\n",
 333 |     "            Y.append(\n",
 334 |     "                np.concatenate(\n",
 335 |     "                    [self.ds[v].isel(time=itime, lat=ilat, lon=ilon).values for v in self.output_vars]\n",
 336 |     "                )\n",
 337 |     "            )\n",
 338 |     "\n",
 339 |     "        return np.array(X), np.array(Y)"
 340 |    ]
 341 |   },
 342 |   {
 343 |    "cell_type": "markdown",
 344 |    "metadata": {},
 345 |    "source": [
 346 |     "### Multi-file dataset\n",
 347 |     "\n",
 348 |     "Let's start by using the split dataset `sample_SPCAM_1.nc` and `sample_SPCAM_2.nc`."
 349 |    ]
 350 |   },
 351 |   {
 352 |    "cell_type": "code",
 353 |    "execution_count": 58,
 354 |    "metadata": {},
 355 |    "outputs": [],
 356 |    "source": [
 357 |     "gen = DataGenerator(DATADIR + 'sample_SPCAM_[1-2].nc')"
 358 |    ]
 359 |   },
 360 |   {
 361 |    "cell_type": "code",
 362 |    "execution_count": 59,
 363 |    "metadata": {},
 364 |    "outputs": [],
 365 |    "source": [
 366 |     "# This is how we get one batch of inputs and corresponding outputs\n",
 367 |     "x, y = gen[0]"
 368 |    ]
 369 |   },
 370 |   {
 371 |    "cell_type": "code",
 372 |    "execution_count": 60,
 373 |    "metadata": {},
 374 |    "outputs": [
 375 |     {
 376 |      "data": {
 377 |       "text/plain": [
 378 |        "((128, 60), (128, 60))"
 379 |       ]
 380 |      },
 381 |      "execution_count": 60,
 382 |      "metadata": {},
 383 |      "output_type": "execute_result"
 384 |     }
 385 |    ],
 386 |    "source": [
 387 |     "x.shape, y.shape"
 388 |    ]
 389 |   },
 390 |   {
 391 |    "cell_type": "code",
 392 |    "execution_count": 61,
 393 |    "metadata": {},
 394 |    "outputs": [],
 395 |    "source": [
 396 |     "# A little test function to check the timing.\n",
 397 |     "def test(g, n):\n",
 398 |     "    for i in range(n):\n",
 399 |     "        x, y = g[i]"
 400 |    ]
 401 |   },
 402 |   {
 403 |    "cell_type": "code",
 404 |    "execution_count": 64,
 405 |    "metadata": {},
 406 |    "outputs": [
 407 |     {
 408 |      "name": "stdout",
 409 |      "output_type": "stream",
 410 |      "text": [
 411 |       "CPU times: user 13.3 s, sys: 1.34 s, total: 14.6 s\n",
 412 |       "Wall time: 14.3 s\n"
 413 |      ]
 414 |     }
 415 |    ],
 416 |    "source": [
 417 |     "%%time\n",
 418 |     "test(gen, 10)"
 419 |    ]
 420 |   },
 421 |   {
 422 |    "cell_type": "code",
 423 |    "execution_count": 65,
 424 |    "metadata": {},
 425 |    "outputs": [
 426 |     {
 427 |      "name": "stdout",
 428 |      "output_type": "stream",
 429 |      "text": [
 430 |       "CPU times: user 12.5 s, sys: 1.28 s, total: 13.8 s\n",
 431 |       "Wall time: 13.5 s\n"
 432 |      ]
 433 |     }
 434 |    ],
 435 |    "source": [
 436 |     "# does shuffling make a big difference\n",
 437 |     "gen = DataGenerator(DATADIR + 'sample_SPCAM_[1-2].nc', shuffle=True)\n",
 438 |     "%time test(gen, 10)"
 439 |    ]
 440 |   },
 441 |   {
 442 |    "cell_type": "markdown",
 443 |    "metadata": {},
 444 |    "source": [
 445 |     "So it takes more than one second to read one batch. This is way too slow to train a neural network in a reasonable amount of time. Shuffling doesn't seem to be a huge problem, but even without shuffling I am probably accessing the data in a different order than saved on disc. \n",
 446 |     "\n",
 447 |     "Let's check what actually takes that long."
 448 |    ]
 449 |   },
 450 |   {
 451 |    "cell_type": "code",
 452 |    "execution_count": 66,
 453 |    "metadata": {},
 454 |    "outputs": [
 455 |     {
 456 |      "name": "stdout",
 457 |      "output_type": "stream",
 458 |      "text": [
 459 |       "The line_profiler extension is already loaded. To reload it, use:\n",
 460 |       "  %reload_ext line_profiler\n"
 461 |      ]
 462 |     }
 463 |    ],
 464 |    "source": [
 465 |     "%load_ext line_profiler"
 466 |    ]
 467 |   },
 468 |   {
 469 |    "cell_type": "code",
 470 |    "execution_count": 67,
 471 |    "metadata": {},
 472 |    "outputs": [],
 473 |    "source": [
 474 |     "%lprun -f gen.__getitem__ test(gen, 10)"
 475 |    ]
 476 |   },
 477 |   {
 478 |    "cell_type": "markdown",
 479 |    "metadata": {},
 480 |    "source": [
 481 |     "Output:\n",
 482 |     "\n",
 483 |     "```\n",
 484 |     "Timer unit: 1e-06 s\n",
 485 |     "\n",
 486 |     "Total time: 24.5229 s\n",
 487 |     "File: <ipython-input-57-78b9d254df3b>\n",
 488 |     "Function: __getitem__ at line 18\n",
 489 |     "\n",
 490 |     "Line #      Hits         Time  Per Hit   % Time  Line Contents\n",
 491 |     "==============================================================\n",
 492 |     "    18                                               def __getitem__(self, index):\n",
 493 |     "    19        10         17.0      1.7      0.0          time_indices, lat_indices, lon_indices = np.unravel_index(\n",
 494 |     "    20        10        267.0     26.7      0.0              self.indices[index*self.batch_size:(index+1)*self.batch_size], (self.ntime, self.nlat, self.nlon)\n",
 495 |     "    21                                                   )\n",
 496 |     "    22                                                   \n",
 497 |     "    23        10         10.0      1.0      0.0          X, Y = [], []\n",
 498 |     "    24      1290       4642.0      3.6      0.0          for itime, ilat, ilon in zip(time_indices, lat_indices, lon_indices):\n",
 499 |     "    25      1280       1399.0      1.1      0.0              X.append(\n",
 500 |     "    26      1280       1721.0      1.3      0.0                  np.concatenate(\n",
 501 |     "    27      1280   12256070.0   9575.1     50.0                      [self.ds[v].isel(time=itime, lat=ilat, lon=ilon).values for v in self.input_vars]\n",
 502 |     "    28                                                           )\n",
 503 |     "    29                                                       )\n",
 504 |     "    30      1280       2393.0      1.9      0.0              Y.append(\n",
 505 |     "    31      1280       1750.0      1.4      0.0                  np.concatenate(\n",
 506 |     "    32      1280   12253415.0   9573.0     50.0                      [self.ds[v].isel(time=itime, lat=ilat, lon=ilon).values for v in self.output_vars]\n",
 507 |     "    33                                                           )\n",
 508 |     "    34                                                       )\n",
 509 |     "    35                                           \n",
 510 |     "    36        10       1218.0    121.8      0.0          return np.array(X), np.array(Y)\n",
 511 |     "```"
 512 |    ]
 513 |   },
 514 |   {
 515 |    "cell_type": "markdown",
 516 |    "metadata": {},
 517 |    "source": [
 518 |     "### Using the concatenated dataset\n",
 519 |     "\n",
 520 |     "Let's see whether it makes a difference to use the pre-concatenated dataset."
 521 |    ]
 522 |   },
 523 |   {
 524 |    "cell_type": "code",
 525 |    "execution_count": 74,
 526 |    "metadata": {},
 527 |    "outputs": [
 528 |     {
 529 |      "name": "stdout",
 530 |      "output_type": "stream",
 531 |      "text": [
 532 |       "CPU times: user 5.93 s, sys: 984 ms, total: 6.91 s\n",
 533 |       "Wall time: 6.91 s\n"
 534 |      ]
 535 |     }
 536 |    ],
 537 |    "source": [
 538 |     "ds = xr.open_dataset(f'{DATADIR}sample_SPCAM_concat.nc')\n",
 539 |     "gen = DataGenerator(ds, shuffle=True)\n",
 540 |     "%time test(gen, 10)"
 541 |    ]
 542 |   },
 543 |   {
 544 |    "cell_type": "code",
 545 |    "execution_count": 76,
 546 |    "metadata": {},
 547 |    "outputs": [
 548 |     {
 549 |      "name": "stdout",
 550 |      "output_type": "stream",
 551 |      "text": [
 552 |       "CPU times: user 11.5 s, sys: 1.25 s, total: 12.8 s\n",
 553 |       "Wall time: 12.5 s\n"
 554 |      ]
 555 |     }
 556 |    ],
 557 |    "source": [
 558 |     "ds = xr.open_mfdataset(f'{DATADIR}sample_SPCAM_concat.nc')\n",
 559 |     "gen = DataGenerator(ds, shuffle=True)\n",
 560 |     "%time test(gen, 10)"
 561 |    ]
 562 |   },
 563 |   {
 564 |    "cell_type": "markdown",
 565 |    "metadata": {},
 566 |    "source": [
 567 |     "So yes, it approximately halves the time but only if the single dataset is NOT opened with `open_mfdataset`."
 568 |    ]
 569 |   },
 570 |   {
 571 |    "cell_type": "markdown",
 572 |    "metadata": {},
 573 |    "source": [
 574 |     "### With h5py engine\n",
 575 |     "\n",
 576 |     "Let's see whether using the h5py backend makes a difference"
 577 |    ]
 578 |   },
 579 |   {
 580 |    "cell_type": "code",
 581 |    "execution_count": 77,
 582 |    "metadata": {},
 583 |    "outputs": [],
 584 |    "source": [
 585 |     "import h5netcdf"
 586 |    ]
 587 |   },
 588 |   {
 589 |    "cell_type": "code",
 590 |    "execution_count": 79,
 591 |    "metadata": {},
 592 |    "outputs": [],
 593 |    "source": [
 594 |     "ds.close()"
 595 |    ]
 596 |   },
 597 |   {
 598 |    "cell_type": "code",
 599 |    "execution_count": 80,
 600 |    "metadata": {},
 601 |    "outputs": [],
 602 |    "source": [
 603 |     "ds = xr.open_dataset(f'{DATADIR}sample_SPCAM_concat.nc', engine='h5netcdf')\n",
 604 |     "gen = DataGenerator(ds)"
 605 |    ]
 606 |   },
 607 |   {
 608 |    "cell_type": "code",
 609 |    "execution_count": 81,
 610 |    "metadata": {},
 611 |    "outputs": [
 612 |     {
 613 |      "name": "stdout",
 614 |      "output_type": "stream",
 615 |      "text": [
 616 |       "CPU times: user 6.97 s, sys: 972 ms, total: 7.94 s\n",
 617 |       "Wall time: 7.8 s\n"
 618 |      ]
 619 |     }
 620 |    ],
 621 |    "source": [
 622 |     "%%time\n",
 623 |     "test(gen, 10)"
 624 |    ]
 625 |   },
 626 |   {
 627 |    "cell_type": "markdown",
 628 |    "metadata": {},
 629 |    "source": [
 630 |     "Doesn't seem to speed it up"
 631 |    ]
 632 |   },
 633 |   {
 634 |    "cell_type": "code",
 635 |    "execution_count": 83,
 636 |    "metadata": {},
 637 |    "outputs": [],
 638 |    "source": [
 639 |     "ds.close()"
 640 |    ]
 641 |   },
 642 |   {
 643 |    "cell_type": "markdown",
 644 |    "metadata": {},
 645 |    "source": [
 646 |     "### Using plain h5py\n",
 647 |     "\n",
 648 |     "Let's write a version of the data generator that uses plain h5py for data loading."
 649 |    ]
 650 |   },
 651 |   {
 652 |    "cell_type": "code",
 653 |    "execution_count": 82,
 654 |    "metadata": {},
 655 |    "outputs": [],
 656 |    "source": [
 657 |     "class DataGeneratorH5(object):\n",
 658 |     "    def __init__(self, fn, batch_size=128, input_vars=['TAP', 'QAP'], output_vars=['TPHYSTND', 'PHQ'], shuffle=True):\n",
 659 |     "        self.ds = xr.open_dataset(fn)\n",
 660 |     "        self.batch_size = batch_size\n",
 661 |     "        self.input_vars = input_vars\n",
 662 |     "        self.output_vars = output_vars\n",
 663 |     "        self.ntime, self.nlat, self.nlon = self.ds.time.size, self.ds.lat.size, self.ds.lon.size\n",
 664 |     "        self.ntot = self.ntime * self.nlat * self.ntime\n",
 665 |     "        self.n_batches = self.ntot // batch_size\n",
 666 |     "        self.indices = np.arange(self.ntot)\n",
 667 |     "        if shuffle:\n",
 668 |     "            self.indices = np.random.permutation(self.indices)\n",
 669 |     "            \n",
 670 |     "        # Close xarray dataset and open h5py object\n",
 671 |     "        self.ds.close()\n",
 672 |     "        self.ds = h5py.File(fn, 'r')\n",
 673 |     "        \n",
 674 |     "    def __getitem__(self, index):\n",
 675 |     "        time_indices, lat_indices, lon_indices = np.unravel_index(\n",
 676 |     "            self.indices[index*self.batch_size:(index+1)*self.batch_size], (self.ntime, self.nlat, self.nlon)\n",
 677 |     "        )\n",
 678 |     "        \n",
 679 |     "        X, Y = [], []\n",
 680 |     "        for itime, ilat, ilon in zip(time_indices, lat_indices, lon_indices):\n",
 681 |     "            X.append(\n",
 682 |     "                np.concatenate(\n",
 683 |     "                    [self.ds[v][itime, :, ilat, ilon] for v in self.input_vars]\n",
 684 |     "                )\n",
 685 |     "            )\n",
 686 |     "            Y.append(\n",
 687 |     "                np.concatenate(\n",
 688 |     "                    [self.ds[v][itime, :, ilat, ilon] for v in self.output_vars]\n",
 689 |     "                )\n",
 690 |     "            )\n",
 691 |     "\n",
 692 |     "        return np.array(X), np.array(Y)"
 693 |    ]
 694 |   },
 695 |   {
 696 |    "cell_type": "code",
 697 |    "execution_count": 84,
 698 |    "metadata": {},
 699 |    "outputs": [],
 700 |    "source": [
 701 |     "gen = DataGeneratorH5(f'{DATADIR}sample_SPCAM_concat.nc')"
 702 |    ]
 703 |   },
 704 |   {
 705 |    "cell_type": "code",
 706 |    "execution_count": 85,
 707 |    "metadata": {},
 708 |    "outputs": [
 709 |     {
 710 |      "name": "stdout",
 711 |      "output_type": "stream",
 712 |      "text": [
 713 |       "CPU times: user 1.78 s, sys: 860 ms, total: 2.64 s\n",
 714 |       "Wall time: 2.61 s\n"
 715 |      ]
 716 |     }
 717 |    ],
 718 |    "source": [
 719 |     "%%time\n",
 720 |     "test(gen, 10)"
 721 |    ]
 722 |   },
 723 |   {
 724 |    "cell_type": "code",
 725 |    "execution_count": 96,
 726 |    "metadata": {},
 727 |    "outputs": [],
 728 |    "source": [
 729 |     "gen.ds.close()"
 730 |    ]
 731 |   },
 732 |   {
 733 |    "cell_type": "markdown",
 734 |    "metadata": {},
 735 |    "source": [
 736 |     "So this is significantly faster than xarray."
 737 |    ]
 738 |   },
 739 |   {
 740 |    "cell_type": "markdown",
 741 |    "metadata": {},
 742 |    "source": [
 743 |     "## Use in a simple neural network\n",
 744 |     "\n",
 745 |     "How would we actually use this data generator for network training...\n",
 746 |     "\n",
 747 |     "Note that this neural network will not actually learn much because we didn't normalize the input data. But we only care about computational performance here, right?"
 748 |    ]
 749 |   },
 750 |   {
 751 |    "cell_type": "code",
 752 |    "execution_count": 87,
 753 |    "metadata": {},
 754 |    "outputs": [],
 755 |    "source": [
 756 |     "import tensorflow as tf\n",
 757 |     "from tensorflow.keras.layers import *\n",
 758 |     "from tensorflow.keras.models import Sequential"
 759 |    ]
 760 |   },
 761 |   {
 762 |    "cell_type": "code",
 763 |    "execution_count": 88,
 764 |    "metadata": {},
 765 |    "outputs": [
 766 |     {
 767 |      "data": {
 768 |       "text/plain": [
 769 |        "'2.1.6-tf'"
 770 |       ]
 771 |      },
 772 |      "execution_count": 88,
 773 |      "metadata": {},
 774 |      "output_type": "execute_result"
 775 |     }
 776 |    ],
 777 |    "source": [
 778 |     "tf.keras.__version__"
 779 |    ]
 780 |   },
 781 |   {
 782 |    "cell_type": "code",
 783 |    "execution_count": 89,
 784 |    "metadata": {},
 785 |    "outputs": [],
 786 |    "source": [
 787 |     "model = Sequential([\n",
 788 |     "    Dense(128, input_shape=(60,), activation='relu'),\n",
 789 |     "    Dense(60),\n",
 790 |     "])"
 791 |    ]
 792 |   },
 793 |   {
 794 |    "cell_type": "code",
 795 |    "execution_count": 90,
 796 |    "metadata": {},
 797 |    "outputs": [
 798 |     {
 799 |      "name": "stdout",
 800 |      "output_type": "stream",
 801 |      "text": [
 802 |       "_________________________________________________________________\n",
 803 |       "Layer (type)                 Output Shape              Param #   \n",
 804 |       "=================================================================\n",
 805 |       "dense (Dense)                (None, 128)               7808      \n",
 806 |       "_________________________________________________________________\n",
 807 |       "dense_1 (Dense)              (None, 60)                7740      \n",
 808 |       "=================================================================\n",
 809 |       "Total params: 15,548\n",
 810 |       "Trainable params: 15,548\n",
 811 |       "Non-trainable params: 0\n",
 812 |       "_________________________________________________________________\n"
 813 |      ]
 814 |     }
 815 |    ],
 816 |    "source": [
 817 |     "model.summary()"
 818 |    ]
 819 |   },
 820 |   {
 821 |    "cell_type": "code",
 822 |    "execution_count": 91,
 823 |    "metadata": {},
 824 |    "outputs": [],
 825 |    "source": [
 826 |     "model.compile('adam', 'mse')"
 827 |    ]
 828 |   },
 829 |   {
 830 |    "cell_type": "code",
 831 |    "execution_count": 99,
 832 |    "metadata": {},
 833 |    "outputs": [],
 834 |    "source": [
 835 |     "# Load the xarray version using the concatenated dataset\n",
 836 |     "ds = xr.open_dataset(f'{DATADIR}sample_SPCAM_concat.nc')\n",
 837 |     "gen = DataGenerator(ds, shuffle=True)"
 838 |    ]
 839 |   },
 840 |   {
 841 |    "cell_type": "code",
 842 |    "execution_count": 101,
 843 |    "metadata": {},
 844 |    "outputs": [
 845 |     {
 846 |      "name": "stdout",
 847 |      "output_type": "stream",
 848 |      "text": [
 849 |       "Epoch 1/1\n",
 850 |       "  37/4608 [..............................] - ETA: 1:04:11 - loss: 1733.6299"
 851 |      ]
 852 |     },
 853 |     {
 854 |      "ename": "KeyboardInterrupt",
 855 |      "evalue": "",
 856 |      "output_type": "error",
 857 |      "traceback": [
 858 |       "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
 859 |       "\u001b[0;31mKeyboardInterrupt\u001b[0m                         Traceback (most recent call last)",
 860 |       "\u001b[0;32m<ipython-input-101-401f100b439b>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[0;32m----> 1\u001b[0;31m \u001b[0mmodel\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mfit_generator\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0miter\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mgen\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0msteps_per_epoch\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mgen\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mn_batches\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
 861 |       "\u001b[0;32m~/miniconda3/envs/tf/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py\u001b[0m in \u001b[0;36mfit_generator\u001b[0;34m(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\u001b[0m\n\u001b[1;32m   2175\u001b[0m         \u001b[0muse_multiprocessing\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0muse_multiprocessing\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   2176\u001b[0m         \u001b[0mshuffle\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mshuffle\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 2177\u001b[0;31m         initial_epoch=initial_epoch)\n\u001b[0m\u001b[1;32m   2178\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   2179\u001b[0m   def evaluate_generator(self,\n",
 862 |       "\u001b[0;32m~/miniconda3/envs/tf/lib/python3.6/site-packages/tensorflow/python/keras/engine/training_generator.py\u001b[0m in \u001b[0;36mfit_generator\u001b[0;34m(model, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)\u001b[0m\n\u001b[1;32m    145\u001b[0m       \u001b[0mbatch_index\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;36m0\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    146\u001b[0m       \u001b[0;32mwhile\u001b[0m \u001b[0msteps_done\u001b[0m \u001b[0;34m<\u001b[0m \u001b[0msteps_per_epoch\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 147\u001b[0;31m         \u001b[0mgenerator_output\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mnext\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0moutput_generator\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    148\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    149\u001b[0m         \u001b[0;32mif\u001b[0m \u001b[0;32mnot\u001b[0m \u001b[0mhasattr\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mgenerator_output\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'__len__'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
 863 |       "\u001b[0;32m~/miniconda3/envs/tf/lib/python3.6/site-packages/tensorflow/python/keras/utils/data_utils.py\u001b[0m in \u001b[0;36mget\u001b[0;34m(self)\u001b[0m\n\u001b[1;32m    823\u001b[0m           \u001b[0;32mraise\u001b[0m \u001b[0mStopIteration\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    824\u001b[0m         \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 825\u001b[0;31m           \u001b[0mtime\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msleep\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mwait_time\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    826\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    827\u001b[0m     \u001b[0;31m# Make sure to rethrow the first exception in the queue, if any\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
 864 |       "\u001b[0;31mKeyboardInterrupt\u001b[0m: "
 865 |      ]
 866 |     }
 867 |    ],
 868 |    "source": [
 869 |     "model.fit_generator(iter(gen), steps_per_epoch=gen.n_batches)"
 870 |    ]
 871 |   },
 872 |   {
 873 |    "cell_type": "markdown",
 874 |    "metadata": {},
 875 |    "source": [
 876 |     "So as you can see, it would take around 1 hour to go through one epoch (i.e. the entire dataset once). This is crazy slow since we only used 2 days of data. The full dataset contains a year of data..."
 877 |    ]
 878 |   },
 879 |   {
 880 |    "cell_type": "markdown",
 881 |    "metadata": {},
 882 |    "source": [
 883 |     "## Pre-processing the dataset\n",
 884 |     "\n",
 885 |     "What I have resorted to to solve this issue is to prestack the data, preshuffle the data and save it to disc conveniently.\n",
 886 |     "\n",
 887 |     "These files contain the exactly same information for the input (features) and output (targets) variables required.\n",
 888 |     "\n",
 889 |     "The files only have two dimensions: sample, which is the shuffled, flattened time, lat and lon dimensions and lev which is the stacked vertical coordinate.\n",
 890 |     "\n",
 891 |     "The preprocessing for these two files only takes a few seconds but for an entire year of data, the preprocessing alone can take around an hour.\n"
 892 |    ]
 893 |   },
 894 |   {
 895 |    "cell_type": "code",
 896 |    "execution_count": 163,
 897 |    "metadata": {},
 898 |    "outputs": [
 899 |     {
 900 |      "name": "stdout",
 901 |      "output_type": "stream",
 902 |      "text": [
 903 |       "--2019-02-07 15:42:32--  https://zenodo.org/record/2559313/files/preproc_features.nc\n",
 904 |       "Resolving zenodo.org (zenodo.org)... 137.138.76.77\n",
 905 |       "Connecting to zenodo.org (zenodo.org)|137.138.76.77|:443... connected.\n",
 906 |       "HTTP request sent, awaiting response... 200 OK\n",
 907 |       "Length: 205465847 (196M) [application/octet-stream]\n",
 908 |       "Saving to: ‘/local/S.Rasp/tmp/preproc_features.nc.2’\n",
 909 |       "\n",
 910 |       "preproc_features.nc 100%[===================>] 195.95M  10.8MB/s    in 15s     \n",
 911 |       "\n",
 912 |       "2019-02-07 15:42:48 (13.0 MB/s) - ‘/local/S.Rasp/tmp/preproc_features.nc.2’ saved [205465847/205465847]\n",
 913 |       "\n",
 914 |       "--2019-02-07 15:42:48--  https://zenodo.org/record/2559313/files/preproc_targets.nc\n",
 915 |       "Resolving zenodo.org (zenodo.org)... 137.138.76.77\n",
 916 |       "Connecting to zenodo.org (zenodo.org)|137.138.76.77|:443... connected.\n",
 917 |       "HTTP request sent, awaiting response... 200 OK\n",
 918 |       "Length: 205465846 (196M) [application/octet-stream]\n",
 919 |       "Saving to: ‘/local/S.Rasp/tmp/preproc_targets.nc.1’\n",
 920 |       "\n",
 921 |       "preproc_targets.nc. 100%[===================>] 195.95M  9.98MB/s    in 9.5s    \n",
 922 |       "\n",
 923 |       "2019-02-07 15:42:58 (20.6 MB/s) - ‘/local/S.Rasp/tmp/preproc_targets.nc.1’ saved [205465846/205465846]\n",
 924 |       "\n"
 925 |      ]
 926 |     }
 927 |    ],
 928 |    "source": [
 929 |     "!wget -P $DATADIR https://zenodo.org/record/2559313/files/preproc_features.nc\n",
 930 |     "!wget -P $DATADIR https://zenodo.org/record/2559313/files/preproc_targets.nc"
 931 |    ]
 932 |   },
 933 |   {
 934 |    "cell_type": "code",
 935 |    "execution_count": 104,
 936 |    "metadata": {},
 937 |    "outputs": [
 938 |     {
 939 |      "name": "stdout",
 940 |      "output_type": "stream",
 941 |      "text": [
 942 |       "-rw-r--r-- 1 S.Rasp ls-craig 196M Feb  7 13:57 /local/S.Rasp/tmp//preproc_features.nc\r\n",
 943 |       "-rw-r--r-- 1 S.Rasp ls-craig 196M Feb  7 13:57 /local/S.Rasp/tmp//preproc_targets.nc\r\n"
 944 |      ]
 945 |     }
 946 |    ],
 947 |    "source": [
 948 |     "!ls -lh $DATADIR/preproc*"
 949 |    ]
 950 |   },
 951 |   {
 952 |    "cell_type": "code",
 953 |    "execution_count": 105,
 954 |    "metadata": {},
 955 |    "outputs": [],
 956 |    "source": [
 957 |     "ds = xr.open_dataset(f'{DATADIR}preproc_features.nc')"
 958 |    ]
 959 |   },
 960 |   {
 961 |    "cell_type": "code",
 962 |    "execution_count": 106,
 963 |    "metadata": {},
 964 |    "outputs": [
 965 |     {
 966 |      "data": {
 967 |       "text/plain": [
 968 |        "<xarray.Dataset>\n",
 969 |        "Dimensions:        (feature_lev: 60, sample: 778240)\n",
 970 |        "Coordinates:\n",
 971 |        "  * feature_lev    (feature_lev) int64 0 1 2 3 4 5 6 7 ... 53 54 55 56 57 58 59\n",
 972 |        "    time           (sample) int64 ...\n",
 973 |        "    lat            (sample) float64 ...\n",
 974 |        "    lon            (sample) float64 ...\n",
 975 |        "    feature_names  (feature_lev) object ...\n",
 976 |        "Dimensions without coordinates: sample\n",
 977 |        "Data variables:\n",
 978 |        "    features       (sample, feature_lev) float32 ...\n",
 979 |        "Attributes:\n",
 980 |        "    log:      \\n    Time: 2019-02-07T13:57:24\\n\\n    Executed command:\\n\\n   ..."
 981 |       ]
 982 |      },
 983 |      "execution_count": 106,
 984 |      "metadata": {},
 985 |      "output_type": "execute_result"
 986 |     }
 987 |    ],
 988 |    "source": [
 989 |     "ds"
 990 |    ]
 991 |   },
 992 |   {
 993 |    "cell_type": "code",
 994 |    "execution_count": 129,
 995 |    "metadata": {},
 996 |    "outputs": [],
 997 |    "source": [
 998 |     "# Write a new data generator\n",
 999 |     "class DataGeneratorPreproc(object):\n",
1000 |     "    \"\"\"\n",
1001 |     "    Data generator that randomly (if shuffle = True) picks columns from the dataset and returns them in \n",
1002 |     "    batches. For each column the input variables and output variables will be stacked.\n",
1003 |     "    \"\"\"\n",
1004 |     "    def __init__(self, feature_fn, target_fn, batch_size=128, shuffle=True, engine='netcdf4'):\n",
1005 |     "        self.feature_ds = xr.open_dataset(feature_fn, engine=engine)\n",
1006 |     "        self.target_ds = xr.open_dataset(target_fn, engine=engine)\n",
1007 |     "        self.batch_size = batch_size\n",
1008 |     "        self.ntot = self.feature_ds.sample.size\n",
1009 |     "        self.n_batches = self.ntot // batch_size\n",
1010 |     "        self.indices = np.arange(self.ntot)\n",
1011 |     "        if shuffle:\n",
1012 |     "            self.indices = np.random.permutation(self.indices)\n",
1013 |     "    def __getitem__(self, index):\n",
1014 |     "        batch_indices = self.indices[index*self.batch_size:(index+1)*self.batch_size]\n",
1015 |     "        \n",
1016 |     "        X = self.feature_ds.features.isel(sample=batch_indices)\n",
1017 |     "        Y = self.target_ds.targets.isel(sample=batch_indices)\n",
1018 |     "\n",
1019 |     "        return X, Y"
1020 |    ]
1021 |   },
1022 |   {
1023 |    "cell_type": "code",
1024 |    "execution_count": 130,
1025 |    "metadata": {},
1026 |    "outputs": [],
1027 |    "source": [
1028 |     "gen = DataGeneratorPreproc(f'{DATADIR}preproc_features.nc', f'{DATADIR}preproc_targets.nc')"
1029 |    ]
1030 |   },
1031 |   {
1032 |    "cell_type": "code",
1033 |    "execution_count": 131,
1034 |    "metadata": {},
1035 |    "outputs": [],
1036 |    "source": [
1037 |     "x, y = gen[0]"
1038 |    ]
1039 |   },
1040 |   {
1041 |    "cell_type": "code",
1042 |    "execution_count": 132,
1043 |    "metadata": {},
1044 |    "outputs": [
1045 |     {
1046 |      "data": {
1047 |       "text/plain": [
1048 |        "((128, 60), (128, 60))"
1049 |       ]
1050 |      },
1051 |      "execution_count": 132,
1052 |      "metadata": {},
1053 |      "output_type": "execute_result"
1054 |     }
1055 |    ],
1056 |    "source": [
1057 |     "x.shape, y.shape"
1058 |    ]
1059 |   },
1060 |   {
1061 |    "cell_type": "code",
1062 |    "execution_count": 133,
1063 |    "metadata": {},
1064 |    "outputs": [
1065 |     {
1066 |      "name": "stdout",
1067 |      "output_type": "stream",
1068 |      "text": [
1069 |       "CPU times: user 84 ms, sys: 0 ns, total: 84 ms\n",
1070 |       "Wall time: 81.6 ms\n"
1071 |      ]
1072 |     }
1073 |    ],
1074 |    "source": [
1075 |     "%%time\n",
1076 |     "test(gen, 10)"
1077 |    ]
1078 |   },
1079 |   {
1080 |    "cell_type": "code",
1081 |    "execution_count": 134,
1082 |    "metadata": {},
1083 |    "outputs": [],
1084 |    "source": [
1085 |     "gen = DataGeneratorPreproc(f'{DATADIR}preproc_features.nc', f'{DATADIR}preproc_targets.nc', shuffle=False)"
1086 |    ]
1087 |   },
1088 |   {
1089 |    "cell_type": "code",
1090 |    "execution_count": 135,
1091 |    "metadata": {},
1092 |    "outputs": [
1093 |     {
1094 |      "name": "stdout",
1095 |      "output_type": "stream",
1096 |      "text": [
1097 |       "CPU times: user 84 ms, sys: 0 ns, total: 84 ms\n",
1098 |       "Wall time: 83.9 ms\n"
1099 |      ]
1100 |     }
1101 |    ],
1102 |    "source": [
1103 |     "%%time\n",
1104 |     "test(gen, 10)"
1105 |    ]
1106 |   },
1107 |   {
1108 |    "cell_type": "code",
1109 |    "execution_count": 152,
1110 |    "metadata": {},
1111 |    "outputs": [],
1112 |    "source": [
1113 |     "gen.feature_ds.close(); gen.target_ds.close()"
1114 |    ]
1115 |   },
1116 |   {
1117 |    "cell_type": "code",
1118 |    "execution_count": 139,
1119 |    "metadata": {},
1120 |    "outputs": [],
1121 |    "source": [
1122 |     "gen = DataGeneratorPreproc(f'{DATADIR}preproc_features.nc', f'{DATADIR}preproc_targets.nc', engine='h5netcdf')"
1123 |    ]
1124 |   },
1125 |   {
1126 |    "cell_type": "code",
1127 |    "execution_count": 140,
1128 |    "metadata": {},
1129 |    "outputs": [
1130 |     {
1131 |      "name": "stdout",
1132 |      "output_type": "stream",
1133 |      "text": [
1134 |       "CPU times: user 84 ms, sys: 0 ns, total: 84 ms\n",
1135 |       "Wall time: 80.6 ms\n"
1136 |      ]
1137 |     }
1138 |    ],
1139 |    "source": [
1140 |     "%%time\n",
1141 |     "test(gen, 10)"
1142 |    ]
1143 |   },
1144 |   {
1145 |    "cell_type": "markdown",
1146 |    "metadata": {},
1147 |    "source": [
1148 |     "So these are the sort of times that are required for training a neural network."
1149 |    ]
1150 |   },
1151 |   {
1152 |    "cell_type": "markdown",
1153 |    "metadata": {},
1154 |    "source": [
1155 |     "### Pure h5py version"
1156 |    ]
1157 |   },
1158 |   {
1159 |    "cell_type": "code",
1160 |    "execution_count": 158,
1161 |    "metadata": {},
1162 |    "outputs": [],
1163 |    "source": [
1164 |     "class DataGeneratorPreprocH5(object):\n",
1165 |     "    \"\"\"\n",
1166 |     "    Data generator that randomly (if shuffle = True) picks columns from the dataset and returns them in \n",
1167 |     "    batches. For each column the input variables and output variables will be stacked.\n",
1168 |     "    \"\"\"\n",
1169 |     "    def __init__(self, feature_fn, target_fn, batch_size=128):\n",
1170 |     "        self.feature_ds = xr.open_dataset(feature_fn)\n",
1171 |     "        self.target_ds = xr.open_dataset(target_fn)\n",
1172 |     "        self.batch_size = batch_size\n",
1173 |     "        self.ntot = self.feature_ds.sample.size\n",
1174 |     "        self.n_batches = self.ntot // batch_size\n",
1175 |     "        \n",
1176 |     "        # Close xarray dataset and open h5py object\n",
1177 |     "        self.feature_ds.close()\n",
1178 |     "        self.feature_ds = h5py.File(feature_fn, 'r')\n",
1179 |     "        self.target_ds.close()\n",
1180 |     "        self.target_ds = h5py.File(target_fn, 'r')\n",
1181 |     "        \n",
1182 |     "    def __getitem__(self, index):\n",
1183 |     "        \n",
1184 |     "        X = self.feature_ds['features'][index*self.batch_size:(index+1)*self.batch_size, :]\n",
1185 |     "        Y = self.target_ds['targets'][index*self.batch_size:(index+1)*self.batch_size, :]\n",
1186 |     "\n",
1187 |     "        return X, Y"
1188 |    ]
1189 |   },
1190 |   {
1191 |    "cell_type": "code",
1192 |    "execution_count": 159,
1193 |    "metadata": {},
1194 |    "outputs": [],
1195 |    "source": [
1196 |     "gen.feature_ds.close(); gen.target_ds.close()"
1197 |    ]
1198 |   },
1199 |   {
1200 |    "cell_type": "code",
1201 |    "execution_count": 160,
1202 |    "metadata": {},
1203 |    "outputs": [],
1204 |    "source": [
1205 |     "gen = DataGeneratorPreprocH5(f'{DATADIR}preproc_features.nc', f'{DATADIR}preproc_targets.nc')"
1206 |    ]
1207 |   },
1208 |   {
1209 |    "cell_type": "code",
1210 |    "execution_count": 161,
1211 |    "metadata": {},
1212 |    "outputs": [
1213 |     {
1214 |      "name": "stdout",
1215 |      "output_type": "stream",
1216 |      "text": [
1217 |       "CPU times: user 8 ms, sys: 0 ns, total: 8 ms\n",
1218 |       "Wall time: 6.61 ms\n"
1219 |      ]
1220 |     }
1221 |    ],
1222 |    "source": [
1223 |     "%%time\n",
1224 |     "test(gen, 10)"
1225 |    ]
1226 |   },
1227 |   {
1228 |    "cell_type": "markdown",
1229 |    "metadata": {},
1230 |    "source": [
1231 |     "So again, the pure h5py version is an order of magnitude faster than the xarray version."
1232 |    ]
1233 |   },
1234 |   {
1235 |    "cell_type": "markdown",
1236 |    "metadata": {},
1237 |    "source": [
1238 |     "## End"
1239 |    ]
1240 |   },
1241 |   {
1242 |    "cell_type": "code",
1243 |    "execution_count": null,
1244 |    "metadata": {},
1245 |    "outputs": [],
1246 |    "source": []
1247 |   }
1248 |  ],
1249 |  "metadata": {
1250 |   "kernelspec": {
1251 |    "display_name": "Python 3",
1252 |    "language": "python",
1253 |    "name": "python3"
1254 |   },
1255 |   "language_info": {
1256 |    "codemirror_mode": {
1257 |     "name": "ipython",
1258 |     "version": 3
1259 |    },
1260 |    "file_extension": ".py",
1261 |    "mimetype": "text/x-python",
1262 |    "name": "python",
1263 |    "nbconvert_exporter": "python",
1264 |    "pygments_lexer": "ipython3",
1265 |    "version": "3.6.5"
1266 |   },
1267 |   "toc": {
1268 |    "base_numbering": 1,
1269 |    "nav_menu": {},
1270 |    "number_sections": true,
1271 |    "sideBar": false,
1272 |    "skip_h1_title": true,
1273 |    "title_cell": "Table of Contents",
1274 |    "title_sidebar": "Contents",
1275 |    "toc_cell": false,
1276 |    "toc_position": {},
1277 |    "toc_section_display": true,
1278 |    "toc_window_display": false
1279 |   }
1280 |  },
1281 |  "nbformat": 4,
1282 |  "nbformat_minor": 2
1283 | }
1284 | 


--------------------------------------------------------------------------------