{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "%matplotlib inline" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "\n# 7. HPO Concat\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "import site\nsite.addsitedir(r\"E:\\AA\\AI4Water\")\nsite.addsitedir(r\"E:\\AA\\easy_mpl\")\n\nfrom ai4water import Model\n\nimport os\nSEP = os.sep\nimport numpy as np\nimport pandas as pd\nimport math\nfrom typing import Union\n\nfrom skopt.plots import plot_objective\nimport matplotlib.pyplot as plt\nplt.rcParams[\"font.family\"] = \"Times New Roman\"\n\nfrom SeqMetrics import RegressionMetrics\n\nfrom ai4water.utils.utils import jsonize\nfrom ai4water.utils.utils import dateandtime_now\nfrom ai4water.hyperopt import Categorical, Real, Integer, HyperOpt\n\nfrom utils import prepare_data, eval_model" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "PREFIX = f\"hpo_lstm_concat_{dateandtime_now()}\"\nITER = 0\nnum_iterations = 200\nlookback = 12\ntarget = 'TP'" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "MONITOR = {\"rmse\": [], \"nse\": [], \"r2\": [], \"mape\": []}" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "def objective_fn(\n prefix: str = None,\n return_model: bool = False,\n epochs:int = 500,\n verbosity: int = 0,\n predict : bool = False,\n target = target,\n **suggestions\n )->Union[float, Model]:\n\n suggestions = jsonize(suggestions)\n global ITER\n\n tr_x, tr_y, val_x, val_y, test_x, test_y = prepare_data(\n target=target, lookback=lookback, batch_size=suggestions[\"batch_size\"])\n\n x_mean = tr_x[0].mean()\n x_std = tr_x[0].std()\n x_mean_st = tr_x[1].mean()\n x_std_st = tr_x[1].std()\n\n y_mean = tr_y.mean()\n y_std = tr_y.std()\n\n tr_x[0] = (tr_x[0] - x_mean) / x_std\n tr_x[1] = (tr_x[1] - x_mean_st) / x_std_st\n tr_y = (tr_y - y_mean) / y_std\n #tr_y = transformer.fit_transform(tr_y)\n\n val_x[0] = (val_x[0] - x_mean) / x_std\n val_x[1] = (val_x[1] - x_mean_st) / x_std_st\n val_y = (val_y - y_mean) / y_std\n #val_y = transformer.transform(val_y)\n\n test_x[0] = (test_x[0] - x_mean) / x_std\n test_x[1] = (test_x[1] - x_mean_st) / x_std_st\n test_y = (test_y - y_mean) / y_std\n #test_y = transformer.transform(test_y)\n\n num_dyn_inputs = tr_x[0].shape[-1]\n num_static_inputs = tr_x[1].shape[-1]\n\n layers = {\n \"Input_dyn\": {\"batch_shape\": (suggestions[\"batch_size\"], lookback, num_dyn_inputs)},\n \"Input_cat\": {\"batch_shape\": (suggestions[\"batch_size\"], num_static_inputs)},\n \"LSTM\": {\"config\": {\"units\": suggestions[\"lstm_units\"]}, \"inputs\": \"Input_dyn\"},\n \"Dense_cat\": {\"config\": {\"units\": suggestions[\"dense_units\"]}, \"inputs\": \"Input_cat\"},\n \"Concatenate\": {\"config\": {}, \"inputs\": [\"LSTM\", \"Dense_cat\"]},\n \"Dense\": 1\n }\n\n # build model\n _model = Model(model = {\"layers\": layers},\n batch_size=suggestions[\"batch_size\"],\n epochs=epochs,\n lr=suggestions[\"lr\"],\n prefix=prefix or PREFIX,\n verbosity=verbosity)\n\n # train model\n _model.fit(x=tr_x, y=tr_y, validation_data=(val_x, val_y))\n\n # evaluate model\n t, p = _model.predict(x=val_x, y=val_y, return_true=True,\n process_results=False)\n metrics = RegressionMetrics(t, p)\n val_score = metrics.rmse()\n\n for metric in MONITOR.keys():\n val = getattr(metrics, metric)()\n MONITOR[metric].append(val)\n\n # here we are evaluating model with respect to mse, therefore\n # we don't need to subtract it from 1.0\n if not math.isfinite(val_score):\n val_score = 9999\n\n print(f\"{ITER} {val_score}\")\n\n ITER += 1\n\n if return_model:\n return _model\n\n return val_score" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "param_space = [\n Integer(30, 100, name=\"lstm_units\"),\n Integer(30, 100, name=\"dense_units\"),\n Real(0.00001, 0.01, name=\"lr\"),\n Categorical([\"relu\", \"elu\", \"tanh\", \"sigmoid\"], name=\"activation\"),\n Categorical([8, 12, 16, 24, 32, 48, 64, 128], name=\"batch_size\")\n ]" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "x0 = [30, 30, 0.001, 'relu', 8]" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "optimizer = HyperOpt(\n algorithm=\"bayes\",\n objective_fn=objective_fn,\n param_space=param_space,\n x0=x0,\n num_iterations=num_iterations,\n process_results=False, # we can turn it False if we want post-processing of results\n opt_path=f\"results{SEP}{PREFIX}\"\n)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "#\n# results = optimizer.fit()\n#\n# best_iteration = optimizer.best_iter()\n#\n# print(f\"optimized parameters are \\n{optimizer.best_paras()} at {best_iteration}\")\n#\n# for key in ['rmse', 'mape']:\n# print(key, np.nanmin(MONITOR[key]), np.nanargmin(MONITOR[key]))\n#\n# for key in ['r2', 'nse']:\n# print(key, np.nanmax(MONITOR[key]), np.nanargmax(MONITOR[key]))\n#\n# model = objective_fn(prefix=f\"{PREFIX}{SEP}best\",\n# return_model=True,\n# epochs=500,\n# verbosity=1,\n# predict=True,\n# **optimizer.best_paras())\n#\n# tr_x, tr_y, val_x, val_y, test_x, test_y = prepare_data(\n# target=target, lookback=lookback,\n# batch_size=optimizer.best_paras()['batch_size']\n# )\n#\n# x_mean = tr_x[0].mean()\n# x_std = tr_x[0].std()\n# x_mean_st = tr_x[1].mean()\n# x_std_st = tr_x[1].std()\n#\n# y_mean = tr_y.mean()\n# y_std = tr_y.std()\n#\n# tr_x[0] = (tr_x[0] - x_mean) / x_std\n# tr_x[1] = (tr_x[1] - x_mean_st) / x_std_st\n# tr_y = (tr_y - y_mean) / y_std\n# # tr_y = transformer.fit_transform(tr_y)\n#\n# val_x[0] = (val_x[0] - x_mean) / x_std\n# val_x[1] = (val_x[1] - x_mean_st) / x_std_st\n# val_y = (val_y - y_mean) / y_std\n# # val_y = transformer.transform(val_y)\n#\n# test_x[0] = (test_x[0] - x_mean) / x_std\n# test_x[1] = (test_x[1] - x_mean_st) / x_std_st\n# test_y = (test_y - y_mean) / y_std\n# # test_y = transformer.transform(test_y)\n#\n# model.evaluate(x=tr_x, y=tr_y, metrics=['r2', 'nse', 'rmse'])\n# model.evaluate(x=val_x, y=val_y, metrics=['r2', 'nse', 'rmse'])\n# model.evaluate(x=test_x, y=test_y, metrics=['r2', 'nse', 'rmse'])\n#\n# # tr_t, tr_p = model.predict(x=tr_x, y=tr_y, return_true=True)\n# #\n# # val_t, val_p = model.predict(x=val_x, y=val_y, return_true=True)\n#\n# t, p = model.predict(x=test_x, y=test_y, return_true=True)\n#\n# print(f'r2 = {RegressionMetrics(t, p).r2()}')\n# print(f'nse = {RegressionMetrics(t, p).nse()}')\n# print(f'rmse = {RegressionMetrics(t, p).rmse()}')\n#\n# optimizer._plot_convergence(save=True)\n#\n# optimizer.plot_importance(save=True)\n# plt.tight_layout()\n# plt.show()\n#\n# _ = plot_objective(results)\n# plt.show()\n#\n# optimizer.save_iterations_as_xy()\n#\n# optimizer._plot_parallel_coords(figsize=(14, 8), save=True)\n#\n# monitor = pd.DataFrame.from_dict(MONITOR)\n# a = os.path.split(model.path)\n# a = os.path.split(a[0])\n#\n# monitor.to_csv (r'monitor.csv', index = True, header=True)" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.7.9" } }, "nbformat": 4, "nbformat_minor": 0 }