{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "## General Information" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "This notebook is used to train a simple neural network model to predict the chemistry in the barite benchmark (50x50 grid). The training data is stored in the repository using **git large file storage** and can be downloaded after the installation of git lfs using the `git lfs pull` command.\n", "\n", "It is then recommended to create a Python environment using miniconda. The necessary dependencies are contained in `environment.yml` and can be installed using `conda env create -f environment.yml`.\n", "\n", "The data set is divided into a design and result part and consists of the iterations of a reference simulation. The design part of the data set contains the chemical concentrations at time $t$ and the result part at time $t+1$, which are to be learned by the model." ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Setup Libraries" ] }, { "cell_type": "code", "execution_count": 197, "metadata": {}, "outputs": [], "source": [ "import keras\n", "import tensorflow as tf\n", "import h5py\n", "import numpy as np\n", "import pandas as pd\n", "import time\n", "import sklearn.model_selection as sk\n", "import matplotlib.pyplot as plt\n", "from sklearn.cluster import KMeans\n", "from sklearn.pipeline import Pipeline, make_pipeline\n", "from sklearn.preprocessing import StandardScaler, MinMaxScaler\n", "from imblearn.over_sampling import SMOTE\n", "from imblearn.under_sampling import RandomUnderSampler\n", "from imblearn.over_sampling import RandomOverSampler\n", "from collections import Counter\n", "import os\n", "from preprocessing import *\n", "from sklearn import set_config\n", "from importlib import reload\n", "set_config(transform_output = \"pandas\")" ] }, { "cell_type": "code", "execution_count": 136, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "The autoreload extension is already loaded. To reload it, use:\n", " %reload_ext autoreload\n" ] } ], "source": [ "%load_ext autoreload\n", "%autoreload 2" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Define parameters" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "dtype = \"float32\"\n", "activation = \"relu\"\n", "\n", "lr = 0.001\n", "batch_size = 512\n", "epochs = 50 # default 400 epochs\n", "\n", "lr_schedule = keras.optimizers.schedules.ExponentialDecay(\n", " initial_learning_rate=lr,\n", " decay_steps=2000,\n", " decay_rate=0.9,\n", " staircase=True\n", ")\n", "\n", "optimizer_simple = keras.optimizers.Adam(learning_rate=lr_schedule)\n", "optimizer_large = keras.optimizers.Adam(learning_rate=lr_schedule)\n", "optimizer_paper = keras.optimizers.Adam(learning_rate=lr_schedule)\n", "\n", "\n", "loss = keras.losses.Huber()\n", "\n", "sample_fraction = 0.8" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Setup the model" ] }, { "cell_type": "code", "execution_count": 150, "metadata": {}, "outputs": [ { "data": { "text/html": [ "
Model: \"sequential_7\"\n",
       "
\n" ], "text/plain": [ "\u001b[1mModel: \"sequential_7\"\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n",
       "┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n",
       "┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n",
       "│ dense_27 (Dense)                │ (None, 128)            │         1,280 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_28 (Dense)                │ (None, 128)            │        16,512 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_29 (Dense)                │ (None, 9)              │         1,161 │\n",
       "└─────────────────────────────────┴────────────────────────┴───────────────┘\n",
       "
\n" ], "text/plain": [ "┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n", "┃\u001b[1m \u001b[0m\u001b[1mLayer (type) \u001b[0m\u001b[1m \u001b[0m┃\u001b[1m \u001b[0m\u001b[1mOutput Shape \u001b[0m\u001b[1m \u001b[0m┃\u001b[1m \u001b[0m\u001b[1m Param #\u001b[0m\u001b[1m \u001b[0m┃\n", "┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n", "│ dense_27 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m128\u001b[0m) │ \u001b[38;5;34m1,280\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_28 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m128\u001b[0m) │ \u001b[38;5;34m16,512\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_29 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m9\u001b[0m) │ \u001b[38;5;34m1,161\u001b[0m │\n", "└─────────────────────────────────┴────────────────────────┴───────────────┘\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Total params: 18,953 (74.04 KB)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Total params: \u001b[0m\u001b[38;5;34m18,953\u001b[0m (74.04 KB)\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Trainable params: 18,953 (74.04 KB)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Trainable params: \u001b[0m\u001b[38;5;34m18,953\u001b[0m (74.04 KB)\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Non-trainable params: 0 (0.00 B)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Non-trainable params: \u001b[0m\u001b[38;5;34m0\u001b[0m (0.00 B)\n" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# small model\n", "model_simple = keras.Sequential(\n", " [\n", " keras.Input(shape = (9,), dtype = \"float32\"),\n", " keras.layers.Dense(units = 128, activation = \"relu\", dtype = \"float32\"),\n", " keras.layers.Dense(units = 128, activation = \"relu\", dtype = \"float32\"),\n", " keras.layers.Dense(units = 9, dtype = \"float32\")\n", " ]\n", ")\n", "\n", "model_simple.compile(optimizer=optimizer_simple, loss = loss)\n", "model_simple.summary()" ] }, { "cell_type": "code", "execution_count": 161, "metadata": {}, "outputs": [ { "data": { "text/html": [ "
Model: \"sequential_8\"\n",
       "
\n" ], "text/plain": [ "\u001b[1mModel: \"sequential_8\"\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n",
       "┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n",
       "┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n",
       "│ dense_30 (Dense)                │ (None, 512)            │         5,120 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_31 (Dense)                │ (None, 1024)           │       525,312 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_32 (Dense)                │ (None, 512)            │       524,800 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_33 (Dense)                │ (None, 9)              │         4,617 │\n",
       "└─────────────────────────────────┴────────────────────────┴───────────────┘\n",
       "
\n" ], "text/plain": [ "┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n", "┃\u001b[1m \u001b[0m\u001b[1mLayer (type) \u001b[0m\u001b[1m \u001b[0m┃\u001b[1m \u001b[0m\u001b[1mOutput Shape \u001b[0m\u001b[1m \u001b[0m┃\u001b[1m \u001b[0m\u001b[1m Param #\u001b[0m\u001b[1m \u001b[0m┃\n", "┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n", "│ dense_30 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m512\u001b[0m) │ \u001b[38;5;34m5,120\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_31 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m1024\u001b[0m) │ \u001b[38;5;34m525,312\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_32 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m512\u001b[0m) │ \u001b[38;5;34m524,800\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_33 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m9\u001b[0m) │ \u001b[38;5;34m4,617\u001b[0m │\n", "└─────────────────────────────────┴────────────────────────┴───────────────┘\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Total params: 1,059,849 (4.04 MB)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Total params: \u001b[0m\u001b[38;5;34m1,059,849\u001b[0m (4.04 MB)\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Trainable params: 1,059,849 (4.04 MB)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Trainable params: \u001b[0m\u001b[38;5;34m1,059,849\u001b[0m (4.04 MB)\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Non-trainable params: 0 (0.00 B)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Non-trainable params: \u001b[0m\u001b[38;5;34m0\u001b[0m (0.00 B)\n" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# large model\n", "model_large = keras.Sequential(\n", " [keras.layers.Input(shape=(9,), dtype=dtype),\n", " keras.layers.Dense(512, activation='relu', dtype=dtype),\n", " keras.layers.Dense(1024, activation='relu', dtype=dtype),\n", " keras.layers.Dense(512, activation='relu', dtype=dtype),\n", " keras.layers.Dense(9, dtype=dtype)\n", " ])\n", "\n", "model_large.compile(optimizer=optimizer_large, loss = loss)\n", "model_large.summary()\n" ] }, { "cell_type": "code", "execution_count": 140, "metadata": {}, "outputs": [ { "data": { "text/html": [ "
Model: \"sequential_5\"\n",
       "
\n" ], "text/plain": [ "\u001b[1mModel: \"sequential_5\"\u001b[0m\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n",
       "┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n",
       "┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n",
       "│ dense_19 (Dense)                │ (None, 128)            │         1,664 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_20 (Dense)                │ (None, 256)            │        33,024 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_21 (Dense)                │ (None, 512)            │       131,584 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_22 (Dense)                │ (None, 256)            │       131,328 │\n",
       "├─────────────────────────────────┼────────────────────────┼───────────────┤\n",
       "│ dense_23 (Dense)                │ (None, 12)             │         3,084 │\n",
       "└─────────────────────────────────┴────────────────────────┴───────────────┘\n",
       "
\n" ], "text/plain": [ "┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n", "┃\u001b[1m \u001b[0m\u001b[1mLayer (type) \u001b[0m\u001b[1m \u001b[0m┃\u001b[1m \u001b[0m\u001b[1mOutput Shape \u001b[0m\u001b[1m \u001b[0m┃\u001b[1m \u001b[0m\u001b[1m Param #\u001b[0m\u001b[1m \u001b[0m┃\n", "┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n", "│ dense_19 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m128\u001b[0m) │ \u001b[38;5;34m1,664\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_20 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m256\u001b[0m) │ \u001b[38;5;34m33,024\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_21 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m512\u001b[0m) │ \u001b[38;5;34m131,584\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_22 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m256\u001b[0m) │ \u001b[38;5;34m131,328\u001b[0m │\n", "├─────────────────────────────────┼────────────────────────┼───────────────┤\n", "│ dense_23 (\u001b[38;5;33mDense\u001b[0m) │ (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m12\u001b[0m) │ \u001b[38;5;34m3,084\u001b[0m │\n", "└─────────────────────────────────┴────────────────────────┴───────────────┘\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Total params: 300,684 (1.15 MB)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Total params: \u001b[0m\u001b[38;5;34m300,684\u001b[0m (1.15 MB)\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Trainable params: 300,684 (1.15 MB)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Trainable params: \u001b[0m\u001b[38;5;34m300,684\u001b[0m (1.15 MB)\n" ] }, "metadata": {}, "output_type": "display_data" }, { "data": { "text/html": [ "
 Non-trainable params: 0 (0.00 B)\n",
       "
\n" ], "text/plain": [ "\u001b[1m Non-trainable params: \u001b[0m\u001b[38;5;34m0\u001b[0m (0.00 B)\n" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# model from paper\n", "# (see https://doi.org/10.1007/s11242-022-01779-3 model for the complex chemistry)\n", "model_paper = keras.Sequential(\n", " [keras.layers.Input(shape=(12,), dtype=dtype),\n", " keras.layers.Dense(128, activation='relu', dtype=dtype),\n", " keras.layers.Dense(256, activation='relu', dtype=dtype),\n", " keras.layers.Dense(512, activation='relu', dtype=dtype),\n", " keras.layers.Dense(256, activation='relu', dtype=dtype),\n", " keras.layers.Dense(12, dtype=dtype)\n", " ])\n", "\n", "model_paper.compile(optimizer=optimizer_paper, loss = loss)\n", "model_paper.summary()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Define transformer functions" ] }, { "cell_type": "code", "execution_count": 141, "metadata": {}, "outputs": [], "source": [ "def Safelog(val):\n", " # get range of vector\n", " if val > 0:\n", " return np.log10(val)\n", " elif val < 0:\n", " return -np.log10(-val)\n", " else:\n", " return 0\n", "\n", "def Safeexp(val):\n", " if val > 0:\n", " return -10 ** -val\n", " elif val < 0:\n", " return 10 ** val\n", " else:\n", " return 0" ] }, { "cell_type": "code", "execution_count": 142, "metadata": {}, "outputs": [], "source": [ "# ? Why does the charge is using another logarithm than the other species\n", "\n", "func_dict_in = {\n", " \"H\" : np.log1p,\n", " \"O\" : np.log1p,\n", " \"Charge\" : Safelog,\n", " \"H_0_\" : np.log1p,\n", " \"O_0_\" : np.log1p,\n", " \"Ba\" : np.log1p,\n", " \"Cl\" : np.log1p,\n", " \"S_2_\" : np.log1p,\n", " \"S_6_\" : np.log1p,\n", " \"Sr\" : np.log1p,\n", " \"Barite\" : np.log1p,\n", " \"Celestite\" : np.log1p,\n", "}\n", "\n", "func_dict_out = {\n", " \"H\" : np.expm1,\n", " \"O\" : np.expm1,\n", " \"Charge\" : Safeexp,\n", " \"H_0_\" : np.expm1,\n", " \"O_0_\" : np.expm1,\n", " \"Ba\" : np.expm1,\n", " \"Cl\" : np.expm1,\n", " \"S_2_\" : np.expm1,\n", " \"S_6_\" : np.expm1,\n", " \"Sr\" : np.expm1,\n", " \"Barite\" : np.expm1,\n", " \"Celestite\" : np.expm1,\n", "}\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Read data from `.h5` file and convert it to a `pandas.DataFrame`" ] }, { "cell_type": "code", "execution_count": 143, "metadata": {}, "outputs": [], "source": [ "# os.chdir('/mnt/beegfs/home/signer/projects/model-training')\n", "data_file = h5py.File(\"Barite_50_Data_training.h5\")\n", "\n", "design = data_file[\"design\"]\n", "results = data_file[\"result\"]\n", "\n", "df_design = pd.DataFrame(np.array(design[\"data\"]).transpose(), columns = np.array(design[\"names\"].asstr()))\n", "df_results = pd.DataFrame(np.array(results[\"data\"]).transpose(), columns = np.array(results[\"names\"].asstr()))\n", "\n", "data_file.close()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Preprocess Data\n", "\n", "The data are preprocessed in the following way:\n", "\n", "1. Label data points in the `design` dataset with `reactive` and `non-reactive` labels using kmeans clustering\n", "2. Transform `design` and `results` data set into log-scaled data.\n", "3. Split data into training and test sets.\n", "4. Learn scaler on training data for `design` and `results` together (option `global`) or individual (option `individual`).\n", "5. Transform training and test data.\n", "6. Split training data into training and validation dataset." ] }, { "cell_type": "code", "execution_count": 182, "metadata": {}, "outputs": [], "source": [ "species_columns = [\"H\", \"O\", \"Charge\", \"Ba\", \"Cl\", \"S_6_\", \"Sr\", \"Barite\", \"Celestite\"]" ] }, { "cell_type": "code", "execution_count": 183, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/home/signer/bin/miniconda3/envs/training/lib/python3.11/site-packages/sklearn/base.py:1473: ConvergenceWarning: Number of distinct clusters (1) found smaller than n_clusters (2). Possibly due to duplicate points in X.\n", " return fit_method(estimator, *args, **kwargs)\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "Amount class 0 before: 0.9879169719169719\n", "Amount class 1 before: 0.012083028083028084\n" ] } ], "source": [ "X_train, X_val, X_test, y_train, y_val, y_test, scaler_X, scaler_y = preprocessing_training(df_design[species_columns], df_results[species_columns], func_dict_in, func_dict_out, \"off\", 'global', 0.1)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Custom Loss function" ] }, { "cell_type": "code", "execution_count": 146, "metadata": {}, "outputs": [], "source": [ "def custom_loss_H20(df_design_log, df_result_log, data_min_log, data_max_log, func_dict_out, postprocess):\n", " df_result = postprocess(df_result_log, func_dict_out, data_min_log, data_max_log) \n", " return keras.losses.Huber + np.sum(((df_result['H'] / df_result['O']) - 2)**2)\n", "\n", "def loss_wrapper(data_min_log, data_max_log, func_dict_out, postprocess):\n", " def loss(df_design_log, df_result_log):\n", " return custom_loss_H20(df_design_log, df_result_log, data_min_log, data_max_log, func_dict_out, postprocess)\n", " return loss" ] }, { "cell_type": "code", "execution_count": 199, "metadata": {}, "outputs": [], "source": [ "def custom_loss(scaler_X, scaler_y, FuncTransform, dict_in, dict_out, columns):\n", " def loss(results, predicted):\n", " \n", " predicted = pd.DataFrame(scaler_X.inverse_transform(predicted), columns = columns)\n", " results = pd.DataFrame(scaler_y.inverse_transform(results), columns = columns)\n", " \n", " predicted = FuncTransform(dict_in, dict_out).inverse_transform(predicted)\n", " results = FuncTransform(dict_in, dict_out).inverse_transform(results)\n", " \n", " dBa = tf.keras.backend.abs((predicted[\"Ba\"] + predicted[\"Barite\"]) - (results[\"Ba\"] + results[\"Barite\"]))\n", " dSr = tf.keras.backend.abs((predicted[\"Sr\"] + predicted[\"Celestite\"]) - (results[\"Sr\"] + results[\"Celestite\"]))\n", " total_loss = keras.loss.Huber(results, predicted) + 0.1 * dBa + 0.1 * dSr\n", " \n", " return total_loss\n", "\n", " return loss" ] }, { "cell_type": "code", "execution_count": 200, "metadata": {}, "outputs": [], "source": [ "model_simple.compile(optimizer=optimizer_simple, loss=custom_loss(scaler_X, scaler_y, FuncTransform, func_dict_in, func_dict_out, species_columns))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Train the model" ] }, { "cell_type": "code", "execution_count": 188, "metadata": {}, "outputs": [], "source": [ "# measure time\n", "def model_training(model):\n", " start = time.time()\n", " callback = keras.callbacks.EarlyStopping(monitor='loss', patience=3)\n", " history = model.fit(X_train.iloc[:, X_train.columns != \"Class\"], \n", " y_train.iloc[:, y_train.columns != \"Class\"], \n", " batch_size = batch_size, \n", " epochs = 20, \n", " validation_data = (X_val.iloc[:, X_val.columns != \"Class\"], y_val.iloc[:, y_val.columns != \"Class\"]),\n", " callbacks = [callback])\n", "\n", " end = time.time()\n", "\n", " print(\"Training took {} seconds\".format(end - start))" ] }, { "cell_type": "code", "execution_count": 203, "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
HOChargeH_0_O_0_BaClS_2_S_6_SrBariteCelestite
0111.01243455.506219-7.748371e-091.541375e-130.000000e+001.916106e-024.082138e-021.440295e-206.403132e-070.0012500.0022500.999244
1111.01243455.506220-4.672647e-090.000000e+006.197818e-149.637903e-032.189133e-020.000000e+007.449873e-070.0013090.0023080.999244
2111.01243455.508129-1.325997e-091.416456e-130.000000e+005.839037e-075.887491e-042.981540e-174.781732e-040.0007720.0012940.999847
3111.01243455.508676-1.223549e-094.449984e-140.000000e+004.536952e-072.183551e-053.803886e-196.148996e-040.0006250.0010100.999995
4111.01243455.508699-1.216518e-090.000000e+004.098861e-144.496372e-078.098352e-070.000000e+006.205560e-040.0006210.0010001.000000
.......................................
2502495111.01243455.5074883.573897e-090.000000e+001.374167e-109.953042e-072.266555e-030.000000e+003.178142e-040.0014500.0010001.000014
2502496111.01243455.5075013.494199e-090.000000e+001.378928e-109.814307e-072.217997e-030.000000e+003.210423e-040.0014290.0010001.000010
2502497111.01243455.5075123.429947e-090.000000e+001.376072e-109.704342e-072.179066e-030.000000e+003.236905e-040.0014120.0010001.000006
2502498111.01243455.5075203.381818e-090.000000e+001.368903e-109.632999e-072.149820e-030.000000e+003.257170e-040.0014000.0010001.000004
2502499111.01243455.5075253.349044e-090.000000e+001.378174e-109.563975e-072.129912e-030.000000e+003.271123e-040.0013910.0010001.000001
\n", "

2502500 rows × 12 columns

\n", "
" ], "text/plain": [ " H O Charge H_0_ O_0_ \\\n", "0 111.012434 55.506219 -7.748371e-09 1.541375e-13 0.000000e+00 \n", "1 111.012434 55.506220 -4.672647e-09 0.000000e+00 6.197818e-14 \n", "2 111.012434 55.508129 -1.325997e-09 1.416456e-13 0.000000e+00 \n", "3 111.012434 55.508676 -1.223549e-09 4.449984e-14 0.000000e+00 \n", "4 111.012434 55.508699 -1.216518e-09 0.000000e+00 4.098861e-14 \n", "... ... ... ... ... ... \n", "2502495 111.012434 55.507488 3.573897e-09 0.000000e+00 1.374167e-10 \n", "2502496 111.012434 55.507501 3.494199e-09 0.000000e+00 1.378928e-10 \n", "2502497 111.012434 55.507512 3.429947e-09 0.000000e+00 1.376072e-10 \n", "2502498 111.012434 55.507520 3.381818e-09 0.000000e+00 1.368903e-10 \n", "2502499 111.012434 55.507525 3.349044e-09 0.000000e+00 1.378174e-10 \n", "\n", " Ba Cl S_2_ S_6_ Sr \\\n", "0 1.916106e-02 4.082138e-02 1.440295e-20 6.403132e-07 0.001250 \n", "1 9.637903e-03 2.189133e-02 0.000000e+00 7.449873e-07 0.001309 \n", "2 5.839037e-07 5.887491e-04 2.981540e-17 4.781732e-04 0.000772 \n", "3 4.536952e-07 2.183551e-05 3.803886e-19 6.148996e-04 0.000625 \n", "4 4.496372e-07 8.098352e-07 0.000000e+00 6.205560e-04 0.000621 \n", "... ... ... ... ... ... \n", "2502495 9.953042e-07 2.266555e-03 0.000000e+00 3.178142e-04 0.001450 \n", "2502496 9.814307e-07 2.217997e-03 0.000000e+00 3.210423e-04 0.001429 \n", "2502497 9.704342e-07 2.179066e-03 0.000000e+00 3.236905e-04 0.001412 \n", "2502498 9.632999e-07 2.149820e-03 0.000000e+00 3.257170e-04 0.001400 \n", "2502499 9.563975e-07 2.129912e-03 0.000000e+00 3.271123e-04 0.001391 \n", "\n", " Barite Celestite \n", "0 0.002250 0.999244 \n", "1 0.002308 0.999244 \n", "2 0.001294 0.999847 \n", "3 0.001010 0.999995 \n", "4 0.001000 1.000000 \n", "... ... ... \n", "2502495 0.001000 1.000014 \n", "2502496 0.001000 1.000010 \n", "2502497 0.001000 1.000006 \n", "2502498 0.001000 1.000004 \n", "2502499 0.001000 1.000001 \n", "\n", "[2502500 rows x 12 columns]" ] }, "execution_count": 203, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df_results" ] }, { "cell_type": "code", "execution_count": 202, "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
HOChargeH_0_O_0_BaClS_2_S_6_SrBariteCelestite
0111.01243455.508192-7.779554e-092.697041e-262.210590e-152.041069e-024.082138e-020.000000e+000.0004940.0004940.0011.000000
1111.01243455.508427-4.736083e-091.446346e-262.473481e-151.094567e-022.189133e-020.000000e+000.0005530.0005530.0011.000000
2111.01243455.508691-1.311169e-093.889826e-282.769320e-152.943745e-045.887491e-040.000000e+000.0006190.0006190.0011.000000
3111.01243455.508698-1.220023e-091.442658e-292.777193e-151.091776e-052.183551e-050.000000e+000.0006200.0006200.0011.000000
4111.01243455.508699-1.216643e-095.350528e-312.777485e-154.049176e-078.098352e-070.000000e+000.0006200.0006200.0011.000000
.......................................
2502495111.01243455.5074883.573728e-095.424062e-1451.375204e-109.953520e-072.266555e-035.509534e-1490.0003180.0014500.0011.000014
2502496111.01243455.5075013.494007e-092.011675e-1461.377139e-109.817216e-072.217997e-032.043375e-1500.0003210.0014290.0011.000010
2502497111.01243455.5075123.429764e-097.460897e-1481.377819e-109.706451e-072.179066e-037.578467e-1520.0003240.0014120.0011.000006
2502498111.01243455.5075203.381745e-092.767237e-1491.371144e-109.621074e-072.149820e-032.810844e-1530.0003260.0014000.0011.000004
2502499111.01243455.5075253.348864e-095.321610e-1511.376026e-109.564401e-072.129912e-035.405468e-1550.0003270.0013910.0011.000001
\n", "

2502500 rows × 12 columns

\n", "
" ], "text/plain": [ " H O Charge H_0_ O_0_ \\\n", "0 111.012434 55.508192 -7.779554e-09 2.697041e-26 2.210590e-15 \n", "1 111.012434 55.508427 -4.736083e-09 1.446346e-26 2.473481e-15 \n", "2 111.012434 55.508691 -1.311169e-09 3.889826e-28 2.769320e-15 \n", "3 111.012434 55.508698 -1.220023e-09 1.442658e-29 2.777193e-15 \n", "4 111.012434 55.508699 -1.216643e-09 5.350528e-31 2.777485e-15 \n", "... ... ... ... ... ... \n", "2502495 111.012434 55.507488 3.573728e-09 5.424062e-145 1.375204e-10 \n", "2502496 111.012434 55.507501 3.494007e-09 2.011675e-146 1.377139e-10 \n", "2502497 111.012434 55.507512 3.429764e-09 7.460897e-148 1.377819e-10 \n", "2502498 111.012434 55.507520 3.381745e-09 2.767237e-149 1.371144e-10 \n", "2502499 111.012434 55.507525 3.348864e-09 5.321610e-151 1.376026e-10 \n", "\n", " Ba Cl S_2_ S_6_ Sr \\\n", "0 2.041069e-02 4.082138e-02 0.000000e+00 0.000494 0.000494 \n", "1 1.094567e-02 2.189133e-02 0.000000e+00 0.000553 0.000553 \n", "2 2.943745e-04 5.887491e-04 0.000000e+00 0.000619 0.000619 \n", "3 1.091776e-05 2.183551e-05 0.000000e+00 0.000620 0.000620 \n", "4 4.049176e-07 8.098352e-07 0.000000e+00 0.000620 0.000620 \n", "... ... ... ... ... ... \n", "2502495 9.953520e-07 2.266555e-03 5.509534e-149 0.000318 0.001450 \n", "2502496 9.817216e-07 2.217997e-03 2.043375e-150 0.000321 0.001429 \n", "2502497 9.706451e-07 2.179066e-03 7.578467e-152 0.000324 0.001412 \n", "2502498 9.621074e-07 2.149820e-03 2.810844e-153 0.000326 0.001400 \n", "2502499 9.564401e-07 2.129912e-03 5.405468e-155 0.000327 0.001391 \n", "\n", " Barite Celestite \n", "0 0.001 1.000000 \n", "1 0.001 1.000000 \n", "2 0.001 1.000000 \n", "3 0.001 1.000000 \n", "4 0.001 1.000000 \n", "... ... ... \n", "2502495 0.001 1.000014 \n", "2502496 0.001 1.000010 \n", "2502497 0.001 1.000006 \n", "2502498 0.001 1.000004 \n", "2502499 0.001 1.000001 \n", "\n", "[2502500 rows x 12 columns]" ] }, "execution_count": 202, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df_design" ] }, { "cell_type": "code", "execution_count": 206, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "0 0.324673\n", "1 0.307932\n", "2 -2.590695\n", "3 -1.376848\n", "4 0.045476\n", " ... \n", "2502495 -0.000021\n", "2502496 -0.000129\n", "2502497 -0.000094\n", "2502498 0.000537\n", "2502499 -0.000019\n", "Length: 2502500, dtype: float64" ] }, "execution_count": 206, "metadata": {}, "output_type": "execute_result" } ], "source": [ "(np.log10(df_results['Ba']) + np.log10(df_results['Barite'])) - (np.log10(df_design['Ba']) + np.log10(df_design['Barite']))" ] }, { "cell_type": "code", "execution_count": 201, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Epoch 1/20\n" ] }, { "ename": "NotImplementedError", "evalue": "Cannot convert a symbolic tf.Tensor (sequential_7_1/dense_29_1/Add:0) to a numpy array. This error may indicate that you're trying to pass a Tensor to a NumPy call, which is not supported.", "output_type": "error", "traceback": [ "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m", "\u001b[0;31mNotImplementedError\u001b[0m Traceback (most recent call last)", "Cell \u001b[0;32mIn[201], line 1\u001b[0m\n\u001b[0;32m----> 1\u001b[0m model_training(model_simple)\n", "Cell \u001b[0;32mIn[188], line 5\u001b[0m, in \u001b[0;36mmodel_training\u001b[0;34m(model)\u001b[0m\n\u001b[1;32m 3\u001b[0m start \u001b[38;5;241m=\u001b[39m time\u001b[38;5;241m.\u001b[39mtime()\n\u001b[1;32m 4\u001b[0m callback \u001b[38;5;241m=\u001b[39m keras\u001b[38;5;241m.\u001b[39mcallbacks\u001b[38;5;241m.\u001b[39mEarlyStopping(monitor\u001b[38;5;241m=\u001b[39m\u001b[38;5;124m'\u001b[39m\u001b[38;5;124mloss\u001b[39m\u001b[38;5;124m'\u001b[39m, patience\u001b[38;5;241m=\u001b[39m\u001b[38;5;241m3\u001b[39m)\n\u001b[0;32m----> 5\u001b[0m history \u001b[38;5;241m=\u001b[39m model\u001b[38;5;241m.\u001b[39mfit(X_train\u001b[38;5;241m.\u001b[39miloc[:, X_train\u001b[38;5;241m.\u001b[39mcolumns \u001b[38;5;241m!=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mClass\u001b[39m\u001b[38;5;124m\"\u001b[39m], \n\u001b[1;32m 6\u001b[0m y_train\u001b[38;5;241m.\u001b[39miloc[:, y_train\u001b[38;5;241m.\u001b[39mcolumns \u001b[38;5;241m!=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mClass\u001b[39m\u001b[38;5;124m\"\u001b[39m], \n\u001b[1;32m 7\u001b[0m batch_size \u001b[38;5;241m=\u001b[39m batch_size, \n\u001b[1;32m 8\u001b[0m epochs \u001b[38;5;241m=\u001b[39m \u001b[38;5;241m20\u001b[39m, \n\u001b[1;32m 9\u001b[0m validation_data \u001b[38;5;241m=\u001b[39m (X_val\u001b[38;5;241m.\u001b[39miloc[:, X_val\u001b[38;5;241m.\u001b[39mcolumns \u001b[38;5;241m!=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mClass\u001b[39m\u001b[38;5;124m\"\u001b[39m], y_val\u001b[38;5;241m.\u001b[39miloc[:, y_val\u001b[38;5;241m.\u001b[39mcolumns \u001b[38;5;241m!=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mClass\u001b[39m\u001b[38;5;124m\"\u001b[39m]),\n\u001b[1;32m 10\u001b[0m callbacks \u001b[38;5;241m=\u001b[39m [callback])\n\u001b[1;32m 12\u001b[0m end \u001b[38;5;241m=\u001b[39m time\u001b[38;5;241m.\u001b[39mtime()\n\u001b[1;32m 14\u001b[0m \u001b[38;5;28mprint\u001b[39m(\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mTraining took \u001b[39m\u001b[38;5;132;01m{}\u001b[39;00m\u001b[38;5;124m seconds\u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;241m.\u001b[39mformat(end \u001b[38;5;241m-\u001b[39m start))\n", "File \u001b[0;32m~/bin/miniconda3/envs/training/lib/python3.11/site-packages/keras/src/utils/traceback_utils.py:122\u001b[0m, in \u001b[0;36mfilter_traceback..error_handler\u001b[0;34m(*args, **kwargs)\u001b[0m\n\u001b[1;32m 119\u001b[0m filtered_tb \u001b[38;5;241m=\u001b[39m _process_traceback_frames(e\u001b[38;5;241m.\u001b[39m__traceback__)\n\u001b[1;32m 120\u001b[0m \u001b[38;5;66;03m# To get the full stack trace, call:\u001b[39;00m\n\u001b[1;32m 121\u001b[0m \u001b[38;5;66;03m# `keras.config.disable_traceback_filtering()`\u001b[39;00m\n\u001b[0;32m--> 122\u001b[0m \u001b[38;5;28;01mraise\u001b[39;00m e\u001b[38;5;241m.\u001b[39mwith_traceback(filtered_tb) \u001b[38;5;28;01mfrom\u001b[39;00m \u001b[38;5;28;01mNone\u001b[39;00m\n\u001b[1;32m 123\u001b[0m \u001b[38;5;28;01mfinally\u001b[39;00m:\n\u001b[1;32m 124\u001b[0m \u001b[38;5;28;01mdel\u001b[39;00m filtered_tb\n", "Cell \u001b[0;32mIn[199], line 4\u001b[0m, in \u001b[0;36mcustom_loss..loss\u001b[0;34m(results, predicted)\u001b[0m\n\u001b[1;32m 2\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mloss\u001b[39m(results, predicted):\n\u001b[0;32m----> 4\u001b[0m predicted \u001b[38;5;241m=\u001b[39m pd\u001b[38;5;241m.\u001b[39mDataFrame(scaler_X\u001b[38;5;241m.\u001b[39minverse_transform(predicted), columns \u001b[38;5;241m=\u001b[39m columns)\n\u001b[1;32m 5\u001b[0m results \u001b[38;5;241m=\u001b[39m pd\u001b[38;5;241m.\u001b[39mDataFrame(scaler_y\u001b[38;5;241m.\u001b[39minverse_transform(results), columns \u001b[38;5;241m=\u001b[39m columns)\n\u001b[1;32m 7\u001b[0m predicted \u001b[38;5;241m=\u001b[39m FuncTransform(dict_in, dict_out)\u001b[38;5;241m.\u001b[39minverse_transform(predicted)\n", "File \u001b[0;32m~/bin/miniconda3/envs/training/lib/python3.11/site-packages/sklearn/preprocessing/_data.py:566\u001b[0m, in \u001b[0;36mMinMaxScaler.inverse_transform\u001b[0;34m(self, X)\u001b[0m\n\u001b[1;32m 562\u001b[0m check_is_fitted(\u001b[38;5;28mself\u001b[39m)\n\u001b[1;32m 564\u001b[0m xp, _ \u001b[38;5;241m=\u001b[39m get_namespace(X)\n\u001b[0;32m--> 566\u001b[0m X \u001b[38;5;241m=\u001b[39m check_array(\n\u001b[1;32m 567\u001b[0m X,\n\u001b[1;32m 568\u001b[0m copy\u001b[38;5;241m=\u001b[39m\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mcopy,\n\u001b[1;32m 569\u001b[0m dtype\u001b[38;5;241m=\u001b[39m_array_api\u001b[38;5;241m.\u001b[39msupported_float_dtypes(xp),\n\u001b[1;32m 570\u001b[0m force_writeable\u001b[38;5;241m=\u001b[39m\u001b[38;5;28;01mTrue\u001b[39;00m,\n\u001b[1;32m 571\u001b[0m force_all_finite\u001b[38;5;241m=\u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mallow-nan\u001b[39m\u001b[38;5;124m\"\u001b[39m,\n\u001b[1;32m 572\u001b[0m )\n\u001b[1;32m 574\u001b[0m X \u001b[38;5;241m-\u001b[39m\u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mmin_\n\u001b[1;32m 575\u001b[0m X \u001b[38;5;241m/\u001b[39m\u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mscale_\n", "File \u001b[0;32m~/bin/miniconda3/envs/training/lib/python3.11/site-packages/sklearn/utils/validation.py:1012\u001b[0m, in \u001b[0;36mcheck_array\u001b[0;34m(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_writeable, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, estimator, input_name)\u001b[0m\n\u001b[1;32m 1010\u001b[0m array \u001b[38;5;241m=\u001b[39m xp\u001b[38;5;241m.\u001b[39mastype(array, dtype, copy\u001b[38;5;241m=\u001b[39m\u001b[38;5;28;01mFalse\u001b[39;00m)\n\u001b[1;32m 1011\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[0;32m-> 1012\u001b[0m array \u001b[38;5;241m=\u001b[39m _asarray_with_order(array, order\u001b[38;5;241m=\u001b[39morder, dtype\u001b[38;5;241m=\u001b[39mdtype, xp\u001b[38;5;241m=\u001b[39mxp)\n\u001b[1;32m 1013\u001b[0m \u001b[38;5;28;01mexcept\u001b[39;00m ComplexWarning \u001b[38;5;28;01mas\u001b[39;00m complex_warning:\n\u001b[1;32m 1014\u001b[0m \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mValueError\u001b[39;00m(\n\u001b[1;32m 1015\u001b[0m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mComplex data not supported\u001b[39m\u001b[38;5;130;01m\\n\u001b[39;00m\u001b[38;5;132;01m{}\u001b[39;00m\u001b[38;5;130;01m\\n\u001b[39;00m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;241m.\u001b[39mformat(array)\n\u001b[1;32m 1016\u001b[0m ) \u001b[38;5;28;01mfrom\u001b[39;00m \u001b[38;5;21;01mcomplex_warning\u001b[39;00m\n", "File \u001b[0;32m~/bin/miniconda3/envs/training/lib/python3.11/site-packages/sklearn/utils/_array_api.py:745\u001b[0m, in \u001b[0;36m_asarray_with_order\u001b[0;34m(array, dtype, order, copy, xp, device)\u001b[0m\n\u001b[1;32m 743\u001b[0m array \u001b[38;5;241m=\u001b[39m numpy\u001b[38;5;241m.\u001b[39marray(array, order\u001b[38;5;241m=\u001b[39morder, dtype\u001b[38;5;241m=\u001b[39mdtype)\n\u001b[1;32m 744\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[0;32m--> 745\u001b[0m array \u001b[38;5;241m=\u001b[39m numpy\u001b[38;5;241m.\u001b[39masarray(array, order\u001b[38;5;241m=\u001b[39morder, dtype\u001b[38;5;241m=\u001b[39mdtype)\n\u001b[1;32m 747\u001b[0m \u001b[38;5;66;03m# At this point array is a NumPy ndarray. We convert it to an array\u001b[39;00m\n\u001b[1;32m 748\u001b[0m \u001b[38;5;66;03m# container that is consistent with the input's namespace.\u001b[39;00m\n\u001b[1;32m 749\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m xp\u001b[38;5;241m.\u001b[39masarray(array)\n", "\u001b[0;31mNotImplementedError\u001b[0m: Cannot convert a symbolic tf.Tensor (sequential_7_1/dense_29_1/Add:0) to a numpy array. This error may indicate that you're trying to pass a Tensor to a NumPy call, which is not supported." ] } ], "source": [ "model_training(model_simple)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Test Mass Balance" ] }, { "cell_type": "code", "execution_count": 164, "metadata": {}, "outputs": [], "source": [ "def mass_balance(model, X, scaler_X, scaler_y, func_dict_in, func_dict_out):\n", " \n", " # predict the chemistry\n", " columns = X.iloc[:, X.columns != \"Class\"].columns\n", " prediction = pd.DataFrame(model.predict(X[columns]), columns=columns)\n", " # backtransform min/max\n", " X = pd.DataFrame(scaler_X.inverse_transform(X.iloc[:, X.columns != \"Class\"]), columns=columns)\n", " prediction = pd.DataFrame(scaler_y.inverse_transform(prediction), columns=columns)\n", " \n", " # backtransform log\n", " X = FuncTransform(func_dict_in, func_dict_out).inverse_transform(X)\n", " prediction = FuncTransform(func_dict_in, func_dict_out).inverse_transform(prediction)\n", " \n", " # calculate mass balance\n", " dBa = np.abs((prediction[\"Ba\"] + prediction[\"Barite\"]) - (X[\"Ba\"] + X[\"Barite\"]))\n", " dSr = np.abs((prediction[\"Sr\"] + prediction[\"Celestite\"]) - (X[\"Sr\"] + X[\"Celestite\"]))\n", " \n", " return dBa + dSr" ] }, { "cell_type": "code", "execution_count": 85, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "\u001b[1m7821/7821\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m3s\u001b[0m 328us/step\n" ] } ], "source": [ "columns = X_test.columns[X_test.columns != \"Class\"]\n", "prediction = pd.DataFrame(model_simple.predict(X_test[columns]), columns = columns)" ] }, { "cell_type": "code", "execution_count": 167, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "\u001b[1m7821/7821\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m9s\u001b[0m 1ms/step\n" ] } ], "source": [ "mass_balance_results = mass_balance(model_large, X_test, scaler_X, scaler_y, func_dict_in, func_dict_out)" ] }, { "cell_type": "code", "execution_count": 176, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "0.022793206793206792" ] }, "execution_count": 176, "metadata": {}, "output_type": "execute_result" } ], "source": [ "len(mass_balance_results[mass_balance_results < 1e-5]) / len(mass_balance_results)" ] }, { "cell_type": "code", "execution_count": 169, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "4 0.000006\n", "46 0.000010\n", "157 0.000010\n", "160 0.000003\n", "189 0.000009\n", " ... \n", "250016 0.000007\n", "250022 0.000007\n", "250086 0.000007\n", "250139 0.000003\n", "250223 0.000006\n", "Length: 5704, dtype: float64" ] }, "execution_count": 169, "metadata": {}, "output_type": "execute_result" } ], "source": [ "mass_balance_results[mass_balance_results < 1e-5]" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Percentage of cells that pass the mass balance condition\n", "\n", "small_modell_20_epochs = 0.0031088911088911087\n", "\n", "large_modell_20_epochs = 0.022793206793206792" ] }, { "cell_type": "code", "execution_count": 178, "metadata": {}, "outputs": [ { "ename": "TypeError", "evalue": "argument of type 'int' is not iterable", "output_type": "error", "traceback": [ "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m", "\u001b[0;31mTypeError\u001b[0m Traceback (most recent call last)", "Cell \u001b[0;32mIn[178], line 11\u001b[0m\n\u001b[1;32m 7\u001b[0m y_results \u001b[38;5;241m=\u001b[39m []\n\u001b[1;32m 8\u001b[0m y_differences \u001b[38;5;241m=\u001b[39m []\n\u001b[0;32m---> 11\u001b[0m df_design_transformed_scaled \u001b[38;5;241m=\u001b[39m scaler_X\u001b[38;5;241m.\u001b[39mtransform(FuncTransform(func_dict_in, func_dict_out)\u001b[38;5;241m.\u001b[39mfit_transform(df_design[species]))\n\u001b[1;32m 12\u001b[0m df_results_transformed_scaled \u001b[38;5;241m=\u001b[39m scaler_X\u001b[38;5;241m.\u001b[39mtransform(FuncTransform(func_dict_in, func_dict_out)\u001b[38;5;241m.\u001b[39mfit_transform(df_results[species]))\n\u001b[1;32m 14\u001b[0m \u001b[38;5;28;01mfor\u001b[39;00m i \u001b[38;5;129;01min\u001b[39;00m \u001b[38;5;28mrange\u001b[39m(\u001b[38;5;241m0\u001b[39m,iterations):\n", "File \u001b[0;32m~/Documents/model-training/preprocessing.py:63\u001b[0m, in \u001b[0;36mFuncTransform.fit_transform\u001b[0;34m(self, X, y)\u001b[0m\n\u001b[1;32m 61\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mfit_transform\u001b[39m(\u001b[38;5;28mself\u001b[39m, X, y\u001b[38;5;241m=\u001b[39m\u001b[38;5;28;01mNone\u001b[39;00m):\n\u001b[1;32m 62\u001b[0m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mfit(X)\n\u001b[0;32m---> 63\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mtransform(X, y)\n", "File \u001b[0;32m~/Documents/model-training/preprocessing.py:57\u001b[0m, in \u001b[0;36mFuncTransform.transform\u001b[0;34m(self, X, y)\u001b[0m\n\u001b[1;32m 55\u001b[0m X \u001b[38;5;241m=\u001b[39m X\u001b[38;5;241m.\u001b[39mcopy()\n\u001b[1;32m 56\u001b[0m \u001b[38;5;28;01mfor\u001b[39;00m key \u001b[38;5;129;01min\u001b[39;00m X\u001b[38;5;241m.\u001b[39mkeys(): \n\u001b[0;32m---> 57\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mClass\u001b[39m\u001b[38;5;124m\"\u001b[39m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;129;01min\u001b[39;00m key:\n\u001b[1;32m 58\u001b[0m X[key] \u001b[38;5;241m=\u001b[39m X[key]\u001b[38;5;241m.\u001b[39mapply(\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mfunc_transform[key])\n\u001b[1;32m 59\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m X\n", "\u001b[0;31mTypeError\u001b[0m: argument of type 'int' is not iterable" ] } ], "source": [ "import matplotlib.pyplot as plt\n", "\n", "species = \"Barite\"\n", "iterations = 1000\n", "cell_offset = 9\n", "y_design = []\n", "y_results = []\n", "y_differences = []\n", "\n", "\n", "df_design_transformed_scaled = scaler_X.transform(FuncTransform(func_dict_in, func_dict_out).fit_transform(df_design[species]))\n", "df_results_transformed_scaled = scaler_X.transform(FuncTransform(func_dict_in, func_dict_out).fit_transform(df_results[species]))\n", "\n", "for i in range(0,iterations):\n", " idx = i*50*50 + cell_offset -1\n", " y_design.append(df_design_transformed_scaled.iloc[idx, :])\n", " y_results.append(df_results_transformed_scaled.iloc[idx,:])\n", " \n", "y_design = pd.DataFrame(y_design)\n", "y_results = pd.DataFrame(y_results)\n", "# plt.plot(np.arange(0,iterations), y_design[species], label = \"Design\")\n", "plt.plot(np.arange(0,iterations), y_results[species], label = \"Results\")\n", "\n", "prediction = model_simple.predict(y_design.iloc[:, y_design.columns != \"Class\"])\n", "prediction = pd.DataFrame(prediction, columns = y_results.columns)\n", "\n", "y_results_back = FuncTransform(func_dict_in, func_dict_out).inverse_transform(pd.DataFrame(scaler_X.inverse_transform(y_results), columns=df_results.columns))\n", "prediction_back = FuncTransform(func_dict_in, func_dict_out).inverse_transform(pd.DataFrame(scaler_X.inverse_transform(prediction), columns=df_results.columns))\n", "\n", "plt.plot(np.arange(0,iterations), prediction[species], label = \"Prediction\")\n", "plt.xlabel('Iteration')\n", "plt.ylabel(species)\n", "plt.title(species+' Concentration over Iterations')\n", "plt.legend()\n", "plt.show()\n", "\n", "# plt.plot(np.arange(0,iterations), y_results_back[species], label = \"Results\")\n", "# plt.plot(np.arange(0,iterations), prediction_back[species], label = \"Prediction\")\n", "# plt.legend()\n", "\n", "# plt.show()\n", "\n", "timestep = 1000\n", "plt.imshow(np.array(df_results[\"Barite\"][(timestep*2500):(timestep*2500+2500)]).reshape(50,50), origin='lower')" ] }, { "cell_type": "code", "execution_count": 25, "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
HOChargeH_0_O_0_BaClS_2_S_6_SrBariteCelestite
0111.01243455.508700-1.216415e-090.02.217711e-134.495355e-071.532249e-120.00.0006210.0006200.0010001.000000
1111.01243455.508700-1.222504e-090.01.312902e-124.500359e-071.044603e-080.00.0006210.0006200.0010001.000000
2111.01243455.508699-1.220407e-090.01.614098e-124.500563e-074.907802e-070.00.0006210.0006200.0010001.000000
3111.01243455.508695-1.216831e-090.02.293739e-124.504482e-074.772370e-060.00.0006200.0006220.0010001.000000
4111.01243455.508679-1.216842e-090.02.641545e-124.534070e-072.200220e-050.00.0006160.0006260.0010001.000000
.......................................
995111.01243455.5064102.170844e-080.04.777415e-111.550089e-049.784038e-020.00.0000480.0488140.1494760.846327
996111.01243455.5064102.166504e-080.04.808612e-111.559012e-049.785750e-020.00.0000480.0488210.1517080.844093
997111.01243455.5064092.162167e-080.04.811205e-111.567226e-049.787459e-020.00.0000480.0488290.1539450.841856
998111.01243455.5064092.157995e-080.04.815004e-111.574812e-049.789167e-020.00.0000480.0488360.1561850.839614
999111.01243455.5064092.153938e-080.04.815067e-111.581835e-049.790872e-020.00.0000480.0488440.1584280.837370
\n", "

1000 rows × 12 columns

\n", "
" ], "text/plain": [ " H O Charge H_0_ O_0_ Ba \\\n", "0 111.012434 55.508700 -1.216415e-09 0.0 2.217711e-13 4.495355e-07 \n", "1 111.012434 55.508700 -1.222504e-09 0.0 1.312902e-12 4.500359e-07 \n", "2 111.012434 55.508699 -1.220407e-09 0.0 1.614098e-12 4.500563e-07 \n", "3 111.012434 55.508695 -1.216831e-09 0.0 2.293739e-12 4.504482e-07 \n", "4 111.012434 55.508679 -1.216842e-09 0.0 2.641545e-12 4.534070e-07 \n", ".. ... ... ... ... ... ... \n", "995 111.012434 55.506410 2.170844e-08 0.0 4.777415e-11 1.550089e-04 \n", "996 111.012434 55.506410 2.166504e-08 0.0 4.808612e-11 1.559012e-04 \n", "997 111.012434 55.506409 2.162167e-08 0.0 4.811205e-11 1.567226e-04 \n", "998 111.012434 55.506409 2.157995e-08 0.0 4.815004e-11 1.574812e-04 \n", "999 111.012434 55.506409 2.153938e-08 0.0 4.815067e-11 1.581835e-04 \n", "\n", " Cl S_2_ S_6_ Sr Barite Celestite \n", "0 1.532249e-12 0.0 0.000621 0.000620 0.001000 1.000000 \n", "1 1.044603e-08 0.0 0.000621 0.000620 0.001000 1.000000 \n", "2 4.907802e-07 0.0 0.000621 0.000620 0.001000 1.000000 \n", "3 4.772370e-06 0.0 0.000620 0.000622 0.001000 1.000000 \n", "4 2.200220e-05 0.0 0.000616 0.000626 0.001000 1.000000 \n", ".. ... ... ... ... ... ... \n", "995 9.784038e-02 0.0 0.000048 0.048814 0.149476 0.846327 \n", "996 9.785750e-02 0.0 0.000048 0.048821 0.151708 0.844093 \n", "997 9.787459e-02 0.0 0.000048 0.048829 0.153945 0.841856 \n", "998 9.789167e-02 0.0 0.000048 0.048836 0.156185 0.839614 \n", "999 9.790872e-02 0.0 0.000048 0.048844 0.158428 0.837370 \n", "\n", "[1000 rows x 12 columns]" ] }, "execution_count": 25, "metadata": {}, "output_type": "execute_result" } ], "source": [ "FuncTransform(func_dict_in, func_dict_out).inverse_transform(pd.DataFrame(scaler_X.inverse_transform(y_results), columns=df_results.columns))" ] }, { "cell_type": "code", "execution_count": 26, "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
HOChargeH_0_O_0_BaClS_2_S_6_SrBariteCelestite
0111.01242855.508682-1.193141e-093.397941e-152.128961e-13-0.0000120.000021-1.191799e-170.0006200.0006300.0009851.000231
1111.01242855.508682-1.202185e-092.918474e-151.019832e-12-0.0000160.000008-7.920033e-180.0006200.0006300.0009461.000121
2111.01242855.508682-1.203471e-091.785468e-152.398433e-12-0.000015-0.000013-9.158671e-180.0006200.0006270.0009130.999977
3111.01242855.508682-1.199235e-091.746077e-152.357316e-12-0.000016-0.000011-9.246642e-180.0006190.0006290.0009160.999936
4111.01242855.508682-1.197043e-091.533956e-152.670053e-12-0.0000190.000003-9.144569e-180.0006150.0006350.0009430.999769
.......................................
995111.01242855.5064162.189995e-08-3.792428e-154.785985e-110.0002790.0976421.182626e-160.0000510.0487380.1490160.844585
996111.01242855.5064162.188781e-08-3.910681e-154.822730e-110.0002790.0976651.169111e-160.0000510.0487510.1512410.842272
997111.01242855.5064162.184875e-08-3.749360e-154.824932e-110.0002790.0976871.146696e-160.0000510.0487610.1534770.840015
998111.01242855.5064162.180415e-08-3.500642e-154.816323e-110.0002790.0977101.119656e-160.0000510.0487700.1557200.837773
999111.01242855.5064162.177183e-08-3.375572e-154.822685e-110.0002790.0977321.095921e-160.0000500.0487800.1579620.835504
\n", "

1000 rows × 12 columns

\n", "
" ], "text/plain": [ " H O Charge H_0_ O_0_ \\\n", "0 111.012428 55.508682 -1.193141e-09 3.397941e-15 2.128961e-13 \n", "1 111.012428 55.508682 -1.202185e-09 2.918474e-15 1.019832e-12 \n", "2 111.012428 55.508682 -1.203471e-09 1.785468e-15 2.398433e-12 \n", "3 111.012428 55.508682 -1.199235e-09 1.746077e-15 2.357316e-12 \n", "4 111.012428 55.508682 -1.197043e-09 1.533956e-15 2.670053e-12 \n", ".. ... ... ... ... ... \n", "995 111.012428 55.506416 2.189995e-08 -3.792428e-15 4.785985e-11 \n", "996 111.012428 55.506416 2.188781e-08 -3.910681e-15 4.822730e-11 \n", "997 111.012428 55.506416 2.184875e-08 -3.749360e-15 4.824932e-11 \n", "998 111.012428 55.506416 2.180415e-08 -3.500642e-15 4.816323e-11 \n", "999 111.012428 55.506416 2.177183e-08 -3.375572e-15 4.822685e-11 \n", "\n", " Ba Cl S_2_ S_6_ Sr Barite Celestite \n", "0 -0.000012 0.000021 -1.191799e-17 0.000620 0.000630 0.000985 1.000231 \n", "1 -0.000016 0.000008 -7.920033e-18 0.000620 0.000630 0.000946 1.000121 \n", "2 -0.000015 -0.000013 -9.158671e-18 0.000620 0.000627 0.000913 0.999977 \n", "3 -0.000016 -0.000011 -9.246642e-18 0.000619 0.000629 0.000916 0.999936 \n", "4 -0.000019 0.000003 -9.144569e-18 0.000615 0.000635 0.000943 0.999769 \n", ".. ... ... ... ... ... ... ... \n", "995 0.000279 0.097642 1.182626e-16 0.000051 0.048738 0.149016 0.844585 \n", "996 0.000279 0.097665 1.169111e-16 0.000051 0.048751 0.151241 0.842272 \n", "997 0.000279 0.097687 1.146696e-16 0.000051 0.048761 0.153477 0.840015 \n", "998 0.000279 0.097710 1.119656e-16 0.000051 0.048770 0.155720 0.837773 \n", "999 0.000279 0.097732 1.095921e-16 0.000050 0.048780 0.157962 0.835504 \n", "\n", "[1000 rows x 12 columns]" ] }, "execution_count": 26, "metadata": {}, "output_type": "execute_result" } ], "source": [ "FuncTransform(func_dict_in, func_dict_out).inverse_transform(pd.DataFrame(scaler_X.inverse_transform(prediction), columns=prediction.columns))" ] }, { "cell_type": "code", "execution_count": 25, "metadata": {}, "outputs": [ { "data": { "image/png": "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", "text/plain": [ "
" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "plt.plot(history.history[\"loss\"], \"o-\", label = \"Training Loss\")\n", "plt.xlabel(\"Epoch\")\n", "# plt.yscale('log')\n", "plt.ylabel(\"Loss (Huber)\")\n", "plt.grid('on')\n", "\n", "plt.savefig(\"loss_all.png\", dpi=300)\n" ] }, { "cell_type": "code", "execution_count": 14, "metadata": {}, "outputs": [ { "data": { "image/png": "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", "text/plain": [ "
" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "plt.plot(history.history[\"loss\"][1:], \"o-\", label = \"Training Loss\")\n", "plt.xlabel(\"Epoch\")\n", "# plt.yscale('log')\n", "plt.ylabel(\"Loss (Huber)\")\n", "plt.grid('on')\n", "plt.savefig(\"loss_1_to_end.png\", dpi=300)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Test the model" ] }, { "cell_type": "code", "execution_count": 184, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "\u001b[1m7821/7821\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m9s\u001b[0m 1ms/step - loss: 7.4076e-07\n" ] }, { "data": { "text/plain": [ "1.0149801710213069e-06" ] }, "execution_count": 184, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# test on all test data\n", "model_large.evaluate(X_test[species_columns], y_test[species_columns])" ] }, { "cell_type": "code", "execution_count": 185, "metadata": {}, "outputs": [ { "ename": "IndexError", "evalue": ".iloc requires numeric indexers, got ['H' 'O' 'Charge' 'Ba' 'Cl' 'S_6_' 'Sr' 'Barite' 'Celestite']", "output_type": "error", "traceback": [ "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m", "\u001b[0;31mIndexError\u001b[0m Traceback (most recent call last)", "Cell \u001b[0;32mIn[185], line 2\u001b[0m\n\u001b[1;32m 1\u001b[0m \u001b[38;5;66;03m# test on non-reactive data\u001b[39;00m\n\u001b[0;32m----> 2\u001b[0m model_large\u001b[38;5;241m.\u001b[39mevaluate(X_test[X_test[\u001b[38;5;124m'\u001b[39m\u001b[38;5;124mClass\u001b[39m\u001b[38;5;124m'\u001b[39m] \u001b[38;5;241m==\u001b[39m \u001b[38;5;241m0\u001b[39m]\u001b[38;5;241m.\u001b[39miloc[:,species_columns], y_test[X_test[\u001b[38;5;124m'\u001b[39m\u001b[38;5;124mClass\u001b[39m\u001b[38;5;124m'\u001b[39m] \u001b[38;5;241m==\u001b[39m \u001b[38;5;241m0\u001b[39m]\u001b[38;5;241m.\u001b[39miloc[:,species_columns])\n", "File \u001b[0;32m~/bin/miniconda3/envs/training/lib/python3.11/site-packages/pandas/core/indexing.py:1184\u001b[0m, in \u001b[0;36m_LocationIndexer.__getitem__\u001b[0;34m(self, key)\u001b[0m\n\u001b[1;32m 1182\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_is_scalar_access(key):\n\u001b[1;32m 1183\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mobj\u001b[38;5;241m.\u001b[39m_get_value(\u001b[38;5;241m*\u001b[39mkey, takeable\u001b[38;5;241m=\u001b[39m\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_takeable)\n\u001b[0;32m-> 1184\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_getitem_tuple(key)\n\u001b[1;32m 1185\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m 1186\u001b[0m \u001b[38;5;66;03m# we by definition only have the 0th axis\u001b[39;00m\n\u001b[1;32m 1187\u001b[0m axis \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39maxis \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;241m0\u001b[39m\n", "File \u001b[0;32m~/bin/miniconda3/envs/training/lib/python3.11/site-packages/pandas/core/indexing.py:1690\u001b[0m, in \u001b[0;36m_iLocIndexer._getitem_tuple\u001b[0;34m(self, tup)\u001b[0m\n\u001b[1;32m 1689\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21m_getitem_tuple\u001b[39m(\u001b[38;5;28mself\u001b[39m, tup: \u001b[38;5;28mtuple\u001b[39m):\n\u001b[0;32m-> 1690\u001b[0m tup \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_validate_tuple_indexer(tup)\n\u001b[1;32m 1691\u001b[0m \u001b[38;5;28;01mwith\u001b[39;00m suppress(IndexingError):\n\u001b[1;32m 1692\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_getitem_lowerdim(tup)\n", "File \u001b[0;32m~/bin/miniconda3/envs/training/lib/python3.11/site-packages/pandas/core/indexing.py:966\u001b[0m, in \u001b[0;36m_LocationIndexer._validate_tuple_indexer\u001b[0;34m(self, key)\u001b[0m\n\u001b[1;32m 964\u001b[0m \u001b[38;5;28;01mfor\u001b[39;00m i, k \u001b[38;5;129;01min\u001b[39;00m \u001b[38;5;28menumerate\u001b[39m(key):\n\u001b[1;32m 965\u001b[0m \u001b[38;5;28;01mtry\u001b[39;00m:\n\u001b[0;32m--> 966\u001b[0m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_validate_key(k, i)\n\u001b[1;32m 967\u001b[0m \u001b[38;5;28;01mexcept\u001b[39;00m \u001b[38;5;167;01mValueError\u001b[39;00m \u001b[38;5;28;01mas\u001b[39;00m err:\n\u001b[1;32m 968\u001b[0m \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mValueError\u001b[39;00m(\n\u001b[1;32m 969\u001b[0m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mLocation based indexing can only have \u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m 970\u001b[0m \u001b[38;5;124mf\u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m[\u001b[39m\u001b[38;5;132;01m{\u001b[39;00m\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_valid_types\u001b[38;5;132;01m}\u001b[39;00m\u001b[38;5;124m] types\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m 971\u001b[0m ) \u001b[38;5;28;01mfrom\u001b[39;00m \u001b[38;5;21;01merr\u001b[39;00m\n", "File \u001b[0;32m~/bin/miniconda3/envs/training/lib/python3.11/site-packages/pandas/core/indexing.py:1608\u001b[0m, in \u001b[0;36m_iLocIndexer._validate_key\u001b[0;34m(self, key, axis)\u001b[0m\n\u001b[1;32m 1606\u001b[0m \u001b[38;5;66;03m# check that the key has a numeric dtype\u001b[39;00m\n\u001b[1;32m 1607\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m is_numeric_dtype(arr\u001b[38;5;241m.\u001b[39mdtype):\n\u001b[0;32m-> 1608\u001b[0m \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mIndexError\u001b[39;00m(\u001b[38;5;124mf\u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m.iloc requires numeric indexers, got \u001b[39m\u001b[38;5;132;01m{\u001b[39;00marr\u001b[38;5;132;01m}\u001b[39;00m\u001b[38;5;124m\"\u001b[39m)\n\u001b[1;32m 1610\u001b[0m \u001b[38;5;66;03m# check that the key does not exceed the maximum size of the index\u001b[39;00m\n\u001b[1;32m 1611\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;28mlen\u001b[39m(arr) \u001b[38;5;129;01mand\u001b[39;00m (arr\u001b[38;5;241m.\u001b[39mmax() \u001b[38;5;241m>\u001b[39m\u001b[38;5;241m=\u001b[39m len_axis \u001b[38;5;129;01mor\u001b[39;00m arr\u001b[38;5;241m.\u001b[39mmin() \u001b[38;5;241m<\u001b[39m \u001b[38;5;241m-\u001b[39mlen_axis):\n", "\u001b[0;31mIndexError\u001b[0m: .iloc requires numeric indexers, got ['H' 'O' 'Charge' 'Ba' 'Cl' 'S_6_' 'Sr' 'Barite' 'Celestite']" ] } ], "source": [ "# test on non-reactive data\n", "model_large.evaluate(X_test[X_test['Class'] == 0].iloc[:,species_columns], y_test[X_test['Class'] == 0].iloc[:,species_columns])" ] }, { "cell_type": "code", "execution_count": 17, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "\u001b[1m94/94\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 4ms/step - loss: 4.0710e-05\n" ] } ], "source": [ "mass_balance = mass_balance(model_simple, X_test, scaler_X, func_dict_in, func_dict_out)" ] }, { "cell_type": "code", "execution_count": 116, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "4.047931361128576e-05" ] }, "execution_count": 17, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# test on reactive data\n", "model_large.evaluate(X_test[X_test['Class'] == 1].iloc[:,:-1], y_test[X_test['Class'] == 1].iloc[:, :-1])" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Save the model" ] }, { "cell_type": "code", "execution_count": 53, "metadata": {}, "outputs": [], "source": [ "# Save the model\n", "model.save(\"Barite_50_Model_additional_species.keras\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Legacy Code" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "def log_scale(df_design, df_result, func_dict):\n", " \n", " df_design = df_design.copy()\n", " df_result = df_result.copy()\n", " \n", " for key in df_design.keys():\n", " if key != \"Class\":\n", " df_design[key] = np.vectorize(func_dict[key])(df_design[key])\n", " df_result[key] = np.vectorize(func_dict[key])(df_result[key])\n", " \n", " return df_design, df_result\n", "\n", "# Get minimum and maximum values for each column\n", "def get_min_max(df_design, df_result):\n", " \n", " min_vals_des = df_design.min()\n", " max_vals_des = df_design.max()\n", " \n", " min_vals_res = df_result.min()\n", " max_vals_res = df_result.max()\n", "\n", " # minimum of input and output data to get global minimum/maximum\n", " data_min = np.minimum(min_vals_des, min_vals_res).to_dict()\n", " data_max = np.maximum(max_vals_des, max_vals_res).to_dict()\n", "\n", " return data_min, data_max\n", "\n", "df_design_log, df_results_log = log_scale(df_design, df_results, func_dict_in)\n", "data_min_log, data_max_log = get_min_max(df_design_log, df_design_log)\n", "\n", "train_min_log, train_max_log = get_min_max(X_train_log, y_train_log)\n", "test_min_log, test_max_log = get_min_max(X_test_log, y_test_log)\n", "\n", "X_train_preprocess = preprocess(X_train_log, func_dict_in, train_min_log, train_max_log)\n", "y_train_preprocess = preprocess(y_train_log, func_dict_in, train_min_log, train_max_log)\n", "\n", "X_test_preprocess = preprocess(X_test_log, func_dict_in, test_min_log, test_max_log)\n", "y_test_preprocess = preprocess(y_test_log, func_dict_in, test_min_log, test_max_log)\n", "\n", "X_train_log, y_train_log = log_scale(X_train, y_train, func_dict_in)\n", "X_test_log, y_test_log = log_scale(X_test, y_test, func_dict_in)\n", "\n", "\n", "def preprocess(data, func_dict, data_min, data_max):\n", " data = data.copy()\n", " for key in data.keys():\n", " if key != \"Class\":\n", " data[key] = (data[key] - data_min[key]) / (data_max[key] - data_min[key])\n", "\n", " return data\n", "\n", "def postprocess(data, func_dict, data_min, data_max):\n", " data = data.copy()\n", " for key in data.keys():\n", " if key != \"Class\":\n", " data[key] = data[key] * (data_max[key] - data_min[key]) + data_min[key]\n", " data[key] = np.vectorize(func_dict[key])(data[key])\n", " return data\n", "\n", "X_train, X_val, y_train, y_val = sk.train_test_split(X_train_preprocess, y_train_preprocess, test_size = 0.1)\n", "\n", "pp_design = preprocess(df_design_log, func_dict_in, data_min_log, data_max_log)\n", "pp_results = preprocess(df_results_log, func_dict_in, data_min_log, data_max_log)" ] } ], "metadata": { "kernelspec": { "display_name": "training", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.11.11" } }, "nbformat": 4, "nbformat_minor": 2 }