WeatherPrediction.cs

using System;
using System.Text;
using static Tensorflow.Binding;
using static Tensorflow.KerasApi;
using static PandasNet.PandasApi;
using System.IO;
using Tensorflow;
using SciSharp.Models;
using SciSharp.Models.TimeSeries;
using PandasNet;

namespace TensorFlowNET.Examples;

public class WeatherPrediction : SciSharpExample, IExample
{
    ITimeSeriesTask task;
    IDatasetV2 training_ds, val_ds, test_ds;
    public ExampleConfig InitConfig()
        => Config = new ExampleConfig
        {
            Name = "Weather Prediction",
            Enabled = true
        };

    public bool Run()
    {
        var wizard = new ModelWizard();
        // Change to RnnModel for RNN
        task = wizard.AddTimeSeriesTask<RnnModel>(new TaskOptions
        {
            WeightsPath = @"timeseries_linear_v1\saved_weights.h5"
        });
        task.SetModelArgs(new TimeSeriesModelArgs
        {
            InputWidth = 3,
            LabelWidth = 1,
            LabelColumns = new[] { "T (degC)" }
        });
        (training_ds, val_ds, test_ds) = task.GenerateDataset(PrepareData);

        Train();
        //Test();
        //Predict();

        return true;
    }

    public override void Train()
    {
        task.Train(new TrainingOptions
        {
            Epochs = 3,
            Dataset = (training_ds, val_ds)
        });
    }

    public override void Test()
    {
        var result = task.Test(new TestingOptions
        {
            Dataset = test_ds
        });
    }

    public override void Predict()
    {
        foreach (var (input, label) in test_ds.take(10))
        {
            var result = task.Predict(input);
        }
    }

    new DataFrame PrepareData()
    {
        var zip_path = keras.utils.get_file("jena_climate_2009_2016.csv.zip",
            "https://storage.googleapis.com/tensorflow/tf-keras-datasets/jena_climate_2009_2016.csv.zip",
            cache_subdir: "jena_climate_2009_2016",
            extract: true);

        var df = pd.read_csv(Path.Combine(zip_path, "jena_climate_2009_2016.csv"));
        // deal with hourly predictions, so start by sub-sampling the data from 10-minute intervals to one-hour intervals:
        df = df[new Slice(5, step: 6)];
        var date_time_string = df.pop("Date Time");
        var date_time = pd.to_datetime(date_time_string, "dd.MM.yyyy HH:mm:ss");
        print(df.head());

        // plot featuers
        /*var plot_cols = new string[] { "T (degC)", "p (mbar)", "rho (g/m**3)" };
        var plot_features = df[plot_cols];
        plot_features.index = date_time_string;
        plot_features.plot();*/

        print(df.describe().transpose());

        // Wind velocity
        var wv = df["wv (m/s)"];
        var bad_wv = wv == -9999.0f;
        wv[bad_wv] = 0.0f;

        var max_wv = df["max. wv (m/s)"];
        var bad_max_wv = max_wv == -9999.0f;
        max_wv[bad_max_wv] = 0.0f;

        // The above inplace edits are reflected in the DataFrame
        print(df["wv (m/s)"].min());

        // convert the wind direction and velocity columns to a wind vector
        wv = df.pop("wv (m/s)");
        max_wv = df.pop("max. wv (m/s)");

        // Convert to radians.
        var wd_rad = df.pop("wd (deg)") * pd.pi / 180;

        // Calculate the wind x and y components.
        df["Wx"] = wv * pd.cos(wd_rad);
        df["Wy"] = wv * pd.sin(wd_rad);

        // Calculate the max wind x and y components.
        df["max Wx"] = max_wv * pd.cos(wd_rad);
        df["max Wy"] = max_wv * pd.sin(wd_rad);

        var timestamp_s = date_time.map<DateTime, float>(pd.timestamp);

        var day = 24 * 60 * 60;
        var year = 365.2425f * day;
        df["Day sin"] = pd.sin(timestamp_s * (2 * pd.pi / day));
        df["Day cos"] = pd.cos(timestamp_s * (2 * pd.pi / day));
        df["Year sin"] = pd.sin(timestamp_s * (2 * pd.pi / year));
        df["Year cos"] = pd.cos(timestamp_s * (2 * pd.pi / year));

        return df;
    }
}