machine_learning_learning

Running through a bunch of tutorials as found on

https://www.tensorflow.org/tutorials/keras/classification

Machine Learning Types

This is mainly taken from the time series tutorial as that is one of the more difficult tutorials in the list and goes through a LOT of different types.

Terms involved:

window: width of the input data.
label_width: width of the ouput data.
offset: how offset the label data is from the input

Single Step Models

These models take a single input and then make a prediction. For example, it may use today's data to predict tomorroww's values. IE: Input = t0, output = t1. To create an input for this, you would create window=1, label_width=1, and offset=1

The problem with this for time series data is that is has no historical data, only the previous reading.

Single Step Linear

This is the simplest type of model. The cool part is that you can see how the input values affect the output. The linear model is just a dense network with no activation set. Example:

    linear = tf.keras.Sequential([
        tf.keras.layers.Dense(units=1)
    ])

Multi-layer, Single Step Dense Model

Instead of a single linear layer, this will have multiple layers with activation functions that are not linear. The reason for non-linear activation function is outside the scope of this write up, but it allow for more complicated predictions than linear activation would allow

    dense = tf.keras.Sequential([
        tf.keras.layers.Dense(units=64, activation='relu'),
        tf.keras.layers.Dense(units=64, activation='relu'),
        tf.keras.layers.Dense(units=1)
    ])

Multi-Step, Multi-Step Dense Model

Look at that fancy, fancy name. Multi-step models means that instead of just a single time step, multiple steps are given to the model. This is inputs: t1, t2...tn are used to predict: tn+1. The input data should look like: window=$step_size, label_width=1, and offset=1.

In order to do this, we need to flatten the input data. The inputs data is columns x step_size, and this will flatten that data into a single list.

    multi_step_dense = tf.keras.Sequential([
        # Shape: (time, features) => (time*features)
        tf.keras.layers.Flatten(),
        tf.keras.layers.Dense(units=32, activation='relu'),
        tf.keras.layers.Dense(units=32, activation='relu'),
        tf.keras.layers.Dense(units=1),
        # Add back the time dimension.
        # Shape: (outputs) => (1, outputs)
        tf.keras.layers.Reshape([1, -1]),
    ])

Convolution Neural Network

This is basically the previous model but in a better way. This model is the same as the above.

    conv_model = tf.keras.Sequential([
        tf.keras.layers.Conv1D(filters=32,
                               kernel_size=(CONV_WIDTH,),
                               activation='relu'),
        tf.keras.layers.Dense(units=32, activation='relu'),
        tf.keras.layers.Dense(units=1),
    ])

Recurrent Neural Network

Instead of manually feeding the previous data in, you can have the previous data automatically retained.

BOOM! Done. Easy.

lstm_model = tf.keras.models.Sequential([
    # Shape [batch, time, features] => [batch, time, lstm_units]
    tf.keras.layers.LSTM(32, return_sequences=True),
    # Shape => [batch, time, features]
    tf.keras.layers.Dense(units=1)
])

Multi-Output Models

Instead of just outputting one value - such as temperature - the models can output many outputs. To do this, you just need to have multiple labels. I.... I don't really know how to do this, because I've only done a single output prediction.

    single_step_window = WindowGenerator(
        # `WindowGenerator` returns all features as labels if you 
        # don't set the `label_columns` argument.
        input_width=1, label_width=1, shift=1)

    wide_window = WindowGenerator(
        input_width=24, label_width=24, shift=1)

    for example_inputs, example_labels in wide_window.train.take(1):
      print(f'Inputs shape (batch, time, features): {example_inputs.shape}')
      print(f'Labels shape (batch, time, features): {example_labels.shape}')

Multi-step Predictions

Instead of making just one prediction, we could make a whole bunch of predictions. Kind of like running weather a model and forecasting into the future. I also don't really know how to do this. But we'll leave this here to remind it's possible later

acherrera / machine_learning_learning