Predicting Eye Open/Closed State Using Keras LSTM

eye_eeg.py

import pandas as pd
from keras import Sequential
from keras.layers import LSTM, Dense, Dropout, Activation
import numpy as np
from matplotlib import pyplot


def get_data():
    data = pd.read_csv('eeg_data.csv')
    return data[data.apply(lambda x: np.abs(x - x.median()) / x.std() < 4).all(axis=1)]


def show(df: pd.DataFrame):
    # create a subplot for each time series
    fig, plots = pyplot.subplots(len(df.columns), squeeze=False)
    for i, name in enumerate(df.columns):
        subplot = plots[i][0]
        subplot.set_title(name, loc='right', x=1.05)
        subplot.plot(df[name])
    fig.subplots_adjust(bottom=0.2)
    pyplot.show()


def do_model(data):
    batch_size = 1
    training_split = 0.25
    num_fields = 14
    y_col = 'eyeDetection'
    num_rows = len(data)
    xs = data.drop(y_col, axis=1).values.astype(np.float32).reshape(num_rows, 1, num_fields)
    ys = data[y_col].values.astype(np.float32)
    model = Sequential()
    model.add(LSTM(128,
                   stateful=True,
                   return_sequences=True,
                   batch_input_shape=(batch_size, 1, num_fields)))
    model.add(Dropout(0.1))
    model.add(LSTM(128, stateful=True))
    model.add(Dropout(.1))
    model.add(Dense(1, activation='sigmoid' ))
    model.compile(loss='mse', optimizer='rmsprop', metrics=['accuracy'])
    model.fit(xs, ys, batch_size=batch_size, validation_split=training_split, epochs=5, shuffle=False)
    data['Predicted EyeState'] = model.predict(xs, batch_size=batch_size)


if __name__ == "__main__":
    filtered = get_data()
    do_model(filtered)
    show(filtered)

What's the window size in your model?

I think the window size is 1, since batch shape is (1,1, 14). Ie each batch is 1 point in time with 14 fields.