In this work, we consider the calibration of the Heston model for European put or call options on a single or a basket of financial assets. We use the high level environment Keras of Google in Python. The calibration is done directly with prices corresponding to an array of values for the strike and the maturity or the initial values of the asset and its volatility. The calibration is for the three parameters of the Heston model or the correlation between the asset and the stochastic volatility. It turns out to be a rather easy programming exercise but a large and computer-intensive generation of the synthetic data is necessary to calibrate the Neural Network. A simple network with one hidden layer seems to be appropriate, yet the precision stalls upon a problem dependent threshold beyond which it seems difficult to go.