Data leakage in Next-Frame Video Prediction with Convolutional LSTMs

[lugi7]

trafficstars

In the Keras tutorial "Next-Frame Video Prediction with Convolutional LSTMs" (https://keras.io/examples/vision/conv_lstm/) I've noticed that there seems to be a data leakage problem. Namely, the 3D convolution has direct access to the frame it is trying to predict, so the results are unreasonably good. The correct way to implement this would be to use causal padding, which would inhibit the model's ability to "cheat" by not providing him directly with the value that it should predict. Let me know if I haven't missed something.