4 ms·

Quoting: " If the model is trained for more than an epoch, the weights of the RNN Estimator Block also need to be stored, and would count in the compression siz

Quoting: " If the model is trained for more than an epoch, the weights of the RNN Estimator Block also need to be stored, and would count in the compression size (MDL Principle [14])"

YetAnotherNick 9y ago
After another read, they are only storing seed and not the weights, needing entire data to be decoded, and so the compressed format cannot be transferred without all the data. > Only random seed is Stored: The model is initialized using a stored random seed, which is also communicated to the decoder. Thus, the effectively the model weights do not contribute to the total compressed data size > Single Pass: As we do not store the weights of the model, we perform a single pass (1 epoch) through the data