Model Architecture related question

[srikar2097]

Model Architecture related question

[maodouy]

There is a quantitative training part in the code, but I have not seen the part of the entropy estimate. Therefore, the code cannot control the rate, and only one network model corresponds to a bit rate. For your first question: the autoencoded's encoder outputs the compressed feature, which is compression. Entropy coding is lossless compression. The second problem: the calculation of the code rate is the output of the encoder except the source image. As far as I know, there are currently two rate control methods: one is the RD curve and the other is the iteration. As far as I know, there are currently two rate control methods: one is the RD curve (the difficulty is the rate estimate, which is the estimate of the entropy. You need to add this part to the loss function of the control), and the other is the iteration (using the RNN). It's easier than using CNN).

[alexandru-dinu]

Moving discussion to #15.