Network based image filtering for video coding
A method and an apparatus for hybrid training of neural networks for video coding are provided. The method includes: obtaining, in an offline training stage, an offline trained network by training a neural network offline; and refining, in an online training stage, a plurality of neural network layers with constraint on a plurality of parameters of the plurality of neural network layers, where the plurality of neural network layers may include at least one neural network layer in the offline trained network or in a simple neural network connected to the offline trained network.
1 . A method for video coding, comprising:
obtaining, in an offline training stage, an offline trained network by training a neural network offline;
refining, in an online training stage, a plurality of neural network layers with constraint on a plurality of parameters of the plurality of neural network layers, wherein the plurality of neural network layers comprise at least one neural network layer in the offline trained network or in a simple neural network connected to the offline trained network; and
performing video coding using the plurality of refined neural network layers,
wherein refining the plurality of neural network layers with the constraint on the plurality of parameters comprises:
classifying the plurality of parameters into different categories and assigning the different categories into different bands; and
updating one or more parameters in the different bands differently.
2 . The method of claim 1 , wherein classifying the plurality of parameters into different categories comprises:
classifying the plurality of parameters based on one of followings: intensity values of the plurality of parameters, or roles of neurons.
3 . The method of claim 1 , wherein the different bands comprise a first band and a second band, one or more parameters in the first band are updated using a first dynamic range and one or more parameters in the second band are updated using a second dynamic range, and the first dynamic range is greater than the second dynamic range.
4 . The method of claim 1 , further comprising:
skipping updating one or more parameters in one or more bands.
5 . The method of claim 1 , wherein updating one or more parameters in the different bands differently comprises:
dividing a band into a first sub-band and a second sub-band; and
updating one or more parameters in the first sub-band without updating one or more parameters in the second sub-band, wherein the one or more parameters in the first sub-band have higher values than the one or more parameters in the second sub-band.
6 . The method of claim 1 , further comprising:
applying post-processing based constraint on the plurality of parameters after the online training stage.
7 . The method of claim 6 , wherein applying the post-processing based constraint on the plurality of parameters after the online training stage comprises:
approximating an update difference to a predefined value, wherein the update difference is a difference between a first value of at least one parameter of the at least one neural network layer in the offline trained network before the online training and a second value of the at least one parameter after the online training; and
updating the at least one parameter according to the updated difference.
8 . The method of claim 7 , wherein approximating the update difference to a predefined value comprises:
approximating the update difference by quantizing the update difference.
9 . The method of claim 7 , wherein approximating the update difference to a predefined value comprises:
in response to determining that the update difference is smaller than a predefined threshold, determining that the update difference is 0.
10 . The method of claim 6 , wherein applying the post-processing based constraint on the plurality of parameters after the online training stage comprises:
approximating a parameter of the at least one neural network layer in the simple neural network connected to the offline trained network to a predefined value.
11 . The method of claim 10 , wherein approximating the parameter of the at least one neural network layer in the simple neural network connected to the offline trained network to a predefined value comprises:
approximating the parameter by quantizing the parameter.
12 . The method of claim 10 , wherein approximating the parameter of the at least one neural network layer in the simple neural network connected to the offline trained network to a predefined value comprises:
in response to determining that a value of the parameter is smaller than a predefined threshold, determining that the value of the parameter is 0.
13 . An apparatus for video coding, comprising:
one or more processors; and
a memory configured to store instructions executable by the one or more processors,
wherein the one or more processors, upon execution of the instructions, are configured to perform a method comprising:
obtaining, in an offline training stage, an offline trained network by training a neural network offline;
refining, in an online training stage, a plurality of neural network layers with constraint on a plurality of parameters of the plurality of neural network layers, wherein the plurality of neural network layers comprise at least one neural network layer in the offline trained network or in a simple neural network connected to the offline trained network; and
performing video coding using the plurality of refined neural network layers,
wherein refining the plurality of neural network layers with the constraint on the plurality of parameters comprises:
classifying the plurality of parameters into different categories and assigning the different categories into different bands; and
updating one or more parameters in the different bands differently.
14 . A non-transitory computer-readable storage medium storing a bitstream generated or to be decoded by using a method comprising:
obtaining, in an offline training stage, an offline trained network by training a neural network offline;
refining, in an online training stage, a plurality of neural network layers with constraint on a plurality of parameters of the plurality of neural network layers, wherein the plurality of neural network layers comprise at least one neural network layer in the offline trained network or in a simple neural network connected to the offline trained network; and
performing video coding using the plurality of refined neural network layers,
wherein refining the plurality of neural network layers with the constraint on the plurality of parameters comprises:
classifying the plurality of parameters into different categories and assigning the different categories into different bands; and
updating one or more parameters in the different bands differently.