Method, an apparatus and a computer program product for neural network compression
The embodiments relate to a method for encoding two or more tensors. The method comprises processing the two or more tensors having respective dimensions so that the dimensions of said two or more sensors have the same number ( 510 ); identifying which axis of each individual tensor is swappable to result in concatenable tensors around an axis of concatenation ( 520 ); reshaping the tensors so that the dimensions are modified based on the swapped axis ( 530 ); concatenating the tensors around the axis of concatenation to result in concatenated tensor ( 540 ); compressing the concatenated tensor ( 550 ); generating syntax structures for carrying concatenation and axis swapping information ( 560 ); and generating a bitstream by combining the syntax structures and the compressed concatenated tensor ( 570 ). The embodiments also relate to a method for decoding, and to apparatuses for implementing the methods.
1. An apparatus for encoding two or more tensors, the apparatus comprising at least one processor, memory including computer program code, the memory and the computer program code configured to, with the at least one processor, cause the apparatus to perform at least the following:
processing the two or more tensors having respective dimensions so that the dimensions of said two or more tensors have the same number;
identifying which axis of each individual tensor is swappable to result in concatenable tensors around an axis of concatenation;
reshaping the tensors so that the dimensions are modified based on the swapped axis;
concatenating the tensors around the axis of concatenation to result in a concatenated tensor;
compressing the concatenated tensor;
generating syntax structures for carrying concatenation and axis swapping information; and
generating a bitstream by combining the syntax structures and the compressed concatenated tensor.
2. The apparatus according to claim 1 , wherein the apparatus is further configured to perform combining or flattening dimensions of a tensor so that said two or more tensors have the same number of dimensions.
3. The apparatus according to claim 1 , wherein the bitstream is a compressed neural network bitstream.
4. The apparatus according to claim 1 , wherein the apparatus is further caused to perform: signaling swapped dimension indexes in a syntax element present in a compressed data unit header.
5. The apparatus according to claim 1 , wherein the apparatus is further caused to perform: signaling a dimension index swapping difference in a syntax element present in a compressed data unit header.
6. The apparatus according to claim 5 , wherein the dimension index swapping difference comprises non-zero indexes.
7. An apparatus for decoding, the apparatus comprising at least one processor, memory including computer program code, the memory and the computer program code configured to, with the at least one processor, cause the apparatus to perform at least the following:
receiving a bitstream comprising a bitstream of a compressed tensor;
processing the bitstream and identify from syntax structures that the bitstream comprises a compressed concatenated tensor;
identifying dimensions of individual tensors generating the concatenated tensor;
identifying from the bitstream an axis swapping information indicating whether axis swapping has been applied;
decompressing the tensor into a decompressed tensor;
splitting the decompressed tensor into individual tensors based on the identified dimensions of the individual tensors;
swapping axis of the individual tensors based on the axis swapping information; and
decomposing the individual tensors so that their final dimensions match with the identified dimensions of the individual tensors.
8. The apparatus according to claim 7 , wherein the bitstream is a compressed neural network bitstream.
9. The apparatus according to claim 7 , wherein the apparatus is further caused to perform: determining swapped dimension indexes from a syntax element present in a compressed data unit header.
10. The apparatus according to claim 7 , wherein the apparatus is further caused to perform: determining a dimension index swapping difference from a syntax element present in a compressed data unit header.
11. The apparatus according to claim 10 , wherein the dimension index swapping difference comprises non-zero indexes.
12. A method, comprising:
processing two or more tensors having respective dimensions so that the dimensions of said two or more tensors have the same number;
identifying which axis of each individual tensor is swappable to result in concatenable tensors around an axis of concatenation;
reshaping the tensors so that the dimensions are modified based on the swapped axis;
concatenating the tensors around the axis of concatenation to result in a concatenated tensor;
compressing the concatenated tensor;
generating syntax structures for carrying concatenation and axis swapping information; and
generating a bitstream by combining the syntax structures and the compressed concatenated tensor.
13. The method according to claim 12 , wherein the processing comprises combining or flattening dimensions of a tensor so that said two or more tensors have the same number of dimensions.
14. The method according to claim 12 , wherein the bitstream is a compressed neural network bitstream.
15. The method according to claim 12 , further comprising: signaling swapped dimension indexes in a syntax element present in a compressed data unit header.
16. The method according to claim 12 , further comprising: signaling a dimension index swapping difference in a syntax element present in a compressed data unit header.
17. The method according to claim 16 , wherein the dimension index swapping difference comprises non-zero indexes.
18. A method, comprising:
receiving a bitstream comprising a bitstream of a compressed tensor;
processing the bitstream and identifying from syntax structures that the bitstream comprises a compressed concatenated tensor;
identifying dimensions of individual tensors generating the concatenated tensor;
identifying from the bitstream an axis swapping information indicating whether axis swapping has been applied;
decompressing the tensor into a decompressed tensor;
splitting the decompressed tensor into individual tensors based on the identified dimensions of the individual tensors;
swapping axis of the individual tensors based on the axis swapping information; and
decomposing the individual tensors so that their final dimensions match with the identified dimensions of the individual tensors.
19. The method according to claim 18 , further comprising: determining swapped dimension indexes from a syntax element present in a compressed data unit header.
20. The method according to claim 18 , further comprising: determining a dimension index swapping difference from a syntax element present in a compressed data unit header.