System and method for compression of data stream
Procedures, methods, architectures, apparatuses, systems, devices, and computer program products, according to a first aspect, for compressing data including encoding at least one information representative of a use, during the compression, of a compressed sparse format. Procedures, methods, architectures, apparatuses, systems, devices, and computer program products, according to a second aspect, for decompressing input data comprising obtaining information representative of zero or non-zero values in at least a part of the input data, and using only the non-zero values of the zero or non-zero values for a further processing of the part of the input data, based on the representative information.
1 . A method for decoding a bitstream comprising, coded data representative of a weight matrix of at least one Deep Neural Network (DNN), wherein a not compressed sparse representation of the weight matrix is coded with Context-Adaptive Binary Adaptive Coding (CABAC), the method comprising:
decoding a respective significance flag (SigFlag) for a plurality of coefficients of the weight matrix in the not compressed sparse representation, wherein the respective SigFlags for each of the plurality of coefficients of the weight matrix are included in a significance map, wherein each respective SigFlags is decoded by parsing the significance map, and wherein the significance map is decoded separately from one or more bins representing weight values in the bitstream;
determining whether the respective SigFlag for each of the plurality of coefficients is a zero value or a non-zero value; and
storing a compressed sparse representation of the weight matrix by determining a first part of the compressed sparse representation of the weight matrix based on the respective SigFlags for each of the plurality of coefficients, wherein the first part of the compressed sparse representation comprises numbers of consecutive zero-coefficients preceding a non-zero coefficient.
2 . The method of claim 1 , wherein the compressed sparse representation of the weight matrix is determined without using coefficients of the plurality of coefficients that have been determined to have a SigFlag with a zero value.
3 . The method of claim 1 , wherein the compressed sparse representation of the weight matrix is determined without loading a full tensor associated with the bitstream.
4 . The method of claim 1 , wherein determining the compressed sparse representation of the weight matrix comprises building a sparse tensor having a compressed sparse format based on information obtained for weights of the at least one layer of at least one DNN.
5 . The method of claim 1 , wherein the compressed sparse representation is a Compressed Sparse Row (CSR).
6 . The method of claim 1 , further comprising receiving the bitstream comprising the coded data.
7 . The method of claim 6 , wherein the compressed sparse representation is updated as coefficients of the weight matrix of the received bitstream are decoded.
8 . The method of claim 1 , further comprising decoding a layer parameter set prior to decoding the respective SigFlags.
9 . The method of claim 1 , wherein the first part of the compressed sparse representation is stored in a first vector and a second part of the compressed sparse representation is stored in a second vector, the second part of the compressed sparse representation comprising non-zero coefficients, and wherein the first vector and second vector have equal lengths.
10 . A decoding device for decoding a bitstream comprising coded data representative of a weight matrix of at least one Deep Neural Network (DNN), wherein a not compressed sparse representation of the weight matrix is coded with Context-Adaptive Binary Adaptive Coding (CABAC), the decoding device comprising:
a processor configured to:
decode one or more respective significance flag (SigFlag) for a plurality of coefficients of the weight matrix, wherein the respective SigFlags for each of the plurality of coefficients of the weight matrix are included in a significance map, wherein the respective SigFlags is decoded by parsing the significance map, and wherein the significance map is decoded separately from one or more bins representing weight values in the bitstream;
determine whether the respective SigFlag for each of the plurality of coefficients is a zero value or a non-zero value; and
store a compressed sparse representation of the weight matrix by determining a first part of the compressed sparse representation of the weight matrix based on the respective SigFlags for each of the plurality of coefficients, wherein the first part of the compressed sparse representation comprises numbers of consecutive zero-coefficients preceding a non-zero coefficient.
11 . The decoding device of claim 10 , wherein the compressed sparse representation of the weight matrix is determined without using coefficients of the plurality of coefficients that have been determined to have a SigFlag with a zero value.
12 . The decoding device of claim 10 , wherein the compressed sparse representation of the weight matrix is determined without loading a full tensor associated with the bitstream.
13 . The decoding device of claim 10 , wherein being configured to determine the compressed sparse representation of the weight matrix comprises being configured to build a sparse tensor having a compressed sparse format based on information obtained for weights of the at least one layer of at least one Deep Neural Network.
14 . The decoding device of claim 10 , wherein the compressed sparse representation is a Compressed Sparse Row (CSR).
15 . The decoding device of claim 10 , wherein the processor is further configured to receive the bitstream that comprises the coded data.
16 . The decoding device of claim 15 , wherein the compressed representation is updated as coefficients of the weight matrix of the received bitstream are decoded.
17 . The decoding device of claim 10 , wherein the processor is further configured to decode a layer parameter set prior to decoding the respective SigFlags.
18 . The decoding device of claim 10 , wherein the first part of the compressed sparse representation is stored in a first vector and a second part of the compressed sparse representation is stored in a second vector, the second part of the compressed sparse representation comprising non-zero coefficients, and wherein the first vector and second vector have equal lengths.