A structure preserving feature encoding method and system based on double-stream cepstrum

Abstract

The application discloses a structure-preserving feature encoding method and system based on double-flow pressure expansion, and the method comprises the following steps: performing structure-preserving preprocessing on original intermediate features of a basic large model; decomposing the preprocessed features into a bounded main flow component and a complementary residual flow component; applying a differentiable nonlinear pressure expansion function to the main flow component and the residual flow component respectively, mapping them to a bounded interval, and splicing the two components after pressure expansion in the channel dimension to obtain encoder input features; quantizing and entropy encoding the encoder input features by using a deep learning encoder to generate a bit stream, and decoding the bit stream to obtain reconstructed pressure expansion features; and segmenting the reconstructed pressure expansion features, and recovering the reconstructed intermediate features through an inverse pressure expansion function and an inverse flow merging operation; the application solves the encoding difficulty caused by the long-tail distribution of large model features, and realizes efficient end-to-end feature compression without damaging the feature structure.

Description

Technical Field

[0001] This invention belongs to the field of data compression and artificial intelligence interdisciplinary technology, and in particular relates to a structure-preserving feature encoding method and system based on dual-stream companding. Background Technology

[0002] In recent years, large models have demonstrated unprecedented capabilities in natural language processing and computer vision. However, with the continuous expansion of model size, computational costs, storage overhead, and privacy issues have become increasingly prominent. Distributed deployment, as a solution, requires efficient transmission of intermediate features between different devices or servers, making feature encoding a critical performance bottleneck. Unlike traditional image / video encoding, the intermediate features of large models exhibit extremely high heterogeneity and typically display a "spike-and-long-tail" statistical distribution. For example, in large language models and visual Transformers, activation values ​​are usually concentrated near zero, but there are sparse and extremely large outliers, which often carry crucial semantic information. Existing feature encoding methods typically rely on manually designed preprocessing procedures, such as truncation, uniform quantization, or rearranging features into two-dimensional "pseudo-images" to adapt to traditional image codecs. However, simple truncation loses crucial long-tail information, leading to performance degradation in downstream tasks; while rearranging features into pseudo-images introduces artificial boundaries and high-frequency artifacts, disrupting the local structure of the features. Furthermore, existing learning codecs are highly sensitive to input distribution, and directly processing long-tailed distribution features can lead to unstable entropy model training and low encoding efficiency. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a structure-preserving feature encoding method and system based on dual-stream companding that can improve feature compression efficiency.

[0004] Technical solution: The present invention provides a structure-preserving feature encoding method based on dual-stream companding, comprising:

[0005] Step 1: Perform structure-preserving preprocessing on the original intermediate features of the basic large model to maintain the original tensor structure of the features and obtain the preprocessed features.

[0006] Step 2: Calculate the channel-level centering parameters of the preprocessed features, and decompose the preprocessed features into bounded mainstream components and complementary residual flow components.

[0007] Step 3: Apply differentiable nonlinear companding functions to the main stream component and the residual stream component respectively, map them to bounded intervals, and then concatenate the two companded components in the channel dimension to obtain the encoder input features;

[0008] Step 4: Quantize and entropy encode the encoder input features using a deep learning encoder to generate a bitstream, and decode the reconstructed companded features from the bitstream;

[0009] Step 5: Segment the reconstructed companding features and recover the reconstructed intermediate features through inverse companding function and inverse flow merging operation.

[0010] Further, step 1 includes:

[0011] For the sequence features of commonsense reasoning tasks, maintain as or In this form, without two-dimensional image rearrangement, where, For the number of tokens, Number of channels;

[0012] For multi-cropping features in semantic segmentation tasks, each cropping is treated as an independent single-channel sample, without padding or splicing.

[0013] For multi-channel features in text-based image tasks, preserve the original... Tensor layout, in which Indicates the feature map height. Indicates the width of the feature map.

[0014] Further, step 2 includes: firstly, centering each channel of the preprocessed features:

[0015] ;

[0016] in, For channel Learnable centralized bias, For the channels of preprocessed features ;

[0017] Features of centralized processing Decomposed into mainstream components and residual flow components :

[0018] ;

[0019] ;

[0020] in, and For learnable asymmetric truncation thresholds, residual flow components It is non-zero only when the feature value exceeds the threshold range.

[0021] Furthermore, step 3 includes: based on the hyperbolic tangent function Introducing a learnable scaling factor and ; Calculate the mainstream component after companding and the residual flow component after compression and diffusion :

[0022] ;

[0023] ;

[0024] in, As the mainstream component, For residual flow components;

[0025] Will and Concatenating along the channel dimension yields the input range within... The encoder input features between.

[0026] Furthermore, step 5 includes: segmenting the reconstructed compressive features into reconstructed mainstream compressive components in the channel dimension. and the reconstructed residual companding component Using the inverse hyperbolic tangent function Perform the inverse mapping and multiply by the corresponding scaling factor to obtain the recovered mainstream components. and the recovered residual flow component ; Calculate the intermediate features of the reconstruction :

[0027] ;

[0028] in, For channel Learnable central bias.

[0029] Based on the same inventive concept, the present invention also provides a structure-preserving feature encoding system based on dual-stream companding, comprising:

[0030] The structure-preserving preprocessing module is used to perform structure-preserving preprocessing on the original intermediate features of a large model, preserving the original tensor structure of the features.

[0031] The decomposition module is used to calculate the channel-level centralization parameters of the features, decomposing the preprocessed features into bounded mainstream components and complementary residual flow components.

[0032] The companding module is used to apply differentiable nonlinear companding functions to the main stream component and the residual stream component respectively, map them to a bounded interval, and splice the two companded components in the channel dimension to obtain the encoder input features.

[0033] The neural decoding module is used to quantize and entropy encode the encoder input features using a neural feature codec to generate a bitstream, and decode the reconstructed companded features from the bitstream.

[0034] The inverse companding reconstruction module is used to segment the reconstructed companding features and recover the intermediate reconstructed features through the inverse companding function and the inverse flow merging operation.

[0035] Furthermore, the structure preservation preprocessing module includes:

[0036] For the sequence features of commonsense reasoning tasks, maintain as or In this form, without two-dimensional image rearrangement, where, For the number of tokens, Number of channels;

[0037] For multi-cropping features in semantic segmentation tasks, each cropping is treated as an independent single-channel sample, without padding or splicing.

[0038] For multi-channel features in text-based image tasks, preserve the original... Tensor layout, in which Indicates the feature map height. Indicates the width of the feature map.

[0039] Furthermore, the decomposition module includes: firstly, centering each channel of the preprocessed features:

[0040] ;

[0041] in, For channel Learnable centralized bias, For the channels of preprocessed features ;

[0042] Features of centralized processing Decomposed into mainstream components and residual flow components :

[0043] ;

[0044] ;

[0045] in, and For learnable asymmetric truncation thresholds, residual flow components It is non-zero only when the feature value exceeds the threshold range.

[0046] Furthermore, the companding module includes: based on the hyperbolic tangent function Introducing a learnable scaling factor and ; Calculate the mainstream component after companding and the residual flow component after compression and diffusion :

[0047] ;

[0048] ;

[0049] in, As the mainstream component, For residual flow components;

[0050] Will and Concatenating along the channel dimension yields the input range within... The encoder input features between.

[0051] Furthermore, the inverse companding reconstruction module includes: segmenting the reconstructed companding features into reconstructed mainstream companding components in the channel dimension. and the reconstructed residual companding component Using the inverse hyperbolic tangent function Perform the inverse mapping and multiply by the corresponding scaling factor to obtain the recovered mainstream components. and the recovered residual flow component ; Calculate the intermediate features of the reconstruction :

[0052] ;

[0053] in, For channel Learnable central bias.

[0054] Beneficial effects: Compared with existing technologies, this invention directly processes the original tensor through a structure preservation strategy, avoiding structural damage caused by pseudo-image rearrangement. On the other hand, a dual-stream companding module is designed to decompose features into bounded mainstream and sparse residual streams. The mainstream captures the main distribution of features, while the residual stream specifically handles long-tailed outliers. By performing differentiable nonlinear mappings (such as Tanh) on the two streams respectively, high dynamic range features are mapped to a finite interval friendly to the codec. This invention successfully solves the problem of the difficulty in balancing "typical values" and "outliers" in feature encoding of large models. Through the dual-stream mechanism, not only are sparse activation values ​​that are crucial to downstream tasks effectively preserved, but the data distribution input to the entropy model is also made more stable, improving compression efficiency. Attached Figure Description

[0055] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0057] Example 1 presents a structure-preserving feature encoding method based on dual-stream companding (DSC-SPFC). Its core architecture integrates a lightweight, learnable, and reversible dual-stream companding front-end with a modern neural image codec (such as Hyperprior or ELIC). The system directly processes the original feature tensor, mapping it to a bounded representation through dual-stream companding, and then processes it via a neural codec.

[0058] like Figure 1 As shown, the multi-region fault handling method based on shared control resources in this embodiment includes:

[0059] Step 1: Perform structure-preserving preprocessing on the original intermediate features of the basic large model to maintain the original tensor structure of the features and obtain the preprocessed features.

[0060] Step 2: Calculate the channel-level centering parameters of the preprocessed features, and decompose the preprocessed features into bounded mainstream components and complementary residual flow components.

[0061] Step 3: Apply differentiable nonlinear companding functions to the main stream component and the residual stream component respectively, map them to bounded intervals, and then concatenate the two companded components in the channel dimension to obtain the encoder input features;

[0062] Step 4: Quantize and entropy encode the encoder input features using a deep learning encoder to generate a bitstream, and decode the reconstructed companded features from the bitstream;

[0063] Step 5: Segment the reconstructed companding features and recover the reconstructed intermediate features through inverse companding function and inverse flow merging operation.

[0064] Specifically, in step 1, the raw intermediate features generated by the large model... Preprocessing is performed. Unlike traditional methods that reshape features into square images, this invention preserves the natural tensor structure of the features: for commonsense reasoning (CSR) tasks, the features are preserved as... In this format, no padding or rearrangement is performed; for semantic segmentation (Seg) tasks, each crop is treated as an independent single-channel sample. To avoid artificial boundaries; for text-based graph (TTI) tasks, to preserve the multi-channel tensor structure. And apply subsequent companding parameters independently to each channel.

[0065] In step 2, a dual-stream companding module is constructed to first centralize the preprocessed input features: ,in For channel The learnable centering bias is then utilized. Next, a learnable asymmetric truncation threshold is employed. and The features are decomposed into mainstream and residual flow: mainstream component This component is limited to a threshold range and contains most of the data energy; residual flow component This component is non-zero only when the feature value exceeds the threshold (i.e., long tail / outlier value), and has high sparsity.

[0066] In step 3, to adapt to the input range typically assumed by neural codecs, differentiable methods are applied to the two decomposed streams. Compressive expansion: ,as well as ,in and It is a learnable scaling factor. After companding, both signals are mapped to... Within the interval. Finally, and The input features are concatenated along the channel dimension to generate encoder input features with twice the number of channels as the original. .

[0067] In step 4, The input is fed into a deep learning encoder (such as a Hyperprior model based on variational autoencoders or an ELIC model). The encoder-decoder transforms, quantizes, and arithmetic-encodes the features to generate a compressed bitstream, and recovers the reconstructed compressed features at the decoding end. .

[0068] In step 5, at the decoding end, Segmented into reconstructed mainstream companding components and reconstruction residual compressive expansion component First, perform the inverse hyperbolic tangent transform. To prevent numerical overflow, a truncation operation is performed first: , Then the mainstream content was restored: And recovering the residual flow components: Finally, the two streams are merged and the mean is added back to obtain the final reconstructed features: .

[0069] Based on the same inventive concept, this embodiment also provides a structure-preserving feature encoding system based on dual-stream companding, comprising:

[0070] The structure-preserving preprocessing module is used to perform structure-preserving preprocessing on the original intermediate features of a large model, preserving the original tensor structure of the features.

[0071] The decomposition module is used to calculate the channel-level centralization parameters of the features, decomposing the preprocessed features into bounded mainstream components and complementary residual flow components.

[0072] The companding module is used to apply differentiable nonlinear companding functions to the main stream component and the residual stream component respectively, map them to a bounded interval, and splice the two companded components in the channel dimension to obtain the encoder input features.

[0073] The neural decoding module is used to quantize and entropy encode the encoder input features using a neural feature codec to generate a bitstream, and decode the reconstructed companded features from the bitstream.

[0074] The inverse companding reconstruction module is used to segment the reconstructed companding features and recover the intermediate reconstructed features through the inverse companding function and the inverse flow merging operation.

[0075] Furthermore, the structure preservation preprocessing module includes:

[0076] For the sequence features of commonsense reasoning tasks, maintain as or In this form, without two-dimensional image rearrangement, where, For the number of tokens, Number of channels;

[0077] For multi-cropping features in semantic segmentation tasks, each cropping is treated as an independent single-channel sample, without padding or splicing.

[0078] For multi-channel features in text-based image tasks, preserve the original... Tensor layout, in which Indicates the feature map height. Indicates the width of the feature map.

[0079] Furthermore, the decomposition module includes: firstly, centering each channel of the preprocessed features:

[0080] ;

[0081] in, For channel Learnable centralized bias, For the channels of preprocessed features ;

[0082] Features of centralized processing Decomposed into mainstream components and residual flow components :

[0083] ;

[0084] ;

[0085] in, and For learnable asymmetric truncation thresholds, residual flow components It is non-zero only when the feature value exceeds the threshold range.

[0086] Furthermore, the companding module includes: based on the hyperbolic tangent function Introducing a learnable scaling factor and ; Calculate the mainstream component after companding and the residual flow component after compression and diffusion :

[0087] ;

[0088] ;

[0089] in, As the mainstream component, For residual flow components;

[0090] Will and Concatenating along the channel dimension yields the input range within... The encoder input features between.

[0091] Furthermore, the inverse companding reconstruction module includes: segmenting the reconstructed companding features into reconstructed mainstream companding components in the channel dimension. and the reconstructed residual companding component Using the inverse hyperbolic tangent function Perform the inverse mapping and multiply by the corresponding scaling factor to obtain the recovered mainstream components. and the recovered residual flow component ; Calculate the intermediate features of the reconstruction :

[0092] ;

[0093] in, For channel Learnable central bias.

[0094] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Therefore, any modifications made to the technical solution of the present invention based on the technical concept proposed in this invention shall fall within the scope of protection of the present invention.

Claims

1. A structure-preserving feature encoding method based on dual-stream companding, characterized in that, include: Step 1: Perform structure-preserving preprocessing on the original intermediate features of the basic large model to maintain the original tensor structure of the features and obtain the preprocessed features. Step 2: Calculate the channel-level centering parameters of the preprocessed features, and decompose the preprocessed features into bounded mainstream components and complementary residual flow components. Step 3: Apply differentiable nonlinear companding functions to the main stream component and the residual stream component respectively, map them to bounded intervals, and then concatenate the two companded components in the channel dimension to obtain the encoder input features; Step 4: Quantize and entropy encode the encoder input features using a deep learning encoder to generate a bitstream, and decode the reconstructed companded features from the bitstream; Step 5: Segment the reconstructed companding features and recover the reconstructed intermediate features through inverse companding function and inverse flow merging operation.

2. The structure-preserving feature encoding method based on dual-stream companding according to claim 1, characterized in that, Step 1 includes: For the sequence features of commonsense reasoning tasks, maintain as or In this form, without two-dimensional image rearrangement, where, For the number of tokens, Number of channels; For multi-cropping features in semantic segmentation tasks, each cropping is treated as an independent single-channel sample, without padding or splicing. For multi-channel features in text-based image tasks, preserve the original... Tensor layout, in which Indicates the feature map height. Indicates the width of the feature map.

3. The structure-preserving feature encoding method based on dual-stream companding according to claim 1, characterized in that, Step 2 includes: firstly, centering each channel of the preprocessed features: ； in, For channel Learnable centralized bias, For the channels of preprocessed features ; Features of centralized processing Decomposed into mainstream components and residual flow components : ； ； in, and For learnable asymmetric truncation thresholds, residual flow components It is non-zero only when the feature value exceeds the threshold range.

4. The structure-preserving feature encoding method based on dual-stream companding according to claim 1, characterized in that, Step 3 includes: based on the hyperbolic tangent function Introducing a learnable scaling factor and ; Calculate the mainstream component after companding and the residual flow component after compression and diffusion : ； ； in, As the mainstream component, For residual flow components; Will and Concatenating along the channel dimension yields the input range within... The encoder input features between.

5. The structure-preserving feature encoding method based on dual-stream companding according to claim 1, characterized in that, Step 5 includes: segmenting the reconstructed compressive features into reconstructed mainstream compressive components along the channel dimension. and the reconstructed residual companding component Using the inverse hyperbolic tangent function Perform the inverse mapping and multiply by the corresponding scaling factor to obtain the recovered mainstream components. and the recovered residual flow component ; Calculate the intermediate features of the reconstruction : ； in, For channel Learnable central bias.

6. A structure-preserving feature encoding system based on dual-stream companding, characterized in that, include: The structure-preserving preprocessing module is used to perform structure-preserving preprocessing on the original intermediate features of a large model, preserving the original tensor structure of the features. The decomposition module is used to calculate the channel-level centralization parameters of the features, decomposing the preprocessed features into bounded mainstream components and complementary residual flow components. The companding module is used to apply differentiable nonlinear companding functions to the main stream component and the residual stream component respectively, map them to a bounded interval, and splice the two companded components in the channel dimension to obtain the encoder input features. The neural decoding module is used to quantize and entropy encode the encoder input features using a neural feature codec to generate a bitstream, and decode the reconstructed companded features from the bitstream. The inverse companding reconstruction module is used to segment the reconstructed companding features and recover the intermediate reconstructed features through the inverse companding function and the inverse flow merging operation.

7. The structure-preserving feature coding system based on dual-stream companding according to claim 6, characterized in that, The structure preservation preprocessing module includes: For the sequence features of commonsense reasoning tasks, maintain as or In this form, without two-dimensional image rearrangement, where, For the number of tokens, Number of channels; For multi-cropping features in semantic segmentation tasks, each cropping is treated as an independent single-channel sample, without padding or splicing. For multi-channel features in text-based image tasks, preserve the original... Tensor layout, in which Indicates the feature map height. Indicates the width of the feature map.

8. The structure-preserving feature coding system based on dual-stream companding according to claim 6, characterized in that, The decomposition module includes: firstly, centering each channel of the preprocessed features: ； in, For channel Learnable centralized bias, For the channels of preprocessed features ; Features of centralized processing Decomposed into mainstream components and residual flow components : ； ； in, and For learnable asymmetric truncation thresholds, residual flow components It is non-zero only when the feature value exceeds the threshold range.

9. The structure-preserving feature coding system based on dual-stream companding according to claim 6, characterized in that, The companding module includes: based on the hyperbolic tangent function. Introducing a learnable scaling factor and ; Calculate the mainstream component after companding and the residual flow component after compression and diffusion : ； ； in, As the mainstream component, For residual flow components; Will and Concatenating along the channel dimension yields the input range within... The encoder input features between.

10. The structure-preserving feature coding system based on dual-stream companding according to claim 6, characterized in that, The inverse companding reconstruction module includes: segmenting the reconstructed companding features into reconstructed mainstream companding components along the channel dimension. and reconstruction residual compressive expansion component Using the inverse hyperbolic tangent function Perform the inverse mapping and multiply by the corresponding scaling factor to obtain the recovered mainstream components. and the recovered residual flow component ; Calculate the intermediate features of the reconstruction : ； in, For channel Learnable central bias.

Images