An adaptive lightweight AI model system and method

Abstract

The application discloses a kind of self-adapting lightweight AI model system and method, it is related to artificial intelligence technical field, the method is by reading input sample and by leading detection path obtains initial response graph, extracts main response area and response structure record in combination with area division, and introduce reference sample to form original response record set;Through the statistical analysis and channel state evaluation of identifying high response unit, generate channel active atlas;Temporary channel scheduling table is constructed, and according to layer configuration lightweight inference link is selectively forward calculated to participating channel;Through the operator running instruction of executing to channel, form end layer fusion feature tensor and output task result vector and feature vector;Calculate comprehensive check value Z and carry out check evaluation, determine to use lightweight inference result or back to full-channel inference path.The method constructs scheduling basis around channel level response behavior and structure information, so that model inference process has the ability of self-adapting adjustment according to sample.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an adaptive lightweight AI model system and method. Background Technology

[0002] As deep learning applications expand in visual recognition, semantic understanding, and industrial inspection, traditional AI models have gradually formed complex computational graph systems centered on multi-layer convolutional or attention structures. These models typically rely on the participation of all channels in computation to maintain output stability. Building on this, lightweight AI models have emerged to address scenarios with limited edge computing power or high real-time response requirements. These models reduce computational scale through pruning, quantization, or structural reconstruction. However, such lightweight processing is usually based on fixed strategies or offline statistical results, making it difficult to adjust in real-time to the dynamic feature changes of different input samples. This has led to the development of adaptive lightweight AI models, which dynamically schedule channel participation during the inference phase, allowing the model to temporarily construct inference paths based on the response characteristics of input samples while maintaining its existing structure.

[0003] Existing adaptive lightweight AI models often rely on single response strengths or simple statistics to select channels during channel scheduling, lacking a joint characterization of spatial distribution structure, cross-layer connections, and regional response continuity. This results in channel selection only reflecting local response amplitudes, failing to describe the organizational relationships of responses in spatial dimensions and hierarchical transmission. Furthermore, in the inference path construction stage, some solutions directly generate channel subsets based on threshold selection results, without constraining the scope of channel influence in different regions or explicitly recording cross-layer connections. This leads to structural breaks or redundancy in the generated inference links. Moreover, in the inference result evaluation stage, judgments are often made using single output errors or class confidence scores, lacking collaborative verification of feature layer expression consistency, resulting in insufficient criteria for determining the reliability of lightweight inference results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an adaptive lightweight AI model system and method, which solves the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an adaptive lightweight AI model method, comprising the following steps: S1. Read the input sample, output the initial response map of each target scheduling layer channel through the leader detection path, perform regional division on the input sample, extract the main response region and response structure record, and form the original response record set by combining with the reference sample; S2. Based on the original response record set, extract the absolute value set of the responses of the input sample and the reference sample to identify high response units, and perform statistical analysis and channel status evaluation on the high response units to generate a channel activity map. S3. After reading the channel activity graph and retaining the channel table, generate a temporary channel scheduling table by layer, and configure the lightweight inference link for this round for each layer to be scheduled according to the temporary channel scheduling table. S4. Input the input sample into the lightweight inference link, execute the operator running instructions for the channel according to the temporary channel scheduling table, obtain the final layer fusion feature tensor, and generate the task output vector and the final layer feature vector. S5. Read the reference task output vector and the reference last layer feature vector, calculate the comprehensive verification value Z, and perform verification evaluation. Determine the final task result based on the verification evaluation result.

[0006] Preferably, S1 includes S11, S12 and S13; S11. Read input samples from the batch input stream and input the input samples into the pre-probe path of the deployed AI model. Before formal inference, call the sampling branches corresponding to the input layer, pre-feature layer and each target scheduling layer of the AI ​​model, and output the initial response map corresponding to each channel in each target scheduling layer. S12. Divide the input sample into equal regions according to the input size of the input sample to obtain several region blocks. Then, perform response sorting processing on the initial response map corresponding to each channel in each target scheduling layer. Perform connectivity merging on the first and spatially adjacent response units to obtain the main response region. Perform response statistics on the initial response map and the main response region to obtain the response structure record set. The response structure record set includes the number of main response units, the total number of units, the coordinates of the response centroid, the number of main response connected islands, and the number of receiving positions; S13. Extract the first sample, median sample and last sample from the current batch input stream as reference samples, and record the set of absolute values ​​of the response of the reference samples in the same layer, the output vector of the last layer and the feature vector of the last layer, to form the reference cache corresponding to the current sample. Based on the initial response map, response structure record set, and reference buffer of each channel in each target scheduling layer, the original response record set of the input sample is generated.

[0007] Preferably, S2 includes S21; S21. Read the original response record set and extract the absolute value set of the input sample's response at the corresponding target scheduling layer and the absolute value set of the reference sample's response at the corresponding target scheduling layer in the reference buffer for each target scheduling layer. Denote the absolute value set of the input sample's response at the l-th target scheduling layer as... Let the set of absolute values ​​of the responses of the reference samples at the l-th target scheduling layer be denoted as And calculate the dynamic trigger threshold T of the current l-th target scheduling layer; Response units in each channel of the target scheduling layer whose response values ​​are greater than or equal to the dynamic trigger threshold T are identified as high response units.

[0008] Preferably, S2 further includes S22; S22. Statistical analysis of high-response units and construction of channel scheduling value G for channel status evaluation, specifically including S221, S222 and S223; S221. Perform statistical analysis on the high-response unit to obtain the channel pulse occupancy rate P, response centroid migration distance M, response island segment number B, and continuation coverage rate E.

[0009] Preferably, in step S222, the channel scheduling value G for each channel is calculated based on the channel pulse occupancy rate P, the response centroid migration distance M, the number of response island segments B, and the continuation coverage rate E. S223. Form a set {G} for the channel scheduling values ​​G corresponding to all channels of each target scheduling layer, and calculate the intra-layer scheduling threshold U of each target scheduling layer; The channel scheduling value G is compared with the intra-layer scheduling threshold U, and the channel status is evaluated based on the comparison results, as follows; When the channel scheduling value G is less than the intra-layer scheduling threshold U, the current channel is marked as a silent channel; When the channel scheduling value G is greater than or equal to the intra-layer scheduling threshold U, the current channel is marked as a reserved channel; Then, record the layer number, channel number, channel scheduling value, retention flag, and silent flag corresponding to each channel, and generate the channel activity map corresponding to the input sample.

[0010] Preferably, S3 includes S31 and S32; S31. Read the channel activity map and extract the layer number, channel number, channel scheduling value, reservation flag and silent flag corresponding to each channel in each layer to be scheduled. If the current channel corresponds to a reserved channel, write the current channel into the reserved channel table. If the current channel corresponds to a silent channel, write the current channel into the silent channel table. At the same time, read the area block number, cross-layer acceptance position record and local supplementary calculation permission flag corresponding to the channel, and generate a temporary channel scheduling table by layer. The temporary channel scheduling table includes layer number, channel number, channel scheduling value, reservation flag, silent flag, area block number, cross-layer acceptance position, and local supplementary calculation permission flag; S32. Based on the temporary channel scheduling table, execute the channel call configuration for each layer of the deployed AI model to be scheduled in this round. Specifically, read the layer number, channel number, reserved flag, silent flag, continuous region block number, cross-layer acceptance position and local supplementary calculation permission flag corresponding to each channel in the temporary channel scheduling table layer by layer. Determine the channel corresponding to the reserved flag as the forward calculation channel in this round, and determine the channel corresponding to the silent flag as the non-participation channel in this round. Channels that have been marked with a local supplementary calculation permission and have not entered the reserved channel table are written into the region supplementary table. Based on the region block number of the current channel, each region block and its start and end target scheduling layer range are determined. Then, based on the cross-layer connection position of the channel, the inter-layer connection relationship between the current channel and the output of the adjacent preceding target scheduling layer and the input of the adjacent subsequent target scheduling layer within the start and end target scheduling layer range is determined. Channels with inter-layer connection are restricted to participate in the calculation only in the forward propagation process of the target scheduling layer within the region block, and the calculation path of the current channel is closed outside the region block, forming a lightweight inference link for the input sample.

[0011] Preferably, S4 includes S41 and S42; S41. Input the input sample into the input layer of the lightweight inference link and pass it sequentially along the original forward hierarchy of the deployed AI model. When passing it to each layer to be scheduled, execute the operator running instructions for the channel according to the temporary channel scheduling table, as follows; The current layer retention operator is executed on the retained channel to obtain the output feature map of the retained channel; Execute the shutdown operator on the silent channel; For the channel that is written to the region padding table, the current layer operator is executed only within the region block, and the region calculation result is backfilled to the corresponding position of the feature map of the current layer to be scheduled, so as to obtain the local output feature fragment; Based on the temporary channel scheduling table, the retained channel output feature map and local output feature fragments are combined to form the output tensor of the layer to be scheduled. The output tensor of the layer to be scheduled is then input into the subsequent layers to be scheduled after assembly, and passed layer by layer until the final layer fused feature tensor is obtained. S42. Based on the hierarchical connection relationship located after the final layer output feature tensor in the original computation graph deployed in the AI ​​model, continue to perform forward propagation on the final layer output feature tensor until the task output node response tensor is obtained. Read the response values ​​corresponding to each task output node in the task output node response tensor and arrange them in the order of task output node number to generate the task output vector corresponding to the input sample. At the same time, perform average calculation on the final layer output feature tensor along the spatial dimension by channel to obtain the compressed response value of each channel and arrange them in the order of channel number to generate the final layer feature vector of the input sample. The original computation graph of the deployed AI model represents the hierarchical connection and forward propagation relationships of the AI ​​model when channel scheduling processing has not been performed.

[0012] Preferably, S5 includes S51; S51. Read the reference task output vector and reference final layer feature vector corresponding to the input sample from the reference cache, and calculate the comprehensive verification value Z of the input sample based on the reserved channel set of each scheduling layer and the corresponding channel scheduling value G.

[0013] Preferably, S5 further includes S52; S52. Read the comprehensive verification value Z sequence {Z} recorded within the historical stable operating range. ref} Calculate the lower bound value Z0 of the comprehensive verification, and then verify and evaluate it with the real-time acquired comprehensive verification value Z, as follows; When the comprehensive verification value Z < the comprehensive verification lower bound Z0, it means that the lightweight inference link of the input sample does not meet the output conditions. At this time, the task result of the lightweight inference link is terminated as the final task result, and the original full-channel inference path is called to perform iterative analysis until the final task result is obtained. When the comprehensive verification value Z ≥ the comprehensive verification lower bound Z0, it means that the lightweight inference link of the input sample meets the output condition. At this time, the task result of the lightweight inference link is determined as the final task result of the input sample.

[0014] An adaptive lightweight AI model system includes an original response construction module, a channel activity determination module, a lightweight link assembly module, a lightweight inference output module, and a comprehensive verification and decision module; The original response construction module is used to read input samples, output the initial response map of each target scheduling layer channel through the leading detection path, perform regional division on the input samples, extract the main response region and response structure record, and form an original response record set by combining with the reference sample; The channel activity determination module is used to extract the set of absolute response values ​​of input samples and reference samples based on the original response record set to identify high response units, and to perform statistical analysis and channel status evaluation on high response units to generate a channel activity map. The lightweight link assembly module is used to read the channel activity graph and retain the channel table, generate a temporary channel scheduling table by layer, and configure the lightweight inference link for this round for each layer to be scheduled according to the temporary channel scheduling table. The lightweight inference output module is used to input the input sample into the lightweight inference link, execute operator running instructions on the channel according to the temporary channel scheduling table, obtain the final layer fusion feature tensor, and generate the task output vector and the final layer feature vector. The comprehensive verification decision module is used to read the reference task output vector and the reference last layer feature vector to calculate the comprehensive verification value Z and perform verification evaluation, and determine the final task result based on the verification evaluation result.

[0015] This invention provides an adaptive lightweight AI model system and method. It has the following beneficial effects: (1) This method introduces a leading probe path before formal inference to obtain the initial response of each target scheduling layer channel, and extracts the main response region and its structure record by combining region division and connectivity merging methods, so that the channel response differences that were originally only implicit in the inference process are explicitly quantified and expressed. On this basis, by introducing reference samples to form a comparison relationship, the response behavior of the input sample is no longer judged in isolation, but is characterized in the relative position in the distribution of the same batch of data, completing the transformation from "single sample response" to "batch response structure", and providing a unified data foundation for channel screening and scheduling.

[0016] (2) In the channel screening and scheduling stage, this method constructs a set of absolute response values ​​and generates a dynamic trigger threshold by combining the median value and interquartile range, thereby achieving adaptive identification of high response units. Furthermore, based on parameters such as channel pulse occupancy rate, response centroid migration distance, number of response island segments and continuation coverage, channel scheduling values ​​are constructed to quantify the participation status of each channel, forming a channel activity map, and generating a temporary channel scheduling table and lightweight inference link accordingly. Compared with the processing method of directly pruning channels based on fixed rules or static pruning, this process forms differentiated scheduling paths between different input samples, so that the channel participation range is consistent with the sample response structure, and completes the transformation of the inference path from a fixed structure to a dynamic structure.

[0017] (3) In the inference execution and result verification stages, the method completes feature transfer and outputs task results through a lightweight inference link. At the same time, it constructs a comprehensive verification value Z by combining the reference output vector and the reference feature vector to evaluate the consistency of the lightweight inference results. The verification lower bound value generated in the historical stable interval is used as the judgment basis to realize the adaptive switching between the lightweight path and the original full-channel path. This mechanism enables the lightweight inference results to have a measurable judgment basis. When the conditions are met, it is directly output as the final result. When the conditions are not met, it falls back to the full inference path to continue calculation. A computable constraint relationship is established between inference efficiency and result stability, and the processing process from "single-path inference output" to "multi-path collaborative judgment output" is completed. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of an adaptive lightweight AI model method according to the present invention; Figure 2 This is a schematic diagram of the process of an adaptive lightweight AI model system according to the present invention; Figure 3 This is a logic block diagram for an adaptive lightweight AI model method according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, please refer to Figure 1 This invention provides an adaptive lightweight AI model method. To achieve the above objectives, this invention utilizes the following technical solution, comprising the following steps: S1. Read the input sample, output the initial response map of each target scheduling layer channel through the leader detection path, perform regional division on the input sample, extract the main response region and response structure record, and form the original response record set by combining with the reference sample; S2. Based on the original response record set, extract the absolute value set of the responses of the input sample and the reference sample to identify high response units, and perform statistical analysis and channel status evaluation on the high response units to generate a channel activity map. S3. After reading the channel activity graph and retaining the channel table, generate a temporary channel scheduling table by layer, and configure the lightweight inference link for this round for each layer to be scheduled according to the temporary channel scheduling table. S4. Input the input sample into the lightweight inference link, execute the operator running instructions for the channel according to the temporary channel scheduling table, obtain the final layer fusion feature tensor, and generate the task output vector and the final layer feature vector. S5. Read the reference task output vector and the reference last layer feature vector, calculate the comprehensive verification value Z, and perform verification evaluation. Determine the final task result based on the verification evaluation result.

[0021] In this embodiment, by introducing a leading probe path before inference, the initial responses of each target scheduling layer channel are acquired in advance. Combined with region partitioning and main response region extraction, the original feature responses are transformed from discrete distributions into response structure records with spatial structural constraints. Simultaneously, reference samples are introduced to construct a comparison relationship, enabling the response features of input samples to be positionally characterized within the same batch. This completes the structured modeling of input data before formal inference, allowing subsequent channel determination to no longer rely on single-point response values ​​but rather on the overall response structure. During channel selection and path construction, the absolute value set of responses is processed and high-response units are identified. Combined with statistical analysis results, the channel status is quantitatively evaluated, forming a channel activity map. Based on this, a temporary channel scheduling table and lightweight inference link are constructed, ensuring that each input sample corresponds to an independent channel participation path. Compared to fixed channel structures or uniform pruning strategies, this process adjusts the channel participation range according to the response distribution of the input samples, achieving sample-based adaptive configuration of the inference path and establishing a correspondence between the channels participating in the calculation and the actual response requirements. During the inference execution and result determination phase, feature calculation is completed and task results are output through a lightweight inference link. At the same time, a comprehensive verification value is constructed by combining the reference output vector and the reference feature vector to evaluate the consistency of the current lightweight inference result. Based on the evaluation result, it is decided whether to adopt the result or roll back to the original full-channel path to continue calculation. This processing method enables the inference result to have quantifiable judgment criteria, forms a path selection mechanism under different input conditions, establishes a correspondence between the calculation process and the result determination, and forms a controllable balance between computational overhead control and result stability.

[0022] Example 2, please refer to Figure 3 Specifically: S1 includes S11, S12 and S13; S11. Read input samples from the batch input stream and input the input samples into the pre-probe path of the deployed AI model. Before formal inference, call the sampling branches corresponding to the input layer, pre-feature layer and each target scheduling layer of the AI ​​model, and output the initial response map corresponding to each channel in each target scheduling layer. Wherein, the input sample represents the current data unit to be processed that is sent into the deployed AI model via the existing data input link, the leading probe path represents the probe execution path set from the deployed AI model, and the probe execution path is used to output the intermediate response results of each channel of each target scheduling layer; S12. Divide the input sample into equal regions according to the input size of the input sample to obtain several region blocks. Then, perform response sorting processing on the initial response map corresponding to each channel in each target scheduling layer. Perform connectivity merging on the first and spatially adjacent response units to obtain the main response region. Perform response statistics on the initial response map and the main response region to obtain the response structure record set. The response structure record set includes the number of main response units, the total number of units, the coordinates of the response centroid, the number of main response connected islands, and the number of receiving positions; The response statistics are as follows; Count the number of marked response units in the main response region to obtain the number of main response units; Calculate the total number of response units based on the number of rows and columns of the initial response diagram to obtain the total number of units; The coordinate ranges of the main response region and each region block are superimposed one by one. The average coordinates of the x-coordinate and y-coordinate of the marked response units falling into each region block are calculated to obtain the centroid coordinates of the corresponding region block. Perform connectivity numbering on the main response region labeling graph, count the number of independent connectivity regions, and use this as the number of main response connected islands; Based on the convolution stride, kernel size, and padding relationship between the current layer and the next layer, the coordinates of each labeled response unit in the main response region are converted into the corresponding sensing position range in the next layer, and the number of next layer positions that are positionally related to the sensing position range is counted as the number of receiving positions. S13. Extract the first sample, median sample and last sample from the current batch input stream as reference samples, and record the set of absolute values ​​of the response of the reference samples in the same layer, the output vector of the last layer and the feature vector of the last layer, to form the reference cache corresponding to the current sample. Based on the initial response map, response structure record set, and reference buffer of each channel in each target scheduling layer, the original response record set of the input sample is generated.

[0023] In this embodiment, input samples are read from the batch input stream, and the pre-deployed AI model's probing path is input. Without entering formal inference, the sampling branches of the input layer, pre-feature layer, and each target scheduling layer are invoked to obtain the initial response map of each layer and each channel. Subsequently, based on the size of the input sample, equal regions are divided. The high response units in the initial response map of each channel are sorted by response value and spatially connected and merged to extract the main response region. The number of main response units, the total number of response units, the response centroid coordinates within each region block, the number of main response connected islands, and the number of receiving positions obtained by cross-layer mapping are calculated around this region to form a response structure record set. Finally, the first sample, the median sample, and the last sample in the current batch are selected as reference samples. Their absolute response value set in the corresponding layer and the output and feature vector of the last layer are recorded to construct a reference cache. The initial response map, the response structure record set, and the reference cache are integrated to generate the original response record set corresponding to the input sample.

[0024] Example 3, please refer to Figure 3 Specifically: S2 includes S21; S21. Read the original response record set and extract the absolute value set of the input sample's response at the corresponding target scheduling layer and the absolute value set of the reference sample's response at the corresponding target scheduling layer in the reference buffer for each target scheduling layer. Denote the absolute value set of the input sample's response at the l-th target scheduling layer as... Let the set of absolute values ​​of the responses of the reference samples at the l-th target scheduling layer be denoted as And calculate the dynamic trigger threshold T of the current l-th target scheduling layer, as follows: ; in, Indicates the median value. Indicates the interquartile range. and These represent the median values ​​of the input sample and the reference sample in the set of absolute values ​​of the target scheduling layer response at layer l, respectively. It represents the interquartile range value in the set of absolute values ​​of the reference sample response of the l-th target scheduling layer, and is used to represent the fluctuation span of the middle main interval in the current target scheduling layer reference response data; Response units in each channel of the target scheduling layer whose response values ​​are greater than or equal to the dynamic trigger threshold T are identified as high response units.

[0025] S2 further includes S22; S22. Statistical analysis of high-response units and construction of channel scheduling value G for channel status evaluation, specifically including S221, S222 and S223; S221. Perform statistical analysis on the high-response unit to obtain the channel pulse occupancy rate P, response centroid migration distance M, number of response island segments B, and continuity coverage rate E, respectively. The statistical analysis is as follows: The channel pulse occupancy rate P is obtained by comparing the number of high-response units in each channel with the total number of response units in each channel. Specifically: ,in, This represents the channel pulse occupancy rate of the c-th channel in the l-th target scheduling layer. This represents the number of high-response units in the c-th channel of the l-th target scheduling layer. This represents the total number of response units in the c-th channel of the l-th target scheduling layer; The formula for obtaining the response centroid migration distance M is: ,in, Represents the response centroid migration distance of the c-th channel in the l-th target scheduling layer, where R represents the total number of region blocks, (x l，c，r y l，c，r () represents the centroid coordinates of the high-response unit of the c-th channel in the l-th target scheduling layer of the r-th region block. and These represent the row length and column length of the input sample, respectively. The formula for obtaining the number of response island segments B is: ,in, This represents the number of main response connected islands in the c-th channel of the l-th target scheduling layer within the r-th region block; The formula for obtaining the continuity coverage rate E is: ,in, This represents the continuity coverage of the c-th channel in the l-th target scheduling layer; This represents the number of units that form a positional correspondence with the response region of the c-th channel in the l-th target scheduling layer after the high-response region is mapped to the next layer.

[0026] S222. Based on the channel pulse occupancy rate P, the response centroid migration distance M, the number of response island segments B, and the continuation coverage rate E, calculate the channel scheduling value G for each channel, specifically as follows: ,in, This represents the channel scheduling value of the c-th channel in the l-th target scheduling layer; S223. Form a set {G} for the channel scheduling values ​​G corresponding to all channels of each target scheduling layer, and calculate the intra-layer scheduling threshold U of each target scheduling layer, specifically: U=Med({G})+Iqr({G}); The channel scheduling value G is compared with the intra-layer scheduling threshold U, and the channel status is evaluated based on the comparison results, as follows; When the channel scheduling value G is less than the intra-layer scheduling threshold U, the current channel is marked as a silent channel; When the channel scheduling value G is greater than or equal to the intra-layer scheduling threshold U, the current channel is marked as a reserved channel; Then, record the layer number, channel number, channel scheduling value, retention flag, and silent flag corresponding to each channel, and generate the channel activity map corresponding to the input sample.

[0027] In this embodiment, the original response record set is read, and the median and interquartile range of the absolute response value sets of the input sample and the reference sample in each target scheduling layer are statistically analyzed to generate the dynamic trigger threshold T for the corresponding layer. The response units in each channel are then screened one by one using this dynamic trigger threshold T to identify high-response units. Subsequently, based on the distribution of high-response units in each region block, the channel pulse occupancy rate P, response centroid migration distance M, response island segment number B, and continuation coverage rate E are calculated to represent the occupancy density, cross-region migration trajectory, spatial dispersion, and spatial connection with the next layer of the current channel under the current sample. After obtaining the above statistics, each channel is uniformly measured using the channel scheduling value G, and the division of silent channels and reserved channels is completed by combining the intra-layer scheduling threshold U. Finally, a channel activity map is generated layer by layer. In actual implementation, for the dynamic trigger threshold T, the absolute response value set of the input sample is first extracted from the l-th target scheduling layer. and the set of absolute values ​​of the responses of the reference samples ,in, This indicates the middle position of the current input sample's response intensity in this layer, reflecting the overall response level of the current sample in this layer; This indicates the intermediate response level of the reference sample within the same stratum, used to provide a baseline control for samples from the same batch. This represents the discrete span of the central main interval of the reference sample, used to express the natural fluctuation range of the reference response itself in this layer. All three are included in the calculation of T because using only the median value of the current sample response would cause the threshold to only follow the fluctuations of the current sample itself, lacking batch control; using only the reference sample statistics would fail to reflect the immediate response position of the current sample; while incorporating the interquartile range of the reference sample into the calculation allows the current sample position, the batch main level, and the batch fluctuation boundary to be simultaneously included in the same judgment formula. Therefore, this threshold is not a fixed value, but rather an intra-layer triggering benchmark that varies with the layer, batch, and sample. After screening out high-response units using the dynamic triggering threshold T, the channel pulse occupancy rate P is then calculated, where... It is the number of high-response units in the c-th channel of the l-th target scheduling layer. This represents the total number of response units in the current channel. The reason for using the ratio is that different channels may have the same spatial size, but the coverage ratio of high-response units may not be the same. Using the percentage rather than the absolute number can directly indicate how many spatial locations of the channel are active in the current sample. When the channel pulse occupancy rate P is large, it indicates that the effective response distribution range of the channel is wide. When the channel pulse occupancy rate P is small, it indicates that the channel only responds in a few local locations. Next, the centroid migration distance M of the response is calculated, where R is the total number of regions, (x l，c，r y l，c，rLet ) be the centroid coordinates of the high-response unit of the channel within the r-th region block. The formula accumulates the lateral and longitudinal differences of the centroids between adjacent regions blocks, and then uses (R-1)(G) w -G h Normalization is performed to measure whether the high-response region of the channel experiences continuous positional shifts across multiple regions. The centroid difference is used because individual high-response points are easily affected by local spikes, while the centroid coordinates are closer to the main response position of the region, representing the channel's spatial migration trajectory. A large centroid migration distance M indicates that the response center jumps more between different regions, indicating weaker channel stability; a small centroid migration distance M indicates stronger continuity of the response center. Subsequently, the number of response island segments B is calculated, where K... l，c，r This represents the number of connected islands in the current channel main response within the r-th region block, using... The reason for this approach is that each region block can be considered basically continuous when it has at least one main connected region. Only when the number of connected islands exceeds 1 does it indicate that the response within that region block has splitted. Therefore, subtracting 1 and then accumulating the results gives the number of redundant segments. The larger this number, the more fragmented the high-response region of the channel and the more dispersed the spatial structure. Then, the continuation coverage rate E is calculated, in the formula... This represents the number of units that, after being mapped to the next layer, still maintain a positional correspondence with the response region of the next layer. The denominator still takes... The purpose is to measure how many of the high-response regions identified in the current layer can continue to correspond to valid response locations in the next layer after propagation across layers. If this ratio is high, it indicates that the active regions in the current layer of the channel do not terminate in isolation but continue to propagate to subsequent layers, still having significance for subsequent inference. Finally, the channel scheduling value G: By incorporating the above quantities into the same criterion, the calculation aligns with engineering logic: a channel suitable for subsequent lightweight inference should simultaneously possess high spatial activity coverage and good cross-layer continuity capability, without being accompanied by excessive positional jumps and too many segment distributions. Therefore, the channel scheduling value G actually represents the scheduling priority of the current channel under the combined effects of activity level, structural continuity, and cross-layer continuity. The higher the value, the more suitable the channel is to participate in subsequent inference as a reserved channel; the lower the value, the closer it is to the processing direction of a silent channel. This implementation method completes the continuous processing task from "response unit identification - channel statistical analysis - channel status determination - active graph generation". It ensures that whether a channel participates in subsequent lightweight inference is no longer determined solely by fixed thresholds or preset pruning rules, but rather by the actual response pattern, spatial connectivity, and cross-layer connections of the current input sample within its layer. Compared to uniform pruning or static retention techniques, this process ensures that the channel selection criteria correspond to the sample content. The resulting channel active graph can be directly used as the input basis for subsequent temporary channel scheduling tables and lightweight inference link construction, providing clear quantitative criteria for sample-level dynamic scheduling. It also creates a hierarchical distinction between irrelevant channels and channels that need to participate, thereby making the subsequent inference path construction closer to the actual response distribution of the current sample and providing a clearer foundation of preceding data for the final output determination.

[0028] Example 4, please refer to Figure 3 Specifically: S3 includes S31 and S32; S31. Read the channel activity map and extract the layer number, channel number, channel scheduling value, reservation flag and silent flag corresponding to each channel in each layer to be scheduled. If the current channel corresponds to a reserved channel, write the current channel into the reserved channel table. If the current channel corresponds to a silent channel, write the current channel into the silent channel table. At the same time, read the area block number, cross-layer acceptance position record and local supplementary calculation permission flag corresponding to the channel, and generate a temporary channel scheduling table by layer. The temporary channel scheduling table includes layer number, channel number, channel scheduling value, reservation flag, silent flag, area block number, cross-layer acceptance position, and local supplementary calculation permission flag; S32. Based on the temporary channel scheduling table, execute the channel call configuration for each layer of the deployed AI model to be scheduled in this round. Specifically, read the layer number, channel number, reserved flag, silent flag, continuous region block number, cross-layer acceptance position and local supplementary calculation permission flag corresponding to each channel in the temporary channel scheduling table layer by layer. Determine the channel corresponding to the reserved flag as the forward calculation channel in this round, and determine the channel corresponding to the silent flag as the non-participation channel in this round. Channels that have been marked as having local supplementary calculation permission and have not entered the reserved channel table are written into the region supplementary table. Based on the region block number of the current channel, each region block and its starting and ending target scheduling layer range are determined. Then, based on the cross-layer connection position of the channel, the inter-layer connection relationship between the current channel and the output of the adjacent preceding target scheduling layer and the input of the adjacent subsequent target scheduling layer within the starting and ending target scheduling layer range is determined. Channels with inter-layer connection are restricted to participate in the calculation only during the forward propagation of the target scheduling layer within the region block, and the calculation path of the current channel is closed outside the region block, forming a lightweight inference link for the input sample. The lightweight inference link refers to the inference execution path temporarily generated for this round based on the channel activity graph of the input sample.

[0029] In this embodiment, by analyzing the channel activity map layer by layer, the scheduling attributes, spatial affiliation, and inter-layer connection information of each channel are uniformly written into a temporary channel scheduling table. This transforms the channel response states, which were originally scattered across layers, into scheduling criteria with structural constraints. Based on this, the channel participation status is divided according to the retention and silence markers. Combined with the region block number and cross-layer connection position, channels with local supplementary calculation permission are restricted to participate in forward propagation only within the corresponding region block and its associated layer. At the same time, their calculation paths are closed in other regions, thereby constructing a lightweight inference link corresponding to the input sample response structure. Through this process, the channel participation range, spatial response distribution, and inter-layer connection relationships form corresponding constraints, completing the transformation from a unified channel execution mode to a differentiated execution mode limited by region and layer. This enables the inference path to have the characteristic of dynamic generation based on samples, maintaining the continuity of inter-layer feature transmission while controlling the scale of the calculation path. This allows effective responses in local regions to be retained and participate in calculations during cross-layer processes.

[0030] Example 5, please refer to Figure 3 Specifically: S4 includes S41 and S42; S41. Input the input sample into the input layer of the lightweight inference link and pass it sequentially along the original forward hierarchy of the deployed AI model. When passing it to each layer to be scheduled, execute the operator running instructions for the channel according to the temporary channel scheduling table, as follows; The current layer retention operator is executed on the retained channel to obtain the output feature map of the retained channel; Execute the shutdown operator on the silent channel; For the channel that is written to the region padding table, the current layer operator is executed only within the region block, and the region calculation result is backfilled to the corresponding position of the feature map of the current layer to be scheduled, so as to obtain the local output feature fragment; Based on the temporary channel scheduling table, the retained channel output feature map and local output feature fragments are combined to form the output tensor of the layer to be scheduled. The output tensor of the layer to be scheduled is then input into the subsequent layers to be scheduled after assembly, and passed layer by layer until the final layer fused feature tensor is obtained. S42. Based on the hierarchical connection relationship located after the final layer output feature tensor in the original computation graph deployed in the AI ​​model, continue to perform forward propagation on the final layer output feature tensor until the task output node response tensor is obtained. Read the response values ​​corresponding to each task output node in the task output node response tensor and arrange them in the order of task output node number to generate the task output vector corresponding to the input sample. At the same time, perform average calculation on the final layer output feature tensor along the spatial dimension by channel to obtain the compressed response value of each channel and arrange them in the order of channel number to generate the final layer feature vector of the input sample. The original computation graph of the deployed AI model represents the hierarchical connection and forward propagation relationships of the AI ​​model when channel scheduling processing has not been performed.

[0031] In this embodiment, the participation status of each channel in the deployed AI model to be scheduled is controlled layer by layer according to the temporary channel scheduling table: the original layer operator is maintained for reserved channels, the current layer calculation is stopped for silent channels, and the local operator calculation is only performed in the corresponding region block for channels written into the region supplement table. The obtained local output feature fragments are then backfilled to the corresponding positions in the feature map of the layer to be scheduled. The output feature map of the reserved channel is then spliced ​​and assembled with the local output feature fragments to form the output tensor of each layer to be scheduled and continues to be passed forward until the final layer fusion feature tensor is obtained. Subsequently, forward propagation is continued along the existing hierarchical connection relationship after the final layer in the original computation graph to obtain the task output node response tensor, organize it to obtain the task output vector, and at the same time, the final layer fusion is performed. The feature tensor is spatially compressed by channel to form the final layer feature vector, which is then used for the calculation of the subsequent comprehensive verification value Z. This implementation method ensures that the lightweight inference link does not simply delete channels, but rather performs supplementary calculations for regions that still need to participate while retaining the main channel calculation path. This keeps the output of the layer to be scheduled connected to the original hierarchical structure, thus completing the forward inference task of the input sample under the lightweight link and forming output vectors and feature vectors that can be used for result verification. Compared with the processing method of only using fixed pruning or closing the entire channel, this processing path can take into account regional response differences, inter-layer continuity, and consistency of the final layer result reading, so that the lightweight inference result has a basis for subsequent verification and the calculation path fits the response distribution of the current sample.

[0032] Example 6, please refer to Figure 3 Specifically: S5 includes S51; S51. Read the reference task output vector and reference final layer feature vector corresponding to the input sample from the reference cache, and calculate the comprehensive verification value Z of the input sample based on the reserved channel set of each scheduling layer and the corresponding channel scheduling value G, as follows; ; Where L represents the total number of layers to be scheduled, and D out D represents the mean of the normalized absolute differences between the class output vector of the input sample and the reference output vector. feat This represents the mean of the normalized absolute difference between the final feature vector of the input sample and the final feature vector of the reference sample.

[0033] S5 also includes S52; S52. Read the comprehensive verification value Z sequence {Z} recorded within the historical stable operating range. ref The comprehensive verification lower bound value Z0 is calculated as follows: Z0 = Med({Z ref})+Iqr({Z ref Then, it is compared with the real-time acquired comprehensive verification value Z for verification and evaluation, as follows; When the comprehensive verification value Z < the comprehensive verification lower bound Z0, it means that the lightweight inference link of the input sample does not meet the output conditions. At this time, the task result of the lightweight inference link is terminated as the final task result, and the original full-channel inference path is called to perform iterative analysis until the final task result is obtained. When the comprehensive verification value Z ≥ the comprehensive verification lower bound Z0, it means that the lightweight inference link of the input sample meets the output condition. At this time, the task result of the lightweight inference link is determined as the final task result of the input sample.

[0034] In this embodiment, the lightweight inference link, task output vector, final layer feature vector, and the set of reserved channels and channel scheduling values ​​G of each scheduled layer corresponding to the input sample are first read from the reference cache in S51. The reference task output vector and reference final layer feature vector corresponding to the current input sample are then read from the reference cache. The channel scheduling values ​​of the reserved channels in each scheduled layer are accumulated layer by layer and compared with the mean of the normalized absolute difference D between the current task output vector and the reference task output vector. out The mean of the normalized absolute difference D between the current last layer feature vector and the reference last layer feature vector. feat Together, they constitute the comprehensive verification value Z, where the cumulative portion of the reserved channel scheduling value is used to represent the main response acceptance status of the current lightweight inference link at each layer, D. out D is used to indicate the degree of deviation between the current task result and the reference output. featUsed to represent the degree of internal representation deviation between the final layer semantic expression and the reference features, so that the comprehensive verification value Z simultaneously covers three levels: "channel participation state - task output state - final layer feature state"; in S52, the comprehensive verification value sequence {Z} recorded within the historical stable operation interval is read again. ref The comprehensive verification lower bound Z0 is calculated using the median and interquartile range. The real-time comprehensive verification value Z is compared with this lower bound Z0. If Z is lower than Z0, it indicates that a stable correspondence has not been formed between the channel retention structure, output result, and final layer representation of the current lightweight inference link. In this case, the direct use of the lightweight inference result is terminated, and the original full-channel inference path is called to recalculate the final task result. If Z is not lower than Z0, it indicates that the current lightweight inference link is within an acceptable range in terms of both output result and internal representation, and the lightweight inference result can be directly used as the final task result. In actual implementation, the construction of this comprehensive verification value Z is uniformly characterized by three types of information: "channel participation status—output result difference—feature expression difference." First, the set of channels retained in each layer to be scheduled is traversed, and the corresponding channel scheduling value G is... l，c The accumulation is performed layer by layer, and its calculation is based on the statistical results of channel pulse occupancy rate, response centroid migration distance, number of response island segments, and continuation coverage rate in the previous steps. This accumulation process reflects the degree of continuity of the channels actually participating in the calculation in the current lightweight inference link in terms of spatial response structure and inter-layer connection relationship. From an engineering perspective, if an input sample retains channels with stable response structures in multiple layers, then the channel scheduling value G is... l，c The cumulative value will be relatively high, indicating that the current lightweight link has complete transmission capabilities in terms of structure. Secondly, D is introduced. out , which is the mean of the normalized absolute difference between the current task output vector and the reference output vector, is calculated to characterize the deviation at the final decision level, avoiding reliance on intermediate features while ignoring changes in the final output; and D is introduced. featThe term Z is used to measure the difference between the final layer feature vector and the reference final layer feature vector. This term reflects the degree of deviation at the semantic representation level, and the two together constitute the deviation term in the denominator. The "cumulative result of channel scheduling value" is placed in the numerator, and the "degree of deviation between output and feature" is placed in the denominator with an increment of 1 for constraint, forming the calculation structure of the comprehensive verification value Z. This means that when the channel participation structure is stable and the output and feature deviations are small, the comprehensive verification value Z remains in a high range; when the lightweight path causes a shift in the output result or feature expression, the denominator increases, and the comprehensive verification value Z decreases accordingly. From an engineering logic perspective, this calculation method avoids the insufficiency of a single indicator, ensuring that the lightweight inference link not only has structural capacity but also maintains consistency with the reference state in terms of results to obtain a high verification value, providing a quantifiable basis for subsequent decisions on whether to adopt the lightweight inference results. The above implementation adds a verification step that connects to the historical stable state before the lightweight inference result is output. Instead of directly outputting the task conclusion based solely on the channel pruning result, it first determines whether the lightweight path has the conditions to continue as the final output basis. This creates a sequential processing relationship between lightweight path selection, result determination, and full channel rollback. Compared to the current processing method that only relies on static pruning or fixed compression to directly output the result, this solution adds a layer of computable result verification constraints, giving the lightweight inference results under different input samples a clearer judgment basis. At the same time, it also enables the original full channel path to intervene under abnormal conditions.

[0035] Example 7, please refer to Figure 2 An adaptive lightweight AI model system includes an original response construction module, a channel activity determination module, a lightweight link assembly module, a lightweight inference output module, and a comprehensive verification decision module. The original response construction module is used to read input samples, output the initial response map of each target scheduling layer channel through the leading detection path, perform regional division on the input samples, extract the main response region and response structure record, and form an original response record set by combining with the reference sample; The channel activity determination module is used to extract the set of absolute response values ​​of input samples and reference samples based on the original response record set to identify high response units, and to perform statistical analysis and channel status evaluation on high response units to generate a channel activity map. The lightweight link assembly module is used to read the channel activity graph and retain the channel table, generate a temporary channel scheduling table by layer, and configure the lightweight inference link for this round for each layer to be scheduled according to the temporary channel scheduling table. The lightweight inference output module is used to input the input sample into the lightweight inference link, execute operator running instructions on the channel according to the temporary channel scheduling table, obtain the final layer fusion feature tensor, and generate the task output vector and the final layer feature vector. The comprehensive verification decision module is used to read the reference task output vector and the reference last layer feature vector to calculate the comprehensive verification value Z and perform verification evaluation, and determine the final task result based on the verification evaluation result.

[0036] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended technical solutions and their equivalents.

Claims

1. An adaptive lightweight AI model method, characterized in that: Includes the following steps: S1. Read the input sample, output the initial response map of each target scheduling layer channel through the leader detection path, perform regional division on the input sample, extract the main response region and response structure record, and form the original response record set by combining with the reference sample; S2. Based on the original response record set, extract the absolute value set of the responses of the input sample and the reference sample to identify high response units, and perform statistical analysis and channel status evaluation on the high response units to generate a channel activity map. S3. After reading the channel activity graph and retaining the channel table, generate a temporary channel scheduling table by layer, and configure the lightweight inference link for this round for each layer to be scheduled according to the temporary channel scheduling table. S4. Input the input sample into the lightweight inference link, execute the operator running instructions for the channel according to the temporary channel scheduling table, obtain the final layer fusion feature tensor, and generate the task output vector and the final layer feature vector. S5. Read the reference task output vector and the reference last layer feature vector, calculate the comprehensive verification value Z, and perform verification evaluation. Determine the final task result based on the verification evaluation result.

2. The adaptive lightweight AI model method according to claim 1, characterized in that: S1 includes S11, S12 and S13; S11. Read input samples from the batch input stream and input the input samples into the pre-probe path of the deployed AI model. Before formal inference, call the sampling branches corresponding to the input layer, pre-feature layer and each target scheduling layer of the AI ​​model, and output the initial response map corresponding to each channel in each target scheduling layer. S12. Divide the input sample into equal regions according to the input size of the input sample to obtain several region blocks. Then, perform response sorting processing on the initial response map corresponding to each channel in each target scheduling layer. Perform connectivity merging on the first and spatially adjacent response units to obtain the main response region. Perform response statistics on the initial response map and the main response region to obtain the response structure record set. The response structure record set includes the number of main response units, the total number of units, the coordinates of the response centroid, the number of main response connected islands, and the number of receiving positions; S13. Extract the first sample, median sample and last sample from the current batch input stream as reference samples, and record the set of absolute values ​​of the response of the reference samples in the same layer, the output vector of the last layer and the feature vector of the last layer, to form the reference cache corresponding to the current sample. Based on the initial response map, response structure record set, and reference buffer of each channel in each target scheduling layer, the original response record set of the input sample is generated.

3. The adaptive lightweight AI model method according to claim 2, characterized in that: S2 includes S21; S21. Read the original response record set and extract the absolute value set of the input sample's response at the corresponding target scheduling layer and the absolute value set of the reference sample's response at the corresponding target scheduling layer in the reference buffer for each target scheduling layer. Denote the absolute value set of the input sample's response at the l-th target scheduling layer as... Let the set of absolute values ​​of the responses of the reference samples at the l-th target scheduling layer be denoted as And calculate the dynamic trigger threshold T of the current l-th target scheduling layer; Response units in each channel of the target scheduling layer whose response values ​​are greater than or equal to the dynamic trigger threshold T are identified as high response units.

4. The adaptive lightweight AI model method according to claim 3, characterized in that: S2 further includes S22; S22. Statistical analysis of high-response units and construction of channel scheduling value G for channel status evaluation, specifically including S221, S222 and S223; S221. Perform statistical analysis on the high-response unit to obtain the channel pulse occupancy rate P, response centroid migration distance M, response island segment number B, and continuation coverage rate E.

5. The adaptive lightweight AI model method according to claim 4, characterized in that: S222. Calculate the channel scheduling value G for each channel based on the channel pulse occupancy rate P, the response centroid migration distance M, the number of response island segments B, and the continuation coverage rate E. S223. Form a set {G} for the channel scheduling values ​​G corresponding to all channels of each target scheduling layer, and calculate the intra-layer scheduling threshold U of each target scheduling layer; The channel scheduling value G is compared with the intra-layer scheduling threshold U, and the channel status is evaluated based on the comparison results, as follows; When the channel scheduling value G is less than the intra-layer scheduling threshold U, the current channel is marked as a silent channel; When the channel scheduling value G is greater than or equal to the intra-layer scheduling threshold U, the current channel is marked as a reserved channel; Then, record the layer number, channel number, channel scheduling value, retention flag, and silent flag corresponding to each channel, and generate the channel activity map corresponding to the input sample.

6. The adaptive lightweight AI model method according to claim 5, characterized in that: S3 includes S31 and S32; S31. Read the channel activity map and extract the layer number, channel number, channel scheduling value, reservation flag and silent flag corresponding to each channel in each layer to be scheduled. If the current channel corresponds to a reserved channel, write the current channel into the reserved channel table. If the current channel corresponds to a silent channel, write the current channel into the silent channel table. At the same time, read the area block number, cross-layer acceptance position record and local supplementary calculation permission flag corresponding to the channel, and generate a temporary channel scheduling table by layer. The temporary channel scheduling table includes layer number, channel number, channel scheduling value, reservation flag, silent flag, area block number, cross-layer acceptance position, and local supplementary calculation permission flag; S32. Based on the temporary channel scheduling table, execute the channel call configuration for each layer of the deployed AI model to be scheduled in this round. Specifically, read the layer number, channel number, reserved flag, silent flag, continuous region block number, cross-layer acceptance position and local supplementary calculation permission flag corresponding to each channel in the temporary channel scheduling table layer by layer. Determine the channel corresponding to the reserved flag as the forward calculation channel in this round, and determine the channel corresponding to the silent flag as the non-participation channel in this round. Channels that have been marked with a local supplementary calculation permission and have not entered the reserved channel table are written into the region supplementary table. Based on the region block number of the current channel, each region block and its start and end target scheduling layer range are determined. Then, based on the cross-layer connection position of the channel, the inter-layer connection relationship between the current channel and the output of the adjacent preceding target scheduling layer and the input of the adjacent subsequent target scheduling layer within the start and end target scheduling layer range is determined. Channels with inter-layer connection are restricted to participate in the calculation only in the forward propagation process of the target scheduling layer within the region block, and the calculation path of the current channel is closed outside the region block, forming a lightweight inference link for the input sample.

7. The adaptive lightweight AI model method according to claim 6, characterized in that: S4 includes S41 and S42; S41. Input the input sample into the input layer of the lightweight inference link and pass it sequentially along the original forward hierarchy of the deployed AI model. When passing it to each layer to be scheduled, execute the operator running instructions for the channel according to the temporary channel scheduling table, as follows; The current layer retention operator is executed on the retained channel to obtain the output feature map of the retained channel; Execute the shutdown operator on the silent channel; For the channel that is written to the region padding table, the current layer operator is executed only within the region block, and the region calculation result is backfilled to the corresponding position of the feature map of the current layer to be scheduled, so as to obtain the local output feature fragment; Based on the temporary channel scheduling table, the retained channel output feature map and local output feature fragments are combined to form the output tensor of the layer to be scheduled. The output tensor of the layer to be scheduled is then input into the subsequent layers to be scheduled after assembly, and passed layer by layer until the final layer fused feature tensor is obtained. S42. Based on the hierarchical connection relationship located after the final layer output feature tensor in the original computation graph deployed in the AI ​​model, continue to perform forward propagation on the final layer output feature tensor until the task output node response tensor is obtained. Read the response values ​​corresponding to each task output node in the task output node response tensor and arrange them in the order of task output node number to generate the task output vector corresponding to the input sample. At the same time, perform average calculation on the final layer output feature tensor along the spatial dimension by channel to obtain the compressed response value of each channel and arrange them in the order of channel number to generate the final layer feature vector of the input sample. The original computation graph of the deployed AI model represents the hierarchical connection and forward propagation relationships of the AI ​​model when channel scheduling processing has not been performed.

8. The adaptive lightweight AI model method according to claim 7, characterized in that: S5 includes S51; S51. Read the reference task output vector and reference final layer feature vector corresponding to the input sample from the reference cache, and calculate the comprehensive verification value Z of the input sample based on the reserved channel set of each scheduling layer and the corresponding channel scheduling value G.

9. The adaptive lightweight AI model method according to claim 8, characterized in that: S5 also includes S52; S52. Read the comprehensive verification value Z sequence {Z} recorded within the historical stable operating range. ref } Calculate the lower bound value Z0 of the comprehensive verification, and then verify and evaluate it with the real-time acquired comprehensive verification value Z, as follows; When the comprehensive verification value Z < the comprehensive verification lower bound Z0, it means that the lightweight inference link of the input sample does not meet the output conditions. At this time, the task result of the lightweight inference link is terminated as the final task result, and the original full-channel inference path is called to perform iterative analysis until the final task result is obtained. When the comprehensive verification value Z ≥ the comprehensive verification lower bound Z0, it means that the lightweight inference link of the input sample meets the output condition. At this time, the task result of the lightweight inference link is determined as the final task result of the input sample.

10. An adaptive lightweight AI model system, applied to the adaptive lightweight AI model method according to any one of claims 1-9, characterized in that: It includes a raw response construction module, a channel activity determination module, a lightweight link assembly module, a lightweight inference output module, and a comprehensive verification and decision module; The original response construction module is used to read input samples, output the initial response map of each target scheduling layer channel through the leading detection path, perform regional division on the input samples, extract the main response region and response structure record, and form an original response record set by combining with reference samples; The channel activity determination module is used to extract the set of absolute response values ​​of input samples and reference samples based on the original response record set to identify high response units, and to perform statistical analysis and channel status evaluation on high response units to generate a channel activity map. The lightweight link assembly module is used to read the channel activity graph and retain the channel table, generate a temporary channel scheduling table by layer, and configure the lightweight inference link for this round for each layer to be scheduled according to the temporary channel scheduling table. The lightweight inference output module is used to input the input sample into the lightweight inference link, execute operator running instructions on the channel according to the temporary channel scheduling table, obtain the final layer fusion feature tensor, and generate the task output vector and the final layer feature vector. The comprehensive verification decision module is used to read the reference task output vector and the reference last layer feature vector to calculate the comprehensive verification value Z and perform verification evaluation, and determine the final task result based on the verification evaluation result.

Images