46 results about "Uniform quantization" patented technology

The invention relates to the technical field of high-performance computing resource management, and discloses a computing resource allocation method for a distributed supercomputing center. The method comprises the following steps: on the basis of obtaining real-time computing task and supercomputing center resource data and uniformly quantifying, integrally predicting resource requirements of future tasks; constructing a mixed integer linear programming model with the minimization of the total operation cost as a single target, wherein the total operation cost is the sum of the energy cost, the carbon emission cost, the data transmission cost and the SLA default penalty cost; solving the model by taking the time-varying electricity price, the green energy ratio, the resource capacity and the network parameters of each center as constraint conditions to generate an optimal resource allocation scheme; and then, by dynamically monitoring the resource state and the task progress, the model is triggered to resolve when the resource utilization rate is detected to be unbalanced or default risks, so that self-adaptive adjustment is realized. According to the invention, global collaborative resource allocation across super computing centers is realized, and operation economy, environmental sustainability and service reliability are considered.

The invention relates to the technical field of data processing compression and machine learning, and discloses a non-stationary data quantitative processing method based on dynamic volatility perception, which comprises the following steps of: continuously tracking real-time volatility statistical measurement of a photovoltaic data stream by adopting a Welford algorithm, and parallelly calculating a first-order differential sequence of the data stream to capture time dynamic characteristics; constructing a non-uniform quantizer with learnable parameters, setting a non-uniform input threshold by using a quantile initialization strategy, and mapping a non-uniform interval into a fixed and uniform integer output in combination with a non-uniform-to-uniform quantization mechanism; in order to guide learning of quantization parameters, a dynamic composite loss function based on volatility adaptive Beta divergence and time gradient consistency is constructed, the function comprises a Beta divergence distribution loss item dynamically controlled by real-time volatility, and the method can adapt to non-stationary characteristics of energy data such as photovoltaic power and the like. The robustness is remarkably improved while the high signal-to-noise ratio is guaranteed, and the method is suitable for edge computing scenes with limited resources.

The present disclosure provides an end-to-end joint optimization method for dynamic neural radiance field representation and compression, which comprises: representing a dynamic neural radiance field by using a coefficient feature grid and a base feature grid, compensating an error area of the dynamic neural radiance field by using a residual feature grid and the coefficient feature grid, and determining a represented dynamic neural radiance field; performing model training processing and analog quantization operation processing on the represented dynamic neural radiance field; performing optimization processing on the dynamic neural radiance field subjected to the analog quantization operation processing according to a predicted compression data amount of the represented dynamic neural radiance field and a preset loss function, and determining a trained dynamic neural radiance field; performing uniform quantization processing and encoding processing on the trained dynamic neural radiance field, and determining an actual compression data amount of the dynamic neural radiance field. Through the present disclosure, the dynamic neural radiance field is efficiently modeled and compressed, the compression rate of the dynamic neural radiance field is improved, and the distortion degree after restoration is reduced.

The application provides a precision-sensitive fine-grained recognition method based on Posit-driven symbolic pattern shaping super-dimensional calculation, and relates to the technical field of digital integrated circuits and artificial intelligence accelerator design. The pre-activation GAP semantic features of the frozen pre-training CNN are taken as input, and prototype Hamming matching is adopted to complete decision. Under the constraint of target deployment bit width, Posit non-uniform quantization is introduced to match the semantic feature distribution of the center, reduce the low bit width quantization error, and improve the low bit width precision; meanwhile, through the learning bipolar projection of quantization perception, the sign margin and dimension contribution are directly shaped, the Hamming separation between classes is expanded, and the reachable precision saturation level is raised. Since the super-dimensional calculation has the class number decoupling characteristic, when the number of classes is expanded, only the on-chip storage resources and the number of parallel comparisons need to be linearly increased, which has better energy efficiency and resource scaling characteristics compared with the traditional fully connected classification head, and significantly improves the hardware implementation friendliness on FPGA or special accelerators.

The invention discloses a fair resource management method and system for multi-user shared large language model reasoning, and the method is characterized in that the method employs a cold data recognition elimination mechanism and a fair cache distribution mechanism, carries out the unified quantification of the resource consumption of a request in GPU calculation, decoding and cache transmission, and carries out the fair scheduling based on the accumulated CPI value of each user; the system comprises a user request management module, a CPI fair scheduling module, a model reasoning execution module, a resource monitoring module and a cache management module. Compared with the prior art, the method has the advantages that users can share computing and caching resources fairly, efficient operation performance of the system is guaranteed, users with high cache hit rate are prevented from excessively occupying GPU execution opportunities due to low apparent computing overhead, the overall reasoning efficiency and the cache hit rate of the system are remarkably improved, and the user experience is improved. Unified and fair distribution of computing resources and cache resources is achieved, the effect is prominent especially in a multi-tenant high-load scene, and the method has good application prospects and commercial development value.

This invention relates to the field of quantum computing technology, and particularly to a method, system, computer device, and storage medium for optimizing the fidelity of quantum circuits. The method includes: acquiring the hardware parameters of the quantum circuit to be executed and the quantum processing unit cluster; standardizing the quantum circuit, analyzing the type of quantum gate and its active bits, and standardizing the multi-bit gates; constructing a unified fidelity cost model based on the quantum circuit and hardware parameters, and quantizing the cost into an additive cost; spatially partitioning the quantum circuit based on the coupling relationship between the quantum bits, and temporally partitioning the quantum circuit based on the quantum gate sequence to form a set of spatiotemporal sub-blocks; pre-calculating the partitioning cost for each spatiotemporal sub-block, and determining the candidate execution mode for each spatiotemporal sub-block accordingly; using a multi-objective global optimization algorithm to globally optimize the allocation scheme to obtain a set of optimal cooperative execution schemes; and outputting the quantum circuit mapping results.

The invention discloses a non-uniform quantization method and device based on density sensing clustering, equipment and a storage medium, and relates to the technical field of network communication, and the method comprises the steps: obtaining a tap coefficient of an equalizer; performing clustering division on the tap coefficient according to a preset quantization bit width to obtain a plurality of target clusters; determining a quantization parameter corresponding to the target cluster according to the dynamic range of the intra-cluster coefficient in the target cluster, wherein the quantization parameter comprises a scaling factor and a zero offset parameter; and independently and uniformly quantizing the tap coefficient in the target cluster according to the quantization parameter to obtain a recovery coefficient. According to the invention, the coefficient is adaptively divided into a plurality of clusters based on density sensing clustering, and uniform quantization is independently performed on each cluster, so that high-precision quantization of the equalizer coefficient under extremely low bits is completed, and the storage efficiency of coefficient quantization and the performance of the equalizer are improved.

The invention belongs to the technical field of chemical engineering, artificial intelligence and model compression, and discloses a uniform quantification method for a large model in the chemical engineering field. On the premise that the expression capability of a model function is not changed, orthogonal equivalent transformation is introduced, combined smoothing processing is carried out on the weight and activation in a linear layer, and activation distribution approximately obeys uniform distribution in a quantization interval by constructing an orthogonal transformation matrix and systematically remodeling the activation distribution, so that the low-bit uniform quantization error is remarkably reduced. And carrying out joint quantization on the smoothed weight and activation by adopting a uniform low-bit uniform quantizer, further introducing a small number of learnable global orthogonal parameters, and carrying out lightweight optimization on module output errors before and after quantization. According to the method, quantitative perception training or large-scale fine tuning is not needed, calculation and storage overhead is low, engineering implementation is simple, inference performance close to full precision can be kept under the 4-bit or 6-bit quantization condition, and video memory occupation and energy consumption in the model deployment process are remarkably reduced.

The invention provides a post-training quantification method for a medical image segmentation basic model, and belongs to the field of medical images. The method comprises the following steps: acquiring a to-be-quantized medical image segmentation basic model and a pre-training weight; constructing a medical segmentation image standard data set, and performing preprocessing; determining a target bit width, and selecting uniform quantization as a basic quantization mode; obtaining a new weight matrix; alternately solving an integer coding matrix and a scaling factor; and constructing a quantized medical image segmentation basic model, and performing reasoning on the medical segmentation image standard data set. The technical problem that the performance of an existing post-training quantification method is seriously degraded under a low bit width due to the fact that the weight distribution of a medical image segmentation basic model is highly heterogeneous in various medical imaging modes and downstream segmentation tasks is solved; and additional calculation or memory overhead is not introduced in the reasoning stage, so that the deployment feasibility of the quantized medical image segmentation basic model in resource-limited clinical / edge equipment is improved.

The invention provides a low-complexity nonlinear compensation method and system for digital subcarrier multiplexing optical communication, and the method comprises the steps: calculating a nonlinear coefficient of a digital subcarrier multiplexing optical communication system, and carrying out the quantization processing of the nonlinear coefficient through employing a uniform quantization mode; by taking the cross-section length as a processing step length, converting the time domain signal of each subcarrier to a frequency domain for dispersion compensation, and then converting the compensated signal back to the time domain; calculating nonlinear phase noise in the subcarriers based on the time domain signal intensity of each subcarrier by taking the cross-section length as a processing step length, and performing phase rotation compensation; and calculating inter-subcarrier nonlinear phase noise and inter-polarization crosstalk based on the quantized nonlinear coefficient, subcarrier signal strength and inter-polarization crosstalk strength by taking the cross-section length as a processing step length, constructing a nonlinear compensation matrix and performing matrix compensation. According to the invention, the calculation complexity of an optical fiber nonlinear compensation algorithm is reduced, and efficient nonlinear compensation of a digital subcarrier multiplexing optical communication system is realized.

The invention discloses a power distribution network demand response capability dynamic prediction and potential analysis method based on deep learning, and belongs to the field of intelligent power distribution network optimization, the method adopts a CNN-BiLSTM-Transform double-layer model for demand response capability dynamic prediction, and compared with a single traditional model, the method has the advantages that the dynamic prediction efficiency is improved, and the power distribution network demand response capability dynamic prediction and potential analysis efficiency is improved. The complex coupling relationship among the load, the weather and the excitation signal can be described more accurately; multi-dimensional demand response potential calculation is carried out, and a demand response potential source is systematically analyzed from the perspective of a distribution network side and a development trend according to response characteristics of different user types, so that comprehensive identification of adjustable resources is facilitated, and the management precision of a demand side is improved; a long-term and short-term demand response potential quantification mode based on a space envelope domain is provided, the space envelope domain is introduced to quantify the potential, a potential expectation value is obtained, potential changes under different time scales and influence factors can be represented in a unified form, and the defect that a traditional method only carries out qualitative analysis and lacks unified quantitative indexes is overcome.

For example, an apparatus may include an input to receive digital radar Receive (Rx) information corresponding to radar Rx signals, the digital radar Rx information having a first number-of-bits-per-sample; and a noise-shaping quantizer configured to generate quantized radar Rx information by quantizing the digital radar Rx information. For example, the quantized radar Rx information may have a second number-of-bits-per-sample less than the first number-of-bits-per-sample. For example, the noise-shaping quantizer may be configured to generate the quantized radar Rx information having a non-uniform quantization noise spectrum, which has a non-uniform distribution in a frequency domain. For example, the apparatus may include an output to provide the quantized radar Rx information.

The invention relates to the technical field of neural network model quantification, and discloses a neural network model quantification method and device based on mixing precision, and the method comprises the steps: S1, obtaining a floating point model and calibration data, and generating a quantification feature graph according to the weight and activation statistical distribution, the loss disturbance quantity and the hardware computing power storage overhead; a quantitative feature graph is generated by fusing weight statistical distribution, activation statistical distribution, a loss disturbance level and hardware computing power storage overhead in a quantitative feature construction stage, and a non-uniform quantization level and a non-uniform quantization interval of each layer are determined in a difficulty weighting mode under the constraint of a mixed precision template. And error backflow gradient correction driven by a shadow path and an error buffer area and hardware side deployment calibration under a bit configuration fixed condition are combined, so that the effects of keeping the model precision and improving the target hardware reasoning efficiency under a relatively low bit width are achieved, and the dependence of quantitative deployment on artificial experience and multiple rounds of retraining is reduced.

The invention provides a lookup table-based image super-division method and system realized through interpolation and quantization, and the method comprises the steps: carrying out the bilinear interpolation of an input low-resolution image, and obtaining a basic low-frequency image; sequentially inputting the low-resolution image into a plurality of interpolation quantization modules, and finally obtaining a high-resolution residual image through an up-sampling module; adding the high-resolution residual image and the low-frequency image to obtain a final super-resolution image; wherein the input end of the interpolation quantization module passes through the non-uniform quantization module, and the hidden variables are optimized and quantized in a piecewise linear mode; respectively querying a lookup table with low input and output bit widths for two groups of indexes obtained by upward quantization and downward quantization to obtain two paths of outputs, and replacing explicit storage in a double-path fusion interpolation module by utilizing a mode of combining low bit width representation and interpolation; a learnable residual path is introduced into the tail end of the interpolation quantization module, so that the module focuses on high-frequency detail reconstruction, and the quality of a super-resolution result is improved.

The invention discloses a wavelet transform-based gradient quantization and gradient rarefaction fusion method and system for accelerating distributed training, and aims to solve the problems that communication efficiency and model performance are unbalanced and a static strategy is difficult to adapt to a dynamic gradient in an existing gradient compression technology. The core of the method is that a gradient is decomposed into a contour part and a detail part through wavelet transform, a self-adaptive non-uniform quantization algorithm based on KL divergence and mu law compression is adopted for the contour part, and differential quantization strategies are carried out at different layers and different training stages so as to reduce quantization errors; sparsification is carried out on the detail part to retain a key gradient; the bias layer is not compressed to accelerate convergence. Meanwhile, a worker node fault processing mechanism and a calculation-communication time overlapping strategy based on adjacent iteration gradient similarity are designed, and it is ensured that training is stable and efficient. According to the method, the communication traffic is greatly reduced, the training time is shortened, meanwhile, the model accuracy is not lost, and the method is suitable for distributed training of various deep learning models.

This invention provides a method and system for performing noise reduction on the quantized latent representation of an image. [Solution] In system 100, the noise reduction method involves the server system encoding the image into a latent representation in the latent space and performing quantization processing on the quantized latent representation. JPEG2026060879000046.jpg6170 To generate the latent representation y, we perform the quantization process on a uniform noise. The quantization process is then performed on the latent representation y. JPEG2026060879000047.jpg6170 The signal is sent to the receiver. The receiver then uses inverse quantization to recover the latent representation. JPEG2026060879000048.jpg6170 To generate the latent representation, a diffusion model is used that performs denoising over several iterations based on a time step t to remove noise from the reconstructed latent representation. The diffusion model is trained to perform denoising using uniform noise.

The application belongs to the technical field of distributed online optimization and learning, and discloses a quantized distributed online proximal gradient descent optimization method based on Bandit feedback. The method aims to solve the distributed online composite optimization problem under the condition that the communication resource is limited, the network structure is unbalanced, and the gradient information of the loss function is difficult to obtain or even cannot be obtained. The technical scheme comprises the following steps: constructing an unbalanced network graph and determining the node neighborhood relationship, modeling a distributed online composite optimization problem based on Bandit feedback, designing a quantized distributed online proximal gradient descent algorithm combined with an adaptive uniform quantization strategy and a proximal gradient descent technology, and performing static regret analysis on the algorithm.

This invention specifically relates to a quantization method for improving the contrast of SAR images, comprising: calculating the mean of the original SAR image; setting two different brightness quantization factors; calculating the uniform quantization coefficient of the original SAR image; calculating the positions of pixels in the original SAR image that are less than a low brightness threshold; setting separate quantization product factors for high, medium, and low scattering regions; calculating pixels in the original SAR image that are greater than a high brightness threshold and quantizing them using the low brightness quantization coefficient; calculating pixels in the original SAR image that are greater than a high brightness threshold and quantizing them using the high brightness quantization coefficient; and quantizing pixels with intermediate values ​​using a medium uniform quantization method. This method can improve pixel values ​​in strong scattering regions while reducing pixel values ​​in weak scattering regions, thereby improving the overall contrast of the image, enhancing the brightness of strong scattering regions, and laying the foundation for subsequent target interpretation in SAR images.

The invention discloses a bandwidth limited system T-S fuzzy sliding mode control method based on an adaptive WTOD protocol, and belongs to the technical field of industrial control and intelligent manufacturing. The method comprises the following steps: establishing a T-S fuzzy networked system model containing time-varying and time-lag; designing a self-adaptive WTOD transmission protocol integrated with a uniform quantization mechanism; constructing a fuzzy sliding mode controller based on the available system state information; giving out discrimination conditions for ensuring that the indexes of the T-S fuzzy networked system are bounded and the sliding mode surface is reachable in an inequality group form; converting an inequality group discrimination condition containing linear coupling into a minimization problem by adopting cone complementary linearization; and optimizing the final upper bound of the closed-loop system state and the upper bound of the sliding mode surface reaching neighborhood based on a particle swarm algorithm. According to the method, the cooperative improvement of the control performance and the communication efficiency is ensured, and a verifiable and easy-to-implement unified analysis and design framework is provided for the bandwidth-limited networked nonlinear system.

The application discloses a kind of steel tower vibration modal and displacement identification method of video tracking, it is related to engineering asset operation and maintenance management technical field, including the following steps: scale calibration step, based on the calibration parameter of monitoring station and camera, calculate scale calibration coefficient, provide uniform quantitative benchmark for whole link calibration;Input layer scale calibration step, mark point topology optimization layout step, space-time frequency domain interference decoupling step, multi-view heterogeneous feature federal fusion step, low-texture area vibration state reasoning step, micro-vibration signal modal identification step;Closed-loop monitoring control step, incremental model updating and data management step, by dynamically adjusting attention layer receptive field, ensure the accuracy of interference decoupling under different distances, effectively separate camera shake, wind-induced interference and steel tower itself vibration component, solve the problem that micro-vibration information is submerged;At the same time, through scale normalization and dynamic weight adjustment, realize the effective federal fusion of multi-view heterogeneous data.

The application discloses a panoramic video transmission method and device, a storage medium and an electronic device. The method comprises the following steps: receiving a video stream from a sending terminal device, and decomposing the video stream into an image set; dividing the images in the image set into different types of image regions according to the user's visual field, wherein the different types of regions correspond to different visual fields; determining a target type region in the different image regions, encoding the target type region, dividing the encoded target type region into blocks, and obtaining a plurality of block images; compressing the plurality of block images using different quantization steps to obtain a target block image; and sending the target block image and other image regions to a receiving terminal device. The application solves the technical problem of unreasonable occupation of channels and waste of resources in real-time transmission of panoramic video caused by uniform quantization step transmission of panoramic video.

The invention discloses a channel state information coding and decoding method based on double-channel heterogeneity. The method comprises the following steps: firstly, extracting CSI (Channel State Information) features by using a hybrid Transform-CNN (Convolutional Neural Network) compression network by a sending end, and performing unified quantization; according to complexity parameters, quantized symbols are divided into a strong context dependency class and a weak context dependency class, the strong context dependency class is subjected to ASCII discretization and then modeled by a language model and subjected to arithmetic coding, and the weak context dependency class is modeled by a factorization model and subjected to parallel coding. And a receiving end respectively decodes and inversely quantizes the two types of bit streams, and reconstructs the original CSI through a decoder. The method has the advantages of being high in compression rate, low in delay, controllable in structural complexity and the like, is suitable for low-overhead CSI feedback tasks in a large-scale MIMO system, and has wide application prospects.

This invention discloses a temperature semantic-driven method and system for infrared thermal anomaly recognition of unmanned aerial vehicles (UAVs). The method includes: acquiring infrared thermal imaging data containing a physical temperature field and a corresponding visual image; fusing the physical temperature field as an independent encoding channel with the visual image to generate a multimodal feature sequence; calculating the temperature importance weight of spatial locations based on the physical temperature field, and performing partitioned non-uniform quantization compression on the attention feature tensor of a pre-trained temperature-sensing multimodal model to obtain a quantized compression model; classifying hierarchical categories and assigning differentiated computational precision according to the sensitivity of the network layers of the quantized compression model to temperature semantic features to obtain an edge deployment model; and inputting the multimodal feature sequence into the edge deployment model to perform real-time inference to output the thermal anomaly recognition result. This achieves infrared thermal anomaly detection on the UAV's airborne edge side, balancing high recognition accuracy with real-time inference capabilities.

The invention provides a surgical skill intelligent evaluation method based on an endoscope video, and relates to the technical field of artificial intelligence. The invention provides a complete intelligent evaluation process from an original endoscopic surgery video to quantitative skill scoring. The process comprises the following steps: carrying out standardized preprocessing and automatically rejecting irrelevant fragment frames; frame-level feature representations adapting to different operation modes and equipment conditions are obtained through fine tuning in the self-supervision field; on this basis, frame-level identification of various types of anomalies such as bleeding, smoke and poor surgical field exposure is realized by using a multi-scale feature pyramid and channel-space attention, and then time sequence coding is performed on an abnormal process through a time sequence coding network based on a selective state space model so as to extract event-level time sequence features; and finally, carrying out joint modeling on the features and structural metadata such as covering operation types, and outputting a normalized skill score and a corresponding grade. According to the method, the accuracy, generalization and interpretability of intelligent evaluation of the surgical skills are greatly improved through a unified quantization framework and a self-adaptive fusion mechanism.

The invention discloses a time-varying coupling unmanned system multi-formation cooperative control method for communication constraints. Firstly, a discrete time dynamic model and a controller input model of the networked unmanned multi-formation system are established, and formation division and formation reference trajectory definition are carried out at the same time. Secondly, under the condition that the communication bandwidth is limited, a synchronization error dynamic model is constructed, a hierarchical bit rate constraint relation of a system layer, a formation layer and a node layer is established, a uniform quantization coding-decoding mechanism is introduced to quantize a node state, and augmented modeling of an error system is realized by using a Kronecker product. And then, designing a multi-formation cooperative control strategy with fault-tolerant capability, solving controller gain through a linear matrix inequality, and ensuring bounded convergence of synchronization errors. Furthermore, a dynamic bit rate allocation mechanism based on a phototropic growth optimization algorithm is designed, and optimal configuration of communication resources is realized, so that the collaborative consistency and robustness of a multi-formation system under communication limited and complex working conditions are improved.

The application discloses a defect detection method, system, computer device and storage medium, the method comprises the following steps: obtaining the original 3D point cloud data of the target object, and constructing a reference surface model based on the local geometric features of the original 3D point cloud data; based on the reference surface model, the original 3D point cloud data is layered and residual quantization coding is carried out to generate a compressed data stream; the compressed data stream is decoded and reconstructed to obtain the reconstructed 3D point cloud data; based on the reference surface model and the reconstructed 3D point cloud data, the target object is subjected to defect identification, and the defect identification result is output. In the embodiment of the application, the problems of low efficiency of mass data processing and insufficient real-time detection capability are solved through the deep cooperation of reference surface fitting and point cloud compression, the integrity of the key features of the reconstructed point cloud is guaranteed by combining the non-uniform quantization strategy, and the whole-process integrated processing not only improves the defect detection accuracy and system stability, but also meets the demand of industrial online detection on result reliability.

The invention relates to the technical field of large-scale language model reasoning optimization, and discloses a learning-based large-scale model reasoning lookup table construction method, which comprises the following steps of: 1, obtaining an original weight matrix and calibration input data of a large-scale language model, and initializing a quantized value of a lookup table; the lookup table is a two-dimensional array used for storing a pre-calculation product result of the two quantized values. Determining an initial quantized value of a lookup table through a learning-based non-uniform quantization mode, realizing gradient transfer by combining a straight-through gradient estimator, and aggregating weight gradient to optimize the quantized value; meanwhile, a unary function is fused into a lookup table to eliminate intermediate quantization loss, a special lookup table is designed for a softmax function, and an input value is adjusted through mathematical transformation, so that the deviation between quantization model output and original model output is remarkably reduced, and a mean square error loss value between the quantization model output and the original model output is remarkably reduced; and the strict requirement of a large language model on reasoning precision is met.

The application relates to a channel estimation method and system based on a memory-computing integrated device, a processing device and a storage medium, which comprises the following steps: acquiring a pilot structure matrix and a probabilistic graph model of a to-be-tested channel estimation task, and determining the initial probability of the state value of each variable node in the probabilistic graph model according to a pilot signal observed by a receiving end and a variance and of channel uniform quantization; constructing a node state matrix corresponding to each undirected edge in the probabilistic graph model; constructing a driving vector corresponding to each undirected edge in the probabilistic graph model; performing memory-computing integrated device configuration based on the node state matrix corresponding to each undirected edge, so as to obtain a memory-computing integrated device subarray corresponding to each undirected edge; and determining a continuous channel estimation result in the to-be-tested channel estimation task according to the memory-computing integrated device subarray, the driving vector and the confidence of message transmission between each variable node and each check node connected through the undirected edge. The application can be widely used in the technical field of signal processing.

The invention provides a visual model quantification method and device based on non-uniform and spatial perception characteristics, and relates to the technical field of model quantification, and the method comprises the steps: obtaining first activation data and second activation data corresponding to a to-be-quantized initial visual model; quantizing the first activation data by adopting a dynamic self-adaptive logarithm quantizer to obtain first quantized activation data, and updating the dynamic self-adaptive logarithm quantizer; dividing the second activation data into first sensitive area activation data and second sensitive area activation data through the local variance of the second activation data; quantizing the first sensitive area activation data based on the updated dynamic adaptive logarithm quantizer, and quantizing the second sensitive area activation data by adopting a standard uniform quantization strategy to obtain second quantized activation data; and determining a quantized target visual model corresponding to the initial visual model based on the first quantized activation data and the second quantized activation data. According to the method, efficient model compression can be realized while the model precision is ensured.

This invention discloses a distributed XL-MIMO near-field channel estimation method and system based on a cooperative mechanism. In the distributed XL-MIMO near-field system, reference access points and non-reference access points are divided. A unified near-field parameterized expression model for Loss-of-Stake (LoS) and Non-LoS (NLoS) components is constructed by characterizing the coupling relationship between angle and distance parameters. A channel parameter correlation mapping mechanism is established based on the access point deployment topology. LoS component cooperative parameter estimation is completed using a uniform quantization codebook and continuous domain gradient correction. After cancelling the LoS components, a sparse representation model of the NLoS components is constructed, and NLoS component cooperative parameter estimation is achieved through sparse reconstruction and continuous domain gradient correction. Finally, the near-field channel of each access point is reconstructed. This invention reduces parameter dimensionality and pilot overhead, improving the channel estimation accuracy and robustness of the distributed XL-MIMO near-field system.