Patents
Literature
Patsnap Eureka AI that helps you search prior art, draft patents, and assess FTO risks, powered by patent and scientific literature data.

38results about How to "Reduce quantization error" patented technology

Abnormal value-based weight activation quantification method after large model training

The invention discloses a weight activation quantification method based on abnormal value large model training. The weight activation quantification method comprises a weight quantification process and an activation value quantification process during reasoning. In weight quantization, model bit width is converted into low bit representation in an off-line mode, and in weight activation value quantization during reasoning, a quantized weight model is initialized, input of a linear layer is pre-scaled during reasoning operation, and finally 8-bit quantization operation is executed. According to the method, mathematical equality transformation of weight and activation is combined, a strategy of lossless quantization and low-delay execution is provided, and the significant weight is considered in an off-line weight quantization process, so that the influence of a smooth activation abnormal value on the significant weight is reduced, and the quantization error of the abnormal value is reduced in a model reasoning process. According to the method, effective quantification during large-scale model reasoning is realized, the memory required for running a large model is remarkably reduced, and the running speed is increased.
Owner:THE ACAD OF TIANJIN UNIV HEFEI

Large language model weight quantization method and system based on optimal quantization coding and application

The invention discloses a large language model weight quantization method based on optimal quantization coding, and the method comprises the steps: partitioning, normalization, quantization coding generation, quantization and inverse quantization, and achieves the weight quantization of a large language model, and comprises the following steps: 1, dividing a to-be-quantized weight matrix into a plurality of data blocks; step 2, carrying out normalization processing on elements in each data block; step 3, according to the normalized weight matrix value distribution characteristics, generating a quantization target code by adopting a clustering method; step 4, mapping elements in each data block after normalization processing in the step 2 to the index value of the quantization target code in the step 3 to realize element quantization; and 5, searching a floating-point number corresponding to the index value in the quantization target code, and multiplying the floating-point number with the normalized constant of the data block to obtain an inverse quantization floating-point number. The invention further discloses a system and application for implementing the method, and the system and the application have wide application scenes.
Owner:SHANGHAI QUSU CHAOWEI TECHNOLOGY CO LTD

Point cloud coding method based on multilevel ball octree and graph-driven attention entropy model

The invention provides a point cloud coding method based on a multilevel spherical octree and a graph-driven attention entropy model. The method comprises the following steps: coding point cloud data by using an octree entropy model based on multilevel spherical coordinates; constructing an adjacent matrix of the parent graph and the distance graph; a graph convolutional network module is used for assisting context information embedding by utilizing an adjacent matrix; constructing a grouping graph attention module and a cross attention module to learn the correlation between a parent node context and a peer node context; predicting the probability of a placeholder symbol of each octree node; compressing the placeholder sequence into a binary floating-point number sequence and converting the binary floating-point number sequence into a bit stream; and converting the bit stream into a placeholder symbol sequence of the octree, reconstructing the octree and restoring the point cloud. According to the invention, through the combination of the multilevel spherical coordinate octree structure, the graph convolution, the grouping graph attention module and the cross attention module, quantization errors are reduced, the compression efficiency is improved, the calculation complexity and the compression effect are balanced, and the adaptability to high-resolution point cloud data is significantly enhanced.
Owner:SUN YAT SEN UNIV

Calibration data selection method for deep learning model post-training uniform quantization

The invention discloses a calibration data selection method for training uniform quantization after a deep learning model. The method comprises four steps of clustering centroid initialization, clustering centroid updating, training set data sorting and calibration set data selection. Firstly, random data conforming to uniform distribution are sent to a network for forward reasoning, and a plurality of clustering centroids are determined; then, reasoning is performed through training set data, a knowledge distance is calculated, a clustering centroid is updated, and an activation distribution space accurately reflecting training set data distribution is constructed; then, training set data are sent to the network for reasoning again, knowledge distances are calculated, summation is carried out, and scores of all data samples are obtained and sorted; and finally, according to the size of a specified calibration set, selecting training set data with relatively high scores as the calibration set from high to low. According to the method, through reasonably selecting calibration data, the effect of training uniform quantification after the deep learning model can be improved.
Owner:合肥瑞徽人工智能研究院有限公司

Instruction fine tuning method for quantized multi-modal large model, electronic equipment and storage medium

ActiveCN119089967AImprove task performanceSolve problems that cannot be effectively handledBiological modelsData setAlgorithm
The invention discloses an instruction fine tuning method for a quantized multi-mode large model. The instruction fine tuning method comprises the following steps: step 1, scale learning of quantized perception; 2, a multi-mode warm-up strategy used for scale learning; and step 3, carrying out specific task learning by using the quantized multi-modal large model. According to the method, scale learning of quantization perception is adopted, quantization errors of abnormal positions can be minimized in a self-adaptive manner, and particularly, the quantization errors in each quantization group can be effectively reduced under the condition that certain position activation represents abnormal value features; therefore, the problem that quantization errors caused by outliers cannot be effectively processed in the prior art is solved; according to the method, through a uniform quantification method, weight cutting is adopted to relieve quantification difficulty occurring in a language task; according to the method, after iteration reaches a certain number of times, a multi-modal instruction data set is replaced with a mixed data set, so that language data is gradually introduced while multi-modal task performance is maintained, and overfitting is avoided.
Owner:XIAMEN UNIV

Adaptive gain control method and device, storage medium and electronic device

PendingCN119110382AReduce quantization errorSolve the problem of quantization error in automatic gain controlPower managementModulated-carrier systemsComputer hardwareCarrier signal
The embodiment of the invention provides an adaptive gain control method and device, a storage medium and an electronic device, signals from second user equipment and third user equipment are received through first user equipment, the first user equipment and the second user equipment use a first subcarrier interval to communicate with a first type of service scheduling unit, and the first type of service scheduling unit communicates with the third user equipment through a second subcarrier interval. The first user equipment communicates with the third user equipment by using a second subcarrier interval and a second type of service scheduling unit, and the signal comprises the first type of service scheduling unit and the second type of service scheduling unit; and the first user equipment performs one or more times of adaptive gain control on the signal. Through the method and the device, the problem that quantization errors exist when automatic gain control is carried out on the received signals in the prior art is solved, and the effect of reducing the quantization errors when the user equipment carries out automatic gain control on the received signals is achieved.
Owner:ZTE CORP

Battery unit internal resistance detection circuit and detection method

According to the battery unit internal resistance detection circuit and detection method provided by the invention, the working frequency of the first switch is controlled, and the disturbance current with a certain frequency is generated on the battery unit based on the discharge current generation circuit, so that the voltage on the battery generates certain disturbance; the disturbance quantity of the voltage is converted into a first digital signal through a first analog-to-digital converter, and the internal resistance of the corresponding battery is detected based on the first digital signal and the bleeder current or a second digital signal converted from a current sampling signal of the bleeder current. Through the arrangement, compared with the mode that the battery voltage is directly sent to the first analog-to-digital converter to be converted, the difference value between the battery voltage and the reference voltage is sent to the first analog-to-digital converter to be converted, so that the voltage range to be quantized by the first analog-to-digital converter is small, the resolution is improved, the quantization error is reduced, and the detection precision is improved. In this way, the use of a blocking filter capacitor is avoided, the area is saved, and integration and miniaturization are facilitated.
Owner:JOULWATT TECH INC LTD

Dual-polarized 2-bit intelligent metasurface antenna unit

PendingCN119009435AReduce quantization errorImprove pointing accuracyAntenna supports/mountingsAntennas earthing switches associationDielectric substrateControl circuit
The invention relates to a dual-polarized 2-bit intelligent metasurface antenna unit, which belongs to the technical field of wireless communication and comprises a patch, a floor with a coupling slot, two layers of dielectric substrates, two vertically distributed microstrip lines and a metal reflecting plate, the upper layer patch comprises a square patch; an H-shaped groove and an I-shaped groove are etched in the floor; a microstrip line is arranged below each slot, each microstrip line comprises a common main line and four branch delay lines, and the common main line is connected with the branch delay lines through SP4T switches; the working state of the unit is controlled by an SP4T switch, two-bit four different phase state configurations can be provided, two polarizations can be independently regulated and controlled, and the two polarizations work at the same time; according to the invention, the two SP4T switches are utilized to realize dual-polarization 2-bit phase regulation and control, the design difficulty of a control circuit and the antenna cost are reduced, and the pointing precision of an antenna beam is improved.
Owner:THE 54TH RESEARCH INSTITUTE OF CHINA ELECTRONICS TECHNOLOGY GROUP CORPORATION

Model effect evaluation system and method based on quantitative reasoning

The invention discloses a quantitative reasoning-based model effect evaluation system and method, and relates to the technical field of quantitative reasoning. The evaluation system comprises a model loading and preprocessing module, a quantitative processing module, a reasoning process simulation module, a performance evaluation module and a result feedback module. According to the technical key points, dynamic adjustment and optimization strategies are implemented according to conditions, various test sets are simulated for multiple times, and whether the performance of a model is stable or not is accurately judged after calculation; once it is detected that the comprehensive effect estimation exceeds an error range, an adjustment and optimization process is triggered, optimization is carried out for weight distribution, and rationality of quantization parameters is ensured; if the problem still exists, the model precision is further adjusted, and the optimal performance point is gradually approached; in the process, a dynamic adjustment formula is also introduced, the automatic adjustment of the precision and the adjustment amplitude is realized, the timeliness of the adjustment is ensured, and the instability caused by excessive adjustment is also avoided.
Owner:杭州点存科技有限公司

Vector quantization-based low-quality remote sensing image deep hash retrieval method, system and equipment and medium

The invention discloses a low-quality remote sensing image deep hash retrieval method, system and device based on vector quantization and a medium, and the method comprises the steps: firstly obtaining high-quality and low-quality remote sensing images as input, respectively extracting the features of the high-quality and low-quality remote sensing images through a deep convolutional neural network, and inputting the features into a vector quantization module; the method comprises the following steps of: quantifying the image data into an independent discrete space, generating quantized features, generating a hash code for image retrieval through a deep hash network, finally obtaining feature representation, and constraining the feature representation by applying a loss function which comprises Pairwise Loss, reconstruction loss and cross entropy loss. The collaborative learning module is used for ensuring semantic information retention, feature distance constraint and collaborative learning of an encoder, a codebook and a decoder; the system, the equipment and the medium are used for implementing the method. According to the method, the low-quality remote sensing image retrieval precision is improved, the storage space occupation and the calculation overhead are reduced, the robustness and the generalization ability of the model are improved, and the relatively high retrieval performance is ensured.
Owner:XIDIAN UNIV

Pixel merging method and related image readout circuit

ActiveCN113630566BReduce quantization errorRadiologyNuclear medicine
The present invention discloses a pixel merging method for processing pixel data from an image sensor including a pixel array, wherein the pixel merging method comprises executing a first scanning process, wherein the first scanning process is to scan a sensing area of ​​the image sensor to obtain a first amount of pixel data; executing a second scanning process, wherein the second scanning process is to scan the sensing area after the first scanning process is completed to obtain a second amount of pixel data; and performing the pixel merging on the second amount of pixel data according to an offset value, a merging size and an arithmetic value, wherein the offset value and the arithmetic value are determined according to the first amount of pixel data.
Owner:NOVATEK MICROELECTRONICS CORP

Method and device for rapidly and intelligently identifying tunnel diseases and medium

The invention provides a rapid intelligent identification method and device for tunnel diseases and a medium, and relates to the field of target classification, machine vision and intelligent identification, and the method comprises the steps: constructing a sample library through obtained unlabeled ground penetrating radar images; constructing a self-supervised contrast learning network and a disease detection network; training a self-supervised contrast learning network through the sample library, and extracting tunnel ground penetrating radar wave features of the ground penetrating radar image in the sample library to obtain feature extraction network parameters; the feature extraction network parameters are migrated to a backbone feature module of the disease detection network; screening and marking disease radar response data in the sample library, and constructing a tunnel disease ground penetrating radar image set; training the disease detection network according to the tunnel disease ground penetrating radar image set; and obtaining a to-be-detected tunnel disease ground penetrating radar image, inputting the to-be-detected tunnel disease ground penetrating radar image into the trained disease detection network, and identifying a tunnel disease result in the to-be-detected tunnel disease ground penetrating radar image.
Owner:广西南宾公路建设发展有限公司 +2

Big language model binarization quantification method and system based on instructive alternate optimization

PendingCN122065892AHigh precisionReduce quantization errorComputer simulationsLinguistic modelAlgorithm
The invention provides a large language model binarization quantification method and system based on instructive alternate optimization, and relates to the technical field of large language model deploying.The method comprises the steps that a weight matrix is divided into an important area and an unimportant area by evaluating the influence degree of all weight parameters on the performance of a large language model; carrying out binaryzation on all areas of the weight matrix, and alternately optimizing a row vector scaling factor and a column vector scaling factor by adopting a first-order row-column alternate optimization iteration mode to obtain a first-order reconstruction weight matrix; and carrying out binaryzation again on the important region of the first-order reconstruction weight matrix, carrying out optimization by adopting a first-order and second-order row-column alternating optimization iteration mode to obtain a second-order reconstruction weight matrix, and correspondingly taking the second-order reconstruction weight matrix as a weight parameter after the large language model is quantized. The quantization error of the key weight parameter can be effectively reduced, the quantization precision is high, the quantization process completely depends on the internal structure information of the large language model, and the edge deployment compatibility is high.
Owner:GUANGDONG UNIV OF TECH

Carbon emission reduction accounting method and system based on multi-modal uncertainty quantification

The invention provides a carbon emission reduction accounting method and system based on multi-modal uncertainty quantization, and relates to the technical field of carbon emission reduction accounting, and the method comprises the following steps: collecting and preprocessing the multi-modal data of a reconstructed building, and obtaining a preprocessed multi-modal data set; uncertainty modeling is carried out on the preprocessed multi-modal data set, and uncertainty distribution is obtained; based on the preprocessed multi-modal data set and uncertainty distribution, performing inversion by adopting a Bayesian framework to obtain a multi-modal data posterior probability; and performing Monte Carlo simulation on the posterior probability of the multi-modal data to obtain an uncertainty weight, and calculating the carbon emission reduction based on the uncertainty weight. According to the method, quantitative modeling is carried out on the uncertainty of the multi-modal data, the probability transmission and fusion of the uncertainty are realized by using the Bayesian framework, and finally the carbon emission reduction amount accounting result with the confidence interval is output, so that the scientificity and the credibility of the accounting are remarkably improved.
Owner:THE SECOND CONSTR OF CHINA CONSTR EIGHTH ENG DIV

A Label-Enhanced Supervised Multimodal Hash Retrieval Method and System

This invention discloses a supervised multimodal hash retrieval method and system based on label enhancement, belonging to the field of artificial intelligence and multimedia retrieval technology. The technical problem this invention aims to solve is how to better capture the similarity information between multimodal data points and achieve better performance and accuracy in multimodal retrieval tasks. The technical solution includes: data preprocessing: acquiring and organizing public datasets of image and text modalities, and dividing each public dataset into training, testing, and retrieval datasets; extracting deep features: using a pre-trained network model to extract features from the raw data of the public datasets of image and text modalities respectively, obtaining deep features of the image modality and the text modality; offline training; variable update and optimization; and online query. The system includes a data preprocessing unit, a feature extraction unit, an offline training unit, a variable update and optimization unit, and an online query unit.
Owner:QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES)

Quantification method and device for online handwritten character recognition model and storage medium

PendingCN119106714AReduced bit precisionSpeed ​​up the reasoning processBiological modelsAlgorithmReconstruction method
The embodiment of the invention provides a quantization method and device for an online handwritten character recognition model and a storage medium, and the method comprises the steps: inserting a quantizer for calculating quantization parameters into a pre-trained convolutional neural network model, performing quantization processing on the weight and the input activation value of each convolutional layer and full connection layer in the convolutional neural network model to generate a quantization model; and reconstructing the quantitative model by adopting a layer-by-layer reconstruction and / or block-by-block reconstruction method, calculating loss through multiple iterations, and gradually optimizing the quantitative model until a target model is generated. According to the embodiment of the invention, high performance can be maintained while the bit precision of the neural network model is reduced, the reasoning process of the network is accelerated, in addition, the bit width can be reduced, and an effective way is provided for the deep neural network to adapt to different hardware conditions.
Owner:ANHUI UNIVERSITY OF TECHNOLOGY

CSI (Channel State Information) quantification method and device, electronic equipment and computer readable storage medium

PendingCN121966632APreserve local structureReduce quantization errorSpatial transmit diversityChannel state informationComputer engineering
The invention provides a channel state information (CSI) quantization method and device, electronic equipment and a computer readable storage medium. The method comprises the following steps: compressing CSI to be fed back to obtain a first compressed channel vector; dividing the first compression channel vector into M first sub-compression channel vectors; obtaining M codebooks, wherein the M codebooks are in one-to-one correspondence with the M first sub-compression channel vectors; and for each first sub-compression channel vector, quantizing the first sub-compression channel vector based on the codebook of the first sub-compression channel vector to obtain a quantization vector of the first sub-compression channel vector, and splicing the quantization vector to obtain a quantized compression channel vector. According to the embodiment of the invention, a compression channel vector of CSI can be divided into M first sub-compression channel vectors with reduced dimensions, and then each first sub-compression channel vector is independently quantized by adopting a corresponding codebook, so that the local structure of data can be better reserved, the quantization error can be reduced, and the quantization effect is improved.
Owner:CHINA MOBILE COMM LTD RES INST +1

Thick plate structure grain size quantification method based on wake waves

The invention provides a wake wave-based thick plate structure grain size quantification method, and belongs to the field of thick plate structure grain size quantification, and the method comprises the steps: S1, carrying out the pulse signal excitation of a thick plate structure, and collecting a wake wave signal; s2, obtaining a preliminary noise reduction signal based on the collected wake wave signal; s3, variational mode decomposition is carried out on the preliminary noise reduction signal, and a limited bandwidth signal with a specific frequency is extracted; s4, fitting the attenuation coefficient by using the average power envelope of the limited bandwidth signal; and S5, obtaining a grain size quantification result considering uncertainty by using a logarithmic linear quantification model. According to the method, accurate and rapid grain size quantification can be carried out on a thick plate structure with a relatively large frequency thickness product by utilizing the characteristics of wake waves; the overall energy attenuation rule of the signal related to the grain size is effectively described through the average power envelope, the obtained wake wave attenuation coefficient can be used for accurately and reliably quantifying the grain size of the thick plate structure, and the accuracy of the prediction result is ensured.
Owner:BEIHANG UNIV +1

Pose determination method and apparatus, terminal device, and computer-readable storage medium

The present disclosure provides a pose determination method and device, terminal equipment and computer readable storage medium. The method comprises: obtaining a first image corresponding to an object photographed by a first camera and a second image corresponding to the object photographed by a second camera, the photographing angles of the first camera and the second camera being different; determining a plurality of first coordinates and a plurality of second coordinates, the first coordinates being coordinates obtained by projecting and quantifying a preset space coordinate according to camera parameters of the first camera, and the second coordinates being coordinates obtained by projecting and quantifying the preset space coordinate according to camera parameters of the second camera; and determining a pose of the object according to the first image, the second image, the plurality of first coordinates and the plurality of second coordinates. The accuracy of pose determination is improved.
Owner:BEIJING ZITIAO NETWORK TECH CO LTD

A foreground analog calibration method applied to a capacitive bridged DAC

ActiveCN116155293BAdjust bit capacitance ratioReduce proportional mismatchElectric signal transmission systemsDigital-analogue convertorsCapacitanceSoftware engineering
The application discloses a foreground analog calibration method applied to a capacitive bridging type DAC, and comprises the following steps: inputting a sinusoidal signal into a SAR ADC to obtain a current digital code output by the SAR ADC; performing histogram distribution statistics on the current digital code; detecting whether the current digital code satisfies preset monotonicity based on N(i), and determining a connection mode of a calibration capacitor of a current capacitor in an LSB array of the bridging DAC according to a detection result; determining a calibration capacitor to be connected from all calibration capacitors of the current capacitor; returning to the step of inputting the sinusoidal signal into the SAR ADC after connecting the calibration capacitor to be connected to the bridging DAC according to the connection mode, and gradually reducing a quantization error of the DAC, so that INL and DNL satisfy error requirements. Meanwhile, the application takes a protection isolation capacitor filled around the LSB array as the calibration capacitor, adjusts a bit capacitor ratio of the capacitor array through a series-parallel calibration capacitor mode, reduces a proportion mismatch of the capacitor array as a whole, and reduces a non-linear error of the circuit.
Owner:XIDIAN UNIV

Optical measurement system and application method for grinding

The present invention discloses an optical measurement system and application method for grinding. The system includes an illumination light source, a support platform, a projection optical component, a display screen, and a computer. A workpiece to be processed is placed on the support platform. During grinding, the illumination light source illuminates the workpiece to be processed, and the projection optical component projects the image of the workpiece to be processed on the display plane of the display screen. The computer controls the display plane to display a pre-stored workpiece reference scale graphic, and positions it on the positioning baseline, compares the workpiece reference scale graphic with the projection of the workpiece to be processed, adjusts the displacement of the grinding equipment according to the comparison information, and performs grinding. When high-precision measurement is required, the present invention can partially enlarge the boundaries of the workpiece reference scale graphic and the workpiece projection image, and improve the accuracy by subdividing the display pixels, which can well meet the requirements of high-precision workpiece grinding for online detection.
Owner:GUANGZHOU YUXIN OPTOELECTRONIC TECH RES INST CO LTD

Analog-to-digital conversion circuit, control method, chip and electronic device

ActiveCN117335807BImprove conversion accuracyReduce quantization error
The application provides an analog-to-digital conversion circuit, a control method, a chip and an electronic device, and belongs to the technical field of electronics. The analog-to-digital conversion circuit comprises a first analog-to-digital conversion module, a second sigma-delta analog-to-digital conversion module and a control module. The first analog-to-digital conversion module is configured to periodically convert an input signal according to a first conversion period and output a first conversion result. The second sigma-delta analog-to-digital conversion module is configured to convert the input signal in a first number of second conversion periods and output a second conversion result, wherein the first number is an integer greater than 1. The control module is configured to determine a target conversion result of the input signal according to the second conversion result and at least two first conversion results. The application can improve the conversion accuracy of the analog-to-digital conversion circuit.
Owner:CHIPSEA TECH SHENZHEN CO LTD

A method and system for post-training channel-mixed precision quantization of neural networks

PendingCN122088573AReduce quantization errorAvoid large subsequent quantization errorsHardware monitoringBiological modelsEngineeringNetwork model
This invention relates to the field of deep learning technology, and more particularly to a method and system for post-training channel-mixed precision quantization of neural networks. For each weight channel in each layer of the pre-trained model to be quantized, the invention calculates the initial scaling factor of the weight channel at different bit widths based on the weight range of the weight tensor of that weight channel; optimizes the initial scaling factor of each weight channel in each layer of the pre-trained model to be quantized at different bit widths to obtain the target scaling factor of the weight channel at different bit widths; constructs an optimal bit width allocation integer linear programming problem based on the obtained target scaling factor; obtains the optimal bit width allocated to each weight channel by solving the optimal bit width allocation integer linear programming problem; and quantizes the pre-trained model to be quantized based on the obtained optimal bit width of the channels. This invention effectively improves the processing accuracy of the quantized neural network model for data such as text, images, and audio.
Owner:SUZHOU UNIV

Signal processing method, baseband module, antenna module, transmitting device, storage medium and program product

The invention provides a signal processing method, a baseband module, an antenna module, a transmitting device, a storage medium and a program product, and relates to the technical field of communication. According to the scheme, service requirements of different users can be reflected by performing power distribution on each beam, the baseband module acquires the power distribution information of each beam, and the power distribution information can be used for realizing differential distribution of the transmitting power of each beam, so that compared with a mode of performing average power distribution on an antenna module side, the service requirements of different users can be reflected. According to the scheme, a more refined power control strategy can be provided, so that signal transmission is optimized, and the overall performance of the system is improved. Moreover, in the scheme of the invention, the baseband signal and the power distribution information are separated by the baseband module and then are transmitted to the antenna module, so that the quantization error can be effectively reduced on the premise that the quantization bit number is fixed.
Owner:SICHUAN CHUANGZHI LIANHENG TECH CO LTD

Quantization method, device and system of neural network model, electronic device and storage medium

The disclosed embodiments disclose a quantization method, device and system, electronic device and medium of a neural network model, wherein the method comprises: obtaining the weight matrix of any network layer to be quantized in the neural network model; performing matrix transformation on the weight matrix of any network layer to be quantized to obtain the weight matrix to be quantized of any network layer to be quantized; quantizing the weight matrix to be quantized of any network layer to be quantized to obtain the quantized weight matrix of any network layer to be quantized; and obtaining the quantized neural network model based on the quantized weight matrix of any network layer to be quantized in the neural network model. The disclosed embodiments can reduce the distribution difference of weight data of each channel in the weight matrix, can reduce quantization error, and help improve the accuracy of the quantized neural network.
Owner:NANJING HOUMO TECH CO LTD

Cache status report processing method, device and equipment and readable storage medium

The embodiment of the invention provides a cache state report processing method, device and equipment and a readable storage medium, and the method comprises the steps: generating a cache value table according to configuration information; generating a cache state report according to the cache value table; wherein the configuration information comprises first information and second information, the first information comprises the number of segment points, the cache level of each segment point and a cache value corresponding to the cache level of each segment point, and the segment points are points with the fastest first-order derivative change on an accumulative distribution function of the cache of the terminal; the second information comprises at least two of the following items: a cache value corresponding to the lowest cache level, a cache value corresponding to the highest cache level, and a step length between adjacent cache levels.
Owner:CHINA MOBILE COMM LTD RES INST +1

Key value processing method and device, text processing method and device, equipment and storage medium

PendingCN122087626AReduce quantization errorimprove accuracyInference methodsComputation processEngineering
The embodiment of the invention provides a key value processing method and device, a text processing method and device, equipment and a storage medium. The key value processing method comprises the steps that in the key value calculation process of a self-attention mechanism of a large language model, each first key value in a plurality of first key values is converted into a corresponding second key value, the first key value comprises a first key and a first value, the second key value comprises a second key and a second value, and the second key value comprises a first value and a second value; the data range of the plurality of second keys is smaller than the data range of the plurality of first keys, and the data range of the plurality of second values is smaller than the data range of the plurality of first values; the outlier degree of the outliers in the plurality of second keys is smaller than the outlier degree of the corresponding outliers in the plurality of first keys, and the outlier degree of the outliers in the plurality of second values is smaller than the outlier degree of the corresponding outliers in the plurality of first values; and performing numerical quantization on the second key value to obtain a corresponding third key value, and storing the third key value. According to the method, the accuracy of a big language model reasoning result is improved.
Owner:ALIBABA CLOUD COMPUTING CO LTD

A robust optimization method for conditional diffusion model

The present application discloses a method for optimizing the robustness of a conditional diffusion model. The method inputs a noise-added information data set and preset conditions into a conditional neural network for training, and inputs a noise-added information data set into an unconditional neural network for training; obtains the fitting errors generated by the trained conditional neural network and the trained unconditional neural network during the training process, and obtains the quantization errors generated by the trained conditional neural network when the input conditions are wrong in the reverse process of the conditional diffusion model; inputs the fitting errors and quantization errors into a constructed filter to calculate the target parameters; inputs the trained conditional neural network, the trained unconditional neural network and the target parameters into a constructed controller to generate a control signal; inputs the Gaussian distribution and the control signal into a constructed generation model to obtain the generation data output by the generation model. The present application can improve the reliability and effectiveness of the model.
Owner:NAT UNIV OF DEFENSE TECH

A retrieval method, device, equipment and storage medium based on hash code

ActiveCN114417251BReduce quantization errorCosine similarity accurately reflectsStill image data queryingComplex mathematical operationsCosine similarityFeature vector
The embodiment of the present application discloses a retrieval method, device, equipment and storage medium based on hash code, and the relevant embodiments can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, etc., to improve the accuracy of cosine similarity between hash code quantized information. The method of the embodiment of the present application includes: obtaining the first eigenvector corresponding to the information to be processed, using the target projection matrix and the target rotation matrix to calculate the first eigenvector to obtain the first score vector, for each score element in the first score vector, the score element is numerically compared with each score element that appears before the score element to obtain the encoding value corresponding to the score element, wherein the encoding value corresponding to each score element is 0 or 1, and according to the encoding value of each score element in the first score vector, the first target hash code corresponding to the information to be processed is generated, and the first target retrieval information is determined according to the first target hash code.
Owner:TENCENT TECHNOLOGY (SHENZHEN) CO LTD

Image processing model compression method, apparatus, and device for terminal deployment

This application discloses an image processing model compression method, apparatus, and device for terminal deployment, belonging to the field of computer technology. Based on a first quantization step size, the floating-point model parameters of the image processing model to be compressed are quantized and dequantized to obtain dequantized floating-point model parameters; the first quantization step size is mapped according to a target mapping relationship to obtain first trainable parameters; the image processing model configured with the dequantized floating-point model parameters is trained, and during training, the floating-point model parameters and the first trainable parameters are adjusted; the adjusted first trainable parameters are demapped according to the target mapping relationship to obtain a second quantization step size; the adjusted floating-point model parameters are quantized based on the second quantization step size to obtain second integer model parameters; the second integer model parameters are configured for the image processing model, and the second quantization step size is configured into the image processing model, thereby improving the model compression effect.
Owner:BEIJING SANKUAI ONLINE TECH CO LTD