12results about How to "Reduce quantization error" patented technology

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.

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.

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.

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.

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.

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.

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.

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.

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.

This invention provides a method for reducing layer norm quantization error, comprising: S1, calculating feature quantization parameters; S2, saving the features output by each layer in the neural network; S3, calculating scaling and bias parameters: S3.1, obtaining feature data in the layer norm; S3.2, calculating scaling factors (scales); S3.3, calculating error bias: S3.3.1, when calculating error bias, the weight γ needs to be updated, the goal of which is to fuse the calculated scaling factors (scales) into the weight Y, denoted as new_gamma; new_gamma = weight γ × scales; S3.3.2, updating the bias correction bias f_Lf, q_LF: f_LF = f_ LF × weight γ; q_LF = q_LF × new_gamma; S3.3.3, use the updated f_LF and q_LF to calculate the bias: calculate the bias by calculating the average error; S3.4, after the scales and bias are calculated, the error after quantization can be reduced by combining them with the weight γ and biasβ fused into the layer norm; the new biasβ is denoted as new_bias; new_bias = biasβ + bias; the bias calibration of one layer of layer norm is now complete; S3.5, update the new weights, i.e., new_gamma and bias, i.e., new_bias, into the corresponding layer norm layer in the quantization network.

This invention relates to a method for weight quantization of large language models based on statistical distribution, belonging to the field of artificial intelligence technology. This method adaptively selects the quantization threshold by analyzing the skewness, kurtosis, and dynamic range of the statistical weight distribution, achieving high-precision sparse preservation of important weights and low-bit compression of minor weights. Simultaneously, it introduces mean error compensation and variance alignment mechanisms, as well as a block-level parallel quantization strategy, to further reduce quantization errors and improve efficiency. This invention addresses the challenges of existing large language models with their massive parameter scale, requiring high-precision computation and substantial resources for inference, making them difficult to deploy on resource-constrained edge devices; and traditional quantization compression techniques suffer from significant accuracy loss, high computational overhead, and poor adaptability. This method achieves 8-bit and 4-bit quantization without retraining, achieving performance close to FP16 accuracy, significantly reducing model storage space and computational resource consumption, making it suitable for efficient deployment of large language models at edge devices.

A low-quality remote sensing image depth hash retrieval method, system, device and medium based on vector quantization, the method comprising: first, obtaining high-quality and low-quality remote sensing images as input, extracting the features of high-quality and low-quality remote sensing images through a deep convolutional neural network respectively, and inputting them into a vector quantization module to quantize them into independent discrete spaces to generate quantized features, then generating hash codes for image retrieval through a deep hash network, and finally obtaining feature representations and constraining the feature representations by applying a loss function, which includes Pairwise Loss, reconstruction loss and cross-entropy loss, to ensure semantic information retention, feature distance constraint and collaborative learning of the encoder, codebook and decoder; the system, device and medium are used to implement the method; the present application improves the retrieval accuracy of low-quality remote sensing images, reduces the storage space occupation and computing overhead, improves the robustness and generalization ability of the model, and ensures high retrieval performance.