69 results about "Model inference" patented technology

A configuration model dividing and configuration sub-model inference processing system and procedure addresses the issue of configuration model and query complexity by breaking a configuration problem down into a set of smaller problems, solving them individually and recombining the results into a single result that is equivalent to a conventional inference procedure. In one embodiment, a configuration model is divided into configuration sub-models that can respectively be processed using existing data processing resources. A sub-model inference procedure provides a way to scale queries to larger and more complicated configuration models. Thus, the configuration model dividing and configuration sub-model processing system and inference procedure allows processing by a data processing system of configuration models and queries whose collective complexity exceeds the complexity of otherwise unprocessable conventional, consolidated configuration models and queries.

The invention relates to the technical field of computer vision, in particular to a face forgery detection method based on an image domain and frequency domain double-flow network. The detection method comprises a model training stage and a model inference stage. In the model training stage, a server with high computing performance is used for training a network model, network parameters are optimized by reducing a network loss function until the network converges, and a double-flow network model based on an image domain and a frequency domain is obtained; in the model inference stage, the network model obtained in the model training stage is utilized to judge whether the new image is a face counterfeit image or not. Compared with a method for converting face counterfeit detection into a dichotomy problem, the method realizes hierarchical classification, and can achieve better detection precision by using diversified supervision information of counterfeit images.

A configuration model dividing and configuration sub-model inference processing system and procedure addresses the issue of configuration model and query complexity by breaking a configuration problem down into a set of smaller problems, solving them individually and recombining the results into a single result that is equivalent to a conventional inference procedure. In one embodiment, a configuration model is divided into configuration sub-models that can respectively be processed using existing data processing resources. A sub-model inference procedure provides a way to scale queries to larger and more complicated configuration models. Thus, the configuration model dividing and configuration sub-model processing system and inference procedure allows processing by a data processing system of configuration models and queries whose collective complexity exceeds the complexity of otherwise unprocessable conventional, consolidated configuration models and queries.

A method of compressing a pre-trained deep neural network model includes inputting the pre-trained deep neural network model as a candidate model. The candidate model is compressed by increasing sparsity of the candidate, removing at least one batch normalization layer present in the candidate model, and quantizing all remaining weights into fixed-point representation to form a compressed model. Accuracy of the compressed model is then determined utilizing an end-user training and validation data set. Compression of the candidate model is repeated when the accuracy improves. Via the method, the size of the DNN is reduced, implementation requirements, such as memory, hardware, or processing necessities, are also reduced. Inference can be achieved with much increased speed and much decreasedcomputational requirements. Lastly, due to the disclosed method of compression, these benefits occur while guaranteeing accuracy during inference.

The invention discloses an edge computing node distribution and quit method based on a branch neural network, which is mainly divided into three steps in order to improve the safety of an artificial intelligence model under edge computing and accelerate the computing efficiency of the model: training a neural network model, deploying a branch neural network model at an edge computing node and selecting a model exit point: firstly building the neural network model and combining with a model rectification algorithm to train the model; secondly, in the process of distributing edge computing nodes for the trained classification model, selecting proper edge computing nodes by utilizing a minimum delay algorithm; and finally, selecting an appropriate model exit point in a model inference stage to reduce the calculation amount of an edge calculation node, so that the precision and accuracy of classification network prediction are improved, the effects of improving the safety of a neural network model and accelerating the calculation efficiency of the model are finally achieved, and the purposes of defending against samples and improving the calculation efficiency of the model are achieved.

According to a method in the technical field of information, an intelligent epidemiological investigation system is realized. The system is composed of an intelligent epidemiological front-end interaction subsystem, an intelligent epidemiological model inference subsystem and an intelligent epidemiological map database storage subsystem. The intelligent epidemiological front-end interaction subsystem is mainly responsible for modifying and inserting a traffic scheduling event by traffic scheduling personnel and synchronously visualizing traffic scheduling time and space; the intelligent epidemiological model inference subsystem adopts a multi-language MBERT pre-training model and is a mature technology, the pre-training model is used for inferring an entity of a text, and finally an inference result is stored in the intelligent epidemiological graph database storage subsystem. According to the system, the technical effects of automatically extracting character information of ta key event site from a streaming survey interview text or the questionnaire with relatively high efficiency and accelerating the streaming survey speed are realized.

The invention discloses a multi-model reasoning service deployment method and device based on a k8s cluster. The method comprises the following steps: deploying a scheduling service in a minimum scheduling unit of a k8s cluster, and configuring a memory, computing resources and a scheduling strategy for the scheduling service; deploying a plurality of model inference services according to a memoryof a scheduling service, and configuring each model inference service as a computing resource for using the scheduling service and to be associated with the scheduling service; and the scheduling service calls the plurality of model reasoning services according to the scheduling strategy so as to process the reasoning task. According to the scheme of the invention, the capability of sharing the minimum scheduling unit by multiple model inference services is realized, the multi-model inference service can be elastically expanded and contracted along with the service load, and the deployment operation is relatively simple.

Inference results of a machine learning model and associated inputs are collected. An inference request is received. A determination is made whether a request input of the inference request matches at least one collected input of a set of collected inputs. In response to determining that the request input matches at least one collected input in the set of collected inputs, an inference result is determined using one or more collected inference results associated with said one or more matching inputs in the set of collected inputs.

The disclosure describes various embodiments for quantizing a trained neural network model. In one embodiment, a two-stage quantization method is described. In the offline stage, statically generatedmetadata (e.g., weights and bias) of the neural network model is quantized from floating-point numbers to integers of a lower bit width on a per-channel basis for each layer. Dynamically generated metadata (e.g., an input feature map) is not quantized in the offline stage. Instead, a quantization model is generated for the dynamically generated metadata on a per-channel basis for each layer. The quantization models and the quantized metadata can be stored in a quantization meta file, which can be deployed as part of the neural network model to an AI engine for execution. One or more speciallyprogrammed hardware components can quantize each layer of the neural network model based on information in the quantization meta file.

The invention relates to a multi-factor data reliability evaluation method in fog computing, and belongs to the technical field of mobile communication. And data sharing enables users to obtain important information around in time. However, in a data sharing process, due to the mobility and variability of the users, the users do not completely trust each other. Due to sensor faults, virus infection and even selfish reasons, the equipment may spread error information. In order to reduce the influence of the malicious messages on other users, the propagation of false messages must be inhibited. The invention designs a method to quantify the authenticity of the data, so that a data requester can obtain more reliable data in the data sharing process. And the fog node prejudges the authenticity of the data through Bayesian model inference to obtain the credibility value of the data. The requester obtains the satisfaction degree of the data by integrating the experience-based trust value of the provider and historical interaction on the basis of the fog node pre-judgment result, so that the reliability evaluation of the data is more comprehensive.

There is a need for solutions for more efficient predictive data analysis systems. This need can be addressed, for example, by a system configured to obtain, for each predictive task of a plurality of predictive tasks, a plurality of per-model inferences; generate, for each predictive task, a cross-model prediction based on the plurality of per-model inferences for the predictive task; and generate, based on each cross-model prediction associated with a predictive task, a cross-prediction for the particular predictive task, wherein determining the cross-prediction comprises applying one or more probabilistic updates to the cross-model prediction for the particular predictive task and each probabilistic update is determined based on the cross-model prediction for a related predictive task of the one or more related predictive tasks.

The invention discloses a new energy accommodation method based on the power generation frequency limit value adjustment, and the method gives consideration to the improvement of the operation frequency limit value of a power grid in a reasonable range when a system is in normal operation and in a time period of low load, stabilizes the characteristics of randomness, fluctuation and reverse peak regulation of new energy power generation, and improves the accommodation capacity of the system for new energy power generation. The method comprises the steps: constructing a new energy accommodation model inference model based on the power generation frequency limit value adjustment, wherein the system can accommodate more new energy output through primary frequency modulation during normal operation; accommodating the new energy in a low load time period through reducing the technical output of the conventional energy resource, wherein the new energy, which is not completely accommodated, can be used for solving the change transient process of the frequency and a new steady-state value of the system after new energy grid connection through an inference formula after the conventional energy resource has already reached the minimum technical output. According to the embodiment of the invention, the method improves the new energy accommodation capacity of the power grid during normal operation and the low-load time period through the real-time dynamic adjustment of the frequency of the power grid.

The invention discloses a multi-model inference service deployment method and device based on k8s cluster. The method includes: deploying a scheduling service in the smallest scheduling unit of the k8s cluster, and configuring memory, computing resources and scheduling policies for the scheduling service; deploying multiple model inference services according to the memory of the scheduling service, and inferring each model A service is configured to use computing resources of the scheduling service and is configured to be associated with the scheduling service; the scheduling service invokes the plurality of model inference services to process inference tasks according to the scheduling policy. The solution of the present invention realizes the ability of multiple model reasoning services to share the minimum scheduling unit, and the multi-model reasoning service can be elastically scaled with the service load, and the deployment operation is relatively simple.

The invention discloses a nut surface defect classification method based on EfficientNet, comprising step 1: collecting training samples and making a data set; training the EfficientNet model through the data set, after the model training converges, and then classifying the nut surface; using EfficientNet to classify the model The scale is self-adjusted to avoid time-consuming manual adjustment and it is difficult to obtain the optimal solution; Step 2: Collect the target image and crop the target image; Step 3: Modify the loss function to focal loss, adjust the weight of positive and negative samples, and improve the accuracy of the model; Avoid the impact of positive and negative sample imbalance on the model; step 4: image preprocessing, make a template image to rotate and correct the input image, and make a difference to obtain a difference image, and fuse the three into a three-channel image as the model input image; Enhance the adaptability of the model to the production environment and improve the generalization ability of the model. Step five: model inference, use the EfficientNet model to infer the input image; step six: output the inference image to the host computer for display.

The present application relates to a method, data processing method, apparatus, computer equipment and storage medium for constructing a model inference network. The method includes: acquiring a deep learning network, where the deep learning network includes a plurality of network layers; acquiring test data; performing compilation and detection on each network layer according to the test data, obtaining compilation results and detection results, and according to the compilation of each network layer The results and / or detection results determine the resource allocation strategy of each network layer, build a model inference network according to the resource allocation strategy of each network layer, and obtain the data to be processed; input the data to be processed into the model inference network, and pass the resources in the model inference network through the model inference network. The network layer whose configuration strategy is to optimize the configuration strategy processes the data to be processed, and the network layer whose resource allocation strategy is the original configuration strategy through the model inference network processes the data to be processed, and obtains the processing result of the data to be processed, which improves the performance of the entire network. Data processing efficiency.