163results about How to "Improve training efficiency" patented technology

The invention discloses a machine learning model training method and device. The method comprises the following steps: on the basis of initialized first weight and second weight of each sample in a training set, and with features of each sample as granularity, training a machine learning model; on the basis of prediction loss of each sample in the training set, determining a first sample set, where corresponding target variables are predicated inaccurately, and a second sample set, where corresponding target variables are predicated accurately; on the basis of the prediction loss of each sample in the first sample set and the corresponding first weight, determining overall prediction loss of the first sample set; on the basis of the overall prediction loss of the first sample set, improving the first weight and the second weight of each sample in the first sample set; and inputting the updated second weight of each sample in the training set and features of each sample and the target variables into the machine learning model, and with the features of each sample as granularity, training the machine learning model. Through the machine learning model training method and device, prediction accuracy and training efficiency of the machine learning model can be improved.

The invention discloses a method for identifying a Web named entity based on a statistical model. The method comprises the following steps of: representing multiple characteristics of the Web named entity with structure and text characteristics; combining a statistical method with a rule method and adopting an improved MR-GHMM (MR-Generalized Hidden Markov Model) to increase the training efficiency; marking the entity with the improved GHMM, and marking each named entity to realize entity identification; and processing a Web complex named entity identifying process on two layers and performing complex nested entity identification by taking a marking result of a first layer as the input of second layer processing. Compared with an original identifying algorithm, the method has the advantages that: the identifying accuracy of an algorithm used in the method is increased, and the time complexity of model training is lowered greatly. By representing multiple characteristics of the Web named entity and modifying entity characteristics in different fields, named entities in different fields on Web can be identified.

The present application relates to a model training method. The method includes acquiring a first sample set, wherein the first sample set includes a first sample image and behavior information of thefirst sample image, and the behavior information is used to indicate behavior action of a virtual object; training the model through the first sample set to obtain an initial model; acquiring a second sample set including a second sample image and behavior information of the second sample image, the second sample image being a scene picture image when the virtual object is controlled by an initial model; the initial model being retrained by the second sample set to obtain the object control model. It does not need long-time on-line training, nor need to prepare a large number of training samples, only a small number of training samples need to be prepared initially, and the subsequent samples in the training process of the initial model execution results are modified to obtain, so as to greatly improve the training efficiency of the virtual object control machine learning model in the virtual scene.

The invention relates to the work fixture technology, in particular to a multifunctional welding piece fixture dedicated to welding training, comprising a pedestal, an upright post, a transverse arm, a clamping part; the bottom of the upright post and the pedestal are connected into a whole, the transverse arm comprises a transverse beam and a sleeve, one end of the transverse beam is connected with the sleeve, and the other end of the transverse beam is connected with the clamping part, the sleeve is sheathed on the upright post, the outside of the sleeve is provided with a tapped hole penetrating through the pipe, the tapped hole is connected with a sleeve fastening handle by a thread, and the sleeve fastening handle is screwed up to clamp the sleeve on the upright post in a firm way. The invention can adjust the height and angle of the welding piece; a pipe fixture is connected with the transverse arm and the transverse beam by a bolt, a plate fixture is connected with the transverse arm and the transverse beam in a suiting way, thus being capable of changing the welding position of the specimen and realizing simulating multiple welding positions; the clamping operation is convenient and simple, thus improving training efficiency; the clamping is steady, and the specimen is uneasy to sway and fall off, thus being safe and reliable.

The invention discloses an automated training device for behavior learning of non-human primates, wherein a personal computer (PC) is provided with two displays, wherein one of the displays and a camera are respectively positioned in front of an animal seat; a water supply control system comprises a water receiver, a driving control circuit and an electromagnetic valve, wherein the water receiver is connected with the electromagnetic valve through a guide tube, the electromagnetic valve is electrically connected with the driving control circuit, and the driving control circuit is electrically connected with a microcontroller; a swinging rod and a water faucet are fixed on the animal seat, a pressure sensor is fixed on a handle of the swinging rod, the pressure sensor and a potentiometer of the swinging rod are electrically connected with the microcontroller respectively, and the water faucet is connected with the electromagnetic valve of the water supply control system through the guide tube; and the PC is electrically connected with the camera and the microcontroller respectively. The invention can realize a painless automated animal behavior training method, and the camera can be used for displaying and monitoring the animal behavior in real time, thereby greatly improving the training speed and saving human resources.

The invention discloses a mechanical structure for a multi-axis wireless movement method of a flight simulator. The mechanical structure employs a plurality of gyroscope-type ring structures and comprises an outer ring, an inner ring, a central ring, a base, servo motors, a simulator cockpit, a power supply system and a wireless receiving and transmitting apparatus, wherein the base is fixedly connected with the outer ring; an X-direction servo motor and an X-direction support shaft are connected with a Y-direction support shaft together so as to control X-direction rotation; a Y-direction servo motor on the Y-direction support shaft and the Y-direction support shaft are staggered from X direction by 90 degrees, and are internally connected with the central ring so as to control Y-direction rotation; a Z-direction servo motor and a Z-direction support shaft on the central ring are staggered from Y direction by 90 degrees, and are internally connected with the simulator cockpit so as to control Z-direction rotation. The simulator cockpit can rotate around three axes of X, Y and Z in a three-dimensional coordinate, thus movement change situation of an airplane can be simulated. The mechanical structure for the multi-axis wireless movement method of the flight simulator disclosed by the invention has the advantages of improving the training efficiency, saving the training cost, guaranteeing flying security, reducing pollution, being simple in structure and low in cost.

The invention relates to a language recognition method and device based on a language model, a text classification method and device, computer equipment and a storage medium. The method comprises thesteps of obtaining a training word vector corresponding to a training statement, and inputting the training word vector into a to-be-trained first model and a trained second model to obtain a featurematrix output by each first network layer of the first model and a feature matrix output by each second network layer of the second model, wherein the first network layers and the second network layers are in one-to-one correspondence, and the network layer number of the first model is smaller than that of the second model; performing similarity calculation on the feature matrix output by each first network layer and the feature matrix output by the second network layer corresponding to each first network layer; and obtaining each similarity, adjusting model parameters of the first model basedon each similarity until the updated target similarity meets a convergence condition, obtaining a trained first model, and performing language recognition through the first model. The method can improve the model training efficiency.

The invention discloses a dense stacking target detection method based on automatic labeling and transfer learning. The dense stacking target detection method comprises the following steps: establishing a labeled training image set by high-resolution image segmentation; inputting the labeled training image set into a pre-trained target detection model YOLOv3, optimizing a priori frame size and a loss function of the YOLOv3 model, and performing fine adjustment on the model by using the training image set; and finally, inputting a to-be-detected image into the fine-tuned YOLOv3 model, outputting the classification of the target sub-regions and the positions of the sub-regions, splicing output result images to recover the output result images into an original image, and counting a total counting result. The dense stacking target detection method provided by the invention has strong anti-interference performance and robustness, and has low requirements on an image photographer and a photographing illumination condition; through an unsupervised learning method, quasi-automatic annotation of the image is realized, and the workload of manual annotation is greatly reduced, and the model training efficiency is improved; and the dense stacking target detection method can be used for image recognition of a large number of densely stacked targets which are mutually shielded, and is suitable for automatic counting scenes of various densely stacked targets.

The invention relates to a table tennis serving mechanism; it is characterized in that it comprises: a ball storage mechanism for storing table tennis balls, which is connected with a conveying mechanism; a conveying mechanism for conveying table tennis balls, including a wheel shell, a conveying wheel and a conveying mechanism tube; the up and down swing mechanism is used to adjust the radian of the table tennis ball, and it is composed of a screw rod, a nut and a tray; the left and right swing mechanism is used to control the angle of the table tennis ball, and it is composed of a hose, a gear, and a toothed plate; The mechanism is used for sending out table tennis balls and controlling the direction of rotation and strength. It consists of a serving tube and two friction wheels arranged on the serving tube; Power is provided, and each stepper motor is controlled by a control box, which is provided with control buttons for controlling the arc, angle, rotation direction and strength of the ball. The invention can not only improve the serving performance, but also reduce the cost, greatly improve the training efficiency, and is beneficial to popularization.

The invention relates to the technical field of sport equipment, in particular to an auxiliary volleyball training device for sports; the auxiliary volleyball training device for sports comprises a mounting plate, wherein supporting columns are fixedly connected to the bottoms of the left end and the right end of the mounting plate respectively; trundles are arranged at the bottom of the supporting columns; a lifting mechanism fixedly connected with the mounting plate is arranged between the supporting columns on the two sides; a shell is fixedly connected to the top of the mounting plate; a ball releasing mechanism is arranged at the top of the shell; and a ball serving mechanism connected with the ball releasing mechanism is arranged on the inner side of the shell. Due to the arrangementof the ball releasing mechanism, the volleyball balls in a ball releasing bucket can fall into a ball releasing tube one by one, so that the ball serving is more ordered; the ball serving mechanism is arranged, so that the volleyball balls can be automatically sent out during use, the waste of manpower in ball serving training is avoided, ball serving training is can be performed continuously, and the training efficiency of athletes is greatly improved; and by arranging an adjusting mechanism, the ball releasing tube can move up and down, the hit position of the volleyball can be adjusted, and different rotation of the volleyball can be realized.

The invention relates to a fly-by-wire device for a trainer aircraft, which is characterized in that: a longitudinal linkage rod, a transverse linkage rod and a course linkage rod passes through two sides of each of front and back cockpits; and a displacement sensor, an eddy-current damper and an adjustment mechanism are arranged on the front and back cockpits respectively. The device has the advantages that: 1) the mechanism linkage is realized in both the front and back cockpits operated by a longitudinal axis, a transverse axis and a course axis, so the mechanical and electrical properties of the front and back cockpits are completely consistent, which contributes to the improvement of training efficiency and the safe operation of the aircraft; 2) the operating rod systems of the front and back cockpits are changed by a few driving rod systems to pass through from two sides of a control console, so the simple layout requirements of the cockpits are met; and 3) the driving displacement sensor, the eddy-current damper, the adjustment mechanism and the like are distributed in the front and back cockpits on the premise of ensuring the important operating properties of the systems, so the whole driver operation device can make a flexible and high-efficiency use of the space of the front and back cockpits in the front and back cockpits.

The invention provides a method and device for processing power system data, and an electronic device. The method comprises the following steps of obtaining the learning value information in the operation data of a power system on the basis of a pre-trained depth strengthening learning model; determining the running state of the power system according to the obtained learning value information; according to the running state of the power system, determining a control decision, wherein the control decision is used for guaranteeing safe and stable operation of the power system. According to theinvention, starting from the power grid data, the learning value information in the running data of the power system is obtained on the basis of the pre-trained depth strengthening learning model, andthe running state of the power system is determined according to the obtained learning value information and a reasonable control scheme is given; the dependence of the power grid analysis on the physical model can be avoided, the model framework does not need to be readjusted for different scenes and different running modes, the training efficiency is high, and the applicability is high.

The embodiment of the invention provides a method and a device for training a graphic neural network model. The method comprises the following steps: acquiring a target training sample and a corresponding target sample label from a sample set; wherein the target training sample corresponds to a target node in a target relationship network graph, the target node has a target node number, the targetrelationship network graph comprises a plurality of nodes and connecting edges between the nodes, each node has a respective corresponding node number, and each connecting edge has a respective corresponding edge number; querying graph information of a target sub-graph of the target relational network graph from pre-stored graph information of the target relational network graph according to thetarget node number and preset parameters; wherein the target sub-graph takes the target node as a central node, and the hop count between each node in the target sub-graph and the target node is lessthan or equal to a preset parameter; and training the graph neural network model by using the graph information of the target sub-graph and the target sample label. Requirements on a machine can be reduced, and training efficiency can be improved.

The invention provides an AI based low-confidence sample processing method and system of board sorting. Image data of at least one format of a low-confidence sample is obtained; an image of at least one format of the low-confidence sample is presented in a display device; a new class marked by the low-confidence sample is obtained; and a training method is input to the marked low-confidence sample, and a new classification model is obtained via re-training. According to the method and system, the low-confidence sample can be discovered continuously and utilized, so that the classification precision of a machine learning method is improved gradually.

The invention discloses a short text sentiment classification method based on a CNN-BiMGU model, and belongs to the technical field of deep learning and natural language processing. The CNN-BiMGU model mainly comprises an embedding layer, a convolution layer, a pooling layer, a BiMGU layer, an attention mechanism layer, a full connection layer and a classification layer, wherein the embedding layer encodes a data set containing commodity comments into word vectors, the convolution layer extracts text feature matrixes from the word vectors through a CNN channel, the pooling layer performs dimension reduction on the text feature matrixes, and the BiMGU layer is used for obtaining hidden state vectors; wherein the attention mechanism layer is used for strengthening important information, thefull connection layer is used for splicing output of the pooling layer and the attention mechanism layer, and the classification layer is used for obtaining final emotion categories in a classified mode. According to the CNN-BiMGU model, the CNN channel and the BiMGU channel are connected in parallel, the defect that the CNN channel ignores the relation before and after the features is overcome, the attention mechanism is fused into the BiMGU channel, important features in context semantic features are highlighted, and the accuracy of short text emotion classification is improved.

The invention discloses in the technical field of basketball sports training installation and monitoring system, including base board, whose top end is perpendicularly installed a bearing pillar, whose right side joins base board top section through reinforce beam, bearing pillar inside space bottom section is arranged with liquid pressure extendable rod, whose top section is placed with basketball placing board, bearing pillar left wall opens a basketball passage, through slide move component and being started by electric engine, in process of player training shot, the basketball board shifts up and down, adds difficulty to shot, gives convenience in actual high strength fight game, player ensures calm performance of shot, camera zoom and sensing device monitors real time player training image and motion information, and stores, helps analyze and statistically summarizes each player behavior point, increases player training efficiency, help fast development of player ball shot art.

The invention discloses a table tennis ball accompanying training robot with a visual recognition function. The robot comprises a table tennis ball serving module, a table tennis ball recovery storagemodule, a visual module and a main control module; the table tennis ball serving module comprises a holder, a table tennis ball conveying mechanism and a ball serving mechanism; the holder realizes all-directional serving; the table tennis ball conveying mechanism is composed of a conveying pipe and a shifting wheel, and conveys a table tennis ball to the ball serving mechanism from a ball storage box on a machine frame; the ball serving mechanism is composed of a serving pipe and two friction wheels, and realizes the control of the ball speed and the self rotation of the table tennis ball byadjusting the speed of a motor; the visual module comprises a camera and a processor; the camera collects images and transmits the images to the processor, recognizes the orientation of a user through a visual recognition algorithm built in the processor, and transmits orientation information to the main control module; and the main control module processes data transmitted by the visual module through an algorithm, and controls the table tennis ball serving module to perform serving. By simulating an actual table tennis court, the training efficiency is improved.

The invention relates to the field of medical apparatuses, in particular to a convenient vital capacity training device for the respiratory medicine department. The technical purpose is to provide theconvenient vital capacity training device which is for the respiratory medicine department, convenient to carry and can conduct free training at any sites. The convenient vital capacity training device for the respiratory medicine department comprises a mounting frame, a cover plate, a locking mechanism, popping mechanisms, oval glass box bodies, a branch pipe, an air valve, an air spraying pipe,piston balls, a glass test tube, supporting pillars, a rotary fastener, an air delivery pipe, a blowtorch and supporting pillars; the left end of the cover plate is hinged to the top of the mountingframe, and an inserting hole is formed in the right end of the top of the cover plate; the locking mechanism is installed at the lateral part of the mounting frame. The convenient vital capacity training device has the advantages of being convenient to carry and capable of conducting training at any sites is achieved.

The invention relates to the technical field of computers, and provides a model training method and terminal equipment. The method comprises the steps of obtaining a to-be-trained model and training data; wherein the training data comprises a plurality of samples, and each sample corresponds to a probability value; selecting a plurality of samples from the training data according to the probability value of each sample to generate a training data set; training the to-be-trained model through the training data set, and obtaining an recognition result of the to-be-trained model for the trainingdata set; calculating a clustering effect evaluation value of the training data set according to the recognition result and the training data set; and adjusting the probability value corresponding toeach sample in the training data set according to the clustering effect evaluation value. According to the method, model training and training data screening are combined, so that the training efficiency of model training can be improved; in the training iteration process, the effect of difficult-to-recognize samples on model training can be enhanced, and then the robustness and accuracy of the trained model are improved.

The invention discloses a short-term load prediction method considering somatosensory temperature and radiation intensity. The short-term load prediction method comprises the following steps of 1) inquiring sample data such as historical load and weather; 2) calculating historical somatosensory temperature data and a daily load level; 3) selecting an optimal'mode similar day 'from the historical sample data set on the basis of the day type information and meteorological data of the to-be-predicted day, and finally calculating to obtain a normalized load curve; 4) establishing a neural networkprediction model considering the somatosensory temperature and the sunlight intensity to obtain the load level of the to-be-predicted day; and 5) calculating load data of the to-be-predicted day through the normalization curve and the load level. According to the method, the influence on the somatosensory temperature of the load and the sunlight intensity of the distributed photovoltaic power generation are fully considered, the change rule of the historical load is fully considered, the load level and the load mode are separately predicted, the input dimension of the neural network is reduced, the network training load is reduced, and the calculation efficiency is improved.

The invention discloses a special simulation training system for ATP vehicle-mounted equipment. The special simulation training system comprises a vehicle-mounted equipment simulation module for simulating core functions of the ATP vehicle-mounted equipment, an environment simulation module for simulating external ground environment and conditions required by operation of the ATP vehicle-mounted equipment, a train driving simulation module for simulating related operation devices, train basic control logic and traction calculation for the ATP vehicle-mounted equipment, and a teaching assisting module for training mode selecting, training data support and training flow control for the entire system. The special simulation training system for ATP vehicle-mounted equipment is based on complete simulation mode and improves training efficiency of the ATP vehicle-mounted equipment.

The invention relates to an intelligent insertion plate, a rehabilitation training system and a rehabilitation interconnection training system. The intelligent insertion plate comprises a shell, a plurality of induction capacitor modules and a processor; a plurality of blind holes are formed in the shell; the induction capacitor modules correspond to the blind holes one to one; the processor is electrically connected with the induction capacitor modules, and when insertion parts are inserted into the blind holes and the corresponding induction capacitor modules are triggered, the processor acquires state information of the corresponding induction capacitor modules and generates training data according to the state information. According to the intelligent insertion plate, collection of thetraining data of a trainer in the rehabilitation training process is achieved, the trainer can better judge the training condition of the trainer according to the collected training data and give more specific training task information, the problem is solved that traditional training is completely dependent on the trainer so that the efficiency can be low, and the training efficiency is improved.

The invention relates to a professional dance evaluation method for realizing human body posture detection based on deep transfer learning, which comprises the following steps: S1, establishing a special human body posture detection model by utilizing a deep transfer learning principle and combining posture characteristics of professional dance training; s2, collecting a video for demonstrating dancing actions, and inputting the video into the human body posture detection model to obtain a human body key point data flow which changes along with time as a reference standard for evaluation; andS3, obtaining the human body key point information of the dancing action of the testee in the same way, and taking the similarity between the human body key point information and the reference standard as the evaluation of the dancing posture standard degree. According to the invention, the efficiency and accuracy of model training are improved by using deep migration learning.

Provided is a SerDes link parameter automatic debugging method. The method includes steps: selecting corresponding coding and decoding modes, a forward error correction method and a check code for a link clock and a channel characteristic, then determining a pre-emphasis parameter of a TX side of a link through a backtracking method, experiencing a code stream learning process, and determining anadaptive configuration parameter through training at an RX side of the link. According to the method, strict sequential control of the training of the TX side and the RX side is performed, the link parameter configuration deviation caused by interference generated by bidirectional communication of the link is avoided, and the accuracy of the set parameter of the link can be effectively guaranteed;furthermore, the TX side performs backtracking training on the pre-emphasis parameter currently selected by the link according to an error reporting condition of the FEC decoding of the RX side link,the parameter configuration selected by each time of training can be determined according to the characteristic of the link, and the training efficiency is improved. The method is applicable to various router backboards, and automatic debugging of various link parameters is realized for high-rate SerDes links.

The invention provides a network traffic classification method based on feature strong correlation, which comprises the following four steps of: inputting data to be trained, and performing feature extraction on the input data to form a multi-dimensional feature vector set; calculating the correlation between the features by using the mutual information between the features and the response variables; calculating redundancy among the features according to the mutual information, and selecting the feature with the highest score as a final feature vector through iterative calculation; and constructing a network traffic classification model based on feature strong correlation according to the classification target and obtaining a classification result. According to the method, the correlationbetween features can be fully utilized, and the feature with the maximum correlation and the minimum redundancy is extracted in the learner training process. Under the same classification model, theclassification efficiency can be effectively improved on the premise that the classification precision is guaranteed, and the defect caused by the fact that correlation between features is not considered in an existing feature selection method based on heuristic search is overcome.

The invention relates to a dialogue generation method based on near-end strategy optimization and adversarial learning, and belongs to the field of computer natural language processing. The dialogue generation method comprises the steps of firstly pre-training a generation model and a discrimination model of an adversarial generative network; then, adopting a Monte Carlo sampling method to calculate awards corresponding to each word in the generated sentence, wherein the size of an award value represents the quality of word generation; secondly, taking the training process of the adversarial generative network as a reinforcement learning process, and training the adversarial generative network by using a near-end strategy optimization algorithm, so that awards obtained by the discrimination model can guide the generation of the generative model, and dialogues obtained by the generative model can guide the training of the discrimination model; and finally, training the generation modelby using a forced guidance method. According to the dialogue generation method, the training efficiency of the model is improved by controlling self-adaptive multiple iterations of the generation model, and the complexity of the sample is improved through the near-end strategy optimization algorithm, and the dialogue generation quality is further improved, and the dialogue closer to human beings can be generated.

The invention relates to the field of automatic recognition algorithms of deep learning, in particular to a multi-component radar signal intra-pulse modulation mode recognition method. The method comprises the following steps of acquiring time-frequency images of the single-component or overlapped multi-component radar signals of several different intra-pulse modulation modes; preprocessing the radar signal time-frequency image by using an image processing algorithm, and making a training set and a test set by using the signal type contained in the radar signal as a label; designing a pre-training network based on a convolutional neural network to extract radar signal time-frequency image features, and designing a multi-component signal classification network based on reinforcement learning to obtain a classification and recognition result; training, testing and perfecting a network structure and parameters; classification and identification of multi-component signals are realized. Themulti-component radar signal identification algorithm provided by the invention has a wide radar signal type application range and relatively high identification accuracy under the condition of low signal-to-noise ratio, and realizes intra-pulse modulation mode identification of randomly overlapped multi-component radar signals.

The invention provides a neural network training method and device and computer equipment, and the method comprises the steps: respectively inputting an obtained sample image into a first neural network and a plurality of second neural networks for classification, obtaining output data of a first preset network layer of the first neural network and output data of a second preset network layer of each second neural network; determining a weight value corresponding to each second neural network based on a preset target function; weighting the distillation loss value corresponding to each secondneural network by adopting the weight value to obtain a first target loss value; and iteratively adjusting network parameters of the first neural network according to the first target loss value. According to the embodiment of the invention, the weight value of the second neural network is determined through the output data of the second neural network and the first neural network, so that the training efficiency of the first neural network can be improved in a mode of training the neural network according to the first target loss value calculated according to the weight value.

The invention relates to a publication PDF layout analysis and recognition method based on mixing of multiple neural networks, belongs to the technical field of image recognition and PDF layout analysis. The method comprises: using a multi-task training mode to recognize, segmenting and marking PDF layouts including paragraphs, titles and illustrations, locating text lines and then recognizing texts. According to the method, on the aspect of layout recognition, through a multi-task training mode, row and structure recognition annotation is completed at the same time, manual participation is not needed in the whole process, and PDF text structure information is effectively reserved. According to data with PDF text structure information obtained through layout analysis, a common Chinese dictionary for version data is constructed, and a text recognition model is trained in a targeted manner, so that the recognition precision of the model in a PDF printed text recognition task is greatly improved. The recognized text also has structural information, an original PDF layout structure is restored, and subsequent secondary editing, electronic book manufacturing and book content knowledge mining are also facilitated.

The invention provides an industrial robot training system combining vitality and reality and a training method, and relates to the field of industrial robot training. The industrial robot training system includes a virtual experience module, an offline programming and simulation module and a physical training module; the virtual experience module includes a VR device and a graphical workstation for determining the operating location; the offline programming and simulation module includes a workstation for computing and storing data, platform software for receiving the data, simulation software for verifying data results and offline programming software for generating correct execution commands; and the physical training module includes a database for storing correct execution commands, alogic control module and an industrial robot training platform. The industrial robot training system has the beneficial effects that the personal safety and training effect of training personnel are improved, and the efficiency of the training is improved.