CN117852588A - AI chip evaluation parameter determination method based on deep residual neural network - Google Patents
AI chip evaluation parameter determination method based on deep residual neural network Download PDFInfo
- Publication number
- CN117852588A CN117852588A CN202311783926.8A CN202311783926A CN117852588A CN 117852588 A CN117852588 A CN 117852588A CN 202311783926 A CN202311783926 A CN 202311783926A CN 117852588 A CN117852588 A CN 117852588A
- Authority
- CN
- China
- Prior art keywords
- neural network
- layer
- deep residual
- residual neural
- fdnet
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 38
- 238000011156 evaluation Methods 0.000 title claims abstract description 35
- 238000000034 method Methods 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims abstract description 20
- 238000010606 normalization Methods 0.000 claims description 9
- 238000013139 quantization Methods 0.000 claims description 8
- 238000013461 design Methods 0.000 description 7
- 102100024607 DNA topoisomerase 1 Human genes 0.000 description 5
- 101000830681 Homo sapiens DNA topoisomerase 1 Proteins 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 4
- 238000005259 measurement Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 239000010977 jade Substances 0.000 description 1
- 238000003032 molecular docking Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an AI chip evaluation parameter determining method based on a deep residual neural network, which comprises the following steps: designing a deep residual neural network FDNet, training the deep residual neural network FDNet to obtain a floating point model, and obtaining an evaluation parameter under a target classification task; quantifying the floating point model to obtain a fixed point model, deploying the fixed point model on a hardware board, and obtaining an evaluation parameter under a target classification task through reasoning on the hardware board; the obtained evaluation parameters obtained by the floating point model and the fixed point model are used for evaluating the application performance of the AI chip when the deep residual neural network FDNet is used. The method and the device can be applied to various aerospace data sets, and have strong robustness.
Description
Technical Field
The invention belongs to the technical field of integrated circuit evaluation, and particularly relates to an AI chip evaluation parameter determination method based on a deep residual error neural network.
Background
Current artificial intelligence technology is vigorously developed, and big data, deep learning algorithms and hardware on which the technology depends are updated in a daily and monthly manner. In terms of algorithms, various brand-new neural network algorithms are proposed, and most of these algorithms require higher quality data requirements. The ability of hardware is often the bottleneck for such high data volume, high computation-volume tasks. The increasing interest in deep learning has stimulated the design of smart chips to a great extent, and various companies have proposed corresponding hardware designs. With the increasing number of AI chips in aerospace, how to systematically evaluate the suitability of aerospace intelligent algorithms of the chips becomes a research hot spot.
Along with the rapid development of the deep learning network model, the difference between models in different scenes is larger and larger, and the structure of the network model is gradually complicated and diversified from a convolutional neural network to a transfer model. At present, one of the fields with more AI chips is the target classification field, such as planetary detection, space debris removal, satellite on-orbit service, spacecraft intersection and docking and other tasks.
The algorithm models in the target classification are various, the operator differences among the models are larger and larger due to the complicated and diversified structures of the network models, the existing evaluation reference program is not applied to aerospace, the operator types of the aerospace typical algorithm can be contained are not more, and some operators are not commonly used in aerospace, so that the operator performance of the AI chip cannot be well evaluated by using the existing evaluation reference program. There is no unified standard for the use of network models in the current evaluation of the suitability of the intelligent algorithm of the AI chip, and most of the used models are still common models on the ground. Therefore, it is needed to design an AI chip intelligent algorithm suitable for aerospace applications, which is used in the evaluation work of the AI chip for aerospace applications.
Disclosure of Invention
The invention solves the technical problems that: the AI algorithm suitability of the chip for aerospace is evaluated by comparing the application performance, operator performance, quantization performance and the like of a floating point model and a quantization model.
The technical scheme of the invention is as follows: an AI chip evaluation parameter determining method based on deep residual neural network comprises the following steps:
designing a deep residual neural network FDNet, training the deep residual neural network FDNet to obtain a floating point model, and obtaining an evaluation parameter under a target classification task;
quantifying the floating point model to obtain a fixed point model, deploying the fixed point model on a hardware board, and obtaining an evaluation parameter under a target classification task through reasoning on the hardware board;
the obtained evaluation parameters obtained by the floating point model and the fixed point model are used for evaluating the application performance of the AI chip when the deep residual neural network FDNet is used.
The evaluation parameters include: accuracy Top-k, average accuracy mean mAP, and error rate indicator.
The operators used in the deep residual neural network FDNet are all common operators in the field of aerospace.
The deep residual neural network FDNet consists of 4 deep residual modules; the first layer and the third layer of the residual error module respectively use a convolution layer, a batch normalization layer and a rectification linear unit layer, the second layer consists of the convolution layer, the batch normalization layer, the rectification linear unit layer and an attention channel layer, and the fourth layer consists of the convolution layer and the batch normalization layer; the output part of the deep residual neural network FDNet comprises a convolution layer and a full connection layer; the classification layer of the deep residual neural network FDNet is a softmax layer.
And training the deep residual neural network FDNet network through the ImageNet data set to obtain a floating point model.
And when the floating point model is quantized to obtain the fixed point model, an AI chip self-contained software tool is adopted for quantization, and an int8 quantization mode is adopted.
Compared with the prior art, the invention has the beneficial effects that:
1. and the FDNet model is utilized, so that operator performance evaluation is more comprehensive in the evaluation of the space navigation AI chip. The FDNet model contains more space navigation common operators. Operator performance includes diversity of operators and support of operators.
2. The model is good in robustness, and can be suitable for different aerospace data sets and used for training and evaluating the model in the target classification task.
3. The target recognition accuracy is higher, and the FDNet model in the invention can obtain higher target recognition accuracy.
Drawings
Fig. 1 is an evaluation algorithm model of an AI chip for aerospace based on a deep residual neural network.
Fig. 2 is a flowchart of a method for determining an AI chip evaluation parameter based on a deep residual neural network.
Detailed Description
The intelligent algorithm suitability evaluation algorithm of the AI chip is designed in the field of target identification. Deep residual neural networks named FDNet (Feature Denoising Residual Network, FDNet) were designed.
Step one:
the algorithm network in the invention is named as FDNet, is an algorithm model based on a depth residual neural network, and extracts key features by utilizing a multi-layer residual neural network as shown in figure 1. FDNet consists of D blocks of depth residuals, where d=4, capturing the difference between the input and the ground truth value using elemental subtraction. FDNet adopts three-channel design, and related features are extracted from RGB channels of the picture respectively. In this architecture, advanced spatial features in the picture are captured from the input through 4 residual modules, helping to enhance the denoising ability of the model, and also ensuring the accuracy and depth of feature extraction.
The first and third layers of the residual block each use a combination operation including convolution (Conv), batch Normalization (BN), and rectifying linear units (ReLU). The batch normalization not only greatly improves the training speed of the model, but also enhances the stability of the model under various data conditions. The second layer is a channel attention layer, consisting of convolution (Conv), batch Normalization (BN), rectifying linear units (ReLU) and an attention channel layer. The fourth layer is composed of Conv layer and BN layer and is mainly responsible for generating a noise matrix, i.e. a noise matrix is generated before the subsequent element subtraction operation. This design ensures that not only noise is removed but also features are corrected and optimized during denoising, taking into account the various noise disturbances to which the image may be subjected.
In the design of FDNet, the output layer is a core part for realizing image recognition and plays a crucial role. Through the convolution and pooling operations of the previous layers, the local features of the image are extracted in a rich way. The design of the output layer is particularly critical in order to convert these extracted features into an explicit classification result. Specifically, the output portion of the network first includes a convolutional layer. This convolution layer further extracts advanced features of the image and prepares for the next full connection layer. Then, a full connection layer is introduced, and plays a role in the neural network, which can flatten the feature vector extracted by the front layer and map the feature vector to a feature space with a fixed length, so that preparation is made for final classification. In this feature space, each neuron can be considered as a weight corresponding to a particular class, and the result of multiplication with it is the score of that class. Finally, the output of the fully connected layer is translated into a probability distribution for each category using the softmax layer. The softmax layer can translate consecutive values into probability distributions, ensuring that the sum of probabilities for all classes is 1. Network training is performed using an ImageNet dataset, and first ImageNet is screened. To ensure training efficiency and generalization ability of the model, ten thousand representative images were selected. The training process is divided into the following stages:
1. pretreatment: to speed up training and improve network convergence, all training images are normalized, i.e., the mean is subtracted and divided by the standard deviation. And the data set pictures are subjected to random cutting, rotation and overturning so as to enhance the robustness of the model.
2. Training strategies: a random gradient descent (SGD) was used as an optimizer, and the initial learning rate was set to 0.1 and gradually decreased according to a predetermined strategy. To prevent overfitting, L2 regularization was added, and Momentum strategy with Momentum of 0.9 was used to accelerate convergence.
3. And (3) learning rate adjustment: the learning rate is reduced to 1/10 of the original learning rate after a certain number of training cycles.
4. And (3) verification and adjustment: at the end of each training period, the performance of the model is evaluated using the validation set. And adjusting the network structure and parameters according to the verification result.
The operator classes contained in the FDNet model are listed in table 1.
Table 1 operator species contained in FDNet model
The training, testing and verifying work of the FDNet model is completed by using a data set, wherein the adopted data set can be an open source data set or a training sample formed by actually measured pictures in a space environment. After the FDNet model is trained, three indexes of TOP1, TOP5 and error rate of the floating point model are obtained, wherein TOP1 is the highest output of the label of the picture and the model prediction is consistent, TOP5 is the label of the picture in the first five labels of the model prediction, and the error rate is the label of the picture is not in the first five results of the model output.
Step two:
after the floating point model is obtained, the model is deployed to a hardware board by using a software compiling tool chain. The hardware board card can be selected according to the needs, and the European bit jade yulon 810A board card is used for deployment in the patent. The FDNet model is generated through a PyTorch deep learning training framework, so that the network model after FDNet training is imported into YL-ACUITY through ONNX conversion, 8bit quantization is carried out on the network model, and an ovxlib-based C language code is generated through an export function. And testing on the board card to obtain three index results of TOP1, TOP5 and error rate under the actual measurement model.
The invention is further described below with reference to the accompanying drawings.
The invention discloses an AI chip evaluation parameter determination method based on a deep residual neural network, and the flow is shown in fig. 2.
Firstly, training, testing and verifying the FDNet model by using a data set, and obtaining three indexes of TOP1, TOP5 and error rate of the floating point model after the training of the FDNet model is completed. After the floating point model is obtained, the model is deployed to a hardware board by using a software compiling tool chain. And carrying out reasoning test on the board card to obtain three index results of TOP1, TOP5 and error rate under the actual measurement model. Table 2 shows the comparison results, comparing the test results of the FDNet model with the results of the Resnet50 model, and comparing the measured model with the results of the floating point model. It can be seen that the recognition accuracy of the FDNet model is higher than that of the Resnet50, and the target recognition task can be completed well. Meanwhile, the AI chip used for supporting all operators in the FDNet model can support more common operators in the astronavigation target recognition task. And compared with floating point and actual measurement model results, the 8bit quantization precision loss of the AI chip is smaller, and the model can be quantized through the chip, so that the complexity is reduced. The evaluation results of the AI chip show that the algorithm model designed by the invention can be applied to the evaluation work of the AI chip for aerospace.
TABLE 2 comparison of Floating Point model and measured model results
The present invention is not described in detail as being well known to those skilled in the art.
Claims (6)
1. An AI chip evaluation parameter determining method based on a deep residual neural network is characterized by comprising the following steps:
designing a deep residual neural network FDNet, training the deep residual neural network FDNet to obtain a floating point model, and obtaining an evaluation parameter under a target classification task;
quantifying the floating point model to obtain a fixed point model, deploying the fixed point model on a hardware board, and obtaining an evaluation parameter under a target classification task through reasoning on the hardware board;
the obtained evaluation parameters obtained by the floating point model and the fixed point model are used for evaluating the application performance of the AI chip when the deep residual neural network FDNet is used.
2. The AI chip evaluation parameter determination method based on the deep residual neural network of claim 1, wherein the evaluation parameter comprises: accuracy Top-k, average accuracy mean mAP, and error rate indicator.
3. The AI chip evaluation parameter determination method based on the deep residual neural network according to claim 1, wherein operators used in the deep residual neural network FDNet are all operators commonly used in the aerospace field.
4. The AI chip evaluation parameter determination method based on the deep residual neural network according to claim 1, wherein the deep residual neural network FDNet is composed of 4 deep residual modules; the first layer and the third layer of the residual error module respectively use a convolution layer, a batch normalization layer and a rectification linear unit layer, the second layer consists of the convolution layer, the batch normalization layer, the rectification linear unit layer and an attention channel layer, and the fourth layer consists of the convolution layer and the batch normalization layer; the output part of the deep residual neural network FDNet comprises a convolution layer and a full connection layer; the classification layer of the deep residual neural network FDNet is a softmax layer.
5. The AI chip evaluation parameter determination method based on the deep residual neural network according to claim 1, wherein the deep residual neural network FDNet network is trained through an ImageNet data set to obtain a floating point model.
6. The method for determining the evaluation parameters of the AI chip based on the deep residual neural network according to claim 1, wherein when the floating point model is quantized to obtain the fixed point model, an AI chip-carried software tool is adopted for quantization, and an int8 quantization mode is adopted.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311783926.8A CN117852588A (en) | 2023-12-22 | 2023-12-22 | AI chip evaluation parameter determination method based on deep residual neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311783926.8A CN117852588A (en) | 2023-12-22 | 2023-12-22 | AI chip evaluation parameter determination method based on deep residual neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117852588A true CN117852588A (en) | 2024-04-09 |
Family
ID=90537300
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311783926.8A Pending CN117852588A (en) | 2023-12-22 | 2023-12-22 | AI chip evaluation parameter determination method based on deep residual neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117852588A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118393329A (en) * | 2024-06-27 | 2024-07-26 | 中国信息通信研究院 | System for testing AI chip in model training and reasoning performance |
-
2023
- 2023-12-22 CN CN202311783926.8A patent/CN117852588A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118393329A (en) * | 2024-06-27 | 2024-07-26 | 中国信息通信研究院 | System for testing AI chip in model training and reasoning performance |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113641820B (en) | Visual angle level text emotion classification method and system based on graph convolution neural network | |
| CN113221687B (en) | Training method of pressing plate state recognition model and pressing plate state recognition method | |
| CN114743020B (en) | Food identification method combining label semantic embedding and attention fusion | |
| CN111966812B (en) | Automatic question answering method based on dynamic word vector and storage medium | |
| CN117852588A (en) | AI chip evaluation parameter determination method based on deep residual neural network | |
| CN115222950A (en) | Lightweight target detection method for embedded platform | |
| CN114821340A (en) | Land utilization classification method and system | |
| CN115019132B (en) | Multi-target identification method for complex background ship | |
| US20210056353A1 (en) | Joint representation learning from images and text | |
| CN114742211B (en) | Convolutional neural network deployment and optimization method facing microcontroller | |
| CN110991247B (en) | Electronic component identification method based on deep learning and NCA fusion | |
| CN114863407A (en) | Multi-task cold start target detection method based on visual language depth fusion | |
| CN117496124A (en) | Large-area photovoltaic panel detection and extraction method based on deep convolutional neural network | |
| Jang et al. | Decision fusion approach for detecting unknown wafer bin map patterns based on a deep multitask learning model | |
| CN117058716A (en) | Cross-domain behavior recognition method and device based on image pre-fusion | |
| CN111783688A (en) | Remote sensing image scene classification method based on convolutional neural network | |
| CN116883792A (en) | Aurora ovum intensity image modeling method based on FA-GAN | |
| CN109934248B (en) | Multi-model random generation and dynamic self-adaptive combination method for transfer learning | |
| CN115019183B (en) | Remote sensing image model migration method based on knowledge distillation and image reconstruction | |
| CN114357160B (en) | Early rumor detection method and device based on generated propagation structural features | |
| CN115358473A (en) | Power load prediction method and prediction system based on deep learning | |
| CN109543571A (en) | A kind of intelligent recognition and search method of Complex Product abnormity machining feature | |
| CN113743593A (en) | Neural network quantization method, system, storage medium and terminal | |
| CN116883709A (en) | Carbonate fracture-cavity identification method and system based on channel attention mechanism | |
| CN114282585A (en) | Astronomical spectrum data-based outlier celestial body classification method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |