Sample Usage Guide【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/geFunction DescriptionThis sample is based on Onnx ResNet-50 network (single input, single Batch) and implements image classification functionality in multi-batch scenarios.In this sample:First use the script transfer_pic.py provided in the sample to convert 2.jpg images to.bin format, while scaling images from 1024683 resolution to 224224.Parse the resnet50 model, then directly perform inference on 2 images (batch value 2), obtain inference results, and process the inference results to output top5 confidence class labels.Directory Structure├── data │ ├── dog1_1024_683.jpg // Test data, need to obtain test images as instructed and place in data directory │ ├── dog2_1024_683.jpg // Test data, need to obtain test images as instructed and place in data directory ├── model │ ├── resnet50_Opset16.onnx // Onnx ResNet-50 network model file ├── scripts │ ├── build.sh // Sample compilation script │ ├── run.sh // Sample execution script │ ├── transfer_pic.py // Convert *.jpg to *.bin, while scaling images from 1024*683 resolution to 224*224 ├── src │ ├── CMakeLists.txt // Compilation configuration script │ ├── sample_dynamic_batch.cpp // Main function, image classification implementation file ├── CMakeLists.txt // Compilation script, calls CMakeLists file in src directoryEnvironment RequirementsCorrectly installtoolkitandopspackages through installation guide Environment PreparationSet environment variables (assuming package is installed in/usr/local/Ascend/)source /usr/local/Ascend/cann/set_env.shImplementation StepsLogin to development environment as running user.After downloading the code and uploading to the environment, please first enter the examples/gesession/sample_dynamic_batch sample directory under the root directory.Note that the sample directory in the following text refers to the examples/gesession/sample_dynamic_batch directory.Prepare ResNet-50 model.Obtain the ResNet-50 original model.You can obtain the ResNet-50 network model file from the following link and upload it to the sample directory/model directory in the development environment as the running user. If the directory does not exist, create it yourself.ResNet-50 network model file (*.onnx): Click Link to download this file.Prepare test images.Please obtain the input images for this sample from the following links and upload them to the sample directory/data directory in the development environment as the running user. If the directory does not exist, create it yourself.https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/aclsample/dog1_1024_683.jpghttps://obs-9be7.obs.cn-east-2.myhuaweicloud.com/models/aclsample/dog2_1024_683.jpgSwitch to sample directory/data directory and execute transfer_pic.py script to convert.jpg to.bin, while scaling images from 1024683 resolution to 224224. 2 *.bin files will be generated in sample directory/data.python3 ../scripts/transfer_pic.pyIf the script execution reports ModuleNotFoundError: No module named PIL, it means the Pillow library is missing. Please usepip3 install Pillow --usercommand to install the Pillow library.Build and VerificationLogin to development environment as running user.Please first enter the examples/gesession/sample_dynamic_batch sample directory under the root directory.Note that the sample directory in the following text refers to the examples/gesession/sample_dynamic_batch directory.Switch to sample directory/scripts and compile the program.bash build.shRun the programbash run.shExecution resultsAfter successful execution, key prompt messages on the screen are shown below. The index in the prompt messages indicates the class label, and value indicates the maximum confidence for that classification. These values may vary depending on version and environment, please refer to actual conditions:[INFO] SAMPLE start to execute. [INFO] Initialize ge success [INFO] Set device 0 success [INFO] Parse model ../model/resnet50_Opset16.onnx success [INFO] Graph add success [INFO] Graph compile success [INFO] Start to process file:../data/dog1_1024_683.bin [INFO] Start to process file:../data/dog2_1024_683.bin [INFO] Graph run success [INFO] Result of picture 1: [INFO] top 1: index[162] value[xxxxxx] [INFO] top 2: index[161] value[xxxxxx] [INFO] top 3: index[166] value[xxxxxx] [INFO] top 4: index[167] value[xxxxxx] [INFO] top 5: index[163] value[xxxxxx] [INFO] Result of picture 2: [INFO] top 1: index[267] value[xxxxxx] [INFO] top 2: index[266] value[xxxxxx] [INFO] top 3: index[265] value[xxxxxx] [INFO] top 4: index[153] value[xxxxxx] [INFO] top 5: index[99] value[xxxxxx] [INFO] Output data success [INFO] SAMPLE PASSEDNote:The correspondence between class labels and classes is related to the dataset used when training the model. The model used in this sample was trained on the imagenet dataset. You can check the correspondence between imagenet dataset labels and classes on the internet. The correspondence between class labels and classes in the current screen output is: 161: [basset, basset hound], 267: [standard poodle].【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/ge创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考