为何代码训练Create ML图像分类模型比可视化界面慢？
用Create ML可视化界面训练图像分类模型仅需数十秒，但为实现功能迭代自动化改用代码训练时，相同数据与参数下，代码生成的可执行文件训练耗时超20分钟，运行时持续出现CIKernel相关警告，推测与图像处理有关。查阅苹果官方文档及Stack Overflow相关问题后仍无法解决速度问题，也不清楚代码训练与可视化界面训练的差异。
训练代码
class CreateMLManager : NSObject { open var dataPath: String? open var outPath: String? public func training() -> Int { guard let dataPath = dataPath else { return 1 } guard let outPath = outPath else { return 1 } do { let aumentation : MLImageClassifier.ImageAugmentationOptions = [.crop, .rotation, .blur, .exposure, .noise, .flip]; let parameters = MLImageClassifier.ModelParameters(validation: .split(strategy: .automatic),maxIterations: 11,augmentation: aumentation,algorithm: .transferLearning(featureExtractor: .scenePrint(revision: 1), classifier: .logisticRegressor)) let out_path = outPath + "/test.mlmodel" let model = try MLImageClassifier(trainingData: .labeledDirectories(at: URL(fileURLWithPath: dataPath)), parameters:parameters) try model.write(to: URL(fileURLWithPath: out_path)) return 0 } catch { return 1 } } }
运行日志
Current working dir: /Users/wbw/Library/Developer/Xcode/DerivedData/mlmodelTransition-cdfkslvlaufvnagpypknnofdmgkv/Build/Products/Debug
2023-01-10 18:02:09.670712+0800 mlmodelTransition[1435:1935624] 训练数据:/Users/wbw/Desktop/Rashomon/work/company_library/ArWeb/web/ml_platform/temp/xjy
2023-01-10 18:02:09.671074+0800 mlmodelTransition[1435:1935624] 存储路径:/Users/wbw/Desktop/Rashomon/work/company_library/ArWeb/web/ml_platform/resource/training_model/xjy
4 augmented images by 'crop' to be generated.
4 augmented images by 'rotation' to be generated.
4 augmented images by 'blur' to be generated.
1 augmented images by 'exposure' to be generated.
1 augmented images by 'noise' to be generated.
1 augmented images by 'flip' to be generated.
Extracting augmented image features from training data.
Analyzing and extracting image features.
+----------------------+------------------+--------------+------------------+
Raw Images Processed | Augmented Images | Elapsed Time | Percent Complete |
+----------------------+------------------+--------------+------------------+
2023-01-10 18:02:11.132922+0800 mlmodelTransition[1435:1936088] Metal API Validation Enabled
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.212669+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.233004+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.252051+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.276394+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.297000+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.318678+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.340412+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.360359+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.379907+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 63687 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.397403+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 81029 pixels because a CIKernel's ROI function did not allow tiling.
2023-01-10 18:02:11.416585+0800 mlmodelTransition[1435:1936088] [api] -[CIContext(CIRenderDestination) _startTaskToRender:toDestination:forPrepareRender:forClear:error:] Failed to render part of the image because memory requirement of -1 too big
Failed to render 89401 pixels because a CIKernel's ROI function did not allow tiling.
相关截图说明
- Create ML可视化界面参数设置截图
- Create ML可视化界面训练过程截图
内容的提问来源于stack exchange，提问作者汪博文
