train_input.json file is passed to custom learning blocks as the value for the --input-file <file> argument. It contains configuration information for model training options that you may want to use within your training script, if applicable. The specification is shown below.
File structure
type CreateKerasTrainModelOptions = {
classes: string[],
mode: 'classification' | 'regression' | 'object-detection' | 'visual-anomaly' | 'anomaly-gmm';
printHWInfo: boolean | undefined,
inputShapeString: string,
yType: 'npy' | 'structured';
trainTestSplit: number,
stratifiedTrainTest: boolean,
onlineDspConfig: OnlineDspConfig | undefined;
convertInt8: boolean,
profileInt8: boolean,
skipEmbeddingsAndMemory: boolean,
objectDetectionLastLayer: 'mobilenet-ssd' | 'fomo' | 'yolov2-akida' | 'yolov5' | 'yolov5v5-drpai' |
'yolox' | 'yolov7' | 'tao-retinanet' | 'tao-ssd' | 'tao-yolov3' | 'tao-yolov4' | undefined;
objectDetectionAugmentation: boolean | undefined,
objectDetectionBatchSize: number | undefined,
syntiantTarget?: boolean,
maxTrainingTimeSeconds: number,
remainingGpuComputeTimeSeconds: number,
isEnterpriseProject: boolean,
flattenDataset: boolean,
akidaModel: boolean,
akidaEdgeModel: boolean,
skipMemoryProfiling: boolean,
tensorboardLogging: boolean,
customValidationSplit: boolean,
validationMetadataKey?: string,
customVariantsToProfile: CustomVariantInferenceJobModelVariant[] | undefined;
};
File example
{
"classes": [
"idle",
"snake",
"updown",
"wave"
],
"mode": "classification",
"printHWInfo": false,
"inputShapeString": "(39,)",
"yType": "npy",
"trainTestSplit": 0.2,
"stratifiedTrainTest": false,
"convertInt8": true,
"profileInt8": true,
"skipEmbeddingsAndMemory": false,
"objectDetectionAugmentation": false,
"syntiantTarget": false,
"maxTrainingTimeSeconds": 604800,
"remainingGpuComputeTimeSeconds": null,
"isEnterpriseProject": true,
"flattenDataset": false,
"akidaModel": false,
"akidaEdgeModel": false,
"skipMemoryProfiling": false,
"tensorboardLogging": false,
"customValidationSplit": false
}