We have a SSD Mobilenet V1 model trained for regression task. Does DL workbench supports INT8 quantization for regression models ? I want to choose MSE as accuracy check metric. DL tool now shows only Object Detection and Classification type of models from Open Model Zoo.
From Computer Vision perspective can you elaborate on SSD Mobilenet V1 trained for Regression task? Because the broad class of tasks in Vision falls under classification and object detection categories. When you mean regression, is it object detection underhood? With respect to SSD architecture, by using the features in the specified depth or channels of the base net, the image classification to predict the label and the regression task to predict the bounding box for small objects on the very point can be done.
By the way, 2020.1 version of Deep Learning Workbench (DL Workbench) is a feature preview release. The documentation and corresponding functionality within the application are subject to change and can contain errata or be inaccurate. Installing, hosting and using this application is at your own risk.
Thanks for prompt reply. We are doing face landmark detection. But question was generic to regression models category, Like face recognistion model gives face vector output instead of object class. How do I quantize such models ? In DL UI, Can we use default quantization mode instead of AccuracyAware mode to quantize regression model ?
OpenModel Zoo also has similar regression models like landmarks-regression-retail-0009, head-pose-estimation-adas-0001, face-recognition-resnet100-arcface etc. How do I quantize this kind of models ?
I am able to use DL workbench docker and quantize SSD mobilenetv1 COCO from model zoo. This gives 2x gain on CPU after quantization.
As communicated earlier 2020.1 version of Deep Learning Workbench (DL Workbench) is a feature preview release.
OpenVINO's Post Training Optimizer tool can be referred for quantization tasks.