Congratulation on completing the lessons of the this module. Now let us apply the concepts in a build following project. Please complete below task:
Pre-requisites:
• Local installation of DL workbench or
• DL Workbench on Intel® DevCloud for the Edge
• Lessons from this module and previous modules
Tasks:
Create a project with following configuration
a. Model: yolo-v3-tf form open model zoo with FP16 precision
b. Environment: GPU
c. Dataset: Create dataset of 30 images with data augmentation options available in DL workbench
Run Single inference with batch size = 2 and number of parallel streams = 4 and compare the performance and visualize the performance metrics
Create another project with different dataset. Following is the configuration.
a. Model: yolo-v3-tf form open model zoo with FP32 precision
b. Environment: CPU
c. Dataset: Import COCO validation dataset with annotations
Check the accuracy of FP32 model
Run Group Inference with following configuration and compare the performance and visualize the performance metrics.
a. Batch Size =2, 4, 8
b. Streams: 2, 4, 8
Find the best performance configuration in Throughput vs Latency graph for your model assuming a latency threshold requirement.
Optimize the model to INT8 with Default quantization method and check the accuracy
Compare the Accuracy, Latency and Throughput values for INT8 and FP32 precision
Create a deployment package for with optimal model configuration for following target environment
a. OS: Ubuntu 18.04
b. Target: CPU
c. Include Model: Yes
d. Include Python API: Yes
e. Include Install Script: Yes
Download the zip file, analyze the content and deploy the package on the target machine.
