# Tasks: [08-31 to 09-04]

### Title: Creation of Metadata on kids AI cameras

1. Integrated motion detector with yolov4 and written inference code for Kids/Adults classifier to create metadata.
2. Faced an issue with TF and CUDA versions as yolov4 used tf-1.14.0 with CUDA 10.0 and kids/adult classifier needed tf-2.3.0 with CUDA 10.1. So intially ran the kid/adult classifier on TF CPU mode, since CUDA 10.0 was already installed on my m/c.
4. Solved multiple CUDA (10.0, 10.1) and tf (1.14.0, 2.3.0) version issues by installing them on same m/c and sped up kids/adults classsifier inference speed by 2X times with a switch from TF CPU to GPU.
5. Progress of generation of metadata on kids playschool cameras:

| Video-series | Images | Status |
| -------- | -------- | -------- |
| totmate_20200811 | 240,000 | Finished
| totmate_20200812 |350,000 | Finished
| totmate_20200817 | 240,000 | Finished
| totmate_20200818 | 483,000 | Finished
| totmate_20200819 | 545,000 | Finished
| totmate_20200820 | 431,000 | Finished
| totmate_20200821 | NA | Running

5. Till now Detected 2.3 Million people and classified them.
6. Inference time: YOLOv4 - 15 FPS
7. Inference time: Kids/Adult classifier - 10 ms per image.
8. H/W: CPU: i5-9400F, GPU: RTX-2070 Super

#### Previously working on model compression Algorithms.
- Tensorflow model optimization library has model compression algorithms to integrate them easily with tf or Keras code during training [recommended-better control] or compressing model post training.
- I was able to compress the size of BA/customer classification model from 513 MB to 117 MB with 90% test set accuracy. Used model pruning.
- I did not check the inference speed increase after model compression.
- Still need to complete investigation and coding of quantization and weight clustering algorithms on Ba/customer classifier for comparitive study.