11. SSD model training

The train and test datasets including labels in csv format can be downloaded from this link:

https://bit.ly/adl2-ssd

In the top-level folder (that is, Chapter 11, Object Detection), create the dataset folder, copy the downloaded file there, and extract it by running:

mkdir dataset
cp drinks.tar.gz dataset
cd dataset
tar zxvf drinks.tar.gz
cd..

The SSD model is trained for 200 epochs by executing:

python3 ssd-11.6.1.py --train

The default batch size, --batch-size=4, can be adjusted depending on the GPU memory. On 1080Ti, the batch size is 2. On 32GB V100, this could be 4 or 8 per GPU. --train represents model training option.

To support normalization of bounding box offsets, the --normalize option is included. To use improved loss functions, the --improved_loss option is added. If only smooth L1 is desired (no focal loss), use –smooth-l1. To illustrate:

• L1, no normalization:
  • python3 ssd-11.1.1.py...