batch_inference.md

Batch Inference

Batch inference is where you take your trained models (or pre-trained models from the model zoo) and perform predictions using groups of images at a time for faster speed.

Steps

Here's the steps for performing batch inference.

1. SSH into the ml-inferencer machine
2. Copy all the images you want to run inference on to /mnt/disk1/ml-apps/detectron2_web_app/data/inference/ directory.
   1. Only *.JPG, *.PNG, *.jpg and *.png files are picked up (case-sensitive globbing)
3. Copy the Corrosion model weights to /mnt/disk1/ml-apps/detectron2_web_app/model-corrosion.pth
4. Copy the Edges-and-Welds model weights to /mnt/disk1/ml-apps/detectron2_web_app/model-edges-and-welds.pth
5. Kick-off the pipeline by running the following scripts in order:
   1. python predict.py corrosion
   2. python predict.py edges-and-welds
   3. python merge.py
   4. python visualize.py
6. Copy the predictions (/mnt/disk1/ml-apps/detectron2_web_app/data/predictions/corrosion/*.pkl and /mnt/disk1/ml-apps/detectron2_web_app/data/predictions/edges-and-welds/*.pkl) to your long-term storage of choice
7. Copy the visualizations (/mnt/disk1/ml-apps/detectron2_web_app/data/predictions/visualized) to your long-term storage of choice