Batch inference with classifier?

[agentmorris]

What is the appropriate way to do batch inference using a trained classifier? The classifier documentation lists three methods:

1. predict_image.py
2. detect_and_predict_image.py
3. generate_sample_predictions.py

The first appears not to support batch inference. #2 and #3 do batch inference, but running each of them resulted in the entire data set attempting to be loaded into memory. What is the correct way to perform inference on a large number of images using this repo?

---

Issue cloned from Microsoft/CameraTraps, original issue posted by davidwhealey on Apr 27, 2020.

[agentmorris]

Wow, we really worked hard to hide the script that I would actually recommend using in this case, which is none of those three... :)

If you are planning to run both our detector and a classifier trained with our recommended pipeline, I would do this in two stages, with:

• run_tf_detector_batch.py to run the detector
• apply_classifier_single_node.py to run the classifier on the crops produced by the detector

The reason you didn't see the latter script before is that it was all by itself in a folder called "api", I just moved it out of that folder.

This two-step process is exactly what we did, for example, to generate these results;

• Snapshot Serengeti season 7 classification results

Let us know if that helps!

-Dan

---

(Comment originally posted by agentmorris)