Labels: Add support for ONNX-based models

[photoprismci]

As a PhotoPrism user, I want image labels to be generated with ONNX-based models, so that I get better labels from a smaller and faster installation.

PhotoPrism currently classifies images with TensorFlow, while face detection already runs on ONNX Runtime. Adding ONNX support for labels as well is a step toward maintaining a single inference runtime instead of two.

Why

• Smaller installation. The bundled TensorFlow library is large and ships in every Docker image. Once the remaining TensorFlow-based features have been migrated, it can be removed entirely.
• One runtime to maintain. ONNX Runtime is already included for face detection, so labels would reuse what is there.
• Build portability. TensorFlow is not available on every platform. See Feature: build-time disable of tensorflow #5591 and Build: Add notf tag to build without TensorFlow dependency #5609, where PhotoPrism cannot be built while TensorFlow is a hard requirement.
• A better choice of models. Many more models, and more recent ones, are published in ONNX format than for TensorFlow 2. This matters most for users who are comfortable selecting a different model but not converting between formats.
• Quality and speed. Modern classifiers are expected to label more accurately than the current model, and ONNX Runtime may be a better fit for this workload.

Expected Behavior

The label vocabulary stays the same, so existing labels remain valid and no re-indexing is required. The model can be selected through the configuration without rebuilding, and custom models supplied by users continue to work (see #5011).

Acceptance Criteria

• Image classification MUST be able to run through ONNX Runtime without requiring TensorFlow.
• The existing label vocabulary MUST be preserved, so stored labels stay valid and no re-indexing is needed.
• The model MUST be selectable through the configuration without rebuilding PhotoPrism.
• Custom models supplied by users MUST continue to be supported.
• The bundled model SHOULD label at least as accurately as the current one, measured on real photos rather than on published benchmark figures.
• Indexing time and memory use SHOULD NOT increase noticeably.
• The size of the bundled model SHOULD stay the same or get smaller.
• Additional optional models MAY be offered for users who want a different trade-off between accuracy and speed.

Related

• Labels: Allow a model to be trained for better results #448 — request for better labeling results
• Proof-of-concept for scene category classification #175 — scene category classification
• Feature: build-time disable of tensorflow #5591, Build: Add notf tag to build without TensorFlow dependency #5609 — building PhotoPrism without TensorFlow
• AI: Allow configuration of custom classification models #5011 — configuration of custom classification models
• AI: Upgrade ONNX Runtime from 1.26 to 1.29 with optional GPU support #5703 — ONNX Runtime upgrade
• AI: Add optional CoreML acceleration for ONNX on native macOS builds #5704 — optional CoreML acceleration for ONNX on macOS
• AI: Add a CLIP-powered semantic search #1287 — CLIP-powered semantic search, tracked separately and not part of this issue