---

name: data-and-alignment description: "Prepare Facenet class-folder datasets, validate LFW pair inputs, and build safe MTCNN alignment workflows." metadata: disco-role: operating disable-model-invocation: true license: MIT

Facenet Data and Alignment

Use this sub-skill when the user needs to prepare or check face image data before embedding, evaluation, or training. This includes class-folder dataset layout, LFW pair files, image preprocessing assumptions, and MTCNN alignment commands.

When to read

• The user asks how to structure a Facenet dataset or why labels/classes are wrong.
• A workflow needs aligned face thumbnails before compare, classifier training, LFW validation, or training.
• The task mentions align_dataset_mtcnn, MTCNN thresholds, margin, image size, multiple faces, or bounding-box output.
• The user has an LFW-style pairs.txt file and missing-image or skipped-pair issues.

Core conventions

• Facenet datasets use data_dir/class_name/image_file directories. facenet.get_dataset() sorts class directories and assigns integer labels by sorted class order.
• facenet.load_data() reads images, converts grayscale to RGB, optionally prewhitens, then crops/flips to the requested image size.
• MTCNN alignment writes class subdirectories under an output directory and records bounding boxes in bounding_boxes_<random>.txt.
• For pretrained 2018 models, later inference/evaluation often needs fixed image standardization; route to ../evaluation/SKILL.md before LFW evaluation.

Practical workflow

1. Validate source layout with scripts/validate_facenet_dataset.py. It checks class directories, image extensions, minimum images per class, and optional LFW pair references without importing TensorFlow.
2. Read references/data-formats.md for class-folder, LFW pair, learning-rate schedule, and large dataset conversion formats.
3. If images are unaligned, read references/alignment-workflows.md, then build a command with scripts/build_alignment_command.py.
4. Choose --image_size and --margin based on the downstream workflow. README-era examples commonly use 182 aligned output for training and 160 for model input.
5. Decide whether multiple detected faces should be saved. The default behavior picks one centered/large face; --detect_multiple_faces changes output naming and class counts.
6. After alignment, validate the output dataset again before training or classifier workflows.

Decision points

• Unaligned vs aligned input: compare and contributed clustering can align internally, but training and LFW evaluation expect prepared aligned face patches.
• Random order: use --random_order only to help multiple alignment jobs split work; it makes logs harder to compare.
• Multiple faces: enable only when each input image may legitimately contain more than one identity; otherwise it can pollute class folders.
• Skipped images: alignment records output filenames without boxes when no face is found. Treat a high skip rate as detector/quality/data issue, not a training issue.

Output contract

A successful preparation pass should leave:

• a validated source or aligned class-folder directory;
• a recorded image size and margin;
• a bounding-box/no-face log from alignment when alignment was run;
• a second validation report showing class counts after alignment.

Keep raw photos and aligned outputs in separate directories so a failed detector run can be discarded without data loss.

Route onward

• For embedding comparison, classifier training, export, or clustering, continue with ../embeddings-and-classification/SKILL.md.
• For LFW pair metrics, continue with ../evaluation/SKILL.md.
• For softmax/triplet model training, continue with ../training/SKILL.md.
• For model path and tensor-name issues, continue with ../model-export-and-checkpoints/SKILL.md.

Troubleshooting

Read references/troubleshooting.md for invalid class-folder layouts, missing pairs, MTCNN import errors, deprecated SciPy image I/O, and unexpected face counts.