The experiment investigates whether spatial colour features, mild class balancing, test-time augmentation, and validation-tuned consistency thresholds can improve an already strong fashion attribute classifier.

Architecture

The model uses:

CLIP ViT-B/32 image encoder
Seven task-specific classification heads
Global HSV and RGB colour statistics
Spatial colour-grid features
Hierarchical residual connections
Article-type-to-gender residual connection
Validation-selected test-time augmentation
Lightweight consistency correction

Dataset

ashraq/fashion-product-images-small

Split	Percentage
Training	70%
Validation	15%
Test	15%

The held-out test set contains 6,611 images.

Corrected Test Metrics

Metric	Result
Average Accuracy	87.87%
Average Macro F1	67.94%
Average Weighted F1	87.47%
Average Top-3 Accuracy	98.13%
Exact-Match Accuracy	41.75%
Test Samples	6,611

Exact-match accuracy requires all seven attributes to be correct for the same image.

Per-Attribute Results

Attribute	Accuracy	Macro F1	Weighted F1	Top-3 Accuracy
Gender	91.98%	81.11%	91.79%	99.76%
Master Category	99.56%	86.59%	99.47%	99.94%
Subcategory	96.58%	76.25%	96.41%	99.71%
Article Type	88.49%	66.56%	87.81%	98.12%
Base Colour	70.29%	37.05%	69.05%	90.55%
Season	76.06%	77.31%	75.92%	99.02%
Usage	92.10%	50.74%	91.83%	99.83%

Raw Model Metrics

The following metrics were calculated before consistency correction.

Metric	Result
Average Accuracy	87.84%
Average Macro F1	67.61%
Average Weighted F1	87.45%
Average Top-3 Accuracy	98.13%
Exact-Match Accuracy	41.48%

Raw and corrected metrics are reported separately for transparency.

Inference Configuration

Test-time augmentation enabled: True
Consistency thresholds selected using validation data only
Final test data was not used for threshold selection
Single-image average latency: 12.92 ms
Single-image P95 latency: 13.74 ms
Latency measured on an NVIDIA Tesla P100 GPU

Production Decision

This experimental checkpoint achieved measurable but relatively small gains over the existing production checkpoint.

The simpler production model was retained because it provides nearly equivalent predictive performance with:

Lower inference latency
A simpler inference pipeline
Lower deployment complexity
Easier maintenance
Better alignment with the existing application

This repository is preserved as an optimization experiment and reproducible research checkpoint.

Repository Files

text

model.pt
config.json
label_maps.json
consistency_rules.json
consistency_thresholds.json
metrics.json
README.md

Loading

This repository contains a custom PyTorch architecture and cannot be loaded directly with AutoModelForImageClassification.from_pretrained().

The project implementation must:

Load config.json
Load label_maps.json
Initialize AutoCatalogFinalClassifier
Load model.pt
Generate global and spatial colour features
Apply TTA when enabled
Optionally apply the saved consistency thresholds

Intended Use

This checkpoint is intended for:

Fashion attribute-classification experiments
Multi-task learning analysis
Colour-feature research
Test-time augmentation evaluation
Performance-versus-latency comparisons
Reproducible portfolio experimentation

Limitations

Rare colour and usage classes remain difficult
Base-colour labels may be visually ambiguous
TTA increases single-image inference latency
Performance may decrease on marketplace images outside the training distribution
Season and usage are not always directly visible from an image
The model is not intended to infer personal attributes about people

The gender label represents the dataset's product-target category, not the gender of a person.

Version