Fine-tuned SigLIP Image
summary
Mercari fine-tunes SigLIP on proprietary image–text pairs to replace MobileNet for Similar Looks, yielding 1.5x tap rate and 14% purchase lift on item pages. Training uses Vertex AI pipelines, BigQuery for data, Dockerized jobs and Weights & Biases for metrics; deployment runs on L4 GPUs in GKE with Triton Inference Server, supporting ONNX/TensorRT and custom CPU preprocessing. Embeddings undergo PCA for dimensionality reduction, then feed an image-search index for similarity and retrieval.
AI-written summary · about
related talks
20mGenAI synthetic data used to fine tune a domain specific model
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
22mHarnessing image and text data for product matching
20mPragmatic Deep Learning for image labelling
46mML Use Cases at Think Big Analytics
33mMachine Learning in Visual Search
20mProductionizing On-device AI Models
36mLeveraging Machine Learning APIs to Write Better Software