Deploying TensorFlow Vision Models in Hugging Face with TF Serving

Hugging Face · 4 years ago · 46

Hugging Face and external contributors added multiple vision models in TensorFlow to the Transformers library, including Vision Transformer, Masked Autoencoders, RegNet, and ConvNeXt. The article demonstrates deploying a Vision Transformer model locally using TensorFlow Serving by embedding preprocessing and postprocessing operations into the model's computation graph, then exposing it via REST and gRPC endpoints on ports 8501 and 8500 respectively. This approach reduces training-serving skew by baking image normalization, resizing, and label mapping directly into the model before deployment.