ResNet50 Embeddings (ONNX)
This repository provides an ONNX version of the TensorFlow/Keras ResNet50 model configured for feature extraction.
The original model is the ImageNet-pretrained ResNet50 distributed through TensorFlow/Keras.
The model has been converted to ONNX format for lightweight inference using ONNX Runtime.
Features
- ResNet50 backbone
- ImageNet pretrained weights
- Classification head removed (
include_top=False) - Global Average Pooling (
pooling="avg") - 2048-dimensional image embeddings
- ONNX Runtime compatible
Input
- RGB image
- Shape:
(1, 3, 224, 224)
Preprocessing:
- Resize to 224×224
- Convert RGB
- Apply ImageNet normalization
Output
(1, 2048)
The output embedding is intended for:
- Image similarity search
- Product recommendation
- Duplicate image detection
- Visual search
Compatibility
- ONNX Runtime >= 1.18
- Python >= 3.10
Model Origin
Original implementation:
- TensorFlow/Keras ResNet50
- ImageNet pretrained weights
This repository only converts the original model to ONNX format for deployment convenience.
Disclaimer
This repository is not affiliated with TensorFlow or Google.