Last updated: November 17, 2020
The Neural Magic Model Repo includes pre-trained, performance-optimized models ready to use for your machine learning projects
The Model Repo features models sparsified with the latest pruning techniques to deliver exceptional performance on CPUs, and accelerates the process of deploying those models in production.
Currently, teams can choose from a growing library of popular image classification models in the table below. More models are being added regularly. Unlike other repositories, Neural Magic already did the hard work of building, pruning, and re-training the models for immediate use in production.
Existing models from the Neural Magic Model Repo can be transfer-learned with customers’ data, and then recalibrated. This approach makes it easier for teams to use Neural Magic pre-optimized models with their own data, in their own environments.
There are three available types*:
- Base: The baseline model (standard training process).
- Recal: A model that is recalibrated for performance optimization on the Neural Magic Inference Engine. This model achieves ~100% of baseline validation metrics.
- Recal-perf: A model that is recalibrated for performance optimization on the Neural Magic Inference Engine. This model achieves ~99% of baseline validation metrics.
|Architecture||Dataset||Available Types*||Frameworks||Validation Baseline Metric|
|EfficientNet-B0||ImageNet||base, recal-perf||ONNX, PyTorch||77.30%|
|EfficientNet-B4||ImageNet||base, recal-perf||ONNX, PyTorch||83.00%|
|Inception-v3||ImageNet||base, recal, recal-perf||ONNX, PyTorch||77.45%|
|MnistNet||MNIST||base||ONNX, PyTorch, TensorFlow||~99.00%|
|MobileNetV1||ImageNet||base, recal, recal-perf||ONNX, PyTorch, TensorFlow||70.90%|
|MobileNetV2||ImageNet||base||ONNX, PyTorch, TensorFlow||71.88%|
|MobileNetV2-SSDLite||COCO||base||ONNX, PyTorch||35.7% [email protected]|
|MobileNetV2-SSDLite||VOC||base||ONNX, PyTorch||43.5% [email protected]|
|ResNet-18||ImageNet||base, recal||ONNX, PyTorch, TensorFlow||69.80%|
|ResNet-34||ImageNet||base, recal||ONNX, PyTorch, TensorFlow||73.30%|
|ResNet-50||ImageNet||base, recal, recal-perf||ONNX, PyTorch, TensorFlow||76.10%|
|ResNet-50-SSD-300||COCO||base, recal-perf||ONNX, PyTorch||42.70% [email protected]|
|ResNet-50-SSD-300||VOC||base, recal-perf||ONNX, PyTorch||52.20% [email protected]|
|ResNet-50 2xwidth||ImageNet||base||ONNX, PyTorch||78.51%|
|ResNet-101||ImageNet||base, recal, recal-perf||ONNX, PyTorch, TensorFlow||77.37%|
|ResNet-101 2xwidth||ImageNet||base||ONNX, PyTorch||78.84%|
|ResNet-152||ImageNet||base,recal-perf||ONNX, PyTorch, TensorFlow||78.31%|
|VGG-11||ImageNet||base,recal-perf||ONNX, PyTorch, TensorFlow||69.02%|
|VGG-11-BN||ImageNet||base||ONNX, PyTorch, TensorFlow||70.38%|
|VGG-13||ImageNet||base||ONNX, PyTorch, TensorFlow||69.93%|
|VGG-13-BN||ImageNet||base||ONNX, PyTorch, TensorFlow||71.55%|
|VGG-16||ImageNet||base, recal, recal-perf||ONNX, PyTorch, TensorFlow||71.59%|
|VGG-16-BN||ImageNet||base||ONNX, PyTorch, TensorFlow||71.55%|
|VGG-19||ImageNet||base,recal-perf||ONNX, PyTorch, TensorFlow||72.38%|
|VGG-19-BN||ImageNet||base||ONNX, PyTorch, TensorFlow||74.24%|
|YOLOv3||COCO||base,recal-perf||ONNX, PyTorch||68.6% [email protected]|
More models are added regularly. Contact us to learn more and to get access to the Neural Magic Model Repo.