Supported models
Supported model lists
Model Compressor's compatibility is not limited to these model list.
The list shows validated environment and model examples to help you to start with.
You can easily get the model from torchvision.models, tf.keras.applications, or Model Compressor Model Zoo.
PyTorch - torchvision.models
- Supported version: PyTorch, ONNX version β₯ 1.10
- If a model is defined in PyTorch, it should be converted into GraphModule or ONNX format before being uploaded.
- How-to-guide for PyTorch GraphModule conversion is at the GitHub.
- How-to-guide for ONNX conversion is at the GitHub.
Classification Models
ImageNet pre-trained models
Model | Structured Pruning | Filter Decomposition | PyTorch & ONNX Version |
---|---|---|---|
Alexnet | O | O | 1.10.x |
VGG16 | O | O | 1.10.x |
ResNet18 | O | O | 1.10.x |
SqueezeNet1_0 | O | O | 1.10.x |
DenseNet161 | O | O | 1.10.x |
MobileNet_v2 | O | O | 1.10.x |
MobileNet_v3_large | O | O | 1.10.x |
MobileNet_v3_small | O | O | 1.10.x |
Wide ResNet50_2 | O | O | 1.10.x |
MNASNet1_0 | O | O | 1.10.x |
EfficientNet_b0 | O | O | 1.10.x |
ResNext50_32x4d | O | O | 1.10.x |
RegNet_y_400mf | O | O | 1.10.x |
RegNet_x_400mf | O | O | 1.10.x |
Semantic Segmentation Models
COCO2017 pre-trained models
Model | Structured Pruning | Filter Decomposition | PyTorch & ONNX Version |
---|---|---|---|
FCN ResNet50 | O | O | 1.10.x |
FCN ResNet101 | O | O | 1.10.x |
TensorFlow-Keras - tf.keras.applications
- Supported version: TensorFlow 2.3~2.8.
- Custom layer must not be included in Keras H5 format (.h5).
- The model must contain not only weights but also the structure of the model (do not use save_weights).
- If there is a custom layer in the model, please upload with TensorFlow SavedModel format (.zip).
ImageNet pre-trained models
Model | Structured Pruning | Filter Decomposition | TensorFlow Version |
---|---|---|---|
VGG16 | O | O | 2.3.x ~ 2.8.x |
VGG19 | O | O | 2.3.x ~ 2.8.x |
ResNet50 | O | O | 2.3.x ~ 2.8.x |
ResNet101 | O | O | 2.3.x ~ 2.8.x |
ResNet152 | O | O | 2.3.x ~ 2.8.x |
ResNet50V2 | O | O | 2.3.x ~ 2.8.x |
ResNet101V2 | O | O | 2.3.x ~ 2.8.x |
ResNet152V2 | O | O | 2.3.x ~ 2.8.x |
InceptionV3 | O | O | 2.3.x ~ 2.8.x |
MobileNet | O | O | 2.3.x ~ 2.8.x |
MobileNetV2 | O | O | 2.3.x ~ 2.8.x |
DenseNet121 | O | O | 2.3.x ~ 2.8.x |
DenseNet169 | O | O | 2.3.x ~ 2.8.x |
DenseNet201 | O | O | 2.3.x ~ 2.8.x |
EfficientNetB1-B7 | O | O | 2.3.x ~ 2.8.x |
Xception | O | O | 2.3.x ~ 2.8.x |
InceptionResNetV2 | O | X (WIP) | 2.3.x ~ 2.8.x |
NASNet | O | X (WIP) | 2.3.x ~ 2.8.x |
- These models and best practices are provided in NetsPresso Model Compressor Model Zoo.
- You can follow the compression process with Google Colab.
Classification Models
[PyTorch] CIFAR100 pre-trained models
Model | Structured Pruning | Filter Decomposition | PyTorch & ONNX Version |
---|---|---|---|
MobileNetV2 | O | O | 1.10.x |
RepVGG-A1 | O | O | 1.10.x |
ResNet56 | O | O | 1.10.x |
VGG16-BN | O | O | 1.10.x |
[TensorFlow-Keras] CIFAR100 pre-trained models
Model | Structured Pruning | Filter Decomposition | TensorFlow Version |
---|---|---|---|
VGG19 | O | O | 2.3.x ~ 2.8.x |
ResNet50 | O | O | 2.3.x ~ 2.8.x |
MobileNetV1 | O | O | 2.3.x ~ 2.8.x |
Object Detection Models
[PyTorch] COCO pre-trained models
Model | Structured Pruning | Filter Decomposition | PyTorch & ONNX Version |
---|---|---|---|
YOLOX | O | O | 1.10.x |
[TensorFlow-Keras] Pascal VOC pre-trained models
Model | Structured Pruning | Filter Decomposition | TensorFlow Version |
---|---|---|---|
YOLOv4 | O | O | 2.3.x ~ 2.8.x |
Updated about 1 year ago