Model Zoo Models
Deci model zoo is a performance-first deep learning model repository.
The Deci model zoo is a collection of pre-trained, state-of-the-art models - public or optimized for inference. The public models are in the ONNX format, while the optimized can be either ONNX or TRT (For GPU) or Open-Vino (for CPU).
Each model can be optimized and downloaded from the platform as well as get deployed using Deci’s inference engine, INFERY - Deci’s Python packaged inference engine.
To use a model, visit the model zoo in your account, choose a model to clone to your lab in Deci platform, and in Lab, click the Download button on the top right. This will guide you on how to deploy, benchmark, and run inference on this model in your production/testing environment.
Models
Deci model zoo currently holds only public Image classification models.
Image Classification
This collection of models takes images as input, then classifies the major objects in the images into 1000 object categories such as keyboard, mouse, pencil, and many animals.
| Model Architecture | Training Dataset | Description | Top-1 | Batch Size | Framework Version | Input | Output |
|---|---|---|---|---|---|---|---|
| ResNet50 | ImageNet (3, 224, 224) | ResNet models provide very high accuracies with affordable model sizes. They are ideal for cases when high accuracy of classification is required. | 76.206% | Dynamic | ONNX 1.8.1 Opset 11 | All pre-trained models expect input JPEG images normalized in the same way (3, 224, 224). | The model outputs image scores for each of the 1000 classes of ImageNet. |
| ResNet101 | ImageNet (3, 224, 224) | ResNet models provide very high accuracies with affordable model sizes. They are ideal for cases when high accuracy of classification is required. | 77.37% | Dynamic | ONNX 1.8.1 Opset 11 | All pre-trained models expect input JPEG images normalized in the same way (3, 224, 224). | The model outputs image scores for each of the 1000 classes of ImageNet. |
| ResNet34 | ImageNet | ResNet models provide very high accuracies with affordable model sizes. They are ideal for cases when high accuracy of classification is required. | 73.886% | Dynamic | ONNX 1.8.1 Opset 11 | All pre-trained models expect input JPEG images normalized in the same way (3, 224, 224). | The model outputs image scores for each of the 1000 classes of ImageNet. |
| ResNet18 | ImageNet | ResNet models provide very high accuracies with affordable model sizes. They are ideal for cases when high accuracy of classification is required. | 70.538% | Dynamic | ONNX 1.8.1 Opset 11 | All pre-trained models expect input JPEG images normalized in the same way (3, 224, 224). | The model outputs image scores for each of the 1000 classes of ImageNet. |
| DenseNet121 | ImageNet | DensNet is a fully-connected model providing strong gradient flow and more diversified features | 75.522% | Dynamic | ONNX 1.8.1 Opset 11 | All pre-trained models expect input JPEG images normalized in the same way (3, 224, 224). | The model outputs image scores for each of the 1000 classes of ImageNet. |
| DenseNet161 | ImageNet | DensNet is a fully-connected model providing strong gradient flow and more diversified features | 77.37% | Dynamic | ONNX 1.8.1 Opset 11 | All pre-trained models expect input JPEG images normalized in the same way (3, 224, 224). | The model outputs image scores for each of the 1000 classes of ImageNet. |
| DenseNet169 | ImageNet | DensNet is a fully-connected model providing strong gradient flow and more diversified features | 77.088% | Dynamic | ONNX 1.8.1 Opset 11 | All pre-trained models expect input JPEG images normalized in the same way (3, 224, 224). | The model outputs image scores for each of the 1000 classes of ImageNet. |
| DenseNet201 | ImageNet | DensNet is a fully-connected model providing strong gradient flow and more diversified features | 77.2% | Dynamic | ONNX 1.8.1 Opset 11 | All pre-trained models expect input JPEG images normalized in the same way (3, 224, 224). | The model outputs image scores for each of the 1000 classes of ImageNet. |
Usage
In order to use the model for either inference or performance benchmarking of accuracy or run-time efficiency (throughput/latency), you should use our inference engine, INFERY - Deci’s Python packaged inference engine.
Click on the above links to go over the quick-start and understand the benefit of each for your organization.
Updated 3 months ago