This repository contains pre-trained Keras model files for CrossSim. This is not a self-executing repository. Please see the main CrossSim repository for more information about the tool.
The following pre-trained models are available in this repository for use with CrossSim Inference:
- MNIST:
- Two shallow CNNs with different numbers of channels, trained by CrossSim's developers
- CIFAR-10: all models were trained by CrossSim's developers
- Four ResNets of different depths, following the topology specified in the original ResNet paper: ResNet14, ResNet20, ResNet32, ResNet56
- A shallow CNN that uses bounded ReLU (cifar10_cnn_brelu)
- A sequential CNN with binary weights and activations using the Larq package
- A sequential CNN with binary activations using the Whetstone package
- CIFAR-100:
- A ResNet56 network, trained by CrossSim's developers
- ImageNet: all models were obtained from external sources
- ResNet50, InceptionV3, MobileNetV1, and MobileNetV2 from keras.applications
- ResNet50v1.5 and MobileNetV1 quantized (int8) from the benchmark suite in MLPerf Inference
- ResNet50v1.5 quantized (int4) submitted by Nvidia for MLPerf Inference