- Add new models: NCF for recommendation class, DSSD for object detection class
- Nvidia docker has license issue on distribution, users have to download by themselves. Add script to install some dependencies
- Add md5 checksum for some big files to help us spot the download issues
- Add multi-card training for DIN model
- Add accuracy test cases in Caffe CNN models. Inference engine from different vendor could compare not only performance number but also accuracy loss
- Add the trained checkpoint file for googlenet, resnet50, resnet152, densenet121
- Add multi-card training in for CNN-Tensorflow, SSD, MaskRCNN, NMT
- Reorganize the automation workflow to improve the running scripts quality.
- Users can choose run all of application in a few scripts or each application separately.
- Add preprocessing script to extract and save data to csv file.
- Remove Alexnet as it is out of date.
- Remove Vgg16 as it is repeatedly used in SSD test.
- Add TensorRT-5 inference script for Caffe model.
- Add Tensorcore FP16 GEMM in micro tests.
DIN model:
- Change the inference workload to apply 100 items for each user to recommend.
- The inference batch size is based on number of users. It is set to 1, 32, and 64. Iteration of 1000 is applied to minimize the overhead.