Created by: wojtuss

We are working on INT8v2 quantization of the models: GoogleNet, ResNet101, MobileNetV2, VGG16, VGG19, ShuffleNetV2. Their trained parameters are available at https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification#supported-models-and-performances. In order to enable the models in the unit tests, we need the full models (parameters and program descriptors) to be uploaded to the http://paddle-inference-dist.bj.bcebos.com server. Could you please upload your full models there?

Internally we use program descriptors saved using python scripts from the PaddlePaddle/models repository. However, the FP32 accuracy values we get using them differ a little bit from the values reported on the above page with links to the parameters. We do care about getting the right accuracy and it would be very supportive if you could share your full models.