Closed psyhtest closed 5 years ago
I vaguely recall that the CK implementation uses *.ckpt
files, while the reference one needs a *.pb
file. However, the official TF ResNet50 page also mentions *.ckpt
. It appears we need the checkpoint listed as "ResNet-50 v1 (fp32, Accuracy 76.53%)" ("v1" means "v1.5", as explained there).
We should also check if SavedModels is something we might want to support as well (perhaps to be converted to TFLite or to proprietary inference engines).
For reference, this is how program:image-classification-tf-cpp
is used in the MLPerf MobileNet reference.
The package ck-tensorflow:package:tensorflowmodel-resnet50-v1.5-mlperf
has been implemented and program:image-classification-tf-cpp
agrees to take it as input.
The MLPerf ResNet50 reference by @guschmue supports weights both in the TensorFlow and ONNX formats.
@guschmue notes:
In fact, we already have such an implementation wrapped in CK: ck-tensorflow:program:image-classification-tf-cpp. Let's check it works with the ResNet50 weights in the TensorFlow format by providing a CK package.