Hyukjin Jeong
Hyukjin Jeong
The idea looks good to me. Could you show the example of `npy2h5.py`? I'm curious about the format of the post-process script.
Q) Is it necessary to match ONE's HDF5 version with Tizen? ONE builds its own hdf5 library and uses it. It seems that HDF5 file created by 1.8 is compatible...
This is not urgent, but we'll need gelu Op for our backend. https://github.com/tensorflow/tensorflow/commit/ad5e163248a41dc5233840bca3ba2d3d9b1842cb
The pattern may appear when reshape's shape has a negative value as below.  I've made a small (sliced) tflite model converted from the onnx model.  The final result...
I skimmed through the options in circle2circle. As @seanshpark mentioned, `convert_nchw_to_nhwc` should be applied only to the models from onnx. If not, performance will be severely degraded. `make_batchnorm_gamma_positive` should not...
> How do we know whether this model is transformed from onnx? AFAIK, there is no way to check whether a model (pb/tflite/circle) is transformed from onnx. We can guess...
In https://github.com/Samsung/ONE/issues/9369#issuecomment-1172005559, `RemoveQuantDequantSeq` and `RemoveFakeQuant` should not be turned on by default (even for npu). They change the semantics of a model. Those options were implemented for experiment. I think...
The options in https://github.com/Samsung/ONE/issues/9369#issuecomment-1182739358 were introduced for onnx models, but they can be used also for other frontends (tf, tflite). Among the list, I think onnx-only option is only `ConvertNCHWToNHWC`.
I get it. Let's go with that option :)
I think nnpkg and quantization have different formats because they have different purposes. > testDataGenerator create hdf5 input data for nnpkg. So, it is hard to test quantization tools with...