Wei-Sheng Chin comments

Results 37 comments of


                                            Wei-Sheng Chin

encounting the below error, please help me

Dynamic axis means axis of a tensor. `eneities` and `relations` are not tensors so it may throw. If possible, I'd suggest to encode your data as tensors instead of dictionary;...

ONNX Training work for ONNX release 1.8 (and maybe beyond)

Item 4) is to check if the use of Gradient is reasonable. To compute gradient of tensor `Y` w.r.t. tensor `X`, we need to make sure the path from `X`...

MaxPool and AveragePool conversion doesn't need explicit Pad with ONNX 1.2

Consequently, things to do would be (1) replace Pad with the Pool's explicit padding for ONNX > 1.2. Then, no-op Pad and the extra Pads will be naturally removed (2)...

Core ML Tiny-yoloV2 confidence drops significantly after conversion

One possibility is that input format is not expected. It can be either RGB, GBR, and so on. Could you try a constant image (for example, a tensor with all...

[Performance] Model gets slower after Training, when nchwc is used

This is just my guess. Without training, some tensors are initialized to `0` so those operators such as `y=Add(x, 0)` can be eliminated. To confirm, I need to see the...

does nnapi (armv7a or armv8a) support model contain lstm ?

I don't think NNAPI EP supports LSTM. There is no such op in [its op registration code](https://github.com/microsoft/onnxruntime/blob/03181caeaef9343bd2e8d7999386893562ebb23a/onnxruntime/core/providers/nnapi/nnapi_builtin/builders/op_builder.cc#L2633).

ONNX Runtime much slower than PyTorch (2-3x slower)

Can you try running with ORTModule? You can just wrap your nn.Module model via ```python from onnxruntime.training.ortmodule import ORTModule new_model = ORTModule(model) # ORTModule is also nn.Module so just use...

ONNX Runtime much slower than PyTorch (2-3x slower)

Could you run `nsys profile` with your model w/wo onnxruntime? It was easy to me to identify which part is the performance bottleneck when I have profiling result. For example,...

ONNX Runtime much slower than PyTorch (2-3x slower)

Why do you have so many Cast's in [this reply's](https://github.com/microsoft/onnxruntime/issues/12880#issuecomment-1240455413) figure? ORT recently adds support for "strided" tensors, so I expect those Cast's are no-op's. If I am wrong and...

ONNX Runtime much slower than PyTorch (2-3x slower)

@thomas-beznik, sure thing. If #9754 is the blocker, you probably need to [build ORT from source](https://tomwildenhain-microsoft.github.io/onnxruntime/docs/build/training.html). Note that you need a clean machine to avoid dependency interference for a clean...