Cast int64 Tensor to float32 in pre-trained model?

#1

So I’m trying to load a certain ONNX model in OpenCV (started in PyTorch), but it crashes with this error:

OpenCV(4.1.0) Error: Assertion failed (inputs.size()) in cv::dnn::dnn4_v20190122::Layer::getMemoryShapes, file c:\build\master_winpack-build-win64-vc15\opencv\modules\dnn\src\dnn.cpp, line 3616

It gives me this error when I try to convert that ONNX model to Tensorflow:

ValueError: Tensor conversion requested dtype float32 for Tensor with dtype int64: 'Tensor("Mul_3:0", shape=(), dtype=int64)'

I suppose this is why it isn’t working in OpenCV, is there any way to cast the int64 Tensor to a float32 one? I can try to modify the float32 <-> float16 conversion function so that it can also convert int64 to float16, but think this could be quite complex or impossible (I would have to move the weight up by 19 digits?).

So I’m basically asking how I could cast at int64 to a float32 in an existing pre-trained Pytorch model, while getting the same result in the neural network (don’t want to retrain it)

#2

By forcibly type casting the weights and resaving the model?

#3

Is there any tutorial on how to cast it and re-save the model? I don’t need to train the model again for this right?

In OpenCV I get this error now:

OpenCV(4.1.0) Error: Assertion failed (inputs.size()) in cv::dnn::dnn4_v20190122::Layer::getMemoryShapes, file c:\build\master_winpack-build-win64-vc15\opencv\modules\dnn\src\dnn.cpp, line 3616

I see that someone converting the model to Tensorflow somehow solved this issue, but I can’t even manage to convert it to Tensorflow… Although casting and resaving the model should fix it!