opencv dnn unsupported data type float16

[dodler]

opencv-python-4.4.0.46
onnx 1.6.0
onnxmltools 1.7.0
onnxruntime 1.4.0

Hi, I'm using opencv dnn to load fp16 onnx model, but it fails with the following error

cv2.error: OpenCV(4.4.0) /private/var/folders/nz/vv4_9tw56nv9k3tkvyszvwg80000gn/T/pip-req-build-ucld1hvm/opencv/modules/dnn/src/onnx/onnx_graph_simplifier.cpp:527: error: (-210:Unsupported format or combination of formats) Unsupported data type: FLOAT16 in function 'getMatFromTensor'

My code:

import cv2 import sys net = cv2.dnn.readNetFromONNX(sys.argv[1]) img = cv2.cvtColor(cv2.imread("img.jpg"), cv2.COLOR_BGR2RGB) img=cv2.resize(img, (224,224)) blob= cv2.dnn.blobFromImage(img, size=(224,224)) net.setInput(blob) print(net.forward())

It works well with original fp32 model. I get fp16 model using onnxmltools using this
from onnxmltools.utils.float16_converter import convert_float_to_float16

The same problem occurs with int8 model.

[krush11]

@alalek May i work on this issue? Although i would need some help

[alalek]

Try convert FLOAT16 ONNX tensors to FP32 and check if it works (with adding some simple test):

• prepare .onnx model for reproducer (you may want to convert some model from this list or try to generate synthetic models by this script)
• add tensor conversion to FP32
• prepare test with FP16 model

---

Alternative is creating Mat as CV_FP16 (master branch only), but we need to check/fix its usage through whole DNN library implementation (there are many places where FP32 is supported / expected only). This is long way.

[zihaomu]

We have supported ONNX float16 model in #22337.