float 16 inference support

[vsooda]

Is your feature request related to a problem? Please describe.
I try to use c-api as this demo, it is ok. but when I try to quantize squeeze.onnx to float 16, and then inference again, it report:

Unexpected input data type. Actual: (N11onnxruntime11NonOnnxTypeIfEE) , expected: (N11onnxruntime11NonOnnxTypeINS_9MLFloat16EEE)

I use the follow script to convert float32 model to float16:

import onnxmltools
from onnxmltools.utils.float16_converter import convert_float_to_float16

input_onnx_model = 'squeezenet/model.onnx'
output_onnx_model = 'squeezenet/model_fp16.onnx'

onnx_model = onnxmltools.utils.load_model(input_onnx_model)
onnx_model = convert_float_to_float16(onnx_model)
onnxmltools.utils.save_model(onnx_model, output_onnx_model)

can onnxruntime support fp16 inference? any plan?

System information
0.4

Describe the solution you'd like
load fp 16 model, input float 32 data, then get float 32 result.

Describe alternatives you've considered
how to convert float32 input to float16 for inference?

[pranavsharma]

Here's a unit test that demonstrates conversion of float32 input to float16 -

onnxruntime/onnxruntime/test/framework/float_16_test.cc

Line 161 in 2663b9c

values_x.push_back(MLFloat16(math::floatToHalf(i)));

. Let us know if this doesn't work for you.

[vsooda]

@pranavsharma thank you. it works