def converto_fp16():# Update the input name and path for your ONNX model
import onnxmltools
from onnxmltools.utils.float16_converter import convert_float_to_float16
input_onnx_model = '/home/sobey/model.onnx'
# Change this path to the output name and path for your float16 ONNX model
output_onnx_model = '/home/sobey/model_f16.onnx'
# Load your model
onnx_model = onnxmltools.utils.load_model(input_onnx_model)
# Convert tensor float type from your input ONNX model to tensor float16
onnx_model = convert_float_to_float16(onnx_model)
# Save as protobuf
onnxmltools.utils.save_model(onnx_model, output_onnx_model)
print("saved model in ",output_onnx_model)
r_hyps_pad_sos_eos)
ort_outs = ort_session.run(None, ort_inputs)
# check encoder output
test(to_numpy(o0), ort_outs[0], rtol=1e-03, atol=1e-05)
logger.info("export to onnx decoder succeed!")
if args.fp16:
try:
import onnxmltools
from onnxmltools.utils.float16_converter import convert_float_to_float16
except ImportError:
print('Please install onnxmltools!')
sys.exit(1)
encoder_onnx_model = onnxmltools.utils.load_model(encoder_onnx_path)
encoder_onnx_model = convert_float_to_float16(encoder_onnx_model)
encoder_onnx_path = os.path.join(args.output_onnx_dir,
'encoder_fp16.onnx')
onnxmltools.utils.save_model(encoder_onnx_model, encoder_onnx_path)
decoder_onnx_model = onnxmltools.utils.load_model(decoder_onnx_path)
decoder_onnx_model = convert_float_to_float16(decoder_onnx_model)
decoder_onnx_path = os.path.join(args.output_onnx_dir,
'decoder_fp16.onnx')
onnxmltools.utils.save_model(decoder_onnx_model, decoder_onnx_path)
# dump configurations
onnx_config = {
"beam_size": args.beam_size,
"reverse_weight": args.reverse_weight,
"ctc_weight": args.ctc_weight,
"fp16": args.fp16
}
import onnxmltools from onnxmltools.utils.float16_converter import convert_float_to_float16 input_onnx_model = './hfnet_github.onnx' # fp32 onnx model output_onnx_model = './hfnet_github_fp16.onnx' # fp16 onnx model onnx_model = onnxmltools.utils.load_model(input_onnx_model) onnx_model = convert_float_to_float16(onnx_model) # onnx_model.graph.output.pop() # onnx_model.graph.output.pop() # onnx_model.graph.output.pop() onnxmltools.utils.save_model(onnx_model, output_onnx_model)
# usage: python convert_fp16_onnx.py <fp32 ONNX file> <fp16 ONNX file>
#
# This script calls the onnxmltools to convert a fp32 model to fp16.
# Note that for some models, this can take a long time, e.g. measured 8 hours
# to convert resnet50i64 model from torchvision. Other models are done within
# seconds.
#
import onnx
import sys
from onnx import optimizer
from onnxmltools.utils.float16_converter import convert_float_to_float16
if len(sys.argv) == 3:
original_model = onnx.load(sys.argv[1])
new_model = convert_float_to_float16(original_model)
onnx.save_model(new_model,sys.argv[2])