def run_end2end_test(tflite_path, onnx_path, tflite_layout, onnx_layout, tensors):
io_layouts = {t: (tflite_layout, onnx_layout) for t in tensors}
t2o.convert(tflite_path, onnx_path, io_layouts)
m = shrub.tflite.parse(tflite_path, tflite_layout)
m.genInput()
onnx_ret = shrub.onnx.run(onnx_path, m.inputs, onnx_layout)
tflite_ret = shrub.tflite.run(tflite_path, m.inputs, tflite_layout)
assert(shrub.network.cmpTensors(onnx_ret, tflite_ret, useLayout=tflite_layout))
def end2end_test(model_name, use_layout):
cur_dir = os.path.dirname(os.path.abspath(__file__))
tflm_dir = os.path.abspath(cur_dir + '/../assets/tests')
tflm_name = model_name + '.tflite'
onnx_name = model_name + '.onnx'
tflm_path = os.path.join(tflm_dir, tflm_name)
t2o.convert(tflm_path, onnx_name)
m = shrub.tflite.parse(tflm_path)
m.genInput()
onnx_ret = shrub.onnx.run(onnx_name, m.inputs, use_layout)
tflite_ret = shrub.tflite.run(tflm_path, m.inputs)
assert(shrub.network.cmpTensors(onnx_ret, tflite_ret, useLayout=use_layout))
def end2end_test(model_name, use_layout, atol, rtol):
cur_dir = os.path.dirname(os.path.abspath(__file__))
tflm_dir = os.path.abspath(cur_dir + '/../assets/tests')
tflm_name = model_name + '.tflite'
onnx_name = model_name + '.onnx'
tflm_path = os.path.join(tflm_dir, tflm_name)
t2o.convert(tflm_path, onnx_name)
m = shrub.tflite.parse(tflm_path)
m.genInput()
# TFLite model is supposed to be end to end quantized
tflite_ret = shrub.tflite.run(tflm_path, m.inputs)
oquant = shrub.tflite.parseQuantParam(tflm_path, False)[0]
foutputs = list()
for f in tflite_ret:
foutput = copy.deepcopy(f)
foutput.quant = oquant
foutput.dequantize()
foutputs.append(foutput)
# ONNX model is supposed to be only several operators quantized
iquant = shrub.tflite.parseQuantParam(tflm_path, True)[0]
finputs = list()
for q in m.inputs:
finput = copy.deepcopy(q)
finput.quant = iquant
finput.dequantize()
finputs.append(finput)
onnx_ret = shrub.onnx.run(onnx_name, finputs, use_layout)
assert (shrub.network.cmpTensors(foutputs,
onnx_ret,
atol=atol,
rtol=rtol,
useLayout=use_layout))
import tf2onnx import tflite2onnx tflite_path = 'checkpoint/SonyTFLite/tf2_sony_int8.tflite' onnx_path = 'checkpoint/int8_sony' tflite2onnx.convert(tflite_path, onnx_path) """DON'T USE THIS!! USE THE FOLLOWING COMMAND: python -m tf2onnx.convert --opset 13 --tflite checkpoint/SonyTFLite/tf2_sony_float32_pad.tflite --output checkpoint/float32_sony_pad.onnx, where after --tflite argument place your tflite model, and after --output put the filepath to the output onnx model """
import tflite2onnx as to
import logging
logging.basicConfig(level=logging.DEBUG)
to.convert('test.tflite', 'test.onnx')
