Example code for handling TensorRT INT8 calibration file

calib_to_float.py

from pathlib import Path
import struct

lines = Path('calib_yolov4-int8-608.bin').read_text().splitlines()
for line in lines:
    pair = line.split(':')
    if len(pair) != 2:
        continue
    assert len(pair[1]) == 9
    bstr = bytes.fromhex(pair[1][1:])    # convert to byte string
    unit = struct.unpack('>f', bstr)[0]  # convert to float32
    print('%s:\t%f' % (pair[0], unit))

find_concats,py

import tensorrt as trt

TRT_LOGGER = trt.Logger(trt.Logger.VERBOSE)
EXPLICIT_BATCH = [1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)]
builder = trt.Builder(TRT_LOGGER)
network = builder.create_network(*EXPLICIT_BATCH)
parser = trt.OnnxParser(network, TRT_LOGGER)
with open('yolov4-int8-608.onnx', 'rb') as f:
    parser.parse(f.read())
shape = list(network.get_input(0).shape)
shape[0] = 1
network.get_input(0).shape = shape

for layer in network:
    if layer.type == trt.LayerType.CONCATENATION:
        inputs = [layer.get_input(i).name for i in range(layer.num_inputs)]
        outputs = [layer.get_output(i).name for i in range(layer.num_outputs)]
        print('Layer %s' % layer.name)
        print('    inputs:  %s' % str(inputs))
        print('    outputs: %s' % str(outputs))