encode & decode onnx model to qr code

0_qr_encode_file.py

import argparse
import binascii
import tarfile

from io import BytesIO

import qrcode

parser = argparse.ArgumentParser("xx")
parser.add_argument("--input_file", type=str, required=True, help="input file")
args = parser.parse_args()

fp = args.input_file
with open(fp, "rb") as fh:
    onnx_buffer = BytesIO(fh.read())
onnx_buffer.seek(0)

# save to tar buffer
tar_buffer = BytesIO()
# mode = 'w:gz'
mode = 'w:xz'
# mode = 'w:bz2'
tar = tarfile.open(fileobj=tar_buffer, mode=mode)
tarinfo = tarfile.TarInfo(fp)
tarinfo.size = len(onnx_buffer.getvalue())
tar.addfile(tarinfo, onnx_buffer)
tar.close()
tar_buffer.seek(0)

tar_content = tar_buffer.read()
xx_ = binascii.hexlify(tar_content)
print(f"{mode}: {len(xx_)}")

# debug
# b = binascii.a2b_hex(xx_)


MAX_LEN = 2300

for i in range(0, len(xx_), MAX_LEN):
    xx = xx_[i : i + MAX_LEN]
    fp = f'qr_{i}.png'
    print(f"encode to {fp} , len {len(xx)}")
    qrcode.make(xx, version=40).save(fp)

0_qr_encode_onnx.py

import argparse
import binascii
import tarfile
import numpy as np

from io import BytesIO

import onnx
import qrcode


def clean_onnx(fp):
    onnx_model = onnx.load_model(fp)
    for init in onnx_model.graph.initializer:
        num_elements = np.prod(init.dims)
        if num_elements > 10:
            init.ClearField('int32_data')
            init.ClearField('int64_data')
            init.ClearField('float_data')
            init.ClearField('double_data')
            init.ClearField('raw_data')
    return onnx_model


def clean_onnx_no_init(fp):
    onnx_model = onnx.load_model(fp)
    init = onnx_model.graph.initializer
    # TODO: keep axis/step/.. param for Gather/Slice and other Nodes
    for idx in range(len(init)):
        init.pop()
    return onnx_model


parser = argparse.ArgumentParser("xx")
parser.add_argument("--input_file", type=str, required=True, help="input file")
args = parser.parse_args()

# test
# wget https://github.com/onnx/models/raw/main/vision/classification/resnet/model/resnet50-v1-7.onnx
# fp = 'resnet50-v1-7.onnx'

fp = args.input_file
onnx_model = clean_onnx(fp)

# save onnx to buffer
onnx_buffer = BytesIO()
onnx.save_model(onnx_model, onnx_buffer)
onnx_buffer.seek(0)

# save to tar buffer
tar_buffer = BytesIO()
# mode = 'w:gz'
mode = 'w:xz'
# mode = 'w:bz2'
tar = tarfile.open(fileobj=tar_buffer, mode=mode)
tarinfo = tarfile.TarInfo(f"{fp}.clean.onnx")
tarinfo.size = len(onnx_buffer.getvalue())
tar.addfile(tarinfo, onnx_buffer)
tar.close()
tar_buffer.seek(0)

tar_content = tar_buffer.read()
xx_ = binascii.hexlify(tar_content)
print(f"{mode}: {len(xx_)}")

# debug
# b = binascii.a2b_hex(xx_)


MAX_LEN = 2300

for i in range(0, len(xx_), MAX_LEN):
    xx = xx_[i : i + MAX_LEN]
    fp = f'qr_{i}.png'
    print(f"encode to {fp} , len {len(xx)}")
    qrcode.make(xx, version=40).save(fp)

1_qr_decode.py

import binascii
import glob
import os
import tarfile

from io import BytesIO

from PIL import Image
from pyzbar.pyzbar import decode

fps = glob.glob('qr_*.png')
fps = sorted(fps, key=lambda x: int(x.split('_')[1].split('.')[0]))

bytes = b''
for fp in fps:
    x = decode(Image.open(fp))
    xxx = x[0].data
    print(f"dencode from {fp} , len {len(x[0].data)}")
    bytes += xxx

he = binascii.a2b_hex(bytes)
tar_buffer = BytesIO()
tar_buffer.write(he)
tar_buffer.seek(0)

# mode = 'r:gz'
mode = 'r:xz'
# mode = 'r:bz2'

tar = tarfile.open(fileobj=tar_buffer, mode=mode)
tar.extractall('.')

2_restore.py

import onnx
import numpy as np
import onnx.helper
import onnx.numpy_helper
import numpy as np

np.random.seed(1984)



def fill_onnx(fp):
    onnx_model = onnx.load_model(fp)
    for init in onnx_model.graph.initializer:
        num_elements = np.prod(init.dims)
        if num_elements > 10:
            print(f"{init.name}: {init.dims}")
            np_dtype = onnx.helper.tensor_dtype_to_np_dtype(init.data_type)
            array = np.random.uniform(-1, 1, size=init.dims).astype(np_dtype)
            init.raw_data = array.tobytes()
    return onnx_model


fp = 'clean.onnx'

onnx_model = fill_onnx(fp)

fp_restored = f"{fp}.restored.onnx"
onnx.save(onnx_model, fp_restored)

3rds

for encode

onnx

qrcode
https://pypi.org/project/qrcode/
https://github.com/lincolnloop/python-qrcode

for decode

PIL

pyzbar
https://pypi.org/project/pyzbar/
https://github.com/NaturalHistoryMuseum/pyzbar