Metadata TFlite/ Keras to Tensorflow .pb conversion
VanqCoding opened this issue · 5 comments
Hello David,
A post-training-int-quantized tflite tiny yolo3 model is giving me an error "Input tensor has type kTLiteFloat32: it requires specifying NormalizationOptions metadata to preprocess input images" when trying to run it on a Raspberry Pi.
Is there a way to generate a metadata for it?
Second issue: I have no success converting my dumped .h5 model to .pb format. I have tried keras_to_tensorflow.py and keras_to_tensorflow_bk.py
2022-12-15 18:32:51,192 - INFO - Saved the graph definition in ascii format at C:\Desktop\keras-YOLOv3-model-set-master.pbtxt Traceback (most recent call last): File "tools/model_converter/keras_to_tensorflow.py", line 176, in <module> main() File "tools/model_converter/keras_to_tensorflow.py", line 172, in main keras_to_tensorflow(args) File "tools/model_converter/keras_to_tensorflow.py", line 136, in keras_to_tensorflow from tensorflow.tools.graph_transforms import TransformGraph ModuleNotFoundError: No module named 'tensorflow.tools.graph_transforms'
Is this a tensorflow version compatibility issue?
@david8862 I am on TF 2.10.1, using a conda env on Windows 11.
Training/eval and everything else has been working pretty well so far.
here is the full package list:
_tflow_select 2.1.0 gpu
abseil-cpp 20210324.2 hd77b12b_0
absl-py 1.3.0 py37haa95532_0
aiohttp 3.8.1 py37h2bbff1b_1
aiosignal 1.2.0 pyhd3eb1b0_0
appdirs 1.4.4 pypi_0 pypi
arcgispro 3.0 0 esri
astor 0.8.1 py37haa95532_0
astunparse 1.6.3 py_0
async-timeout 4.0.2 py37haa95532_0
asynctest 0.13.0 py_0 conda-forge
attrs 22.1.0 py37haa95532_0
blas 1.0 mkl conda-forge
blinker 1.4 py37haa95532_0
bokeh 2.4.3 pypi_0 pypi
brotlipy 0.7.0 py37h2bbff1b_1003
ca-certificates 2022.10.11 haa95532_0
cachetools 4.2.2 pyhd3eb1b0_0
certifi 2022.9.24 py37haa95532_0
cffi 1.15.1 py37h2bbff1b_0
charset-normalizer 2.0.4 pyhd3eb1b0_0
click 8.0.4 py37haa95532_0
colorama 0.4.5 py37haa95532_0
coloredlogs 15.0.1 pypi_0 pypi
coremltools 6.1 pypi_0 pypi
cryptography 38.0.1 py37h21b164f_0
cudatoolkit 11.3.1 h59b6b97_2
cudnn 8.2.1 cuda11.3_0
cycler 0.11.0 pyhd8ed1ab_0 conda-forge
cython 0.29.32 py37hf2a7229_0 conda-forge
dataclasses 0.8 pyh6d0b6a4_7
dm-tree 0.1.7 pypi_0 pypi
fftw 3.3.9 h2bbff1b_1
fire 0.4.0 pypi_0 pypi
flatbuffers 2.0.7 pypi_0 pypi
freetype 2.10.4 h546665d_1 conda-forge
frozenlist 1.2.0 py37h2bbff1b_0
gast 0.4.0 pyhd3eb1b0_0
giflib 5.2.1 h62dcd97_0
google-auth 2.6.0 pyhd3eb1b0_0
google-auth-oauthlib 0.4.1 py_2 conda-forge
google-pasta 0.2.0 pyhd3eb1b0_0
grpcio 1.42.0 py37hc60d5dd_0
h5py 3.7.0 py37h3de5c98_0
hdf5 1.10.6 h1756f20_1
humanfriendly 10.0 pypi_0 pypi
icc_rt 2022.1.0 h6049295_2
icu 68.1 h6c2663c_0
idna 3.4 py37haa95532_0
imagecorruptions 1.1.2 pypi_0 pypi
imageio 2.22.4 pypi_0 pypi
imgaug 0.4.0 pypi_0 pypi
importlib-metadata 4.11.3 py37haa95532_0
intel-openmp 2021.4.0 haa95532_3556
jbig 2.1 h8d14728_2003 conda-forge
jinja2 3.1.2 pypi_0 pypi
jpeg 9e h2bbff1b_0
keras 2.10.0 pypi_0 pypi
keras-applications 1.0.8 pypi_0 pypi
keras-preprocessing 1.1.2 pyhd3eb1b0_0
keras2onnx 1.7.0 pypi_0 pypi
kiwisolver 1.4.4 py37h8c56517_0 conda-forge
lcms2 2.12 h2a16943_0 conda-forge
lerc 2.2.1 h0e60522_0 conda-forge
libclang 14.0.6 pypi_0 pypi
libcurl 7.85.0 h86230a5_0
libdeflate 1.7 h8ffe710_5 conda-forge
libpng 1.6.37 h2a8f88b_0
libprotobuf 3.17.2 h23ce68f_1
libssh2 1.10.0 hcd4344a_0
libtiff 4.3.0 h0c97f57_1 conda-forge
lz4-c 1.9.3 h8ffe710_1 conda-forge
markdown 3.3.4 py37haa95532_0
markupsafe 2.1.1 pypi_0 pypi
matplotlib 3.4.3 py37_arcgispro_4 [arcgispro] esri
matplotlib-base 3.4.3 py37h4a79c79_2 conda-forge
mkl 2021.4.0 haa95532_640
mkl-service 2.4.0 py37h2bbff1b_0
mkl_fft 1.3.1 py37h277e83a_0
mkl_random 1.2.2 py37hf11a4ad_0
mnn 1.1.0 pypi_0 pypi
mpmath 1.2.1 pypi_0 pypi
multidict 6.0.2 py37h2bbff1b_0
networkx 2.6.3 pypi_0 pypi
numpy 1.21.6 pypi_0 pypi
oauthlib 3.2.1 py37haa95532_0
olefile 0.46 pyh9f0ad1d_1 conda-forge
onnx 1.12.0 pypi_0 pypi
onnxconverter-common 1.13.0 pypi_0 pypi
onnxruntime 1.13.1 pypi_0 pypi
opencv-contrib-python 4.2.0.32 pypi_0 pypi
opencv-python 4.2.0.32 pypi_0 pypi
openjpeg 2.4.0 hb211442_1 conda-forge
openssl 1.1.1n 0 esri
opt_einsum 3.3.0 pyhd3eb1b0_1
packaging 21.3 pypi_0 pypi
pandas 1.1.5 pypi_0 pypi
pillow 8.3.2 py37hd7d9ad0_0 conda-forge
pip 22.2.2 py37haa95532_0
protobuf 3.17.2 py37hd77b12b_0
pyasn1 0.4.8 pyhd3eb1b0_0
pyasn1-modules 0.2.8 py_0
pycocotools 2.0.2 py37h5685391_1 esri
pycparser 2.21 pyhd3eb1b0_0
pyjwt 2.4.0 py37haa95532_0
pyopenssl 22.0.0 pyhd3eb1b0_0
pyparsing 3.0.9 pyhd8ed1ab_0 conda-forge
pyreadline 2.1 pypi_0 pypi
pysocks 1.7.1 py37_1
python 3.7.0 hea74fb7_0
python-dateutil 2.8.2 pyhd8ed1ab_0 conda-forge
python_abi 3.7 2_cp37m conda-forge
pytz 2022.6 pypi_0 pypi
pywavelets 1.3.0 pypi_0 pypi
pyyaml 6.0 pypi_0 pypi
requests 2.28.1 py37haa95532_0
requests-oauthlib 1.3.0 py_0
rsa 4.7.2 pyhd3eb1b0_1
scikit-image 0.19.3 pypi_0 pypi
scikit-learn 0.20.3 pypi_0 pypi
scipy 1.7.3 py37h7a0a035_2
seaborn 0.12.1 pypi_0 pypi
setuptools 65.5.0 py37haa95532_0
shapely 1.8.5.post1 pypi_0 pypi
six 1.16.0 pyhd3eb1b0_1
snappy 1.1.9 h6c2663c_0
sqlite 3.39.3 h2bbff1b_0
sympy 1.10.1 pypi_0 pypi
tensorboard 2.10.1 pypi_0 pypi
tensorboard-data-server 0.6.0 py37haa95532_0
tensorboard-plugin-wit 1.8.1 py37haa95532_0
tensorflow 2.10.1 pypi_0 pypi
tensorflow-addons 0.18.0 pypi_0 pypi
tensorflow-estimator 2.10.0 pypi_0 pypi
tensorflow-gpu 2.10.1 pypi_0 pypi
tensorflow-io-gcs-filesystem 0.27.0 pypi_0 pypi
tensorflow-model-optimization 0.7.3 pypi_0 pypi
termcolor 2.1.0 py37haa95532_0
tf2onnx 1.13.0 pypi_0 pypi
tfcoreml 1.1 pypi_0 pypi
tidecv 1.0.1 pypi_0 pypi
tifffile 2021.11.2 pypi_0 pypi
tk 8.6.12 h8ffe710_0 conda-forge
tornado 6.2 py37hcc03f2d_0 conda-forge
tqdm 4.64.1 pypi_0 pypi
typeguard 2.13.3 pypi_0 pypi
typing-extensions 4.3.0 py37haa95532_0
typing_extensions 4.3.0 py37haa95532_0
urllib3 1.26.12 py37haa95532_0
vc 14.2 h21ff451_1
vs2015_runtime 14.27.29016 h5e58377_2 esri
werkzeug 2.0.3 pyhd3eb1b0_0
wheel 0.35.1 pyhd3eb1b0_0
win_inet_pton 1.1.0 py37haa95532_0
wincertstore 0.2 py37haa95532_2
wrapt 1.14.1 py37h2bbff1b_0
xz 5.2.6 h8d14728_0 conda-forge
yarl 1.8.1 py37h2bbff1b_0
zipp 3.8.0 py37haa95532_0
zlib 1.2.13 h8cc25b3_0
zstd 1.5.0 h6255e5f_0 conda-forge
@MrVanq I guess you're trying to do pose traning quantize for a model, but I didn't use keras_to_tensorflow.py for that. Maybe you can try post_train_quant_convert.py and it should works on TF 2.10.1
@david8862 post_train_quant_convert.py works and converts my model to a post-training-int-quantized .tflite model which I sadly can't use with python on the RaspberryPi with the tflite_support framework, I found out that it only supports SSD models.
So now I am trying to convert the .h5 model to a frozen .pb model so I can atleast run OpenCV cv.dnn.readNet('frozen_model.pb') with it on the RaspberryPi.
Is there a way to convert keras to frozen tensorflow? (with post training quantize it would be ofc better but if not I just want to convert it as it is to a frozen tensorflow .pb)
Any help is appreciated.
@david8862 post_train_quant_convert.py works and converts my model to a post-training-int-quantized .tflite model which I sadly can't use with python on the RaspberryPi with the tflite_support framework, I found out that it only supports SSD models. So now I am trying to convert the .h5 model to a frozen .pb model so I can atleast run OpenCV cv.dnn.readNet('frozen_model.pb') with it on the RaspberryPi. Is there a way to convert keras to frozen tensorflow? (with post training quantize it would be ofc better but if not I just want to convert it as it is to a frozen tensorflow .pb) Any help is appreciated.
Sorry but I didn't dig into the quantized pb model convert. The --quantize option in keras_to_tensorflow.py is inherited from the original repo keras_to_tensorflow