tensorflow: TFlite Model Maker installation successfull - but error on import

Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

Source

binary

TensorFlow version

2.8.2

Custom code

OS platform and distribution

WSL Ubuntu 20.04 on Windows 10

Mobile device

No response

Python version

3.8

Bazel version

No response

GCC/compiler version

No response

CUDA/cuDNN version

No response

GPU model and memory

No response

Current behavior?

I thought I have successfully installed TFLite model maker but I can an error when trying to test code which happens on some imports. It seems to be scann related which is installed at version 1.2.6 by pip which should be compatible. One important thing might be that I installed tensorflow 2.8.2 using anaconda because the pip packages are built requiring AVX which my CPU does not have. I verified that my GPU is used correctly within the anaconda environment with this installed package though.

The following error happens on imports e.g. from tflite_model_maker.config import QuantizationConfig:

python
Python 3.8.18 (default, Sep 11 2023, 13:20:55)
[GCC 11.2.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from tflite_model_maker import object_detector
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/alex/examples/tensorflow_examples/lite/model_maker/pip_package/src/tflite_model_maker/__init__.py", line 51, in <module>
    from tflite_model_maker import searcher
  File "/home/alex/examples/tensorflow_examples/lite/model_maker/pip_package/src/tflite_model_maker/searcher/__init__.py", line 25, in <module>
    from tensorflow_examples.lite.model_maker.core.task.searcher import ExportFormat
  File "/home/alex/examples/tensorflow_examples/lite/model_maker/pip_package/src/tensorflow_examples/lite/model_maker/core/task/searcher.py", line 30, in <module>
    from tensorflow_examples.lite.model_maker.core.utils import ondevice_scann_builder
  File "/home/alex/examples/tensorflow_examples/lite/model_maker/pip_package/src/tensorflow_examples/lite/model_maker/core/utils/ondevice_scann_builder.py", line 17, in <module>
    from scann.proto import scann_pb2
  File "/home/alex/anaconda3/envs/tf_gpu_env/lib/python3.8/site-packages/scann/__init__.py", line 2, in <module>
    from scann.scann_ops.py import scann_ops
  File "/home/alex/anaconda3/envs/tf_gpu_env/lib/python3.8/site-packages/scann/scann_ops/py/scann_ops.py", line 23, in <module>
    _scann_ops_so = tf.load_op_library(
  File "/home/alex/anaconda3/envs/tf_gpu_env/lib/python3.8/site-packages/tensorflow/python/framework/load_library.py", line 54, in load_op_library
    lib_handle = py_tf.TF_LoadLibrary(library_filename)
tensorflow.python.framework.errors_impl.NotFoundError: /home/alex/anaconda3/envs/tf_gpu_env/lib/python3.8/site-packages/scann/scann_ops/cc/_scann_ops.so: undefined symbol: _ZN4absl12lts_2021032420raw_logging_internal21internal_log_functionE

Is it related that I use tensorflow from conda and not from pip?

Standalone code to reproduce the issue

This is what i did in WSL

~$ conda install tensorflow=2.8.2=gpu_py38h75b8afa_0
~$ python
>>> import tensorflow as tf
>>> tf.config.list_physical_devices('GPU')
2023-12-18 01:18:17.785271: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:02:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-12-18 01:18:17.799686: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:02:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-12-18 01:18:17.800511: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:02:00.0/numa_node
Your kernel may have been built without NUMA support.
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

~$ pip install -q pycocotools
~$ pip install -q opencv-python-headless==4.1.2.30
~$ sudo apt -y install libportaudio2

~/examples/tensorflow_examples/lite/model_maker/pip_package$ pip install -e .
~/examples/tensorflow_examples/lite/model_maker/pip_package$ python
>>> from tflite_model_maker.config import QuantizationConfig



### Relevant log output

_No response_

About this issue

Original URL
State: closed
Created 6 months ago
Comments: 15 (4 by maintainers)

Most upvoted comments

Hi @alexw92,

Be careful with motherboard compatibility if you go that route.

Maintenance Status: Is the Model Maker located in the examples directory of the TensorFlow repository still actively maintained? Or has it been superseded by the MediaPipe Model Maker?

Model Maker is currently still active but is running into the problems you have seen, so that’s why we are diverting users to MediaPipe Model Maker. Effectively and practically if you want anything done right now, yes … I can’t say if that will continue to be the case in the future.

Official Recommendation: For developing models, especially with a focus on mobile and edge devices, does Google currently recommend transitioning to the MediaPipe Model Maker?

Yes or TFLite directly, as always, the lower you go in the API the more customizability you have at the cost of developer experience.

This is what I am afraid of. I have no experience in building from source.

Of course you know the best use of your own time, but the only way to get experience is to try 😄. Your experience here will be smoother on linux/WSL. Building from source is generally the most reliable way to support unique configurations/combinations – or see why that configuration/combination is not supported.

pkgoogle on Jan 17, 2024