Package path error in tvm/tests/python/frontend/tflite/test_forward.py

[wangshangsam]

The package path on line 134:

converter = tf.contrib.lite.TFLiteConverter.from_session(
            sess, input_tensors, output_tensors)

Should be

converter = tf.lite.TFLiteConverter.from_session(
            sess, input_tensors, output_tensors)

instead. I'm not sure how TVM's official CI tests manage to pass since then, but it is causing our own version of the CI tests to fail and blocking the testing of our own work. Since the original author has not been responsive, I'll submit a PR to fix it myself.

[wangshangsam]

PR #3427

[wangshangsam]

@tqchen @kevinthesun

[apivovarov]

TVM supports both tensorflow 1.12 and 1.13
Both allow to use tf.contrib.lite.TFLiteConverter.
1.13 will show warning

>>> import tensorflow as tf
>>> tf.__version__
'1.13.1'
>>> m = tf.contrib.lite.TFLiteConverter

WARNING: The TensorFlow contrib module will not be included in TensorFlow 2.0.
For more information, please see:
  * https://github.com/tensorflow/community/blob/master/rfcs/20180907-contrib-sunset.md
  * https://github.com/tensorflow/addons
If you depend on functionality not listed there, please file an issue.

>>> m
<class 'tensorflow.lite.python.lite.TFLiteConverter'>

This explains why CI tests work fine.

tensorflow 1.14 was released 6 days ago and it does not have lite under tf.contrib
we probably need need to switch docker and tflite schema to 1.14
Actually, docker file ubuntu_install_tensorflow.sh uses pip3 install tensorflow which should install 1.14. But Jenkins uses cached docker image (which still has 1.13.1) and the image needs to be manually regenerated.

[wangshangsam]

I'm a bit skeptical about this claim. Our CI test builds docker image with tensorflow 1.13 as well, as I just checked (essentially whichever ubuntu_install_tflite.sh specifies).

Regardless, my new commit should fix it for all versions.

[tqchen]

closed by #3427