Issue with converting Keras .h5 files to .tflite files

[bennyooo]

There seems to be an issue with the above mentioned conversion.
According to this documentation (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/lite/toco/g3doc/cmdline_examples.md#keras) it should be possible to convert a Keras model into a .tflite file. However, when I try to convert the file, the following output is given:

tflite_convert: error: one of the arguments --graph_def_file --saved_model_dir is required.

I made sure that the .h5 contains both the model and the weights. Other conversions like .pb files have worked with this file as well. Also in the help section of the commad there is no reference to --keras_model_file.

Update
I was using the wrong TF version so I got the error message above. However, using the nightly build results in other error messages:

ValueError: Unknown activation function:relu6

This error normally occurs when using load_model but there is the possibility to specify custom_objects (like relu6) to avoid it.

How can this be done with tflite_convert?

[hlgrondijs]

Same issue here

[kumarnikhil936]

Hi @bennyooo
Why is the issue closed? Is there a solution to this yet ?
I have the same problem.

[bennyooo]

I closed it because I was referring to TF 1.9 but the documentation I linked above reffers to the master branch. Anyway I tried the nightly build as well and it still doesn't work (so I'm reopening the issue now). I am currently trying to do this via a frozen graph (.pb) file and will report my findings.

Update
The frozen graph file can also not be converted. The error is:

F tensorflow/contrib/lite/toco/graph_transformations/resolve_tensorflow_switch.cc:95] Check failed: other_op->type == OperatorType::kMerge \nAborted (core dumped)\n'
None

I used the transform_graph tool with the strip_unused_nodes flag beforehand.

[gargn]

@kumarnikhil936 As @bennyooo noted, support for Keras is not available in 1.9. Please use the nightly build. You can do that by either using "pip install --upgrade tf_nightly" or follow the instructions here.

@bennyooo Can you provide your Keras model file and the set of commands (or code) you used to convert your model?

[nitheeshmuthuraj]

i thought I can convert the keras model file h5 -> pb -> tflite .. is that right way to go about ?