TFLite Converter missing functions in TF2.0 alpha

[Aakash Kumar Nain]

Hi,

I was trying to convert a `tf.keras` model to `tflite` but to my surprise I found out that there is no `from_keras_model` function available in the API. The only functions available are `convert` and `from_concrete_function`. Here is the error log:

----> 1 converter = tf.lite.TFLiteConverter.from_keras_model(model)

AttributeError: type object 'TFLiteConverterV2' has no attribute 'from_keras_model'

As per the documentation, it should be there but it isn't. Should I upgrade the version of TF2.0 to some nightly build? 

Regards,

Aakash Nain 

[Paige Bailey]

Adding +Tim.

[Margaret Maynard-Reid]

[Adding tfl...@tensorflow.org to cc]

Hi Aakash,

You may find this helpful - look at "How do I get a concrete function for a tf.Keras model?" under "/Convert a TF 2.0 Model/Generate a concrete function"

https://www.tensorflow.org/lite/r2/convert/concrete_function#how_do_i_get_a_concrete_function_for_a_tfkeras_model

Looks like the TFLite Converter has changed quite a bit Here is my compare/contrast of TFLite 1.13 vs TFLite 2.0:

TensorFlow Lite version	TFLite 1.13	TFLite 2.0
Class	TFLiteConverter	TFLiteConverterV2
Converter methods	4 Methods # Converting a GraphDef from session. from_session(sess, in_tensors, out_tensors) # Converting a GraphDef from file. from_frozen_graph(graph_def_file, input_arrays, output_arrays) # Converting a SavedModel. from_saved_model(saved_model_dir) # Converting a tf.keras model from_keras_model_file(keras_model)	2 methods: # creates a TFLiteConverter class from a ConcreteFunction from_concrete_function(cls, func) #Converts a TensorFlow GraphDef based on instance variables convert(self):
Source code	link	link
API doc	link	link
Guide on tensorflow.org	Convert a model/Python API (link)	Convert a TF 2.0 Model/Generate a concrete function (link)

Margaret

--
You received this message because you are subscribed to the Google Groups "TensorFlow Community Testing" group.
To unsubscribe from this group and stop receiving emails from it, send an email to testing+u...@tensorflow.org.

[Aakash Kumar Nain]

Thanks Margaret. I will look into it. Though first look at this gives me an impression that this is more complicated than converting directly using `from_keras_model` method that was present earlier. 

Regards,

Aakash Nain

[Margaret Maynard-Reid]

Yes I agree. The tf.Keras model to TFLite conversion seems to be more complex with TF 2.0 now.

There is one more step added in the process: train tf.Keras model -> save tf.Keras model -> get concrete function from tf.Keras model --> convert to TFLite.

Tim - any plans to simply? Thanks.

Margaret

[Margaret Maynard-Reid]

Thank you Nupur. I just verified that `TFLiteConverter.from_keras_model` is available in the `tf-nightly-2.0-preview` (which is great!), but not in TF 2.0 alpha. 

I have a suggestion and question:

Suggestion

It's super confusing that from_keras_model() is in 1.13 and TF 2.0 preview, but not in TF 2.0 alpha. As developers are still trying out TF 2.0 alpha, it may be helpful to point out in the guide that if they are trying TF 2.0 alpha they need to write the extra code to generate concrete function from tf.Keras model, before being able to converting to TFLite; or better yet just point them to try the preview instead of TF 2.0 Alpha. 

Question

If we can directly convert from a tf.Keras model to TFLite with from_keras_model(), under what scenario do we need to worry about generating a concrete function from a tf.Keras model?

Thanks.

Margaret

On Wed, Apr 24, 2019 at 8:29 AM Nupur Garg <nupu...@google.com> wrote:

The 2.0 TFLiteConverter API has been updated since alpha release. `TFLiteConverter.from_keras_model` is available in the `tf-nightly-2.0-preview`. Documentation on the API is available here: https://www.tensorflow.org/lite/r2/convert/python_api.

Additionally, documentation on concrete functions is available here: https://www.tensorflow.org/lite/r2/convert/concrete_function.

Hope this helps!

Best,

Nupur

You received this message because you are subscribed to the Google Groups "TensorFlow Lite" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tflite+un...@tensorflow.org.
Visit this group at https://groups.google.com/a/tensorflow.org/group/tflite/.

[Margaret Maynard-Reid]

Hi Nupur,

Sounds good. Thanks for the follow-up (note on 2.0 alpha differs from nightly) and clarification!

Margaret

On Wed, Apr 24, 2019 at 10:40 AM 'Nupur Garg' via TensorFlow Lite <tfl...@tensorflow.org> wrote:

Hi Margaret,

Thanks for the feedback. I will add a note to at the top of the "Converter Python API guide" about the 2.0 alpha differing from the nightly.

When using from_keras_model you do not need to worry about generating a concrete function. The only caveat is when you want to convert only a subset of the model. In 2.0, the only way to convert a portion of the model currently is to wrap only the part of the model you want to convert in a tf.function and convert that function.

Best,

Nupur

[Aakash Kumar Nain]

Hi,

Thanks for the info. I tested the `from_keras_model` in the nightly version today and it works fine. I have one more question though. Now, I have the converted tflite model, how do I save it onto disk. In earlier versions, it was through the `.pb` file. What are the changes in tf2.0 for saving the model onto disk with `.tflite` format? 

Regards,

Aakash Nain

[Aakash Kumar Nain]

I found it. It's simple save and read thing like an ordinary file. Thanks anyways

Regards,

Aakash Nain

[GetMotivated]

Hi bro,

I am new to python, I am facing similar issue you were facing, can you kindly share your code, or atleast how you imported or installed nightly version because I have trouble with that also

Thanks

To unsubscribe from this group and stop receiving emails from it, send an email to tes...@tensorflow.org.

--
You received this message because you are subscribed to the Google Groups "TensorFlow Lite" group.

To unsubscribe from this group and stop receiving emails from it, send an email to tfl...@tensorflow.org.

Visit this group at https://groups.google.com/a/tensorflow.org/group/tflite/.

--
You received this message because you are subscribed to the Google Groups "TensorFlow Lite" group.

To unsubscribe from this group and stop receiving emails from it, send an email to tfl...@tensorflow.org.