Sachin Joglekar

This looks likes a bug with the `TPUEstimator`. As far as I understand [this part of the docs](https://github.com/tensorflow/estimator/blob/master/tensorflow_estimator/python/estimator/estimator.py#L314), the Estimator API handles the `OutofRange` error from the input data function...

@Maratyszcza could you take a look at this?

Hey @ramanps05 we don't have a concrete timeline for this yet, but I will take a look at this next week or so.

It essentially boils down to making the [writer](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/writer/writer.cc#L37) write all subgraphs in the Interpreter, instead of just the `primary_subgraph`. The TFLite `Interpreter` has `subgraph(index)` & `subgraphs_size()` methods that we can...

Hmm. This looks like a bug with the writer. Lemme take a look and get back to you.

Started looking into this. For your Control Flow conversion to work, you need to use the TF2 converter like this: ``` class CFModel(tf.Module): @tf.function( input_signature=( tf.TensorSpec(shape=[5, 1], dtype=tf.float32, name='i'), tf.TensorSpec(shape=[5,...

It shouldn't change how WHILE is represented, its just that the TF1 converter may not use our new MLIR converter that converts WHILE (if it does for you in float,...

Aah, then great. It safer to use `from_saved_model` though, since we usually update that API with newer conversion features (mainly because SavedModel is now TF's recommended model format).

I am currently on vacation, but I have made progress on this. There are some internal usages of this tool, so I have to be careful not to break anyone....

The code is under active review, so should be submitted soon :-)