## Overview

LiteRT supports converting TensorFlow RNN models to LiteRT’s fused LSTM operations. Fused operations exist to maximize the performance of their underlying kernel implementations, as well as provide a higher level interface to define complex transformations like quantizatization.

Since there are many variants of RNN APIs in TensorFlow, our approach has been two fold:

- Provide
**native support for standard TensorFlow RNN APIs**like Keras LSTM. This is the recommended option. - Provide an
**interface****into the conversion infrastructure for****user-defined****RNN implementations**to plug in and get converted to LiteRT. We provide a couple of out of box examples of such conversion using lingvo’s LSTMCellSimple and LayerNormalizedLSTMCellSimple RNN interfaces.

## Converter API

The feature is part of TensorFlow 2.3 release. It is also available through the tf-nightly pip or from head.

This conversion functionality is available when converting to LiteRT via a SavedModel or from the Keras model directly. See example usages.

### From saved model

```
# build a saved model. Here concrete_function is the exported function
# corresponding to the TensorFlow model containing one or more
# Keras LSTM layers.
saved_model, saved_model_dir = build_saved_model_lstm(...)
saved_model.save(saved_model_dir, save_format="tf", signatures=concrete_func)
# Convert the model.
converter = TFLiteConverter.from_saved_model(saved_model_dir)
tflite_model = converter.convert()
```

### From Keras model

```
# build a Keras model
keras_model = build_keras_lstm(...)
# Convert the model.
converter = TFLiteConverter.from_keras_model(keras_model)
tflite_model = converter.convert()
```

## Example

Keras LSTM to LiteRT Colab illustrates the end to end usage with the LiteRT interpreter.

## TensorFlow RNNs APIs supported

### Keras LSTM conversion (recommended)

We support out-of-the-box conversion of Keras LSTM to LiteRT. For details on how this works please refer to the Keras LSTM interface and to the conversion logic here.

Also important is to highlight the LiteRT’s LSTM contract with respect to the Keras operation definition:

- The dimension 0 of the
**input**tensor is the batch size. - The dimension 0 of the
**recurrent_weight**tensor is the number of outputs. - The
**weight**and**recurrent_kernel**tensors are transposed. - The transposed weight, transposed recurrent_kernel and
**bias**tensors are split into 4 equal sized tensors along the dimension 0. These correspond to**input gate, forget gate, cell, and output gate**.

#### Keras LSTM Variants

##### Time major

Users may choose time-major or no time-major. Keras LSTM adds a time-major attribute in the function def attributes. For Unidirectional sequence LSTM, we can simply map to unidirecional_sequence_lstm's time major attribute.

##### BiDirectional LSTM

Bidirectional LSTM can be implemented with two Keras LSTM layers, one for
forward and one for backward, see examples
here.
Once we see the go_backward attribute, we recognize it as backward LSTM, then
we group forward & backward LSTM together. **This is future work.** Currently,
this creates two UnidirectionalSequenceLSTM operations in the LiteRT
model.

### User-defined LSTM conversion examples

LiteRT also provides a way to convert user defined LSTM implementations. Here we use Lingvo’s LSTM as an example of how that can be implemented. For details please refer to the lingvo.LSTMCellSimple interface and the conversion logic here. We also provide an example for another of Lingvo’s LSTM definitions in lingvo.LayerNormalizedLSTMCellSimple interface and its conversion logic here.

## “Bring your own TensorFlow RNN” to LiteRT

If a user's RNN interface is different from the standard supported ones, there are a couple of options:

**Option 1:** Write adapter code in TensorFlow python to adapt the RNN interface
to the Keras RNN interface. This means a tf.function with
tf_implements annotation on
the generated RNN interface’s function that is identical to the one generated by
the Keras LSTM layer. After this, the same conversion API used for Keras LSTM
will work.

**Option 2:** If the above is not possible (e.g. the Keras LSTM is missing some
functionality that is currently exposed by LiteRT’s fused LSTM op like
layer normalization), then extend the LiteRT converter by writing
custom conversion code and plug it into the prepare-composite-functions
MLIR-pass
here.
The function’s interface should be treated like an API contract and should
contain the arguments needed to convert to fused LiteRT LSTM
operations - i.e. input, bias, weights, projection, layer normalization, etc. It
is preferable for the tensors passed as arguments to this function to have known
rank (i.e. RankedTensorType in MLIR). This makes it much easier to write
conversion code that can assume these tensors as RankedTensorType and helps
transform them to ranked tensors corresponding to the fused LiteRT
operator’s operands.

A complete example of such conversion flow is Lingvo’s LSTMCellSimple to LiteRT conversion.

The LSTMCellSimple in Lingvo is defined here. Models trained with this LSTM cell can be converted to LiteRT as follows:

- Wrap all uses of LSTMCellSimple in a tf.function with a tf_implements annotation that is labelled as such (e.g. lingvo.LSTMCellSimple would be a good annotation name here). Make sure the tf.function that is generated matches the interface of the function expected in the conversion code. This is a contract between the model author adding the annotation and the conversion code.
Extend the prepare-composite-functions pass to plug in a custom composite op to LiteRT fused LSTM op conversion. See LSTMCellSimple conversion code.

The conversion contract:

**Weight**and**projection**tensors are transposed.The

**{input, recurrent}**to**{cell, input gate, forget gate, output gate}**are extracted by slicing the transposed weight tensor.The

**{bias}**to**{cell, input gate, forget gate, output gate}**are extracted by slicing the bias tensor.The

**projection**is extracted by slicing the transposed projection tensor.Similar conversion is written for LayerNormalizedLSTMCellSimple.

The rest of the LiteRT conversion infrastructure, including all the MLIR passes defined as well as the final export to LiteRT flatbuffer can be reused.

## Known issues/limitations

- Currently there is support only for converting stateless Keras LSTM (default behavior in Keras). Stateful Keras LSTM conversion is future work.
- It is still possible to model a stateful Keras LSTM layer using the underlying stateless Keras LSTM layer and managing the state explicitly in the user program. Such a TensorFlow program can still be converted to LiteRT using the feature being described here.
- Bidirectional LSTM is currently modelled as two UnidirectionalSequenceLSTM operations in LiteRT. This will be replaced with a single BidirectionalSequenceLSTM op.