Add guide for exporting Keras models to LiteRT by pctablet505 · Pull Request #2260 · keras-team/keras-io

pctablet505 · 2026-01-22T07:17:47Z

This pull request introduces a new guide to the guides list and updates the Keras project version reference in the autogen.py script. The main changes are as follows:

Guides update:

Added a new guide entry with the path litert_export and the title "Exporting Keras models to LiteRT" to the guides configuration in guides_master.py.

Version reference update:

Changed the Keras project URL in the PROJECT_URL dictionary in autogen.py to reference version v3.13.0 instead of v3.13.1.Introduces comprehensive documentation and code examples for exporting Keras models to TensorFlow Lite (LiteRT) format. Includes guides in Jupyter notebook, Python script, and Markdown formats, covering basic export, different model architectures, quantization, dynamic shapes, custom signatures, model validation, and advanced export options. Updates autogen scripts to support the new guide.

Introduces comprehensive documentation and code examples for exporting Keras models to TensorFlow Lite (LiteRT) format. Includes guides in Jupyter notebook, Python script, and Markdown formats, covering basic export, different model architectures, quantization, dynamic shapes, custom signatures, model validation, and advanced export options. Updates autogen scripts to support the new guide.

gemini-code-assist · 2026-01-22T07:18:05Z

Summary of Changes

Hello @pctablet505, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly enhances the documentation by introducing a detailed guide on exporting Keras models to TensorFlow Lite (LiteRT) format. This new resource provides developers with essential information and practical examples for deploying Keras models efficiently on mobile and edge devices, covering various model types, optimization strategies like quantization, and advanced configuration options. Additionally, it includes a minor update to a Keras project version reference in the autogeneration scripts.

Highlights

New Guide for Keras to LiteRT Export: A comprehensive guide titled "Exporting Keras models to LiteRT (TensorFlow Lite)" has been added, providing detailed instructions and examples for deploying Keras models on mobile and edge devices.
Diverse Model Architecture Support: The new guide demonstrates how to export various Keras model architectures, including Sequential, Functional API, Subclassed models, and models from KerasHub, to the LiteRT format.
Quantization Techniques: Detailed explanations and code examples are provided for different quantization strategies (basic, Float16, and dynamic range quantization) to optimize model size and inference speed for edge deployment.
Dynamic Shapes and Custom Signatures: The guide covers advanced topics such as handling dynamic input shapes for flexible model inference and defining custom input signatures for models with complex input/output requirements.
Model Validation and Advanced Options: It includes a section on validating exported LiteRT models against their Keras counterparts and explores advanced export options for fine-tuning the conversion process.
Keras Version Reference Update: The Keras project version reference in the scripts/autogen.py file has been updated from v3.13.1 to v3.13.0.
Guide Index Update: An entry for the new LiteRT export guide has been added to the scripts/guides_master.py file, making it discoverable within the documentation.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces a comprehensive guide for exporting Keras models to LiteRT, covering various model types, quantization, dynamic shapes, and validation. The guide is well-structured and provides clear examples. The autogen.py and guides_master.py updates correctly integrate the new guide and adjust the Keras project version reference.

gemini-code-assist · 2026-01-22T07:20:09Z

guides/litert_export.py

+try:
+    from ai_edge_litert.interpreter import Interpreter
+
+    print("Using ai_edge_litert for inference")
+    litert_available = True
+except ImportError:
+    try:
+        from tensorflow.lite import Interpreter
+
+        print("Using tensorflow.lite for inference")
+        litert_available = True
+    except ImportError:
+        try:
+            import tensorflow as tf
+
+            Interpreter = tf.lite.Interpreter
+
+            print("Using tf.lite.Interpreter for inference")
+            litert_available = True
+        except (ImportError, AttributeError):
+            print("LiteRT interpreter not available. Skipping inference test.")
+            print(
+                "To test inference, install ai_edge_litert: pip install ai-edge-litert"
+            )


The logic for importing the LiteRT interpreter is repeated here and in the validate_tflite_model function. Encapsulating this logic into a helper function would improve code readability and maintainability, adhering to the DRY (Don't Repeat Yourself) principle. This also makes it easier to update the import strategy in the future if needed.

def _get_litert_interpreter_and_status(): """Attempts to import a LiteRT interpreter and returns its status.""" litert_available = False Interpreter_class = None try: from ai_edge_litert.interpreter import Interpreter as Interpreter_class print("Using ai_edge_litert for inference") litert_available = True except ImportError: try: from tensorflow.lite import Interpreter as Interpreter_class print("Using tensorflow.lite for inference") litert_available = True except ImportError: try: import tensorflow as tf Interpreter_class = tf.lite.Interpreter print("Using tf.lite.Interpreter for inference") litert_available = True except (ImportError, AttributeError): print("LiteRT interpreter not available. Skipping inference test.") print("To test inference, install ai_edge_litert: pip install ai-edge-litert") return Interpreter_class, litert_available Interpreter, litert_available = _get_litert_interpreter_and_status()

divyashreepathihalli

Thank you for the PR! Left some comments!

divyashreepathihalli · 2026-01-23T18:17:19Z

guides/litert_export.py

+"""
+## Introduction
+
+TensorFlow Lite (LiteRT) is TensorFlow's solution for running machine learning models


This may make it sound like it is a tensorflow solution.
Maybe reword it to - "LiteRT is a solution for running machine learning models..."

divyashreepathihalli · 2026-01-23T18:18:29Z

guides/litert_export.py

+
+# Generate dummy data for demonstration
+X_train = np.random.random((1000, 28, 28))
+y_train = np.random.randint(0, 10, 1000)


Keep the names consistent
X_train> x_train

divyashreepathihalli · 2026-01-23T18:20:48Z

guides/litert_export.py

+
+# Load and test the exported model
+litert_available = False
+try:


This is a demo colab - you should assume the environment will support the interpreter instead of this nested try except blocks. please remove.

divyashreepathihalli · 2026-01-23T18:22:03Z

guides/litert_export.py

+    output = interpreter.get_tensor(output_details[0]["index"])
+
+    print(f"\nInference successful! Output shape: {output.shape}")
+else:


remove the if else block
and let the block throw an error

divyashreepathihalli · 2026-01-23T18:23:29Z

guides/litert_export.py

+"""
+
+keras_hub_available = False
+try:


remove the try except blocks!

Please add relevant install instructions in the beginning of the Colab - for what is needed to run this guide.

divyashreepathihalli · 2026-01-23T18:24:09Z

guides/litert_export.py

+    print("keras-hub not available. Skipping Keras-Hub example.")
+    print("To run this example, install keras-hub: pip install keras-hub")
+
+if keras_hub_available:


remove the if and try blocks - if there is an error let it throw the error

divyashreepathihalli · 2026-01-23T18:24:28Z

guides/litert_export.py

+        vision_model.export("resnet.tflite", format="litert")
+
+        print("Exported Keras-Hub vision model")
+    except Exception as e:


remove the try and except!

divyashreepathihalli · 2026-01-23T18:25:59Z

guides/litert_export.py

+- **Memory issues**: Reduce model size with quantization
+- **Accuracy drops**: Start with float16 instead of full int8 quantization
+
+## Next Steps


are these next steps guide available? - if not - lets remove

pctablet505 · 2026-01-29T07:44:30Z

@gemini-code-assist review

gemini-code-assist

Code Review

This pull request introduces a comprehensive guide on exporting Keras models to LiteRT (TensorFlow Lite) format. The guide is provided in Jupyter Notebook, Python script, and Markdown formats. The changes also include adding the new guide to the site's master list and updating a Keras version reference in an autogeneration script.

My review focuses on improving the code examples in the new guide to follow Keras best practices, such as using keras.Input for defining model input shapes. I've also suggested a minor improvement for character encoding to ensure cross-platform compatibility. The changes to the build scripts look good.

gemini-code-assist · 2026-01-29T07:46:49Z

guides/ipynb/litert_export.ipynb

+    "# Create a simple MNIST classifier\n",
+    "model = keras.Sequential(\n",
+    "    [\n",
+    "        keras.layers.Flatten(input_shape=(28, 28)),\n",


While using input_shape in the first layer works, the recommended best practice for Sequential models is to use a keras.Input object as the first layer. This makes the model's inputs and their shapes more explicit.

Here's how you could refactor the model definition:

model = keras.Sequential( [ keras.Input(shape=(28, 28)), keras.layers.Flatten(), keras.layers.Dense(128, activation="relu"), keras.layers.Dropout(0.2), keras.layers.Dense(10, activation="softmax"), ] )

gemini-code-assist · 2026-01-29T07:46:49Z

guides/ipynb/litert_export.ipynb

+    "# Create a model for quantization\n",
+    "quantization_model = keras.Sequential(\n",
+    "    [\n",
+    "        keras.layers.Dense(64, activation=\"relu\", input_shape=(784,)),\n",


For clarity and consistency with best practices, it's better to define the input shape of a Sequential model using a keras.Input layer instead of the input_shape argument in the first layer.

Here's how you could refactor it:

quantization_model = keras.Sequential( [ keras.Input(shape=(784,)), keras.layers.Dense(64, activation="relu"), keras.layers.Dense(32, activation="relu"), keras.layers.Dense(10, activation="softmax"), ] )

gemini-code-assist · 2026-01-29T07:46:49Z

guides/ipynb/litert_export.ipynb

+    "dynamic_model = keras.Sequential(\n",
+    "    [\n",
+    "        keras.layers.Dense(64, activation=\"relu\"),\n",
+    "        keras.layers.Dense(10, activation=\"softmax\"),\n",
+    "    ]\n",
+    ")\n",
+    "\n",
+    "# Use None for dynamic dimensions\n",
+    "dynamic_model.build((None, 784))  # None = dynamic batch size\n",


Instead of defining an empty Sequential model and then calling .build(), you can define the input shape directly using a keras.Input layer. This makes the model definition more self-contained and readable. Using batch_size=None in keras.Input explicitly declares a dynamic batch size.

Here's a suggested refactoring:

dynamic_model = keras.Sequential( [ keras.Input(shape=(784,), batch_size=None), # None = dynamic batch size keras.layers.Dense(64, activation="relu"), keras.layers.Dense(10, activation="softmax"), ] )

gemini-code-assist · 2026-01-29T07:46:49Z

guides/litert_export.py

+# Create a simple MNIST classifier
+model = keras.Sequential(
+    [
+        keras.layers.Flatten(input_shape=(28, 28)),


While using input_shape in the first layer works, the recommended best practice for Sequential models is to use a keras.Input object as the first layer. This makes the model's inputs and their shapes more explicit.

Here's how you could refactor the model definition:

model = keras.Sequential( [ keras.Input(shape=(28, 28)), keras.layers.Flatten(), keras.layers.Dense(128, activation="relu"), keras.layers.Dropout(0.2), keras.layers.Dense(10, activation="softmax"), ] )

gemini-code-assist · 2026-01-29T07:46:49Z

guides/litert_export.py

+# Create a model for quantization
+quantization_model = keras.Sequential(
+    [
+        keras.layers.Dense(64, activation="relu", input_shape=(784,)),


For clarity and consistency with best practices, it's better to define the input shape of a Sequential model using a keras.Input layer instead of the input_shape argument in the first layer.

Here's how you could refactor it:

quantization_model = keras.Sequential( [ keras.Input(shape=(784,)), keras.layers.Dense(64, activation="relu"), keras.layers.Dense(32, activation="relu"), keras.layers.Dense(10, activation="softmax"), ] )

gemini-code-assist · 2026-01-29T07:46:49Z

guides/litert_export.py

+
+    # Compare outputs
+    np.testing.assert_allclose(keras_output.numpy(), tflite_output, atol=1e-5)
+    print("✓ Model validation passed!")


Using the unicode character ✓ directly might cause rendering issues on some systems or terminals. It's safer to use its unicode escape sequence \u2713 to ensure consistent display.

Suggested change

print("✓ Model validation passed!")

print("\u2713 Model validation passed!")

gemini-code-assist · 2026-01-29T07:46:49Z

guides/md/litert_export.md

+# Create a simple MNIST classifier
+model = keras.Sequential(
+    [
+        keras.layers.Flatten(input_shape=(28, 28)),


While using input_shape in the first layer works, the recommended best practice for Sequential models is to use a keras.Input object as the first layer. This makes the model's inputs and their shapes more explicit.

Here's how you could refactor the model definition:

model = keras.Sequential( [ keras.Input(shape=(28, 28)), keras.layers.Flatten(), keras.layers.Dense(128, activation="relu"), keras.layers.Dropout(0.2), keras.layers.Dense(10, activation="softmax"), ] )

gemini-code-assist · 2026-01-29T07:46:49Z

guides/md/litert_export.md

+# Create a model for quantization
+quantization_model = keras.Sequential(
+    [
+        keras.layers.Dense(64, activation="relu", input_shape=(784,)),


For clarity and consistency with best practices, it's better to define the input shape of a Sequential model using a keras.Input layer instead of the input_shape argument in the first layer.

Here's how you could refactor it:

quantization_model = keras.Sequential( [ keras.Input(shape=(784,)), keras.layers.Dense(64, activation="relu"), keras.layers.Dense(32, activation="relu"), keras.layers.Dense(10, activation="softmax"), ] )

gemini-code-assist · 2026-01-29T07:46:49Z

guides/md/litert_export.md

+# Create model with dynamic batch size
+dynamic_model = keras.Sequential(
+    [
+        keras.layers.Dense(64, activation="relu"),
+        keras.layers.Dense(10, activation="softmax"),
+    ]
+)
+
+# Use None for dynamic dimensions
+dynamic_model.build((None, 784))  # None = dynamic batch size


Instead of defining an empty Sequential model and then calling .build(), you can define the input shape directly using a keras.Input layer. This makes the model definition more self-contained and readable. Using batch_size=None in keras.Input explicitly declares a dynamic batch size.

Here's a suggested refactoring:

# Create model with dynamic batch size dynamic_model = keras.Sequential( [ keras.Input(shape=(784,), batch_size=None), # None = dynamic batch size keras.layers.Dense(64, activation="relu"), keras.layers.Dense(10, activation="softmax"), ] )

gemini-code-assist · 2026-01-29T07:46:49Z

guides/md/litert_export.md

+
+    # Compare outputs
+    np.testing.assert_allclose(keras_output.numpy(), tflite_output, atol=1e-5)
+    print("✓ Model validation passed!")


Using the unicode character ✓ directly might cause rendering issues on some systems or terminals. It's safer to use its unicode escape sequence \u2713 to ensure consistent display.

Suggested change

print("✓ Model validation passed!")

print("\u2713 Model validation passed!")

pctablet505 requested review from MarkDaoust and fchollet as code owners January 22, 2026 07:17

github-actions bot assigned sachinprasadhs Jan 22, 2026

gemini-code-assist bot reviewed Jan 22, 2026

View reviewed changes

divyashreepathihalli reviewed Jan 23, 2026

View reviewed changes

Address reviewer comments and fix rendering in LiteRT export guide

40d87b0

gemini-code-assist bot reviewed Jan 29, 2026

View reviewed changes

	print("✓ Model validation passed!")
	print("\u2713 Model validation passed!")

Conversation

pctablet505 commented Jan 22, 2026

Uh oh!

gemini-code-assist bot commented Jan 22, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

divyashreepathihalli left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pctablet505 commented Jan 29, 2026

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 29, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants