TFLite Conversion

PyTorch -> ONNX -> TF -> TFLite

Convert PyTorch Models to TFLite and run inference in TFLite Python API.

Tested Environment

• pytorch==1.7.1
• tensorflow==2.4.1
• onnx==1.8.0
• onnx-tf==1.7.0

PyTorch to ONNX

Load the PyTorch Model:

model = Model()
model.load_state_dict(torch.load(pt_model_path, map_location='cpu')).eval()

Prepare the Input:

sample_input = torch.rand((batch_size, channels, height, width))

Export to ONNX format:

torch.onnx.export(
    model,                  # PyTorch Model
    sample_input,                    # Input tensor
    onnx_model_path,        # Output file (eg. 'output_model.onnx')
    opset_version=12,       # Operator support version
    input_names=['input']   # Input tensor name (arbitary)
    output_names=['output'] # Output tensor name (arbitary)
)

opset-version: opset_version is very important. Some PyTorch operators are still not supported in ONNX even if opset_version=12. Default opset_version in PyTorch is 12. Please check official ONNX repo for supported PyTorch operators. If your model includes unsupported operators, convert to supported operators. For example, torch.repeat_interleave() is not supported, it can be converted into supported torch.repeat() + torch.view() to achieve the same function.

output-names: If your model returns more than 1 output, provide exact length of arbitary names. For example, if your model returns 3 outputs, then output_names should be ['output0', 'output1', 'output3']. If you don't provide exact length, although PT-ONNX conversion is successful, ONNX-TFLite conversion will not.

For more information about onnx model conversion, please check ONNX_DETAILS

Verification

You can verify the ONNX protobuf with onnx library.

Install onnx:

pip install onnx

import onnx

# Load the ONNX model
model = onnx.load("model.onnx")

# Check that the IR is well formed
onnx.checker.check_model(model)

# Print a Human readable representation of the graph
onnx.helper.printable_graph(model.graph)

ONNX Model Inference

Install onnxruntime:

pip install onnxruntime

Then run the inference:

import onnxruntime as ort

ort_session = ort.InferenceSession('model.onnx')

outputs = ort_session.run(
    None,
    {'input': np.random.randn(batch_size, channels, height, width).astype(np.float32)}
)

ONNX to TF

You cannot convert ONNX model directly into TFLite model. You must first convert to TensorFlow model.

Use onnx-tensorflow to convert models from ONNX to Tensorflow.

Install as follows:

git clone https://github.com/onnx/onnx-tensorflow.git && cd onnx-tensorflow
pip install -e .

Load the ONNX model:

import onnx

onnx_model = onnx.load(onnx_model_path)

Convert with onnx-tf:

from onnx_tf.backend import prepare

tf_rep = prepare(onnx_model)

Export TF model:

tf_rep.export_graph(tf_model_path)

You will get a Tensorflow model in SavedModel format.

Note: tf_model_path should not contain an extension like .pb.

TF Model Inference

import tensorflow as tf

model = tf.saved_model.load(tf_model_path)
model.trainable = False

input_tensor = tf.random.uniform([batch_size, channels, height, width])
out = model(**{'input': input_tensor})

TF to TFLite

To convert TF SavedModel format into TFLite models, you can use official tf.lite.TFLiteConverter class.

# Convert the model
converter = tf.lite.TFLiteConverter.from_saved_model(tf_model_path)
tflite_model = converter.convert()

# Save the model
with open(tflite_model_path, 'wb') as f:
    f.write(tflite_model)

TFLite Model Inference

import numpy as np
import tensorflow as tf

# Load the TFLite model and allocate tensors
interpreter = tf.lite.Interpreter(model_path="converted_model.tflite")
interpreter.allocate_tensors()

# Get input and output tensors
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# Test the model on random input data
input_shape = input_details[0]['shape']
input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32)
interpreter.set_tensor(input_details[0]['index'], input_data)

interpreter.invoke()

# get_tensor() returns a copy of the tensor data
# use tensor() in order to get a pointer to the tensor
output_data = interpreter.get_tensor(output_details[0]['index'])
print(output_data)

Supported Ops and Limitations

TFlite supports a subset of TF operations with some limitations. For full list of operations and limitations see TF Lite Ops page.

Most TFLite ops target float32 and quantized uint8 or int8 inference, but many ops don't support other types like float16 and strings.

TFLite with TF ops

Since TFLite builtin ops only supports a limited number of TF operators, not every model is convertible.

To allow conversion, usage of certain TF ops can be enabled in TFLite model.

However, running TFLite models with TF Ops requires pulling in the core TF runtime, which increases TFLite interpreter binary size.

TF Ops that can be enabled in TFLite

To convert to TFLite model with additional TF ops:

import tensorflow as tf

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.target_spec.supported_ops = [
    tf.lite.OpsSet.TFLITE_BUILTINS,  # enable TFLite ops
    tf.lite.OpsSet.SELECT_TF_OPS  # enable TF ops
]
tflite_model = converter.convert()

open("converted_model.tflite", "wb").write(tflite_model)

Run Inference

When using a TFLite model that has been converted with support for select TF ops, the client must also use a TFLite runtime that includes the necessary library of TF ops.

You don't need to do extra steps to use this select TF ops in Python. TFLite is automatically installed with that support.

Performance

The following table runs inference on MobileNet with Pixel 2.

Build	Time (ms)	APK Size
Builtin ops	260.7	561KB
Builtin ops + TF ops	264.5	8MB

Model Optimization

TFLite supports optimization via quantization, pruning and clustering.

Quantization

Quantization works by reducing the precision of the numbers used to represent a model's parameters (default, float32). This results in a smaller model size and faster computation.

Technique	Data Requirements	Size Reduction	Accuracy
Post-training float16 quantization	No data	Up tp 50%	Insignificant accuracy loss
Post-training dynamic range quantization	No data	Up to 75%	Accuracy loss
Post-training integer quantization	Unlabelled data	Up to 75%	Smaller accuracy loss
Quantization-aware training	Labelled training data	Up to 75%	Smallest accuracy loss

Post-training float16 quantization

Use this when you are deploying to float16-enabled GPU.

import tensorflow as tf

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types = [tf.float16]
tflite_quant_model = converter.convert()

Post-training dynamic range quantization

Don't use this. Use integer quantization.

import tensorflow as tf

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_quant_model = converter.convert()

Post-training integer quantization

Integer with float fallback (using default float input/output)

The model is in integer but use float operators when they don't have an integer implementation.

A common use case for ARM CPU.

import tensorflow as tf

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]

def representative_dataset_gen():
    for _ in range(num_calibration_steps):
        # get sample input data as numpy array 
        yield [input]

converter.representative_dataset = representative_dataset_gen
tflite_quant_model = converter.convert()

Integer only

A common use case for 8-bit MCU and Coral Edge TPU.

import tensorflow as tf

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]

def representative_dataset_gen():
    for _ in range(num_calibration_steps):
        # get sample input data as numpy array 
        yield [input]

converter.representative_dataset = representative_dataset_gen
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.inference_input_type = tf.int8
converter.inference_output_type = tf.int8
tflite_quant_model = converter.convert()

Pruning

Pruning works by removing parameters within a model that have only a minor impact on its predictions.

Pruned models are the same size on disk, and have the same runtime latency, but can be compressed more effectively. This makes pruning a useful technique for reducing model download size.

Clustering

Clustering works by grouping the weights of each layer in a model into a predefined number of clusters, then sharing the centroid values for the weights belonging to each individual cluster. This reduces the number of unique weight values in a model, thus reducing its complexity.

As a result, clustered models can be compressed more effecitvely, providing deployment benefits similar to pruning.

Note: For pruning and clustering, check out official TFLite Guide for more information.

References

• TFLite Documentation
• TFLiteConverter TF 1.x version
• ONNX-TensorFlow