onnx2tflite

ONNX->Keras and ONNX->TFLite tools

Welcome

If you have some good ideas, welcome to discuss or give project PRs.

Install

git clone https://github.com/MPolaris/onnx2tflite.git
cd onnx2tflite
python setup.py install

from onnx2tflite import onnx_converter
res = onnx_converter(
        onnx_model_path = "./model.onnx",
        need_simplify = True,
        output_path = "./models/",
        target_formats = ['tflite'],
    )

---

CLI Usage

# basic conversion
python -m onnx2tflite --weights "./your_model.onnx"

# save to specific path
python -m onnx2tflite --weights "./your_model.onnx" --outpath "./save_path"

# output formats
python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite"
python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" "keras"

# subgraph extraction (cutoff model by layer names)
python -m onnx2tflite --weights "./your_model.onnx" --input-node-names "layer_in" --output-node-names "layer_out1" "layer_out2"

# quantization
python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --weigthquant     # weight-only int8
python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --fp16              # fp16
python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --int8              # int8 (random calib, low accuracy)
python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --int8              # int8 with calibration
    --calibration-data "./calib_input.npy"

# multi-input model with int8
python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --int8
    --calibration-data "./input0.npy" "./input1.npy"

# direct IR conversion (bypass Keras, faster)
python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --direct-ir

---

Features

• High Consistency. Compare to ONNX outputs, average error less than 1e-5 per elements.
• More Faster. Output tensorflow-lite model 30% faster than onnx_tf.
• Auto Channel Align. Auto convert pytorch format(NCHW) to tensorflow format(NHWC).
• Deployment Support. FP16, INT8 (weight-only, full integer) quantization.
• Two Backends: Keras (stable) and Direct IR (fast, bypasses Keras).
• Multi-Input INT8: calibration data via .npy files, supports models with multiple inputs.

---

INT8 Calibration Data Format

INT8 quantization requires a representative dataset for calibration.
Calibration data is provided as .npy files — one file per model input.

Data format:

• Shape: (N, ...) where N is the number of calibration samples (recommended: 50-200)
• The remaining dimensions must match the model input shape
• Data should be preprocessed (normalized, resized, channel-ordered) before saving

Multi-input models: provide one .npy file per input in the same order as the ONNX model inputs.

import numpy as np

# Single-input model: shape (1, 3, 224, 224)
calib = np.random.randn(100, 3, 224, 224).astype(np.float32)
np.save("./calib_input.npy", calib)

# Multi-input model: two inputs with shapes (1, 3, 224, 224) and (1, 1000)
np.save("./calib_img.npy", np.random.randn(100, 3, 224, 224).astype(np.float32))
np.save("./calib_vec.npy", np.random.randn(100, 1000).astype(np.float32))

onnx_converter(
    onnx_model_path = "./model.onnx",
    int8_model = True,
    calibration_data = ["./calib_img.npy", "./calib_vec.npy"],
)

---

Direct IR Backend

The use_direct_ir=True flag enables a new conversion path that builds TFLite FlatBuffer models directly from the ONNX graph, bypassing Keras entirely.

• Faster conversion (no Keras model building or saving)
• More control over the generated TFLite ops
• Supports most common operators

from onnx2tflite import onnx_converter

onnx_converter("model.onnx", use_direct_ir=True)

The Keras backend (use_direct_ir=False, default) remains available for operators not yet supported by the direct IR path (e.g. Erf).

Python API

from onnx2tflite import onnx_converter

# Basic conversion
onnx_converter("model.onnx", output_path="./out", target_formats=['tflite'])

# INT8 quantization with calibration data
onnx_converter(
    "model.onnx",
    int8_model=True,
    calibration_data=["./calib_input.npy"],
)

# Direct IR backend (bypass Keras)
onnx_converter("model.onnx", use_direct_ir=True)

onnx_converter parameters

Parameter	Type	Default	Description
onnx_model_path	str	required	Path to ONNX model file
output_path	str	None	Output directory (default: same as input)
input_node_names	list	None	Subgraph input names (None = all)
output_node_names	list	None	Subgraph output names (None = all)
need_simplify	bool	True	Run onnx-simplifier before conversion
target_formats	list	['tflite']	Output formats: 'tflite', 'keras'
use_direct_ir	bool	False	Use direct IR backend (bypass Keras)
native_groupconv	bool	False	Use native grouped conv (tflite >= 2.9)
weight_quant	bool	False	Weight-only INT8 quantization
fp16_model	bool	False	FP16 quantization
int8_model	bool	False	Full INT8 quantization
calibration_data	list	None	List of .npy file paths for INT8 calibration

---

Validated models

• SSD
• HRNet
• YOLOX
• YOLOV3
• YOLOV4
• YOLOV5
• YOLOV6
• YOLOV7
• YOLOV10
• MoveNet
• UNet\FPN
• ViT(torchvision)
• SwinTransformerV1
• MLP(custom)
• DCGAN(custom)
• AutoEncoder/VAE
• all torchvision classification models
• some segmation models in torchvision
• 1D or 2D CNN without special operators(custom)

---

Add operator by yourself

When you counter unspported operator, you can choose to add it by yourself or make an issue.

Keras backend

Register a handler class in keras_backend/ops/ (e.g. activation.py):

@OPERATOR.register_operator("HardSigmoid")
class TFHardSigmoid():
    def __init__(self, tensor_graph, node_weights, node_inputs,
                 node_attribute, node_outputs, layout_dict, *args, **kwargs):
        super().__init__()
        self.alpha = node_attribute.get("alpha", 0.2)
        self.beta = node_attribute.get("beta", 0.5)

    def __call__(self, inputs):
        return tf.clip_by_value(self.alpha*inputs+self.beta, 0, 1)

TFLite direct IR backend

Add a handler function in tflite_backend/ops/:

@_register("HardSigmoid")
def _hard_sigmoid(builder, node):
    inp = builder._tensor_map[node.input[0]]
    ...

---

License

This software is covered by Apache-2.0 license.