Skip to main content

The deep learning models converter

Project description


ONNX to Keras deep neural network converter.

GitHub License Python Version Downloads PyPI


TensorFlow 2.0


onnx_to_keras(onnx_model, input_names, input_shapes=None, name_policy=None, verbose=True, change_ordering=False) -> {Keras model}

onnx_model: ONNX model to convert

input_names: list with graph input names

input_shapes: override input shapes (experimental)

name_policy: ['renumerate', 'short', 'default'] override layer names (experimental)

verbose: detailed output

change_ordering: change ordering to HWC (experimental)

Getting started

ONNX model

import onnx
from onnx2keras import onnx_to_keras

# Load ONNX model
onnx_model = onnx.load('resnet18.onnx')

# Call the converter (input - is the main model input name, can be different for your model)
k_model = onnx_to_keras(onnx_model, ['input'])

Keras model will be stored to the k_model variable. So simple, isn't it?

PyTorch model

Using ONNX as intermediate format, you can convert PyTorch model as well.

import numpy as np
import torch
from torch.autograd import Variable
from pytorch2keras.converter import pytorch_to_keras
import torchvision.models as models

if __name__ == '__main__':
    input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
    input_var = Variable(torch.FloatTensor(input_np))
    model = models.resnet18()
    k_model = \
        pytorch_to_keras(model, input_var, [(3, 224, 224,)], verbose=True, change_ordering=True)

    for i in range(3):
        input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
        input_var = Variable(torch.FloatTensor(input_np))
        output = model(input_var)
        pytorch_output =
        keras_output = k_model.predict(np.transpose(input_np, [0, 2, 3, 1]))
        error = np.max(pytorch_output - keras_output)
        print('error -- ', error)  # Around zero :)

Deplying model as frozen graph

You can try using the snippet below to convert your onnx / PyTorch model to frozen graph. It may be useful for deploy for Tensorflow.js / for Tensorflow for Android / for Tensorflow C-API.

import numpy as np
import torch
from pytorch2keras.converter import pytorch_to_keras
from torch.autograd import Variable
import tensorflow as tf
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2

# Create and load model
model = Model()

# Make dummy variables (and checking if the model works)
input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
input_var = Variable(torch.FloatTensor(input_np))
output = model(input_var)

# Convert the model!
k_model = \
    pytorch_to_keras(model, input_var, (3, 224, 224), 
                     verbose=True, name_policy='short',

# Save model to SavedModel format, "./models")

# Convert Keras model to ConcreteFunction
full_model = tf.function(lambda x: k_model(x))
full_model = full_model.get_concrete_function(
    tf.TensorSpec(k_model.inputs[0].shape, k_model.inputs[0].dtype))

# Get frozen ConcreteFunction
frozen_func = convert_variables_to_constants_v2(full_model)

print("-" * 50)
print("Frozen model layers: ")
for layer in [ for op in frozen_func.graph.get_operations()]:

print("-" * 50)
print("Frozen model inputs: ")
print("Frozen model outputs: ")

# Save frozen graph from frozen ConcreteFunction to hard drive,


This software is covered by MIT License.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

onnx2keras-0.0.22.tar.gz (17.8 kB view hashes)

Uploaded Source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page