Description of model - What task does it address (i.e. object detection, image classification)? What is the main advantage or feature of this model's architecture?
All ONNX models must pass the ONNX model checker before contribution. The snippet of code below can be used to perform the check. If any errors are encountered, it implies the check has failed.
import onnx
from onnx import checker
model_proto = onnx.load("path to .onnx file")
checker.check_model(model_proto)
- Create an account on Hugging Face: https://huggingface.co/join
- See list of models left to add to ONNX organization, please refer to the table with the Models list
- Add Gradio Demo under your username, see blog for setting up Gradio Demo on Hugging Face: https://huggingface.co/blog/gradio-spaces
- Request to join ONNX Organization: https://huggingface.co/onnx
- Once approved transfer model from your username to ONNX organization
- Add a badge for model in model table, see examples in Models list
| Model | Download | Download (with sample test data) | ONNX version | Opset version | Accuracy |
|---|---|---|---|---|---|
| Model Name | Relative link to ONNX Model with size | tar file containing ONNX model and synthetic test data (in .pb format) | ONNX version used for conversion | Opset version used for conversion | Accuracy values |
| Example (VGG 19) | 548.1 MB | 508.5 MB | 1.2.1 | 7 | 73.72 |
Please submit new models with Git LFS by committing directly to the repository, and using relative links (i.e. model/vgg19-7.onnx) in the table above. In this file name example, vgg19 is the name of the model and 7 is the opset number.
Source Framework ==> ONNX model
i.e. Caffe2 DenseNet-121 ==> ONNX DenseNet
Step by step instructions on how to use the pretrained model and link to an example notebook/code. This section should ideally contain:
Input to network (Example: 224x224 pixels in RGB)
Preprocessing required
Output of network
Post processing and meaning of output
This section should discuss datasets and any preparation steps if required.
Training details (preprocessing, hyperparameters, resources and environment) along with link to a training notebook (optional).
Also clarify in case the model is not trained from scratch and include the source/process used to obtain the ONNX model.
Validation script/notebook used to obtain accuracy reported above along with details of how to use it and reproduce accuracy. Details of experiments leading to accuracy from the reference paper.
Creating test data for uploaded models can help CI to verify the uploaded models by ONNXRuntime utilties. Please upload the ONNX model with created test data (test_data_set_0) in the .tar.gz.
pip install onnx onnxruntime numpy
git clone https://github.com/onnx/models.git
def create_test_dir(model_path, root_path, test_name,
name_input_map=None, symbolic_dim_values_map=None,
name_output_map=None):
"""
Create a test directory that can be used with onnx_test_runner, onnxruntime_perf_test.
Generates random input data for any missing inputs.
Saves output from running the model if name_output_map is not provided.
:param model_path: Path to the onnx model file to use.
:param root_path: Root path to create the test directory in.
:param test_name: Name for test. Will be added to the root_path to create the test directory name.
:param name_input_map: Map of input names to numpy ndarray data for each input.
:param symbolic_dim_values_map: Map of symbolic dimension names to values to use for the input data if creating
using random data.
:param name_output_map: Optional map of output names to numpy ndarray expected output data.
If not provided, the model will be run with the input to generate output data to save.
:return: None
"""
The input/output .pb files will be produced under temp/examples/test1/test_data_set_0.
import sys
sys.path.append('<onnx/models root dir>/workflow_scripts/')
import ort_test_dir_utils
import numpy as np
# example model with two float32 inputs called 'input' [batch_size, 1, 224, 224])
model_path = '<onnx/models root dir>/vision/super_resolution/sub_pixel_cnn_2016/model/super-resolution-10.onnx'
# If there is no symbolic_dim
ort_test_dir_utils.create_test_dir(model_path, 'temp/examples', 'test1')
# when using the default data generation any symbolic dimension values can be provided
# otherwise the default value for missing symbolic_vals would be 1
symbolic_vals = {'batch_size': 1} # provide value for symbolic dim named 'batch_size'
# let create_test_dir create random input in the (arbitrary) default range of -10 to 10.
# it will create data of the correct type based on the model.
ort_test_dir_utils.create_test_dir(model_path, 'temp/examples', 'test1', symbolic_dim_values_map=symbolic_vals)
# alternatively some or all input can be provided directly. any missing inputs will have random data generated.
batch_size = 64
inputs = {'input': np.random.rand(batch_size, 1, 224, 224).astype(np.float32)}
ort_test_dir_utils.create_test_dir(model_path, 'temp/examples', 'test2', name_input_map=inputs)
https://github.com/microsoft/onnxruntime/blob/master/tools/python/PythonTools.md
Link to paper or references.
Contributors' names
Add license information - on default, Apache 2.0