PyTorch to ONNX Outside Wallaroo

How to convert PyTorch ML models into the ONNX format.

This tutorial and the assets can be downloaded as part of the Wallaroo Tutorials repository.

How to Convert PyTorch to ONNX

The following tutorial is a brief example of how to convert a PyTorth (aka sk-learn) ML model to ONNX. This allows organizations that have trained sk-learn models to convert them and use them with Wallaroo.

This tutorial assumes that you have a Wallaroo instance and are running this Notebook from the Wallaroo Jupyter Hub service. This sample code is based on the guide Convert your PyTorch model to ONNX.

This tutorial provides the following:

  • pytorchbikeshare.pt: a RandomForestRegressor PyTorch model. This model has a total of 58 inputs, and uses the class BikeShareRegressor.

Prerequisites

  • An installed Wallaroo instance.
  • The following Python libraries installed:

Conversion Process

Libraries

The first step is to import our libraries we will be using. For this example, the PyTorth torch library will be imported into this kernel.

# the Pytorch libraries
# Import into this kernel

import torch
import torch.onnx 

Load the Model

To load a PyTorch model into a variable, the model’s class has to be defined. For out example we are using the BikeShareRegressor class as defined below.

class BikeShareRegressor(torch.nn.Module):
    def __init__(self):
        super(BikeShareRegressor, self).__init__()

        
        self.net = nn.Sequential(nn.Linear(input_size, l1),
                                 torch.nn.ReLU(),
                                 torch.nn.Dropout(p=dropout),
                                 nn.BatchNorm1d(l1),
                                 nn.Linear(l1, l2),
                                 torch.nn.ReLU(),
                                 torch.nn.Dropout(p=dropout),                                
                                 nn.BatchNorm1d(l2),                                                                                                   
                                 nn.Linear(l2, output_size))

    def forward(self, x):
        return self.net(x)

Now we will load the model into the variable pytorch_tobe_converted.

# load the Pytorch model
model = torch.load("./pytorch_bikesharingmodel.pt")

Convert_ONNX Inputs

Now we will define our method Convert_ONNX() which has the following inputs:

  • PyTorchModel: the PyTorch we are converting.

  • modelInputs: the model input or tuple for multiple inputs.

  • onnxPath: The location to save the onnx file.

  • opset_version: The ONNX version to export to.

  • input_names: Array of the model’s input names.

  • output_names: Array of the model’s output names.

  • dynamic_axes: Sets variable length axes in the format, replacing the batch_size as necessary:
    {'modelInput' : { 0 : 'batch_size'}, 'modelOutput' : {0 : 'batch_size'}}

  • export_params: Whether to store the trained parameter weight inside the model file. Defaults to True.

  • do_constant_folding: Sets whether to execute constant folding for optimization. Defaults to True.

#Function to Convert to ONNX 
def Convert_ONNX(): 

    # set the model to inference mode 
    model.eval() 

    # Export the model   
    torch.onnx.export(model,         # model being run 
         dummy_input,       # model input (or a tuple for multiple inputs) 
         pypath,       # where to save the model  
         export_params=True,  # store the trained parameter weights inside the model file 
         opset_version=15,    # the ONNX version to export the model to 
         do_constant_folding=True,  # whether to execute constant folding for optimization 
         input_names = ['modelInput'],   # the model's input names 
         output_names = ['modelOutput'], # the model's output names 
         dynamic_axes = {'modelInput' : {0 : 'batch_size'}, 'modelOutput' : {0 : 'batch_size'}} # variable length axes 
    ) 
    print(" ") 
    print('Model has been converted to ONNX') 

Convert the Model

We’ll now set our variables and run our conversion. For out example, the input_size is known to be 58, and the device value we’ll derive from torch.cuda. We’ll also set the ONNX version for exporting to 15.

pypath = "pytorchbikeshare.onnx"

input_size = 58

if torch.cuda.is_available():
    device = 'cuda'
else:
    device = 'cpu'

onnx_opset_version = 15

# Set up some dummy input tensor for the model
dummy_input = torch.randn(1, input_size, requires_grad=True).to(device)

Convert_ONNX()
================ Diagnostic Run torch.onnx.export version 2.0.0 ================
verbose: False, log level: Level.ERROR
======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================

Model has been converted to ONNX

Conclusion

And now our conversion is complete. Please feel free to use this sample code in your own projects.