Open Neural Network Exchange (ONNX) is a powerful and open format built to represent machine learning models. The final outcome of training any machine learning or deep learning algorithm is a model file that represents the mapping of input data to output predictions in an efficient manner. These models are stored in different file formats depending on the framework they were created in .pkl for Scikit-learn, .pb for TensorFlow, .pth for PyTorch, and so on. Therein lies the problem, you can’t take a model created and trained in one framework and use it or deploy it in a different framework.  

The intent behind ONNX is to be like the “USB standard” of the machine learning world. Before the introduction of USB computers and computer peripherals used to have ad hoc proprietary interfaces. Much like the pre-USB era, the present machine learning models have ad hoc formats. It overcomes the problem of framework lock-in by providing a universal intermediary model format that frameworks can easily save to and load from. This allows ML developers to create models in the framework of their choice without worrying about the deployment environment. ONNX also makes it easier to access hardware acceleration provided by different frameworks and runtime environments. 

Installing ONNX and other required libraries

Requirement: Python 3.6 or higher version

1. Install ONNX
pip:
pip install onnx

Conda:
conda install -c conda-forge onnx

2. Install tensorflow and onnx-tensorflow

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

3. Install PyTorch and torchvision
pip install pytorch
pip install torchvision

Converting a PyTorch model to TensorFlow

1. Import required libraries and classes

 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.optim as optim
 from torchvision import datasets, transforms
 from torch.autograd import Variable
 import onnx
 from onnx_tf.backend import prepare
 

2. Define a basic CNN model

 class Net(nn.Module):    
      def __init__(self):
         super(Net, self).__init__()
         self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
         self.conv2 = nn.Conv2d(10, 20, kernel_size=5)
         self.conv2_drop = nn.Dropout2d()
         self.fc1 = nn.Linear(320, 50)
         self.fc2 = nn.Linear(50, 10)
     def forward(self, x):
         x = F.relu(F.max_pool2d(self.conv1(x), 2))
         x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
         x = x.view(-1, 320)
         x = F.relu(self.fc1(x))
         x = F.dropout(x, training=self.training)
         x = self.fc2(x)
         return F.log_softmax(x, dim=1)
 

3. Create the train and test methods

 def train(model, device, train_loader, optimizer, epoch):
     model.train()
     for batch_idx, (data, target) in enumerate(train_loader):
         data, target = data.to(device), target.to(device)
         optimizer.zero_grad()
         output = model(data)
         loss = F.nll_loss(output, target)
         loss.backward()
         optimizer.step()
         if batch_idx % 1000 == 0:
             print('Train Epoch: {} \tLoss: {:.6f}'.format(
                     epoch,  loss.item()))

 def test(model, device, test_loader):
     model.eval()
     test_loss = 0
     correct = 0
     with torch.no_grad():
         for data, target in test_loader:
             data, target = data.to(device), target.to(device)
             output = model(data)
             test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
             pred = output.max(1, keepdim=True)[1] # get the index of the maxlog-probability
             correct += pred.eq(target.view_as(pred)).sum().item()
     test_loss /= len(test_loader.dataset)
     print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
         test_loss, correct, len(test_loader.dataset),
         100. * correct / len(test_loader.dataset))) 

4. Download the train and test datasets, normalize them and create data loaders.

train_loader = torch.utils.data.DataLoader(datasets.MNIST('../data', train=True, download=True,
                    transform=transforms.Compose([
                        transforms.ToTensor(),
                        transforms.Normalize((0.1307,), (0.3081,))
                    ])),
     batch_size=64, shuffle=True)
test_loader = torch.utils.data.DataLoader(
     datasets.MNIST('../data', train=False, transform=transforms.Compose([
                        transforms.ToTensor(),
                        transforms.Normalize((0.1307,), (0.3081,))
                    ])),
     batch_size=1000, shuffle=True)
 

5. Create the model, define the optimitier and train it

 device = torch.device("cuda")
 model = Net().to(device)
 optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)
 for epoch in range(21):
     train(model, device, train_loader, optimizer, epoch)
     test(model, device, test_loader) 

6. Save the trained model
   torch.save(model.state_dict(), 'mnist.pth')
   
7. Load the saved model. Generate and pass random input so the Pytorch exporter can trace the model and save it to an ONNX file.

 trained_model = Net()
 trained_model.load_state_dict(torch.load('mnist.pth'))
 dummy_input = Variable(torch.randn(1, 1, 28, 28))
 torch.onnx.export(trained_model, dummy_input, "mnist.onnx")

8. Load the ONNX file and import it to Tensorflow

 model = onnx.load('mnist.onnx')
 tf_rep = prepare(model) 

9. Run and test the Tensorflow model.

 import numpy as np
 from IPython.display import display
 from PIL import Image
 print('Image 1:')
 img = Image.open('two.png').resize((28, 28)).convert('L')
 display(img)
 output = tf_rep.run(np.asarray(img, dtype=np.float32)[np.newaxis, np.newaxis, :, :])
 print('The digit is classified as ', np.argmax(output))
 print('Image 2:')
 img = Image.open('three.png').resize((28, 28)).convert('L')
 display(img)
 output = tf_rep.run(np.asarray(img, dtype=np.float32)[np.newaxis, np.newaxis, :, :])
 print('The digit is classified as ', np.argmax(output)) 

10. Save the Tensorflow model.
    tf_rep.export_graph('mnist.pb')

• GitHub
• Collab Notebook