Displaying model architectures with Netron

Model files in a file browser are usually opaque. XetHub brings underlying model architectures into view by supporting Netron visualizations. Reference the Netron project page for the most current list of compatible model types.

---

Viewing model architectures

Opening any supported model file from the XetHub UI automatically shows the Netron visualization of the model architecture.

To see this feature in full display, explore the ONNX Model Zoo repository and click through the models to see the Netron views rendered in browser.

Diffing model architectures in XetHub

Viewing a single architecture is cool, but it's even more interesting to see how model architectures change over time. A natural place to show this evolution is through pull requests. You can see before-and-after Netron views on your model files on any pull request where the model has changed by navigating to the Files Changed tab of the pull request and expanding the file.

Diffing model architectures in GitHub

Before-and-after Netron views are also supported through our GitHub integration, which adds Xet large file support to existing GitHub repositories.

To see the differences, create a pull request and click the link to the diff view of the model. Note that code changes can still be reviewed within GitHub.

The link will take you to a side-by-side comparison of your models hosted on XetHub:

Pull request flow

Here's a quick walkthrough to demonstrate how it works with your ML workflow. This flow is ideal for reviewing model changes collaboratively with your team.

1. Create a new repository

   • If using our GitHub integration, follow the instructions to install the app for your new repository

2. Clone the repo and create a main.py file with the following code:

   import tensorflow as tf
   from tensorflow import keras
   from tensorflow.keras import layers
   
   # Load and preprocess the MNIST dataset
   mnist = keras.datasets.mnist
   (train_images, train_labels), (test_images, test_labels) = mnist.load_data()
   
   # Normalize pixel values to be between 0 and 1
   train_images, test_images = train_images / 255.0, test_images / 255.0
   
   # Define the neural network model
   model = keras.Sequential([
       layers.Flatten(input_shape=(28, 28)),  # Flatten the 28x28 input images
       layers.Dense(128, activation='relu'),  # Fully connected layer with 128 units and ReLU activation
       layers.Dropout(0.2),  # Dropout layer to reduce overfitting
       layers.Dense(10)  # Output layer with 10 units (one for each digit)
   ])
   
   # Compile the model
   model.compile(optimizer='adam',
                 loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                 metrics=['accuracy'])
   
   # Train the model
   model.fit(train_images, train_labels, epochs=5)
   
   # Evaluate the model on the test dataset
   test_loss, test_accuracy = model.evaluate(test_images,  test_labels, verbose=2)
   print(f'\nTest accuracy: {test_accuracy}')
   
   # Save the model to a file
   model.save('mnist_model.h5')

3. Create a virtualenv and install TensorFlow:

   python -m venv .venv
   source .venv/bin/activate
   
   pip install tensorflow

4. Train your model locally and push the results to the remote:

   python main.py
   
   git add main.py mnist_model.h5
   git commit -m "Push newly trained mnist model"
   git push

5. Navigate to your latest commit

   • If using the GitHub app, you will find a link posted by the XetData bot to view the file in XetHub

6. Locally, check out a branch:

   git checkout -b more-layers

7. Modify main.py to add some more layers:

   layers.Dense(10),  # Output layer with 10 units (one for each digit)
   layers.Dense(128, activation='relu'),
   layers.Dropout(0.1)

   When you're done with your changes, retrain and push the results again.

   git commit -am "More changes on this branch"
   git push -u origin more-layers

8. Create a pull request from the UI and navigate to the Files Changed tab.

   • If on XetHub, expand the model file to see before-and-after Netron views.
   • If on GitHub, click on the View diff on XetHub link to see the before-and-after view in XetHub.