mnist

MNIST

Run MNIST example on VESSL.

Train Model Locally

Train model with the following command.

# PyTorch
pip install -r pytorch/requirements.txt
python examples/mnist/pytorch/main.py --save-model --save-image 

# Keras
pip install -r examples/mnist/keras/requirements.txt
python examples/mnist/keras/main.py --save-model --save-image 

Hyperparameters

epochs # [defaults: 10]
optimizer # adam, sgd, adadelta [defaults: adam]
batch_size # [defaults: 128]
learning_rate # [defaults: 0.01]

Train Model by VESSL Run

Keras

# PyTorch
vessl run "pip install -r pytorch/requirements.txt; python pytorch/main.py --save-model --save-image"

# Keras
vessl run "pip install -r keras/requirements.txt; python keras/main.py --save-model --save-image"

Saving Model to VESSL model registry (PyTorch)

Save the Model

1. Create a model repository in VESSL model registry if you don't have one.
   • Use model repository name as {VESSL_MODEL_REPO_NAME} in the following steps.
2. Train a PyTorch model with --save-model option.
   • Use {SAVED_MODEL_PATH} as the path to the saved model in the following steps.
3. Run a script to register the model:

# PyTorch
python pytorch/model.py --checkpoint {SAVED_MODEL_PATH} --model-repository {VESSL_MODEL_REPO_NAME}

This will register the model to the model repository and yield the following output:

Successfully registered model: https://vessl.ai/{ORGANIZATION_NAME}/models/{VESSL_MODEL_REPO_NAME}/{NUMBER}

Keep the number in the URL as {MODEL_NUMBER} in the following steps.

Serve Model in Local

Run the following command to serve the model previously registered.

vessl model serve vessl-mnist-example {MODEL_NUMBER} --install-reqs

After deploying the registered model, you can curl HTTP POST request with sample mnist image.

curl -X POST --data-binary @pytorch/sample_img/mnist_7.png localhost:8000

This will yield the following output:

{"result": 7}

Serve Model in remote using VESSL Serving

See VESSL Serving docs for more details.