PyTorch is a supported framework on Truss. To package a PyTorch model, follow the steps below. There is no one-click Colab notebook for this example as it requires multiple files to operate successfully.

Install packages

If you're using a Jupyter notebook, add a line to install the torch, torchvision, and truss packages. Otherwise, ensure the packages are installed in your Python environment.
!pip install --upgrade torch torchvision truss
Truss officially supports torch version 1.9.0 or higher. Especially if you're using an online notebook environment like Google Colab or a bundle of packages like Anaconda, ensure that the version you are using is supported. If it's not, use the --upgrade flag and pip will install the most recent version.

Create an in-memory model

This is the part you want to replace with your own code. Using PyTorch, build a machine learning model and keep it in-memory. The PyTorch model example is a bit more involved code-wise, here it is in a couple of chunks.
First, create a file and add the following:
import torch
import torch.nn as nn
import torch.nn.functional as F
class MNISTNet(nn.Module):
def __init__(self):
super(MNISTNet, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 3, 1)
self.conv2 = nn.Conv2d(32, 64, 3, 1)
self.dropout1 = nn.Dropout(0.25)
self.dropout2 = nn.Dropout(0.5)
self.fc1 = nn.Linear(9216, 128)
self.fc2 = nn.Linear(128, 10)
def forward(self, x):
x = self.conv1(x)
x = F.relu(x)
x = self.conv2(x)
x = F.relu(x)
x = F.max_pool2d(x, 2)
x = self.dropout1(x)
x = torch.flatten(x, 1)
x = self.fc1(x)
x = F.relu(x)
x = self.dropout2(x)
x = self.fc2(x)
output = F.log_softmax(x, dim=1)
return output
Then, you can add the following functions in the notebook:
import torch
import torch.optim as optim
from torch.optim.lr_scheduler import StepLR
from torchvision import datasets, transforms
import torch.nn.functional as F
#pytorch model class must be in a python file
from model import MNISTNet
def train(model, device, train_loader, optimizer, epoch, log_interval=10):
for batch_idx, (data, target) in enumerate(train_loader):
data, target =,
output = model(data)
loss = F.nll_loss(output, target)
if batch_idx % log_interval == 0:
"Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}".format(
batch_idx * len(data),
100.0 * batch_idx / len(train_loader),
def test(model, device, test_loader):
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in test_loader:
data, target =,
output = model(data)
test_loss += F.nll_loss(
output, target, reduction="sum"
).item() # sum up batch loss
pred = output.argmax(
dim=1, keepdim=True
) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
"\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n".format(
100.0 * correct / len(test_loader.dataset),
def train_the_model():
device = "cpu"
epochs = 1
train_kwargs = {"batch_size": 64}
test_kwargs = {"batch_size": 1000}
transform = transforms.Compose(
[transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]
dataset1 = datasets.MNIST("../data", train=True, download=True, transform=transform)
dataset2 = datasets.MNIST("../data", train=False, transform=transform)
train_loader =, **train_kwargs)
test_loader =, **test_kwargs)
model = MNISTNet().to(device)
optimizer = optim.Adadelta(model.parameters(), lr=1.0)
scheduler = StepLR(optimizer, step_size=1, gamma=0.7)
for epoch in range(1, epochs + 1):
train(model, device, train_loader, optimizer, epoch)
test(model, device, test_loader)
return model, train_loader, test_loader
Finally, you can train your model:
model, _ , _ = train_the_model()

Create a Truss

Use the create command to package your model into a Truss.
from truss import create
tr = create(model, target_directory="pytorch_truss")
Check the target directory to see your new Truss!

Serve the model

With just a bit of helper code, we can serve a prediction:
transform = transforms.Compose(
[transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]
inputs = datasets.MNIST("../data", train=False, transform=transform)
dataset =, batch_size=1)
import numpy as np
For information on running the Truss locally, see local development.