A Swift Wrapper for PyTorch and Torchvision.

Getting Started

Note: This project is not affiliated with PyTorch and is still in early development.

Prerequisites:

This library requires Xcode toolchain Swift for Tensorflow release/Swift 5

Installing with Swift Package Manager:

Add the following line to your dependencies in Package.swift file

.package(url: "https://github.com/codeamt/PyTorchSwift.git", .branch("master")),

Usage

Importing:

import Foundation import Python
import PyTorch4Swift let torch = PyTorchSwift()
let torchvision = TorchVisionSwift()

Note: If you have trouble or get errors importing Python (even after upgrading your toolchain), try changing your build system to legacy: File > Project Settings> Build System -> Legacy

Data Preprocessing:

var transforms_list:[PythonObject] = [torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]
let transforms = torchvision.transforms.Compose(transforms_list)
let trainset = torchvision.datasets.CIFAR10(root:"./data", train:true, download:true, transform:transforms)
let trainloader = torch.utils.data.DataLoader(trainset, batch_size:4, shuffle:true, num_workers:2)
let testset = torchvision.datasets.CIFAR10(root:"./data", train:false, download:true, transform:transforms)
let testloader = torch.utils.data.DataLoader(testset, batch_size:4, shuffle:false, num_workers:2)
let labels = ("plane", "car", "bird", "cat","deer", "dog", "frog", "horse", "ship", "truck")

Setting Up Transfer Learning:

let arch = torchvision.models.resnet18(pretrained=True)
if (torch.cuda.is_available() == true){
let device = torch.device("cuda:0")
arch.to(device)
print(device)
} else {
print( "using cpu")
}
let criterion = torch.nn.CrossEntropyLoss()
let optimizer = torch.optim.SGD(arch.parameters(), lr:0.001, momentum:0.9)
var runningLoss = 0.0

Training:

for epoch in 0...50 {
runningLoss = 0.0
let trainIter:PythonObject = Python.enumerate(trainloader)
for idx in trainloader.indices{
let data = trainIter.next()
let inputs = data[0]
let labels = data[1]
// zero the parameter gradients
optimizer.zero_grad()
// forward + backward + optimize
let outputs = arch(inputs)
let loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
// print statistics
runningLoss += Double(loss.item())!
//print every 2000 mini-batches
if (Int(idx)! % 2000 == 1999){
print("[epoch:\(epoch), \(idx + 1)] loss: \(runningLoss / 2000)")
}
}
runningLoss = 0.0
print("Finished Training")
}

Testing:

// Test network
let testiter = Python.enumerate(testloader)
let data = testiter.next()
let images = data[0]
let labels = data[1]
let outputs = arch(images)
let (_, predicted) = torch.`max`(outputs)
print("predicted: ",predicted)
print("test images: ",images)
print("test labels: ",labels)
print("nn outputs: ",outputs)

Additional Notes:

This library was inspired by this repo (an executable project) and learnings from the Fast.ai Live Course (2019) - Part 2 with the intention diversifying the DL/ML libraries available to Swift developers working on ML/DL projects and research. To report bugs please submit an issue. To contribute, please make a pull request!