SwiftAI
Pre-alpha version of Swift for TensorFlow's high-level API, modeled after fastai. The stable branch works with v0.6 of Swift for TensorFlow (s4tf), and the master branch follows the nightlies of s4tf (unstable and might be broken, see status below). For older releases, use one of the s4tf_* branch. To learn the details of what's in this repo, check out lessons 13 and 14 of course.fast.ai. Full docs will be added here as things become more stable.
You can download and train imagenette by typing swift run
at the root of this repo.
SwiftAI is built from the notebooks in nbs/
. Once you have the notebooks working the way you want, run the tools/export_import.ipynb
(we'll be replacing this exporter with a script soon).
Here's a walk-thru of training a model:
import SwiftAI
import TensorFlow
We need both SwiftAI and TensorFlow packages, since SwiftAI is designed to work with TensorFlow, not to replace TensorFlow.
let path = downloadImagenette()
As s4tf and SwiftAI add support for more types of models, we'll be adding lots of datasets; for now just Imagenette and MNIST are provided.
let il = ItemList(fromFolder: path, extensions: ["jpeg", "jpg"])
let sd = SplitData(il) {grandParentSplitter(fName: $0, valid: "val")}
We use the Swift version of the data block API to grab the files we need, and split in to train and validation sets.
var procLabel = CategoryProcessor()
let sld = makeLabeledData(sd, fromFunc: parentLabeler, procLabel: &procLabel)
Processors are (potentially stateful) functions which preprocess data. In this case, CategoryProcessor
gets a list of unique labels from the data and creates a mapping to turn labels into Int
s.
let rawData = sld.toDataBunch(itemToTensor: pathsToTensor, labelToTensor: intsToTensor, bs: 128)
A DataBunch
is a simple object which contains labeled training and validation Dataset
s.
let data = transformData(rawData) { openAndResize(fname: $0, size: 128) }
We can add any lazily-applied transformations we need to convert (for instance) raw file names into data ready for modeling (in this case, images of the same size).
func modelInit() -> XResNet { return xresnet18(cOut: 10) }
let optFunc: (XResNet) -> StatefulOptimizer<XResNet> = adamOpt(lr: 1e-3, mom: 0.9, beta: 0.99, wd: 1e-2, eps: 1e-4)
let learner = Learner(data: data, lossFunc: softmaxCrossEntropy, optFunc: optFunc, modelInit: modelInit)
A Learner
is an object that knows how apply an Optimizer
(in this case, adamOpt
, which defaults to AdamW) to train a model (xresnet18
) based on some DataBunch
, to minimize some differentiable loss function (softmaxCrossEntropy
).
let recorder = learner.makeDefaultDelegates(metrics: [accuracy])
learner.addDelegate(learner.makeNormalize(mean: imagenetStats.mean, std: imagenetStats.std))
learner.addOneCycleDelegates(1e-3, pctStart: 0.5)
Delegates are used to customize training in many ways. In this case, we're adding delegates to:
- Record losses and metrics after each batch, add a progress bar, and move data to the GPU (these are all default delegates in SwiftAI)
- Normalize the data
- Use the 1 cycle policy
try! learner.fit(1)
The fit
method will train and validate your model for as many epochs as you request.
Docker
A Dockerfile has been created to help spin up a CPU based learning and development environment.
# Only Build image
make
# Build image and run Jupyter
make jupyter
# Build image and run shell
make shell
# Sync notebooks to sources (Sources must not have modifications)
make sync-nbs-to-srcs
Status
Master is currently working with the latest nightlies toolchain.