In [ ]:
// %install '.package(path: "$cwd/FastaiNotebook_03_minibatch_training")' FastaiNotebook_03_minibatch_training
%install '.package(path: "$cwd/FastaiNotebook_04_callbacks")' FastaiNotebook_04_callbacks
In [ ]:
let IMAGENETTE_URL = "https://s3.amazonaws.com/fast-ai-imageclas/imagenette.tgz"
let IMAGENETTE_160_URL = "https://s3.amazonaws.com/fast-ai-imageclas/imagenette-160.tgz"
let IMAGENETTE_320_URL = "https://s3.amazonaws.com/fast-ai-imageclas/imagenette-320.tgz"
let IMAGENETTE_FILE = "imagenette.tgz"
// First let us download the dataset.
func downloadImagenetteData(_ url: String, _ filename: String) -> String {
let urllib = Python.import("urllib.request")
let downloadResult = urllib.urlretrieve(url, filename)
print("\(downloadResult[0])")
return String(downloadResult[0])!
}
//downloadImagenetteData(IMAGENETTE_160_URL, IMAGENETTE_FILE)
subprocess.run(["wget", IMAGENETTE_160_URL])
subprocess.run(["tar", "-xzvf", "imagenette-160.tgz"])
subprocess.run(["ls"])
In [ ]:
// Create a list of files
subprocess.run(["find", "imagenette-160/train", "-name", "*.JPEG", "-fprint", "imagenette-train-files.txt"])
subprocess.run(["find", "imagenette-160/val", "-name", "*.JPEG", "-fprint", "imagenette-val-files.txt"])
subprocess.run(["ls"])
In [ ]:
%include "EnableIPythonDisplay.swift"
IPythonDisplay.shell.enable_matplotlib("inline")
import TensorFlow
import Python
//import Foundation
let subprocess = Python.import("subprocess")
let plt = Python.import("matplotlib.pyplot")
let np = Python.import("numpy")
In [ ]:
// Utility to resize an image!
func resizeImage(filename: StringTensor) -> Tensor<Float> {
let imageBytes: StringTensor = Raw.readFile(filename: filename)
let decodedImage = Raw.decodeJpeg(contents: imageBytes, channels: 3, dctMethod: "")
let resizedImages = Raw.resizeBicubic(
images: Tensor<UInt8>([decodedImage]),
size: Tensor<Int32>([32, 32]))
return resizedImages.reshaped(to: TensorShape(32, 32, 3))
// return Raw.cast(resizedImage)
}
let resizedImage = resizeImage(filename: StringTensor("imagenette-160/val/n03028079/ILSVRC2012_val_00003682.JPEG"))
let npa = resizedImage.makeNumpyArray().astype(np.uint8)
plt.imshow(npa)
plt.show()
Out[ ]:
None
In [ ]:
// Some sample data to quickly test things!
let sampleValFiles: [String] = [
"imagenette-160/val/n03028079/ILSVRC2012_val_00034021.JPEG",
"imagenette-160/val/n03028079/ILSVRC2012_val_00026451.JPEG",
"imagenette-160/val/n03028079/ILSVRC2012_val_00006268.JPEG",
"imagenette-160/val/n03028079/ILSVRC2012_val_00004912.JPEG",
"imagenette-160/val/n03028079/ILSVRC2012_val_00009705.JPEG",
"imagenette-160/val/n03000684/ILSVRC2012_val_00000537.JPEG",
"imagenette-160/val/n03000684/ILSVRC2012_val_00004034.JPEG",
"imagenette-160/val/n03000684/ILSVRC2012_val_00004262.JPEG",
"imagenette-160/val/n03000684/ILSVRC2012_val_00005506.JPEG",
"imagenette-160/val/n03000684/ILSVRC2012_val_00006043.JPEG",
"imagenette-160/val/n01440764/ILSVRC2012_val_00000293.JPEG",
"imagenette-160/val/n01440764/ILSVRC2012_val_00002138.JPEG",
"imagenette-160/val/n01440764/ILSVRC2012_val_00003014.JPEG",
"imagenette-160/val/n01440764/ILSVRC2012_val_00006697.JPEG",
"imagenette-160/val/n01440764/ILSVRC2012_val_00007197.JPEG"
]
let sampleValFilesTensor = StringTensor(sampleValFiles)
let sampleTrainFiles: [String] = [
"imagenette-160/train/n01440764/n01440764_10026.JPEG",
"imagenette-160/train/n01440764/n01440764_10027.JPEG",
"imagenette-160/train/n01440764/n01440764_10029.JPEG",
"imagenette-160/train/n01440764/n01440764_10040.JPEG",
"imagenette-160/train/n01440764/n01440764_10042.JPEG",
"imagenette-160/train/n03000684/n03000684_1000.JPEG",
"imagenette-160/train/n03000684/n03000684_1011.JPEG",
"imagenette-160/train/n03000684/n03000684_10131.JPEG",
"imagenette-160/train/n03000684/n03000684_10132.JPEG",
"imagenette-160/train/n03000684/n03000684_10136.JPEG",
"imagenette-160/train/n03028079/n03028079_10006.JPEG",
"imagenette-160/train/n03028079/n03028079_10014.JPEG",
"imagenette-160/train/n03028079/n03028079_10020.JPEG",
"imagenette-160/train/n03028079/n03028079_1002.JPEG",
"imagenette-160/train/n03028079/n03028079_10032.JPEG"
]
let sampleTrainFilesTensor = StringTensor(sampleTrainFiles)
In [ ]:
// Test a basic dataset that reads and resizes images
let ds:Dataset<Tensor<Float>> = Dataset<StringTensor>(
elements: sampleTrainFilesTensor).map(resizeImage)
//.batched(5) // For batching
let rows = 2
let cols = 2
//plt.figure(figsize: [rows * 3, cols * 3])
for (i, d) in ds.enumerated() {
print("Processing image \(i)")
let img = plt.subplot(rows, cols, i + 1)
img.imshow(d.makeNumpyArray().astype(np.uint8))
if (i + 1) >= (rows * cols) { break }
}
plt.show()
Processing image 0 Processing image 1 Processing image 2 Processing image 3
Out[ ]:
None
In [ ]:
// Some data structures from 03 minibatch training
// export
public struct DataBatch<Inputs: Differentiable & TensorGroup, Labels: TensorGroup>: TensorGroup {
public var xb: Inputs
public var yb: Labels
public init(xb: Inputs, yb: Labels){
self.xb = xb
self.yb = yb
}
}
// export
public struct DataBunch<Element> where Element: TensorGroup{
public var train: Dataset<Element>
public var valid: Dataset<Element>
public init(train: Dataset<Element>, valid: Dataset<Element>) {
self.train = train
self.valid = valid
}
}
In [ ]:
// TODO: When TF-421 is fixed, switch this back to Int32 labels.
struct LabeledImagenetteFiles : TensorGroup {
let labels: Tensor<Float>
let fnames: StringTensor
}
// e.g., getLabel(from: "imagenette-160/train/n03028079/ILSVRC2012_val_00034021.JPEG")
func getLabel(from fname: String) -> String {
// let splits = fname.components(separatedBy: "/")
// let splits = fname.pathComponents
let splits = fname.split(separator: "/")
assert(splits.endIndex >= 2, "Filename is not in the relevant form.")
return String(splits[splits.endIndex - 2])
}
func labeledFiles(fnames: [String]) -> LabeledImagenetteFiles {
var labels: [String] = []
for fname in fnames {
labels.append(getLabel(from: fname))
}
var labelsDict: [String: Float] = [:]
for (i, label) in Set(labels).enumerated() {
labelsDict[label] = Float(i)
}
let labelIds = labels.map { labelsDict[$0]! }
return LabeledImagenetteFiles(
labels: Tensor<Float>(labelIds),
fnames: StringTensor(fnames)
)
}
/* TODO: Enable when we can have pre-built toolchain with Foundation
func labeledFiles(from inputFile: String) -> [LabeledImagenetteInput] {
do {
let fileContents = try String(contentsOfFile: inputFile)
let fnamesRaw = fileContents.components(separatedBy: .newlines)
let fnames = fnamesRaw.filter { !$0.trimmingCharacters(in: .whitespaces).isEmpty }
return labeledFiles(fnames: fnames)
} catch {
print("Error labeling files")
return []
}
}
*/
let res = labeledFiles(fnames: sampleValFiles)
print ("Labels: \(res.labels.shape) \n \(res.labels)")
print ("Fnames: \(res.fnames)")
Labels: TensorShape(dimensions: [15]) [1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 2.0, 2.0, 2.0, 2.0, 2.0] Fnames: StringTensor(handle: TensorFlow.TensorHandle<Swift.String>)
In [ ]:
// TODO: Not tested.
func imagenetteDataBatch(
_ input: LabeledImagenetteFiles
) -> DataBatch<Tensor<Float>, Tensor<Float>> {
return DataBatch(
xb: resizeImage(filename: input.fnames),
yb: input.labels)
}
func imagenetteDataset(
_ input: LabeledImagenetteFiles
) -> Dataset<DataBatch<Tensor<Float>, Tensor<Float>>> {
return Dataset<LabeledImagenetteFiles>(
elements: labeledFiles(fnames: sampleValFiles)
).map(imagenetteDataBatch)
}
func imagenetteDataBunch(trainFiles: String, valFiles: String, bs: Int = 64
) -> DataBunch<DataBatch<Tensor<Float>, Tensor<Float>>> {
let trainDS = imagenetteDataset(labeledFiles(fnames: sampleTrainFiles))
let valDS = imagenetteDataset(labeledFiles(fnames: sampleValFiles))
return DataBunch(
train: trainDS.batched(Int64(bs)),
valid: valDS.batched(Int64(bs)))
}
let data = imagenetteDataBunch(
trainFiles: "imagenette-train-files.txt",
valFiles: "imagenette-val-files.txt",
bs:2)
for (i, d) in data.train.enumerated() {
print("Processing batch \(i) \(d.xb.shape)")
}
Processing batch 0 TensorShape(dimensions: [2, 32, 32, 3]) Processing batch 1 TensorShape(dimensions: [2, 32, 32, 3]) Processing batch 2 TensorShape(dimensions: [2, 32, 32, 3]) Processing batch 3 TensorShape(dimensions: [2, 32, 32, 3]) Processing batch 4 TensorShape(dimensions: [2, 32, 32, 3]) Processing batch 5 TensorShape(dimensions: [2, 32, 32, 3]) Processing batch 6 TensorShape(dimensions: [2, 32, 32, 3]) Processing batch 7 TensorShape(dimensions: [1, 32, 32, 3])
In [ ]: