我是iOS和CoreML的新手。我有一个非常简单的UI，带有两个UIImageViews（一个应该是输入，第二个应该是输出）。点击第一个时，图像应由神经网络处理，输出应显示在第二个中。
但是，当我尝试从MLMultiArray输出对象下载图像并从中创建UIImage时，可以将其上传到第二个UIImageView时，我得到EXC_BAD_ACCESS（code = 1）。
我已经减少了根本不调用神经网络处理的问题，只是尝试从MLMultiArray创建新图像。结果是一样的。
之后，我尝试从一个空的缓冲区生成一个UIImage。图像创建正确，但是如果我尝试更新UIImageView以使用它，则会遇到相同的错误。
如果我尝试将第二个UIImageView更新为其他图像（例如输入图像），一切正常。
我认为这是关于我正在创建的UIImage对象的内存管理问题，但我无法弄清楚自己在做什么错
class ViewController: UIViewController {
@IBOutlet weak var out: UIImageView!
override func viewDidLoad() {
super.viewDidLoad()
}
@IBAction func imageTapped(_ sender: UITapGestureRecognizer) {
let imageView = sender.view as? UIImageView
if let imageToAnalyse = imageView?.image {
if let outImg = process(forImage: imageToAnalyse) {
out.image = outImg
}
}
}
func process (forImage inImage:UIImage) -> UIImage? {
let size = CGSize(width: 512, height: 512)
let mlOut = try? MLMultiArray(shape: [1, size.height, size.width] as [NSNumber], dataType: .float32)
let newImage = getSinglePlaneImage(inBuffer: mlOut!, width: Int(size.width), height: Int(size.height))
return newImage
}
func getSinglePlaneImage(inBuffer: MLMultiArray, width: Int, height: Int) -> UIImage
{
var newImage: UIImage
// let floatPtr = inBuffer.dataPointer.bindMemory(to: Float32.self, capacity: inBuffer.count)
// let floatBuffer = UnsafeBufferPointer(start: floatPtr, count: inBuffer.count)
// let pixelValues : [UInt8]? = Array(floatBuffer).map({UInt8( ImageProcessor.clamp( (($0) + 1.0) * 128.0, minValue: 0.0, maxValue: 255.0) ) })
//simulating pixels from MLMultiArray
let pixels : [UInt8]? = Array(repeating: 0, count: 512*512)
var imageRef: CGImage?
if var pixelValues = pixels {
let bitsPerComponent = 8
let bytesPerPixel = 1
let bitsPerPixel = bytesPerPixel * bitsPerComponent
let bytesPerRow = bytesPerPixel * width
let totalBytes = height * bytesPerRow
imageRef = withUnsafePointer(to: &pixelValues, {
ptr -> CGImage? in
var imageRef: CGImage?
let colorSpaceRef = CGColorSpaceCreateDeviceGray()
let bitmapInfo = CGBitmapInfo(rawValue: CGImageAlphaInfo.none.rawValue).union(CGBitmapInfo())
let data = UnsafeRawPointer(ptr.pointee).assumingMemoryBound(to: UInt8.self)
let releaseData: CGDataProviderReleaseDataCallback = {
(info: UnsafeMutableRawPointer?, data: UnsafeRawPointer, size: Int) -> () in
}
if let providerRef = CGDataProvider(dataInfo: nil, data: data, size: totalBytes, releaseData: releaseData) {
imageRef = CGImage(width: width,
height: height,
bitsPerComponent: bitsPerComponent,
bitsPerPixel: bitsPerPixel,
bytesPerRow: bytesPerRow,
space: colorSpaceRef,
bitmapInfo: bitmapInfo,
provider: providerRef,
decode: nil,
shouldInterpolate: false,
intent: CGColorRenderingIntent.defaultIntent)
}
return imageRef
})
}
newImage = UIImage(cgImage: imageRef!)
return newImage
}
}
答案 0 :(得分：1)
似乎您的代码将512x512-float32数组成功转换为512x512-UInt8数组，因此我根据您的代码的未注释版本编写此答案。 （尽管，转换效率不高，还有一些改进的余地。）
以下描述不是解决OP问题的正确解决方案。仅作记录。请跳至此答案底部的更新代码。
首先，代码中最严重的缺陷是以下两行：
imageRef = withUnsafePointer(to: &pixelValues, {
let data = UnsafeRawPointer(ptr.pointee).assumingMemoryBound(to: UInt8.self)
上面的第一行将指针指向[UInt8]?，在Swift中，[UInt8]?（又名Optional<Array<UInt8>>）是一个8字节的结构，而不是像C数组那样的连续区域。 / p>
第二个更危险。 ptr.pointee是[UInt8]?，但不能保证通过指针访问Swift数组。并且将Array传递给UnsafeRawPointer.init(_:)可能会创建一个临时区域，该区域会在调用初始化程序后立即释放。
如您所知，偶尔在有限的条件下访问悬空指针不会造成危害，但随时可能会产生意外结果。
我会这样写：
func getSinglePlaneImage(inBuffer: MLMultiArray, width: Int, height: Int) -> UIImage {
//simulating pixels from MLMultiArray
//...
let pixelValues: [UInt8] = Array(repeating: 0, count: 1*512*512)
let bitsPerComponent = 8
let bytesPerPixel = 1
let bitsPerPixel = bytesPerPixel * 8
let bytesPerRow = bytesPerPixel * width
let totalBytes = height * bytesPerRow
let imageRef = pixelValues.withUnsafeBytes({bytes -> CGImage? in
var imageRef: CGImage?
let colorSpaceRef = CGColorSpaceCreateDeviceGray()
let bitmapInfo: CGBitmapInfo = CGBitmapInfo(rawValue: CGImageAlphaInfo.none.rawValue)
let data = bytes.baseAddress!.assumingMemoryBound(to: UInt8.self)
let releaseData: CGDataProviderReleaseDataCallback = {_,_,_ in }
if let providerRef = CGDataProvider(dataInfo: nil, data: data, size: totalBytes, releaseData: releaseData) {
imageRef = CGImage(width: width,
height: height,
bitsPerComponent: bitsPerComponent,
bitsPerPixel: bitsPerPixel,
bytesPerRow: bytesPerRow,
space: colorSpaceRef,
bitmapInfo: bitmapInfo,
provider: providerRef,
decode: nil,
shouldInterpolate: false,
intent: .defaultIntent)
}
return imageRef
})
let newImage = UIImage(cgImage: imageRef!)
return newImage
}
当您想要一个指向数组起始元素的指针时，请使用withUnsafeBytes并在闭包参数中使用该指针（实际上，它是UnsafeRawBufferPointer）。
再有一个，您的pixels或pixelValues不必是可选的。
否则，您可以为每个像素创建Float32的灰度图像。
func getSinglePlaneImage(inBuffer: MLMultiArray, width: Int, height: Int) -> UIImage {
//simulating pixels from MLMultiArray
//...
let pixelValues: [Float32] = Array(repeating: 0, count: 1*512*512)
let bitsPerComponent = 32 //<-
let bytesPerPixel = 4 //<-
let bitsPerPixel = bytesPerPixel * 8
let bytesPerRow = bytesPerPixel * width
let totalBytes = height * bytesPerRow
let imageRef = pixelValues.withUnsafeBytes({bytes -> CGImage? in
var imageRef: CGImage?
let colorSpaceRef = CGColorSpaceCreateDeviceGray()
let bitmapInfo: CGBitmapInfo = [CGBitmapInfo(rawValue: CGImageAlphaInfo.none.rawValue),
.byteOrder32Little, .floatComponents] //<-
let data = bytes.baseAddress!.assumingMemoryBound(to: Float32.self)
let releaseData: CGDataProviderReleaseDataCallback = {_,_,_ in }
if let providerRef = CGDataProvider(dataInfo: nil, data: data, size: totalBytes, releaseData: releaseData) {
imageRef = CGImage(width: width,
height: height,
bitsPerComponent: bitsPerComponent,
bitsPerPixel: bitsPerPixel,
bytesPerRow: bytesPerRow,
space: colorSpaceRef,
bitmapInfo: bitmapInfo,
provider: providerRef,
decode: nil,
shouldInterpolate: false,
intent: CGColorRenderingIntent.defaultIntent)
}
return imageRef
})
let newImage = UIImage(cgImage: imageRef!)
return newImage
}
两项工作均符合我的测试项目中的预期，但是如果发现有问题，请通知我。
我错过了一个事实，即即使创建了CGDataProvider之后，init(dataInfo:data:size:releaseData:)在用CGImage创建时仍保留指针。因此，withUnsafeBytes的关闭结束后，当无效时，可以引用它。
在这种情况下，您最好使用CGDataProvider.init(data:)。
func getSinglePlaneImage(inBuffer: MLMultiArray, width: Int, height: Int) -> UIImage {
var newImage: UIImage
//let floatPtr = inBuffer.dataPointer.bindMemory(to: Float32.self, capacity: inBuffer.count)
//let floatBuffer = UnsafeBufferPointer(start: floatPtr, count: inBuffer.count)
//let pixelValues: Data = Data((floatBuffer.lazy.map{
// UInt8(ImageProcessor.clamp((($0) + 1.0) * 128.0, minValue: 0.0, maxValue: 255.0))
//})
//simulating pixels from MLMultiArray
//...
let pixelValues = Data(count: 1*512*512) // <- ###
var imageRef: CGImage?
let bitsPerComponent = 8
let bytesPerPixel = 1
let bitsPerPixel = bytesPerPixel * bitsPerComponent
let bytesPerRow = bytesPerPixel * width
let colorSpaceRef = CGColorSpaceCreateDeviceGray()
let bitmapInfo = CGBitmapInfo(rawValue: CGImageAlphaInfo.none.rawValue)
if let providerRef = CGDataProvider(data: pixelValues as CFData) { // <-###
imageRef = CGImage(width: width,
height: height,
bitsPerComponent: bitsPerComponent,
bitsPerPixel: bitsPerPixel,
bytesPerRow: bytesPerRow,
space: colorSpaceRef,
bitmapInfo: bitmapInfo,
provider: providerRef,
decode: nil,
shouldInterpolate: false,
intent: CGColorRenderingIntent.defaultIntent)
}
newImage = UIImage(cgImage: imageRef!)
return newImage
}
据我尝试，即使在实际设备中重复触摸的次数也不会崩溃。请试试。感谢您的耐心等候。