深度学习（六十七）metal forge深度学习库使用

1、设置输入：

let input = Input()

或者

let input = Input(width: 100, height: 100, channels: 3)

2、创建网络：

let output = input
         --> Resize(width: 28, height: 28)
         --> Convolution(kernel: (5, 5), channels: 20, activation: relu, name: "conv1")
         --> Dense(neurons: 10, name: "dense1")
         --> Softmax()

3、链接网络、加载参数

model = Model(input: input, output: output)

let success = model.compile(device: device, inflightBuffers: 3) {
  name, count, type in
  return ParameterLoaderBundle(name: name, count: count,
                               suffix: type == .weights ? "_W" : "_b",
                               ext: "bin")
}

if success {
  print(model.summary())
}

4、预测阶段：

model.encode(commandBuffer: commandBuffer, texture: inputTexture, inflightIndex: i)

let probabilities = model.outputImage(inflightIndex: i).toFloatArray()
let top5 = probabilities.top(k: 5)
let top5Labels = top5.map { x -> (String, Float) in (labels[x.0], x.1) }

巴特西

深度学习（六十七）metal forge深度学习库使用

最新文章

热门文章