前言
之前两篇文章介绍了如何利用Tensorflow与Metal来实现图片个性化处理，其中详细描述了神经网络框架，训练方式以及原理：
本篇文章在以上基础上结合苹果在WWDC2017上提出的CoreML
，介绍一种新的实现思路，大大节省iOS端代码并降低集成难度。
效果图
将模型转换到CoreML
苹果官方给了几种三方模型转换成CoreML model format的转换工具，由于我们的训练模型基于Tensorflow，所以我这里用的TensorFlow converter。具体转换工具的使用可以参看苹果的官方文档。
我这里基于工具写了一个转换工程：
from matplotlib import pyplot
from matplotlib.pyplot import imshow
from PIL import Image
import tfcoreml
import numpy as np
import image_utils
import os
import tensorflow as tf
from coremltools.proto import FeatureTypes_pb2 as _FeatureTypes_pb2
import coremltools
tf_model_path = '../protobuf/frozen.pb'
mlmodel = tfcoreml.convert(
tf_model_path = tf_model_path,
mlmodel_path = '../mlmodel/stylize.mlmodel',
output_feature_names = ['transformer/expand/conv3/conv/Sigmoid:0'],
input_name_shape_dict = {'input:0':[1,512,512,3], 'style:0':[32]})
# test model
newstyle = np.zeros([32], dtype=np.float32)
newstyle[0] = 1
newImage = np.expand_dims(image_utils.load_np_image(os.path.expanduser("../sample.jpg")), 0)
newImage = newImage.reshape((512,512,3))
imshow(newImage)
pyplot.show()
coreml_image_input = np.transpose(newImage, (2,0,1))
# coreml_image_input = Image.open("../sample.jpg")
# imshow(coreml_image_input)
# pyplot.show()
coreml_style_index = newstyle[:,np.newaxis,np.newaxis,np.newaxis,np.newaxis]
coreml_input = {'input__0': coreml_image_input, 'style__0': coreml_style_index}
coreml_out = mlmodel.predict(coreml_input, useCPUOnly = True)['transformer__expand__conv3__conv__Sigmoid__0']
coreml_out = np.transpose(coreml_out, (1,2,0))
imshow(coreml_out)
pyplot.show()
iOS实现类Prisma软件(二)
中介绍了网络结构，并在存储图的时候定义了输入（input&style）与输出的节点名字(transformer/expand/conv3/conv/Sigmoid)，所以在调用tfcoreml.convert的时候直接填写了参数，生成并保存了stylize.mlmodel。
iOS中使用CoreML模型
集成mlmodel很简单，只需要引入进工程，Xcode会自动生成模型的接口方法，方便使用的时候调用。
xcode工程
初始化模型只需要：
#import "stylize.h"
@interface HomeViewController ()///<MTKViewDelegate>
{
stylize *styleModel;
}
@implementation HomeViewController
- (void)viewDidLoad {
[super viewDidLoad];
if (@available(iOS 12.0, *)) {
styleModel = [[stylize alloc] init];
} else {
NSLog(@"Need Run iOS 12.0+");
}
}
使用Predict函数，需要按要求构建输入与输出：
- (void)createStyleImage:(UIImage *)source
{
dispatch_async(dispatch_get_global_queue(0, DISPATCH_QUEUE_PRIORITY_DEFAULT), ^{
MLMultiArray *styleArray = [[MLMultiArray alloc] initWithShape:@[@32,@1,@1,@1,@1] dataType:MLMultiArrayDataTypeDouble error:nil];
for (int i = 0; i < styleArray.count; i++) {
[styleArray setObject:@0 atIndexedSubscript:i];
}
[styleArray setObject:@1 atIndexedSubscript:self->currentStyle];
stylizeInput *input = [[stylizeInput alloc] initWithStyle__0:styleArray input__0:[self getImagePixel:source]];
stylizeOutput *output = [styleModel predictionFromFeatures:input error:nil];
dispatch_async(dispatch_get_main_queue(), ^{
self->_styleImageView.image = [self createImage:output.transformer__expand__conv3__conv__Sigmoid__0];
self->isDone = true;
});
});
}
- (MLMultiArray *)getImagePixel:(UIImage *)image
{
int width = image.size.width;
int height = image.size.height;
CGColorSpaceRef colorSpace = CGColorSpaceCreateDeviceRGB();
unsigned char *rawData = (unsigned char*) calloc(height * width * 4, sizeof(unsigned char));
NSUInteger bytesPerPixel = 4;
NSUInteger bytesPerRow = bytesPerPixel * width;
NSUInteger bitsPerComponent = 8;
CGContextRef context = CGBitmapContextCreate(rawData, width, height,
bitsPerComponent, bytesPerRow, colorSpace,
kCGImageAlphaPremultipliedLast | kCGBitmapByteOrder32Big);
CGColorSpaceRelease(colorSpace);
CGContextDrawImage(context, CGRectMake(0, 0, width, height), image.CGImage);
CGContextRotateCTM(context, M_PI_2);
UIImage *ogImg = [UIImage imageWithCGImage:CGBitmapContextCreateImage(context)];
dispatch_async(dispatch_get_main_queue(), ^{
self->_ogImageView.image = ogImg;
});
CGContextRelease(context);
MLMultiArray *tmpArray = [[MLMultiArray alloc] initWithShape:@[[NSNumber numberWithInt: wanted_input_channels],[NSNumber numberWithInt:wanted_input_height],[NSNumber numberWithInt:wanted_input_width]] dataType:MLMultiArrayDataTypeDouble error:nil];
for (int y = 0; y < wanted_input_height; ++y) {
for (int x = 0; x < wanted_input_width; ++x) {
unsigned char *in_pixel =
rawData + (y * wanted_input_width * bytesPerPixel) + (x * bytesPerPixel);
for (int c = 0; c < wanted_input_channels; ++c) {
[tmpArray setObject:[NSNumber numberWithUnsignedChar:in_pixel[c]] atIndexedSubscript:c*wanted_input_height*wanted_input_width+y*wanted_input_width+x];
}
}
}
free(rawData);
return tmpArray;
}
- (UIImage *)createImage:(MLMultiArray *)pixels
API_AVAILABLE(ios(11.0)){
unsigned char *rawData = (unsigned char*) calloc(wanted_input_height * wanted_input_width * 4, sizeof(unsigned char));
for (int y = 0; y < wanted_input_height; ++y) {
unsigned char *out_row = rawData + (y * wanted_input_width * 4);
for (int x = 0; x < wanted_input_width; ++x) {
int index = x * wanted_input_width + y;
unsigned char *out_pixel = out_row + (x * 4);
for (int c = 0; c < wanted_input_channels; ++c) {
out_pixel[c] = [[pixels objectAtIndexedSubscript:c*wanted_input_height*wanted_input_width+index] floatValue] * 255;
}
out_pixel[3] = UINT8_MAX;
}
}
CGColorSpaceRef colorSpace = CGColorSpaceCreateDeviceRGB();
NSUInteger bytesPerPixel = 4;
NSUInteger bytesPerRow = bytesPerPixel * wanted_input_width;
NSUInteger bitsPerComponent = 8;
CGContextRef context = CGBitmapContextCreate(rawData, wanted_input_width, wanted_input_height,
bitsPerComponent, bytesPerRow, colorSpace,
kCGImageAlphaPremultipliedLast | kCGBitmapByteOrder32Big);
CGColorSpaceRelease(colorSpace);
UIImage *retImg = [UIImage imageWithCGImage:CGBitmapContextCreateImage(context)];
CGContextRelease(context);
free(rawData);
return [UIImage imageWithCGImage:retImg.CGImage scale:1 orientation:UIImageOrientationLeftMirrored];
}
由于我们在转换CoreML模型的时候没有设置图片输入，所以所有数据都是以MLMultiArray格式处理。如果觉得这样比较麻烦，可以通过一下方法转换模型：
mlmodel = tfcoreml.convert(
tf_model_path = tf_model_path,
mlmodel_path = '../mlmodel/stylize.mlmodel',
output_feature_names = ['transformer/expand/conv3/conv/Sigmoid:0'],
input_name_shape_dict = {'input:0':[1,512,512,3], 'style:0':[32]},
image_input_names=['input:0'])
#spec = mlmodel.get_spec()
#output = spec.description.output[0]
#output.type.imageType.colorSpace = #_FeatureTypes_pb2.ImageFeatureType.ColorSpace.Value('RGB')
#output.type.imageType.width = 512
#output.type.imageType.height = 512
#coremltools.models.utils.save_spec(spec, '../mlmodel/stylize.mlmodel')
其中，在调用
tfcoreml.convert时就配置了输入图片的节点image_input_names=['input:0']，输出节点也可以转换成图片，但是我们保存的模型不适用，如果官方的pb文件是可以的。
使用的图片输入后，我们在配置input的时候就可以使用CVPixelBufferRef来封装，不用自己考虑图片转换成bytes后的矩阵转置。
运行效果
直接上图🔥：
效果图
效果图
网友评论