mnnsr.py

import gradio as gr
import os
import MNN
import numpy as np
import cv2
from PIL import Image
import time


# 复制原始modelTest函数中的必要函数
def process_image(data, H, W, C, color='BGR'):
    """
    处理图像数据（灰度或彩色）并转换为指定色彩空间
    
    参数:
        data: 输入数据指针 (numpy数组形式传入)
        H: 高度
        W: 宽度
        C: 通道数 (1 或 3)
        color: 目标色彩空间 ('BGR', 'RGB', 'YCbCr', 'YUV')
    
    返回:
        numpy数组 处理后的图像
    """
    if C == 1:
        # 灰度图像处理
        gray = np.array(data, dtype=np.float32).reshape(H, W)
        gray = (gray * 255).astype(np.uint8)
        result = cv2.cvtColor(gray, cv2.COLOR_GRAY2BGR)
        return result
    else:
        # 彩色图像处理
        # 将数据拆分为各个通道
        channels = [
            np.array(data[i*H*W : (i+1)*H*W], dtype=np.float32).reshape(H, W)
            for i in range(C)
        ]
        if color == 'RGB':
            # RGB -> BGR (OpenCV默认顺序)
            channels[0], channels[2] = channels[2], channels[0]  # 交换R和B通道
            result = cv2.merge(channels)
            result = (result * 255).astype(np.uint8)
        else:
            # 先合并为BGR格式
            rgb = cv2.merge(channels)
            rgb = (rgb * 255).astype(np.uint8)
            # 转换为目标色彩空间
            if color == 'YCbCr':
                result = cv2.cvtColor(rgb, cv2.COLOR_BGR2YCrCb)
            elif color == 'YUV':
                result = cv2.cvtColor(rgb, cv2.COLOR_BGR2YUV)
            else:
                result = rgb  # 默认为BGR
            
        return result

def createTensor(tensor):
    shape = tensor.getShape()
    data = np.ones(shape, dtype=np.float32)
    return MNN.Tensor(shape, tensor.getDataType(), data, tensor.getDimensionType())

def modelTest_for_gradio(modelPath, image_path, tilesize = 0, backend = 3):
    if tilesize<=0:
        tilesize = 128
    model_name = os.path.basename(modelPath)
    if "-Grayscale" in model_name:
        model_channel = 1
    elif "-4ch" in model_name or "RGBA" in model_name:
        model_channel = 4
    else:
        model_channel = 3

    # 记录模型加载开始时间
    load_start_time = time.time()
    # 加载模型（计时范围内）
    net = MNN.Interpreter(modelPath)
    # set 9 for Session_Backend_Auto, Let BackGround Tuning
    net.setSessionMode(9)
    # set 0 for tune_num
    # net.setSessionHint(0, 20)
    config = {}
    # "CPU"或0(默认), "OPENCL"或3,"OPENGL"或6, "VULKAN"或7, "METAL"或1, "TRT"或9, "CUDA"或2, "HIAI"或8
    config['backend'] = backend
    #config['precision'] = "low"
    session = net.createSession(config)
    print("Run on backendtype: %d \n" % net.getSessionInfo(session, 2))
    inputTensor = net.getSessionInput(session)
    net.resizeTensor(inputTensor, (1, model_channel, tilesize, tilesize))
    net.resizeSession(session)
    # 计算模型加载耗时
    load_time = time.time() - load_start_time
    print(f"Load mnn model: {load_time:.4f} sec")
    
    # 读取图像
    image = cv2.imread(image_path)
    if image.ndim == 2:
        # 为了方便处理，先将其扩展为3维数组 (height, width, 1)
        # print("extend dims, image.shape=", image.shape)
        image_channel = 1
        # image = np.expand_dims(image, axis=-1)
    else:
        image_channel = image.shape[2]
    
    image = cv2.resize(image, (tilesize, tilesize))

    # 记录推理开始时间
    infer_start_time = time.time()

    # 处理通道数不匹配的情况
    if image_channel == 3:
        if model_channel == 1:
            image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
        elif model_channel == 3:
            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        elif model_channel == 4:    
            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGBA)
        else:
            print(f"unexpect input: model_channel {model_channel}, image_channel {image_channel}")
    elif image_channel == 4:
        if model_channel == 1:
            image = cv2.cvtColor(image, cv2.COLOR_BGRA2GRAY)
        elif model_channel == 3:
            image = cv2.cvtColor(image, cv2.COLOR_BGRA2RGB)
        elif model_channel == 4:
            image = cv2.cvtColor(image, cv2.COLOR_BGRA2RGBA)
        else:
            print(f"unexpect input: model_channel {model_channel}, image_channel {image_channel}")
    else:
        print(f"unexpect input: model_channel {model_channel}, image_channel {image_channel}")

    # 显示图像(在Gradio中不需要)
    # display(Image(data=cv2.imencode('.jpg', image)[1].tobytes()))
    # print("image.shape=", image.shape)
    image = image/255.0
    if model_channel>=3:
        image = image.transpose((2, 0, 1))
    image = image.astype(np.float32)

    tmp_input = MNN.Tensor((1, model_channel, tilesize, tilesize), MNN.Halide_Type_Float, image, MNN.Tensor_DimensionType_Caffe)
    inputTensor.copyFrom(tmp_input)   

    # 执行推理
    net.runSession(session)
    
    outputTensor = net.getSessionOutput(session)
    outputShape = outputTensor.getShape()
    print("outputShape",outputShape)
    outputHost = createTensor(outputTensor)
    outputTensor.copyToHostTensor(outputHost)
    
    outimage = process_image(outputHost.getData(), outputShape[2], outputShape[3], outputShape[1], color='RGB')
    
    # 计算推理耗时
    infer_time = time.time() - infer_start_time
    print(f"Infer latency time: {infer_time:.4f} sec")
    # 返回图像、加载时间、推理时间
    return outimage, load_time, infer_time

# def gradio_interface(modelPath, input_image):
#     processed_image_np = modelTest_for_gradio(modelPath, input_image)
    
#     processed_image_pil = Image.fromarray(cv2.cvtColor(processed_image_np, cv2.COLOR_BGR2RGB))
    
#     return processed_image_pil