DeepLabV3.py

import numpy as np
import cv2
import migraphx
import argparse
import os

def Preprocessing(pil_img, newW, newH):
    assert newW > 0 and newH > 0, 'Scale is too small' 
    img_nd = cv2.cvtColor(pil_img, cv2.COLOR_BGR2RGB)    # BGR转换为RGB
    img_nd = cv2.resize(img_nd, (newW, newH))            # 将图像尺寸修改为newW x newH
    
    if len(img_nd.shape) == 2:                           # 获取图像的维度信息
        img_nd = np.expand_dims(img_nd, axis=2)          # 如果是2维的 扩充为3维  

    img_trans = img_nd.transpose((2, 0, 1))              # HWC转换为CHW     
    img_trans = np.expand_dims(img_trans, 0)             # CHW扩展为NCHW
    img_trans = np.ascontiguousarray(img_trans)          # 保证内存连续存储
    img_trans = img_trans.astype(np.float32)             # 转换成浮点型数据
    if img_trans.max() > 1:                             
        img = img_trans / 255.0                          # 保证数据处于0-1之间的浮点数

    return img

def AllocateOutputMemory(model):
    outputData = {}
    for key in model.get_outputs().keys():
        outputData[key] = migraphx.allocate_gpu(s=model.get_outputs()[key])

    return outputData    

# 对通道维度执行Softmax
def softmax(arr):

    exp_vals = np.exp(arr - np.max(arr, axis=1, keepdims=True))
    sum_exp = np.sum(exp_vals, axis=1, keepdims=True)
    return exp_vals / sum_exp  

if __name__ == '__main__':

    parser = argparse.ArgumentParser()
    parser.add_argument("loadMode", type=int,help="0:DeepLabV3 Single Image Sample.\t 1：DeepLabV3 Multiple Image Sample.")
    parser.add_argument("--enable_offload_copy", action="store_true")
    precision_group = parser.add_mutually_exclusive_group()
    precision_group.add_argument("--int8",action="store_true")
    precision_group.add_argument("--fp16",action="store_true")

    args = parser.parse_args()
    loadMode = args.loadMode
    useInt8 = args.int8
    useFP16 = args.fp16
    offloadCopy = args.enable_offload_copy

    #加载图片方式
    if loadMode == 0:
        maxInput={"input":[1,3,513,513]}
        img = cv2.imread("../Resource/Images/000001.jpg")
        input_img = Preprocessing(img, 513, 513)
    else:
        maxInput={"input":[3,3,513,513]}
        folder_path = "../Resource/Images/"
        image_extensions = ('.jpg')
        image_list = []
        for filename in os.listdir(folder_path):    
            # 检查文件是否为图片
            if filename.lower().endswith(image_extensions):
                file_path = os.path.join(folder_path, filename)
                img = cv2.imread(file_path)
                image_list.append(Preprocessing(img, 513, 513))
        input_img = np.concatenate(image_list,axis=0)

    # 加载模型
    model = migraphx.parse_onnx("../Resource/Models/deeplabv3_resnet101.onnx", map_input_dims=maxInput)

    # 获取模型输入/输出节点信息
    inputs = model.get_inputs()
    outputs = model.get_outputs()
    inputName = model.get_parameter_names()[0]
    inputShape = inputs[inputName].lens()

    #量化
    if useInt8:
        dic = dict()
        calibrate_folder_path = "../Resource/Images/"
        calibrate_image_extensions = ('.jpg')
        calibrate_image_list = []
        for filename in os.listdir(calibrate_folder_path):    
            # 检查文件是否为图片
            if filename.lower().endswith(calibrate_image_extensions):
                file_path = os.path.join(calibrate_folder_path, filename)
                img = cv2.imread(file_path)
                calibrate_image_list.append(Preprocessing(img, 513, 513))
        calibrate_img = np.concatenate(calibrate_image_list,axis=0)
        dic[inputName] = migraphx.argument(calibrate_img)
        calibration = [dic]
        migraphx.quantize_int8(model, migraphx.get_target("gpu"), calibration)
    if useFP16: 
        migraphx.quantize_fp16(model)
    
    if offloadCopy :
        # 编译模型
        model.compile(migraphx.get_target("gpu"), device_id=0)      # device_id: 设置GPU设备，默认为0号设备
        # 模型推理
        mask = model.run({'input':input_img})      
        result = mask[0]                                            # 得到第一个输出节点的结果    
    else:
        # 编译模型
        model.compile(migraphx.get_target("gpu"),offload_copy=False, device_id=0)      # device_id: 设置GPU设备，默认为0号设备
        modelData = AllocateOutputMemory(model)                                          # 为输出节点分配device内存，用于保存输出数据
        modelData[inputName] = migraphx.to_gpu(migraphx.argument(input_img))
        # 推理
        mask = model.run(modelData)
        result = migraphx.from_gpu(mask[0])                         # 获取第1个输出节点的数据,migraphx.argument类型

    # 对通道维度进行softmax
    softmax_result = softmax(result)                            
    # 计算通道维度最大值对应的索引（即类别索引）      
    max_indices = np.argmax(softmax_result, axis=1)      
    # 使用预设颜色
    color_map = np.array([
        [0, 0, 0], [255, 0, 0], [0, 255, 0], [0, 0, 255],       # 0-3类
        [255, 255, 0], [255, 0, 255], [0, 255, 255], [128, 0, 0],  # 4-7类
        [0, 128, 0], [0, 0, 128], [128, 128, 0], [128, 0, 128],  # 8-11类
        [0, 128, 128], [192, 192, 192], [128, 128, 128], [64, 0, 0],  # 12-15类
        [0, 64, 0], [0, 0, 64], [64, 64, 0], [64, 0, 64],       # 16-19类
        [0, 64, 64]                                             # 20类
    ], dtype=np.uint8)
    for i in range(max_indices.shape[0]):
        flat_index = max_indices[i]
        rgb_image = color_map[flat_index]                          # # 将二维的类别索引图直接转换为三维的 RGB 彩色图像
        fileName  = "Result_"+str(i+1)+".jpg"
        cv2.imwrite(fileName, rgb_image)                            # 保存图像分割结果

    print("Segmentation results have been saved to Python directory")