import torch
import torch.nn as nn
import cv2
import numpy as np

# 檢查 GPU 是否可用
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# 讀取圖像并轉(zhuǎn)換為 PyTorch 張量
image = cv2.imread("input.jpg")  # 讀取 BGR 格式圖像
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)  # 轉(zhuǎn)為 RGB
image_tensor = torch.from_numpy(image).float().permute(2, 0, 1)  # HWC -> CHW
image_tensor = image_tensor.unsqueeze(0).to(device)  # 添加 batch 維度并移至 GPU

# 定義邊緣檢測(cè)卷積核（Sobel算子）
conv_layer = nn.Conv2d(
    in_channels=3,
    out_channels=3,
    kernel_size=3,
    bias=False,
    padding=1
).to(device)

# 手動(dòng)設(shè)置 Sobel 核權(quán)重（示例，僅作用于水平邊緣）
sobel_kernel = torch.tensor([
    [[[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]]],  # Red 通道
    [[[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]]],  # Green 通道
    [[[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]]],  # Blue 通道
], dtype=torch.float32).repeat(3, 1, 1, 1).to(device)

conv_layer.weight.data = sobel_kernel

# 執(zhí)行卷積操作（GPU加速）
with torch.no_grad():
    output_tensor = conv_layer(image_tensor)

# 將結(jié)果轉(zhuǎn)換回 numpy 并保存
output = output_tensor.squeeze(0).permute(1, 2, 0).cpu().numpy()
output = np.clip(output, 0, 255).astype(np.uint8)
cv2.imwrite("edge_detection_gpu.jpg", cv2.cvtColor(output, cv2.COLOR_RGB2BGR))

import cv2
import time

# 檢查 OpenCV 是否支持 CUDA
print("CUDA devices:", cv2.cuda.getCudaEnabledDeviceCount())

# 讀取圖像并上傳到 GPU
image = cv2.imread("input.jpg")
gpu_image = cv2.cuda_GpuMat()
gpu_image.upload(image)

# 創(chuàng)建 GPU 加速的高斯濾波器
gaussian_filter = cv2.cuda.createGaussianFilter(
    cv2.CV_8UC3,  # 輸入類(lèi)型 (8-bit unsigned, 3 channels)
    cv2.CV_8UC3,  # 輸出類(lèi)型
    (15, 15),      # 核大小
    0              # Sigma（自動(dòng)計(jì)算）
)

# 執(zhí)行濾波（重復(fù)多次測(cè)試速度）
start_time = time.time()
for _ in range(100):  # 重復(fù) 100 次模擬大數(shù)據(jù)量
    gpu_blur = gaussian_filter.apply(gpu_image)
end_time = time.time()

# 下載結(jié)果到 CPU 并保存
result = gpu_blur.download()
print(f"GPU Time: {end_time - start_time:.4f} seconds")
cv2.imwrite("blur_gpu.jpg", result)

import cupy as cp
import cv2
import numpy as np
import time

# 讀取圖像并轉(zhuǎn)為灰度
image = cv2.imread("input.jpg", cv2.IMREAD_GRAYSCALE)

# 將 numpy 數(shù)組轉(zhuǎn)為 CuPy 數(shù)組（上傳到 GPU）
image_gpu = cp.asarray(image)

# 快速傅里葉變換（FFT）和逆變換（IFFT）
start_time = time.time()
fft_gpu = cp.fft.fft2(image_gpu)
fft_shift = cp.fft.fftshift(fft_gpu)
magnitude_spectrum = cp.log(cp.abs(fft_shift))
end_time = time.time()

# 將結(jié)果轉(zhuǎn)回 CPU
magnitude_cpu = cp.asnumpy(magnitude_spectrum)
print(f"GPU FFT Time: {end_time - start_time:.4f} seconds")

# 歸一化并保存頻譜圖
magnitude_cpu = cv2.normalize(magnitude_cpu, None, 0, 255, cv2.NORM_MINMAX)
cv2.imwrite("fft_spectrum_gpu.jpg", magnitude_cpu.astype(np.uint8))

from numba import cuda
import numpy as np
import cv2
import time

# 讀取圖像
image = cv2.imread("input.jpg")
height, width, channels = image.shape

# 定義 GPU 核函數(shù)
@cuda.jit
def invert_colors_kernel(image):
    x, y = cuda.grid(2)
    if x < image.shape[0] and y < image.shape[1]:
        for c in range(3):  # 遍歷 RGB 通道
            image[x, y, c] = 255 - image[x, y, c]

# 將圖像上傳到 GPU
image_gpu = cuda.to_device(image)

# 配置線程和塊
threads_per_block = (16, 16)
blocks_per_grid_x = (height + threads_per_block[0] - 1) // threads_per_block[0]
blocks_per_grid_y = (width + threads_per_block[1] - 1) // threads_per_block[1]
blocks_per_grid = (blocks_per_grid_x, blocks_per_grid_y)

# 執(zhí)行核函數(shù)
start_time = time.time()
invert_colors_kernel[blocks_per_grid, threads_per_block](image_gpu)
cuda.synchronize()  # 等待 GPU 完成
end_time = time.time()

# 下載結(jié)果并保存
image_cpu = image_gpu.copy_to_host()
print(f"GPU Invert Time: {end_time - start_time:.6f} seconds")
cv2.imwrite("inverted_gpu.jpg", image_cpu)

import torch
import torchvision.models as models
from torchvision import transforms
from PIL import Image

# 加載預(yù)訓(xùn)練模型到 GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = models.vgg19(pretrained=True).features.to(device).eval()

# 圖像預(yù)處理
preprocess = transforms.Compose([
    transforms.Resize(512),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

# 加載內(nèi)容圖像和風(fēng)格圖像
content_image = Image.open("content.jpg")
style_image = Image.open("style.jpg")

# 將圖像轉(zhuǎn)為張量并移至 GPU
content_tensor = preprocess(content_image).unsqueeze(0).to(device)
style_tensor = preprocess(style_image).unsqueeze(0).to(device)

# 定義風(fēng)格遷移函數(shù)（示例，需完整實(shí)現(xiàn)損失計(jì)算和優(yōu)化）
def style_transfer(model, content_input, style_input, iterations=500):
    # 創(chuàng)建可優(yōu)化圖像
    input_image = content_input.clone().requires_grad_(True)
    
    # 定義優(yōu)化器
    optimizer = torch.optim.LBFGS([input_image])
    
    # 風(fēng)格遷移循環(huán)
    for i in range(iterations):
        def closure():
            optimizer.zero_grad()
            # 提取特征并計(jì)算損失（需實(shí)現(xiàn)具體細(xì)節(jié)）
            # ...
            return total_loss
        
        optimizer.step(closure)
    
    return input_image

# 執(zhí)行風(fēng)格遷移（需補(bǔ)充完整代碼）
output_image = style_transfer(model, content_tensor, style_tensor)

# 后處理并保存結(jié)果
output_image = output_image.squeeze().cpu().detach()
output_image = transforms.ToPILImage()(output_image)
output_image.save("style_transfer_gpu.jpg")

pip install torch torchvision opencv-python-headless cupy numba