【一日速成编程系列】用Pytorch操作图像

dataset = '/content/gdrive/MyDrive/'

# Import libraries we will need
import torch
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image

# We will use PIL to read an image
im_pil = Image.open(dataset + 'scene.jpg')

# We first convert the image to a numpy array
im_np = np.array(im_pil)

# Numpy is the standard matrix library in Python. 
# But PyTorch effectively replaces it together with the functionality needed for deep learning.
# Still, we will enounter Numpy arrays in intermediate stages.
# Similar to PyTorch, we can check the shape of the tensor.
im_np.shape

# Now lets display the image. We can do this with matplotlib
plt.imshow(im_np)   # Plot image
plt.axis('on')     # Just turns on the axis
plt.show()          # Finally show it

# We can easily convert the image from numpy to PyTorch
im_torch = torch.from_numpy(im_np)

# Check shape
print(im_torch.shape)

# By default, PyTorch uses the data order C x H x W for images.
# So we should move the RGB channel dimension to the first dimension
im_torch = im_torch.permute(2, 0, 1)

print(im_torch.shape)

# Lets check some details about the image
print('Type:', im_torch.dtype)
print('Min:', im_torch.min().item())
print('Max:', im_torch.max().item())

# Note that the image is a Byte tensor (uint8)
# For most image operations it is better to first convert it to floating point with values between 0 and 1

im = im_torch.float() / 255

print()
print('Type:', im.dtype)
print('Min:', im.min().item())
print('Max:', im.max().item())

# Read an image with the given name and convert it to torch
def imread(image_file):
    im_pil = Image.open(dataset + image_file)
    im_np = np.array(im_pil, copy=False)
    im_torch = torch.from_numpy(im_np).permute(2, 0, 1)
    return im_torch.float()/255

# Show a PyTorch image tensor
def imshow(im, normalize=False):
    # Fit the image to the [0, 1] range if normalize is True
    if normalize:
        im = (im - im.min()) / (im.max() - im.min())

    # Remove redundant dimensions 
    im = im.squeeze()    # Mini excersize: check in the documentation what this function does

    is_color = (im.dim() == 3)

    # If there is a color channel dimension, move it to the end
    if is_color:
        im = im.permute(1, 2, 0)

    im_np = im.numpy().clip(0,1)    # Convert to numpy and ensure the values in the range [0, 1]
    if is_color:
        plt.imshow(im_np)
    else:
        plt.imshow(im_np, cmap='gray')
    plt.axis('off')
    plt.show()

{# 读取给定名称的图片，并将其转换为PyTorch张量  
def imread(image_file):  
    # 使用PIL库打开数据集路径与给定文件名拼接后的图片文件  
    im_pil = Image.open(dataset + image_file)  
    # 将PIL图片对象转换为NumPy数组，设置copy=False以避免数据复制，提高效率  
    im_np = np.array(im_pil, copy=False)  
    # 将NumPy数组转换为PyTorch张量，并调整维度顺序以匹配PyTorch的CHW（通道、高度、宽度）格式  
    im_torch = torch.from_numpy(im_np).permute(2, 0, 1)  
    # 将张量数据类型转换为浮点型，并除以255将其值归一化到[0, 1]区间  
    return im_torch.float()/255  
  
# 显示一个PyTorch图像张量  
def imshow(im, normalize=False):  
    # 如果normalize为True，则将图像张量的值缩放到[0, 1]区间  
    if normalize:  
        im = (im - im.min()) / (im.max() - im.min())  
  
    # 移除张量中多余的维度（即维度大小为1的维度）  
    im = im.squeeze()   
  
    # 判断图像是否为彩色图  
    is_color = (im.dim() == 3)  
  
    # 如果图像是彩色的，将颜色通道维度移动到末尾，以匹配matplotlib的HWC（高度、宽度、通道）格式  
    if is_color:  
        im = im.permute(1, 2, 0)  
  
    # 将PyTorch张量转换为NumPy数组，并确保其值在[0, 1]区间内  
    im_np = im.numpy().clip(0,1)  
    # 根据图像是否为彩色图，使用不同的imshow参数显示图像  
    if is_color:  
        plt.imshow(im_np)  
    else:  
        plt.imshow(im_np, cmap='gray')  # 灰度图使用灰度颜色映射  
    # 关闭坐标轴显示  
    plt.axis('off')  
    # 显示图像  
    plt.show()  
}

import torch

# 假设的原始张量
im = torch.randn(1, 1, 28, 28)  # 形状为 (1, 1, 28, 28)

# 移除大小为1的维度
im_squeezed = im.squeeze()

# 检查结果
print(im_squeezed.shape)  # 输出应为 torch.Size([28, 28])

import torch
import numpy as np
import matplotlib.pyplot as plt

# 假设的PyTorch张量，其值可能不在[0,1]范围内
im = torch.tensor([[-0.5, 1.5, 0.0], [2.0, -0.2, 0.8]])

# 将PyTorch张量转换为NumPy数组，并使用clip方法限制值在[0,1]范围内
im_np = im.numpy().clip(0,1)

# 现在，im_np是一个NumPy数组，其所有值都在[0,1]范围内
print(im_np)
# 输出可能是：
# [[0.   1.   0. ]
#  [1.   0.   0.8]]

# 使用matplotlib显示图像（注意：这里我们假设im_np是灰度图像的一个通道）
plt.imshow(im_np, cmap='gray')
plt.axis('off')
plt.show()

# Lets try these functions
im = imread('scene.jpg')
print(im)
imshow(im)

# Also check the type
print('Type:', im.dtype)

im = imread('scene.jpg')

# Lets start with a grayscale image.
# Color image can easily be converted to grayscale by simply averaging the color channels.

im_gray = im.mean(dim=0)

imshow(im_gray)

import torch  
from torchvision import transforms  
from PIL import Image  
import matplotlib.pyplot as plt  
  
# 读取图像  
image_pil = Image.open('scene.jpg')  
transform = transforms.ToTensor()  
im = transform(image_pil)  
  
# 转换为灰度图像（注意：这里我们直接使用了torchvision的Grayscale转换）  
transform_gray = transforms.Grayscale(num_output_channels=1)  
im_gray = transform_gray(image_pil)  
im_gray = transform(im_gray)  # 转换为张量  
  
# 显示灰度图像（需要先转换为NumPy数组）  
plt.imshow(im_gray.squeeze().numpy(), cmap='gray')  
plt.axis('off')  
plt.show()