nn.Conv2d

PyTorch创建一个2D卷积层

conv2d = nn.Conv2d(
    in_channels=3,
    out_channels=768,
    kernel_size=16,
    stride=16,
    padding=0
)
image_out_of_conv = conv2d(image.unsqueeze(0))

Input (2D image): (224, 224, 3) -> (height, width, color channels)

• 增加批次维度：image.unsqueeze(0)给图像张量添加一个批次维度，将形状从(channels, height, width)转换为(batch_size, channels, height, width)
  • PyTorch中的nn.Conv2d期望的输入格式是(batch_size, channels, height, width)

广播机制

运算操作可能会为新结果分配内存

before = id(Y)
Y = Y + X
id(Y) == before // FALSE


before = id(X)
X += Y
id(X) == before // TRUE