区分stable diffusion中的通道数与张量维度
前言：通道数与张量外形都在数值3和4之间变换，容易混淆。1.通道数：
1.1 channel = 3
RGB 图像具有 3 个通道（红色、绿色和蓝色）。
1.2 channel = 4
Stable Diffusion has 4 latent channels。
怎样理解卷积神经网络中的通道（channel）
2.张量外形
2.1 3D 张量
外形为 (C, H, W)，此中 C 是通道数，H 是高度，W 是宽度。这实用于单个图像。
2.2 4D 张量
2.2.1 通常
外形为 (B, C, H, W)，此中 B 是批次大小，C 是通道数，H 是高度，W 是宽度。这实用于多个图像（比方，批量处理）。
2.2.2 stable diffusion
在img2img中，将image用vae编码并按照timestep加噪：
# This code copyed from diffusers.pipline_controlnet_img2img.py
# 6. Prepare latent variables
latents = self.prepare_latents(
image,
latent_timestep,
batch_size,
num_images_per_prompt,
prompt_embeds.dtype,
device,
generator,
)
image的dim(维度)是3，而latents的dim为4。
让我们先看text2img的prepare_latents函数：
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
if latents is None:
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
else:
latents = latents.to(device)
# scale the initial noise by the standard deviation required by the scheduler
latents = latents * self.scheduler.init_noise_sigma
return latents
显然，shape已经规定了latents的dim(4)和排列次序。
在img2img中：
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.prepare_latents
def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
raise ValueError(
f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
)
image = image.to(device=device, dtype=dtype)
batch_size = batch_size * num_images_per_prompt
if image.shape == 4:
init_latents = image
else:
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
elif isinstance(generator, list):
init_latents = [
self.vae.encode(image).latent_dist.sample(generator) for i in range(batch_size)
]
init_latents = torch.cat(init_latents, dim=0)
else:
init_latents = self.vae.encode(image).latent_dist.sample(generator)
init_latents = self.vae.config.scaling_factor * init_latents
if batch_size > init_latents.shape and batch_size % init_latents.shape == 0:
# expand init_latents for batch_size
deprecation_message = (
f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape} initial"
" images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
" that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
" your script to pass as many initial images as text prompts to suppress this warning."
)
deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
additional_image_per_prompt = batch_size // init_latents.shape
init_latents = torch.cat( * additional_image_per_prompt, dim=0)
elif batch_size > init_latents.shape and batch_size % init_latents.shape != 0:
raise ValueError(
f"Cannot duplicate `image` of batch size {init_latents.shape} to {batch_size} text prompts."
)
else:
init_latents = torch.cat(, dim=0)
shape = init_latents.shape
noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
# get latents
init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
latents = init_latents
return latents
3.应用
3.1 问题
new_map = texture.permute(1, 2, 0)
RuntimeError: permute(sparse_coo): number of dimensions in the tensor input does not match the length of the desired ordering of dimensions i.e. input.dim() = 4 is not equal to len(dims) = 3
该问题是张量外形的问题，跟通道数毫无关系。
3.2 举例
问：4D 张量：外形为 (B, C, H, W)，此中C可以为3吗？
答：4D 张量的外形为 (B,C,H,W)，此中 C 表示通道数。通常环境下，C 可以为 3，这对应于 RGB 图像的三个颜色通道（红色、绿色和蓝色）。
3.3 张量可以理解为多维可变数组
print("sample:", sample.shape)
print("sample:", sample.shape)
print("sample:", sample.shape)
>>
sample: torch.Size()
sample: torch.Size()
sample: torch.Size()
由此可见，可以将张量外形为torch.size()理解为一个4维可变数组。
3.4 将张量化为list
3.4.1
# sample: torch.Size()
views =
print("views:", views.shape)
>>AttributeError: 'list' object has no attribute 'shape'
此时应该：
print("views:", views.shape)
>>views: torch.Size()
3.4.2
# 方法二
for i, view in enumerate(prev_views):
pred_prev_sample = view
3.5 将list化为张量
3.5.1
# 定义一个Python列表
my_list =
# 将Python列表转换为PyTorch张量
my_tensor = torch.tensor(my_list)
print(my_tensor)
>>tensor()
3.5.2
# 假设你有一个包含多个张量的列表
tensor_list = ), torch.tensor(), torch.tensor()]
# 使用torch.stack将它们堆叠成一个新的张量
stacked_tensor = torch.stack(tensor_list)
print(stacked_tensor)
>>tensor([,
,
])
张量运算时对轴参数的设定非经常见，在 Numpy 中一般是参数axis，在 Pytorch 中一般是参数dim，但它们寄义是一样的。
深度学习中的轴/axis/dim全解
# 默认情况下，它在新的维度（即0维）上堆叠这些张量。
# views is a list,and views.shape is ().
views = torch.stack(views, axis=0) # ()
3.5.3 沿着现有维度拼接/在新的维度上增长维度
import torch
# 假设你有一个包含多个张量的列表
tensor_list = ), torch.tensor(), torch.tensor()]
# 使用 torch.cat 将张量沿着现有维度拼接
concatenated_tensor = torch.cat(tensor_list, dim=0)
# 使用 torch.unsqueeze 在新的维度上增加维度
stacked_tensor = torch.unsqueeze(concatenated_tensor, dim=0)
免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！更多信息从访问主页：qidao123.com:ToB企服之家，中国第一个企服评测及商务社交产业平台。
页:
[1]