AIGC笔记--Stable Diffusion源码剖析之UNetModel

1--前言

        以论文《High-Resolution Image Synthesis with Latent Diffusion Models》  开源的项目为例，剖析Stable Diffusion经典组成部分，巩固学习加深印象。

2--UNetModel

一个可以debug的小demo：SD_UNet​​​​​​​

        以文生图为例，剖析UNetModel核心组成模块。

2-1--Forward总揽

提供的文生图Demo中，实际传入的参数只有x、timesteps和context三个，其中：

        x 表示随机初始化的噪声Tensor（shape: [B*2, 4, 64, 64]，*2表示使用Classifier-Free Diffusion Guidance）。

        timesteps 表示去噪过程中每一轮传入的timestep（shape: [B*2]）。

        context表示经过CLIP编码后对应的文本Prompt（shape: [B*2, 77, 768]）。

    def forward(self, x, timesteps=None, context=None, y=None,**kwargs):"""Apply the model to an input batch.:param x: an [N x C x ...] Tensor of inputs.:param timesteps: a 1-D batch of timesteps.:param context: conditioning plugged in via crossattn:param y: an [N] Tensor of labels, if class-conditional.:return: an [N x C x ...] Tensor of outputs."""assert (y is not None) == (self.num_classes is not None), "must specify y if and only if the model is class-conditional"hs = []t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) # Create sinusoidal timestep embeddings.emb = self.time_embed(t_emb) # MLPif self.num_classes is not None:assert y.shape == (x.shape[0],)emb = emb + self.label_emb(y)h = x.type(self.dtype)for module in self.input_blocks:h = module(h, emb, context)hs.append(h)h = self.middle_block(h, emb, context)for module in self.output_blocks:h = th.cat([h, hs.pop()], dim=1)h = module(h, emb, context)h = h.type(x.dtype)if self.predict_codebook_ids:return self.id_predictor(h)else:return self.out(h)

2-2--timestep embedding生成

        使用函数 timestep_embedding() 和 self.time_embed() 对传入的timestep进行位置编码，生成sinusoidal timestep embeddings。

        其中 timestep_embedding() 函数定义如下，而self.time_embed()是一个MLP函数。

def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):"""Create sinusoidal timestep embeddings.:param timesteps: a 1-D Tensor of N indices, one per batch element.These may be fractional.:param dim: the dimension of the output.:param max_period: controls the minimum frequency of the embeddings.:return: an [N x dim] Tensor of positional embeddings."""if not repeat_only:half = dim // 2freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(device=timesteps.device)args = timesteps[:, None].float() * freqs[None]embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)if dim % 2:embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)else:embedding = repeat(timesteps, 'b -> b d', d=dim)return embedding
self.time_embed = nn.Sequential(linear(model_channels, time_embed_dim),nn.SiLU(),linear(time_embed_dim, time_embed_dim),
)

2-3--self.input_blocks下采样

        在 Forward() 中，使用 self.input_blocks 将输入噪声进行分辨率下采样，经过下采样具体维度变化为：[B*2, 4, 64, 64] > [B*2, 1280, 8, 8]；

        下采样模块共有12个 module，其组成如下：

ModuleList((0): TimestepEmbedSequential((0): Conv2d(4, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(1-2): 2 x TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 320, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=320, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 320, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Identity())(1): SpatialTransformer((norm): GroupNorm(32, 320, eps=1e-06, affine=True)(proj_in): Conv2d(320, 320, kernel_size=(1, 1), stride=(1, 1))(transformer_blocks): ModuleList((0): BasicTransformerBlock((attn1): CrossAttention((to_q): Linear(in_features=320, out_features=320, bias=False)(to_k): Linear(in_features=320, out_features=320, bias=False)(to_v): Linear(in_features=320, out_features=320, bias=False)(to_out): Sequential((0): Linear(in_features=320, out_features=320, bias=True)(1): Dropout(p=0.0, inplace=False)))(ff): FeedForward((net): Sequential((0): GEGLU((proj): Linear(in_features=320, out_features=2560, bias=True))(1): Dropout(p=0.0, inplace=False)(2): Linear(in_features=1280, out_features=320, bias=True)))(attn2): CrossAttention((to_q): Linear(in_features=320, out_features=320, bias=False)(to_k): Linear(in_features=768, out_features=320, bias=False)(to_v): Linear(in_features=768, out_features=320, bias=False)(to_out): Sequential((0): Linear(in_features=320, out_features=320, bias=True)(1): Dropout(p=0.0, inplace=False)))(norm1): LayerNorm((320,), eps=1e-05, elementwise_affine=True)(norm2): LayerNorm((320,), eps=1e-05, elementwise_affine=True)(norm3): LayerNorm((320,), eps=1e-05, elementwise_affine=True)))(proj_out): Conv2d(320, 320, kernel_size=(1, 1), stride=(1, 1))))(3): TimestepEmbedSequential((0): Downsample((op): Conv2d(320, 320, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))))(4): TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 320, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(320, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=640, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 640, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Conv2d(320, 640, kernel_size=(1, 1), stride=(1, 1)))(1): SpatialTransformer((norm): GroupNorm(32, 640, eps=1e-06, affine=True)(proj_in): Conv2d(640, 640, kernel_size=(1, 1), stride=(1, 1))(transformer_blocks): ModuleList((0): BasicTransformerBlock((attn1): CrossAttention((to_q): Linear(in_features=640, out_features=640, bias=False)(to_k): Linear(in_features=640, out_features=640, bias=False)(to_v): Linear(in_features=640, out_features=640, bias=False)(to_out): Sequential((0): Linear(in_features=640, out_features=640, bias=True)(1): Dropout(p=0.0, inplace=False)))(ff): FeedForward((net): Sequential((0): GEGLU((proj): Linear(in_features=640, out_features=5120, bias=True))(1): Dropout(p=0.0, inplace=False)(2): Linear(in_features=2560, out_features=640, bias=True)))(attn2): CrossAttention((to_q): Linear(in_features=640, out_features=640, bias=False)(to_k): Linear(in_features=768, out_features=640, bias=False)(to_v): Linear(in_features=768, out_features=640, bias=False)(to_out): Sequential((0): Linear(in_features=640, out_features=640, bias=True)(1): Dropout(p=0.0, inplace=False)))(norm1): LayerNorm((640,), eps=1e-05, elementwise_affine=True)(norm2): LayerNorm((640,), eps=1e-05, elementwise_affine=True)(norm3): LayerNorm((640,), eps=1e-05, elementwise_affine=True)))(proj_out): Conv2d(640, 640, kernel_size=(1, 1), stride=(1, 1))))(5): TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 640, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=640, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 640, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(640, 640, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Identity())(1): SpatialTransformer((norm): GroupNorm(32, 640, eps=1e-06, affine=True)(proj_in): Conv2d(640, 640, kernel_size=(1, 1), stride=(1, 1))(transformer_blocks): ModuleList((0): BasicTransformerBlock((attn1): CrossAttention((to_q): Linear(in_features=640, out_features=640, bias=False)(to_k): Linear(in_features=640, out_features=640, bias=False)(to_v): Linear(in_features=640, out_features=640, bias=False)(to_out): Sequential((0): Linear(in_features=640, out_features=640, bias=True)(1): Dropout(p=0.0, inplace=False)))(ff): FeedForward((net): Sequential((0): GEGLU((proj): Linear(in_features=640, out_features=5120, bias=True))(1): Dropout(p=0.0, inplace=False)(2): Linear(in_features=2560, out_features=640, bias=True)))(attn2): CrossAttention((to_q): Linear(in_features=640, out_features=640, bias=False)(to_k): Linear(in_features=768, out_features=640, bias=False)(to_v): Linear(in_features=768, out_features=640, bias=False)(to_out): Sequential((0): Linear(in_features=640, out_features=640, bias=True)(1): Dropout(p=0.0, inplace=False)))(norm1): LayerNorm((640,), eps=1e-05, elementwise_affine=True)(norm2): LayerNorm((640,), eps=1e-05, elementwise_affine=True)(norm3): LayerNorm((640,), eps=1e-05, elementwise_affine=True)))(proj_out): Conv2d(640, 640, kernel_size=(1, 1), stride=(1, 1))))(6): TimestepEmbedSequential((0): Downsample((op): Conv2d(640, 640, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))))(7): TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 640, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(640, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=1280, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Conv2d(640, 1280, kernel_size=(1, 1), stride=(1, 1)))(1): SpatialTransformer((norm): GroupNorm(32, 1280, eps=1e-06, affine=True)(proj_in): Conv2d(1280, 1280, kernel_size=(1, 1), stride=(1, 1))(transformer_blocks): ModuleList((0): BasicTransformerBlock((attn1): CrossAttention((to_q): Linear(in_features=1280, out_features=1280, bias=False)(to_k): Linear(in_features=1280, out_features=1280, bias=False)(to_v): Linear(in_features=1280, out_features=1280, bias=False)(to_out): Sequential((0): Linear(in_features=1280, out_features=1280, bias=True)(1): Dropout(p=0.0, inplace=False)))(ff): FeedForward((net): Sequential((0): GEGLU((proj): Linear(in_features=1280, out_features=10240, bias=True))(1): Dropout(p=0.0, inplace=False)(2): Linear(in_features=5120, out_features=1280, bias=True)))(attn2): CrossAttention((to_q): Linear(in_features=1280, out_features=1280, bias=False)(to_k): Linear(in_features=768, out_features=1280, bias=False)(to_v): Linear(in_features=768, out_features=1280, bias=False)(to_out): Sequential((0): Linear(in_features=1280, out_features=1280, bias=True)(1): Dropout(p=0.0, inplace=False)))(norm1): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)(norm2): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)(norm3): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)))(proj_out): Conv2d(1280, 1280, kernel_size=(1, 1), stride=(1, 1))))(8): TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=1280, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Identity())(1): SpatialTransformer((norm): GroupNorm(32, 1280, eps=1e-06, affine=True)(proj_in): Conv2d(1280, 1280, kernel_size=(1, 1), stride=(1, 1))(transformer_blocks): ModuleList((0): BasicTransformerBlock((attn1): CrossAttention((to_q): Linear(in_features=1280, out_features=1280, bias=False)(to_k): Linear(in_features=1280, out_features=1280, bias=False)(to_v): Linear(in_features=1280, out_features=1280, bias=False)(to_out): Sequential((0): Linear(in_features=1280, out_features=1280, bias=True)(1): Dropout(p=0.0, inplace=False)))(ff): FeedForward((net): Sequential((0): GEGLU((proj): Linear(in_features=1280, out_features=10240, bias=True))(1): Dropout(p=0.0, inplace=False)(2): Linear(in_features=5120, out_features=1280, bias=True)))(attn2): CrossAttention((to_q): Linear(in_features=1280, out_features=1280, bias=False)(to_k): Linear(in_features=768, out_features=1280, bias=False)(to_v): Linear(in_features=768, out_features=1280, bias=False)(to_out): Sequential((0): Linear(in_features=1280, out_features=1280, bias=True)(1): Dropout(p=0.0, inplace=False)))(norm1): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)(norm2): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)(norm3): LayerNorm((1280,), eps=1e-05, elementwise_affine=True)))(proj_out): Conv2d(1280, 1280, kernel_size=(1, 1), stride=(1, 1))))(9): TimestepEmbedSequential((0): Downsample((op): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))))(10-11): 2 x TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=1280, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Identity()))
)

        12个 module 都使用了 TimestepEmbedSequential 类进行封装，根据不同的网络层，将输入噪声x与timestep embedding和prompt context进行运算。

class TimestepEmbedSequential(nn.Sequential, TimestepBlock):"""A sequential module that passes timestep embeddings to the children thatsupport it as an extra input."""def forward(self, x, emb, context=None):for layer in self:if isinstance(layer, TimestepBlock):x = layer(x, emb)elif isinstance(layer, SpatialTransformer):x = layer(x, context)else:x = layer(x)return x

2-3-1--Module0

 Module 0 是一个2D卷积层，主要对输入噪声进行特征提取；

# init 初始化
self.input_blocks = nn.ModuleList([TimestepEmbedSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1))]
)# 打印 self.input_blocks[0]
TimestepEmbedSequential((0): Conv2d(4, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)

2-3-2--Module1和Module2

        Module1和Module2的结构相同，都由一个ResBlock和一个SpatialTransformer组成；

# init 初始化
for _ in range(num_res_blocks):layers = [ResBlock(ch,time_embed_dim,dropout,out_channels=mult * model_channels,dims=dims,use_checkpoint=use_checkpoint,use_scale_shift_norm=use_scale_shift_norm,)]ch = mult * model_channelsif ds in attention_resolutions:if num_head_channels == -1:dim_head = ch // num_headselse:num_heads = ch // num_head_channelsdim_head = num_head_channelsif legacy:#num_heads = 1dim_head = ch // num_heads if use_spatial_transformer else num_head_channelslayers.append(AttentionBlock(ch,use_checkpoint=use_checkpoint,num_heads=num_heads,num_head_channels=dim_head,use_new_attention_order=use_new_attention_order,) if not use_spatial_transformer else SpatialTransformer(ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim))self.input_blocks.append(TimestepEmbedSequential(*layers))self._feature_size += chinput_block_chans.append(ch)# 打印 self.input_blocks[1]
TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 320, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=320, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 320, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Identity())(1): SpatialTransformer((norm): GroupNorm(32, 320, eps=1e-06, affine=True)(proj_in): Conv2d(320, 320, kernel_size=(1, 1), stride=(1, 1))(transformer_blocks): ModuleList((0): BasicTransformerBlock((attn1): CrossAttention((to_q): Linear(in_features=320, out_features=320, bias=False)(to_k): Linear(in_features=320, out_features=320, bias=False)(to_v): Linear(in_features=320, out_features=320, bias=False)(to_out): Sequential((0): Linear(in_features=320, out_features=320, bias=True)(1): Dropout(p=0.0, inplace=False)))(ff): FeedForward((net): Sequential((0): GEGLU((proj): Linear(in_features=320, out_features=2560, bias=True))(1): Dropout(p=0.0, inplace=False)(2): Linear(in_features=1280, out_features=320, bias=True)))(attn2): CrossAttention((to_q): Linear(in_features=320, out_features=320, bias=False)(to_k): Linear(in_features=768, out_features=320, bias=False)(to_v): Linear(in_features=768, out_features=320, bias=False)(to_out): Sequential((0): Linear(in_features=320, out_features=320, bias=True)(1): Dropout(p=0.0, inplace=False)))(norm1): LayerNorm((320,), eps=1e-05, elementwise_affine=True)(norm2): LayerNorm((320,), eps=1e-05, elementwise_affine=True)(norm3): LayerNorm((320,), eps=1e-05, elementwise_affine=True)))(proj_out): Conv2d(320, 320, kernel_size=(1, 1), stride=(1, 1)))
)# 打印 self.input_blocks[2]
TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 320, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=320, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 320, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(320, 320, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Identity())(1): SpatialTransformer((norm): GroupNorm(32, 320, eps=1e-06, affine=True)(proj_in): Conv2d(320, 320, kernel_size=(1, 1), stride=(1, 1))(transformer_blocks): ModuleList((0): BasicTransformerBlock((attn1): CrossAttention((to_q): Linear(in_features=320, out_features=320, bias=False)(to_k): Linear(in_features=320, out_features=320, bias=False)(to_v): Linear(in_features=320, out_features=320, bias=False)(to_out): Sequential((0): Linear(in_features=320, out_features=320, bias=True)(1): Dropout(p=0.0, inplace=False)))(ff): FeedForward((net): Sequential((0): GEGLU((proj): Linear(in_features=320, out_features=2560, bias=True))(1): Dropout(p=0.0, inplace=False)(2): Linear(in_features=1280, out_features=320, bias=True)))(attn2): CrossAttention((to_q): Linear(in_features=320, out_features=320, bias=False)(to_k): Linear(in_features=768, out_features=320, bias=False)(to_v): Linear(in_features=768, out_features=320, bias=False)(to_out): Sequential((0): Linear(in_features=320, out_features=320, bias=True)(1): Dropout(p=0.0, inplace=False)))(norm1): LayerNorm((320,), eps=1e-05, elementwise_affine=True)(norm2): LayerNorm((320,), eps=1e-05, elementwise_affine=True)(norm3): LayerNorm((320,), eps=1e-05, elementwise_affine=True)))(proj_out): Conv2d(320, 320, kernel_size=(1, 1), stride=(1, 1)))
)

2-3-3--Module3

        Module3是一个下采样2D卷积层。

# init 初始化
if level != len(channel_mult) - 1:out_ch = chself.input_blocks.append(TimestepEmbedSequential(ResBlock(ch,time_embed_dim,dropout,out_channels=out_ch,dims=dims,use_checkpoint=use_checkpoint,use_scale_shift_norm=use_scale_shift_norm,down=True,)if resblock_updownelse Downsample(ch, conv_resample, dims=dims, out_channels=out_ch)))# 打印 self.input_blocks[3]
TimestepEmbedSequential((0): Downsample((op): Conv2d(320, 320, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)))
)

2-3-4--Module4、Module5、Module7和Module8

        与Module1和Module2的结构相同，都由一个ResBlock和一个SpatialTransformer组成，只有特征维度上的区别；

2-3-4--Module6和Module9

        与Module3的结构相同，是一个下采样2D卷积层。

2-3--5--Module10和Module11

        Module10和Module12的结构相同，只由一个ResBlock组成。

# 打印 self.input_blocks[10]
TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=1280, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Identity())
)# 打印 self.input_blocks[11]
TimestepEmbedSequential((0): ResBlock((in_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(h_upd): Identity()(x_upd): Identity()(emb_layers): Sequential((0): SiLU()(1): Linear(in_features=1280, out_features=1280, bias=True))(out_layers): Sequential((0): GroupNorm32(32, 1280, eps=1e-05, affine=True)(1): SiLU()(2): Dropout(p=0, inplace=False)(3): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))(skip_connection): Identity())
)

本文来自互联网用户投稿，该文观点仅代表作者本人，不代表本站立场。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如若转载，请注明出处：http://www.mzph.cn/news/851194.shtml

如若内容造成侵权/违法违规/事实不符，请联系多彩编程网进行投诉反馈email:809451989@qq.com，一经查实，立即删除！

相关文章

中山大学和字节发布「视频虚拟试穿」扩散模型VITON-DiT，一键生成换装后视频！

中山大学和字节发布「视频虚拟试穿」扩散模型VITON-DiT，一键生成换装后视频！

视频虚拟试穿技术日益受到关注，然而现有的工作局限于将服装图像转移到姿势和背景简单的视频上，对于随意拍摄的视频则效果不佳。最近，Sora 揭示了 Diffusion Transformer (DiT) 在生成具有真实场景的逼真视频方面的可扩展性，可以说…
阅读更多...
Kubernetes入门-大简介

Kubernetes入门-大简介

目录 何为微服务 何为云原生 何为编排器 “Kubernetes”这个名字来自希腊语，意思是“舵手”舵手是一个航海/航行术语，指掌舵的人从本质上说，Kubernetes是云原生微服务(cloud-native microservice)应用的编排器(orchestrator) 何为微服务 …
阅读更多...
大模型日报2024-06-09

大模型日报2024-06-09

大模型日报 2024-06-09 大模型资讯 Qwen2-72B：拥有72亿参数的先进AI模型 摘要: Qwen团队近日发布了最新突破性成果——Qwen2-72B。这款最先进的语言模型具备72亿参数、支持128K令牌、多语言精通及业界领先性能，展示了显著的技术进步。 OpenAI与Anthropic…
阅读更多...
WordPress 高级缓存插件 W3 Total Cache 开启支持 Brotli 压缩算法

WordPress 高级缓存插件 W3 Total Cache 开启支持 Brotli 压缩算法

今天明月给大家分享一下 WordPress 高级缓存插件 W3 Total Cache 开启支持 Brotli 压缩算法的教程，在撰写【WordPress 高级缓存插件 W3 Total Cache Pro 详细配置教程】一文的时候明月就发现 W3 Total Cache 已经支持 Brotli 压缩算法了，可惜的是在安装完…
阅读更多...
ctfshow-web入门-命令执行（web53-web55）

ctfshow-web入门-命令执行（web53-web55）

目录 1、web53 2、web54 3、web55 1、web53 这里的代码有点不一样，说一下这两种的区别： （1）直接执行 system($c); system($c);这种方式会直接执行命令 $c 并将命令的输出直接发送到标准输出（通常是浏览器&#xff…
阅读更多...
【qsort函数】

【qsort函数】

前言 我们要学习qsort函数并利用冒泡函数仿照qsort函数 首先我们要了解一下qsort（快速排序） 这是函数的的基本参数 void qsort (void* base, size_t num, size_t size,int (*compar)(const void*,const void*)); 简单解释一下 base：指向…
阅读更多...
23.在游戏中按下Home键呼出辅助窗口

23.在游戏中按下Home键呼出辅助窗口

上一个内容：22.钩子注入原理 在 22.钩子注入原理 它的代码上进行修改 效果图： 首先在CWndMain.h文件中添加下图红框里的东西 ChangeShowState函数的实现 void CWndMain::ChangeShowState() {UiShow !UiShow;ShowWindow(UiShow); } OnInitDialog函数…
阅读更多...
CISCN2024 初赛 wp 部分复现（Re）

CISCN2024 初赛 wp 部分复现（Re）

Misc 1. 火锅链观光打卡 答题即可 Re 1. asm_re 感谢智谱清言，可以读出大致加密算法 这是输入 这是加密部分 这里判断 找到疑似密文的部分，手动改一下端序 #asm_wp def dec(char):return (((char - 0x1E) ^ 0x4D) - 0x14) // 0x50 #return (ord(cha…
阅读更多...
powerdesigner各种字体设置

powerdesigner各种字体设置

1、设置左侧菜单： 步骤如下： tools —> general options —> fonts —> defalut UI font ,选择字体样式及大小即可，同下图。 2、设置Table的字体大小 Tools------>Display Prefrences------>Table------->Format---------…
阅读更多...
RabbitMQ系列-rabbitmq无法重新加入集群，启动失败的问题

RabbitMQ系列-rabbitmq无法重新加入集群，启动失败的问题

当前存在3个节点：rabbitmq5672、rabbitmq5673、rabbitmq5674 当rabbitmq5673节点掉线之后，重启失败 重启的时候5672节点报错如下： 解决方案 在集群中取消失败节点 rabbitmqctl forget_cluster_node rabbitrabbitmq5673删除失败节点5673的…
阅读更多...
王学岗鸿蒙开发(北向)——————(十)子组件修改父组件的内容与 动画

王学岗鸿蒙开发(北向)——————(十)子组件修改父组件的内容与 动画

子组件修改父组件的内容 使用类似Android的回调，父组件传递给子组件一个函数 import { MyComment } from ./component/MyComment import { MyContent } from ./component/MyComtent import { MyTitleComponent } from ./component/MyTitleComponentEntry Componen…
阅读更多...
【MySQL数据库基础】

【MySQL数据库基础】

🌈个人主页：努力学编程’ ⛅个人推荐：基于java提供的ArrayList实现的扑克牌游戏 |C贪吃蛇详解 ⚡学好数据结构，刷题刻不容缓：点击一起刷题 🌙心灵鸡汤：总有人要赢，为什么不能是我呢 …
阅读更多...
React核心概念、主要特点及组件的生命周期

React核心概念、主要特点及组件的生命周期

在前端开发的世界中，React以其独特的魅力和强大的功能，成为了构建用户界面的首选框架之一。本文将深入探讨React的核心概念、主要特点以及组件生命周期 React简介 React是由Facebook开发并开源的前端JavaScript库，专门用于构建可重用的UI组…
阅读更多...
【CS.OS】堆管理算法：不同的堆分配和管理算法

【CS.OS】堆管理算法：不同的堆分配和管理算法

1000.5.CS.OS.1.3-基础-内存管理-堆管理算法-Created: 2024-06-09.Sunday10:41 文章目录 1 内存分配算法概述1.1 首次适应（First-Fit）1.2 最佳适应（Best-Fit） 2 伙伴系统（Buddy System） 3 总结References …
阅读更多...
Python | 正则表达式

Python | 正则表达式

?:标记?之前的字符为可选. used&#xff1f; d可有可无 *:匹配>0个重复的在*号之前的字符。 ab*c 匹配多个b &#xff1a;匹配>1个重复的号前的字符。&#xff08;至少一个&#xff09; {n,m}&#xff1a;匹配num个大括号之前的字符或字符集 &#xff08;n < num …
阅读更多...
算法：101. 对称二叉树

算法：101. 对称二叉树

对称二叉树 给你一个二叉树的根节点 root &#xff0c; 检查它是否轴对称。 示例 1&#xff1a; 输入&#xff1a;root [1,2,2,3,4,4,3] 输出&#xff1a;true示例 2&#xff1a; 输入&#xff1a;root [1,2,2,null,3,null,3] 输出&#xff1a;false提示&#xff1a; 树中节…
阅读更多...
ThreadCache线程缓存

ThreadCache线程缓存

一.ThreadCache整体结构 1.基本结构 定长内存池利用一个自由链表管理释放回来的固定大小的内存obj。 ThreadCache需要支持申请和释放不同大小的内存块&#xff0c;因此需要多个自由链表来管理释放回来的内存块.即ThreadCache实际上一个哈希桶结构&#xff0c;每个桶中存放的都…
阅读更多...
BGP中MED属性

BGP中MED属性

6、MED——多出口鉴定属性 BGP 协议默认没有度量值&#xff0c;没有cost&#xff1b;所谓的 MED 就是人为的在路由条目中编写一个cost数值&#xff0c;干涉选路&#xff1b; 可用于干涉 EBGP/IBGP 关系下的选路&#xff1b;最常用于干涉 EBGP 关系选路&#xff1b; 常常用…
阅读更多...
目标检测（R-CNN）系列（Pytorch 26）

目标检测（R-CNN）系列（Pytorch 26）

一 R-CNN 除了之前描述的单发多框检测之外&#xff0c;区域卷积神经网络&#xff08;region‐based CNN或regions with CNN features&#xff0c; R‐CNN&#xff09;(Girshick et al., 2014)也是将深度模型应用于目标检测的开创性工作之一。下面介绍R‐CNN及其一 系列改进方法…
阅读更多...
架构设计-web项目中跨域问题涉及到的后端和前端配置

架构设计-web项目中跨域问题涉及到的后端和前端配置

WEB软件项目中经常会遇到跨域问题&#xff0c;解决方案早已是业内的共识&#xff0c;简要记录主流的处理方式&#xff1a; 跨域感知session需要解决两个问题&#xff1a; 1. 跨域问题 2. 跨域cookie传输问题 跨域问题 解决跨域问题有很多种方式&#xff0c;如使用springboot…
阅读更多...
最新文章

Copyright @ 2022~2023