# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py import torch import torch.nn as nn import torch.nn.functional as F from diffusers.models.attention_processor import AttnProcessor2_0 from diffusers.models.attention_processor import XFormersAttnProcessor try: import xformers import xformers.ops xformers_available = True except Exception as e: xformers_available = False class RegionControler(object): def __init__(self) -> None: self.prompt_image_conditioning = [] region_control = RegionControler() class AttnProcessor(nn.Module): r""" Default processor for performing attention-related computations. """ def __init__( self, hidden_size=None, cross_attention_dim=None, ): super().__init__() def forward( self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, ): residual = hidden_states if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) input_ndim = hidden_states.ndim if input_ndim == 4: batch_size, channel, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states elif attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) query = attn.head_to_batch_dim(query) key = attn.head_to_batch_dim(key) value = attn.head_to_batch_dim(value) attention_probs = attn.get_attention_scores(query, key, attention_mask) hidden_states = torch.bmm(attention_probs, value) hidden_states = attn.batch_to_head_dim(hidden_states) # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: hidden_states = hidden_states + residual hidden_states = hidden_states / attn.rescale_output_factor return hidden_states class IPAttnProcessor(nn.Module): r""" Attention processor for IP-Adapater. Args: hidden_size (`int`): The hidden size of the attention layer. cross_attention_dim (`int`): The number of channels in the `encoder_hidden_states`. scale (`float`, defaults to 1.0): the weight scale of image prompt. num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16): The context length of the image features. """ def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4): super().__init__() self.hidden_size = hidden_size self.cross_attention_dim = cross_attention_dim self.scale = scale self.num_tokens = num_tokens self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False) self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False) def forward( self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, ): residual = hidden_states if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) input_ndim = hidden_states.ndim if input_ndim == 4: batch_size, channel, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states else: # get encoder_hidden_states, ip_hidden_states end_pos = encoder_hidden_states.shape[1] - self.num_tokens encoder_hidden_states, ip_hidden_states = encoder_hidden_states[:, :end_pos, :], encoder_hidden_states[:, end_pos:, :] if attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) query = attn.head_to_batch_dim(query) key = attn.head_to_batch_dim(key) value = attn.head_to_batch_dim(value) if xformers_available: hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask) else: attention_probs = attn.get_attention_scores(query, key, attention_mask) hidden_states = torch.bmm(attention_probs, value) hidden_states = attn.batch_to_head_dim(hidden_states) # for ip-adapter ip_key = self.to_k_ip(ip_hidden_states) ip_value = self.to_v_ip(ip_hidden_states) ip_key = attn.head_to_batch_dim(ip_key) ip_value = attn.head_to_batch_dim(ip_value) if xformers_available: ip_hidden_states = self._memory_efficient_attention_xformers(query, ip_key, ip_value, None) else: ip_attention_probs = attn.get_attention_scores(query, ip_key, None) ip_hidden_states = torch.bmm(ip_attention_probs, ip_value) ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states) # region control if len(region_control.prompt_image_conditioning) == 1: region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None) if region_mask is not None: h, w = region_mask.shape[:2] ratio = (h * w / query.shape[1]) ** 0.5 mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1]) else: mask = torch.ones_like(ip_hidden_states) ip_hidden_states = ip_hidden_states * mask hidden_states = hidden_states + self.scale * ip_hidden_states # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: hidden_states = hidden_states + residual hidden_states = hidden_states / attn.rescale_output_factor return hidden_states def _memory_efficient_attention_xformers(self, query, key, value, attention_mask): # TODO attention_mask query = query.contiguous() key = key.contiguous() value = value.contiguous() hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask) # hidden_states = self.reshape_batch_dim_to_heads(hidden_states) return hidden_states class AttnProcessor2_0(torch.nn.Module): r""" Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). """ def __init__( self, hidden_size=None, cross_attention_dim=None, ): super().__init__() if not hasattr(F, "scaled_dot_product_attention"): raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") def forward( self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, ): residual = hidden_states if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) input_ndim = hidden_states.ndim if input_ndim == 4: batch_size, channel, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) if attention_mask is not None: attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) # scaled_dot_product_attention expects attention_mask shape to be # (batch, heads, source_length, target_length) attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states elif attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) inner_dim = key.shape[-1] head_dim = inner_dim // attn.heads query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # the output of sdp = (batch, num_heads, seq_len, head_dim) # TODO: add support for attn.scale when we move to Torch 2.1 hidden_states = F.scaled_dot_product_attention( query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False ) hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) hidden_states = hidden_states.to(query.dtype) # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: hidden_states = hidden_states + residual hidden_states = hidden_states / attn.rescale_output_factor return hidden_states class IPAttnProcessor2_0(torch.nn.Module): r""" Attention processor for IP-Adapater for PyTorch 2.0. Args: hidden_size (`int`): The hidden size of the attention layer. cross_attention_dim (`int`): The number of channels in the `encoder_hidden_states`. scale (`float`, defaults to 1.0): the weight scale of image prompt. num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16): The context length of the image features. """ def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4): super().__init__() if not hasattr(F, "scaled_dot_product_attention"): raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") self.hidden_size = hidden_size self.cross_attention_dim = cross_attention_dim self.scale = scale self.num_tokens = num_tokens self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False) self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False) def __call__( self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, ): # 保存输入的 hidden_states，用于最后的残差连接。 residual = hidden_states # 检查是否有 空间归一化 (spatial normalization) if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) # hidden_states 可能是一个 4D 张量（比如图像数据），也可能是一个 3D 张量（比如文本数据） input_ndim = hidden_states.ndim if input_ndim == 4: batch_size, channel, height, width = hidden_states.shape # 调整其形状为 (batch_size, channel, height * width) hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) # 选择 encoder_hidden_states 如果有的话，否则使用 hidden_states 作为输入。sequence_length 表示序列长度，通常是时间步或图像的像素数量。 batch_size, sequence_length, _ = (hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape) # 处理并调整注意力掩码 (attention mask)，使其符合 scaled_dot_product_attention 函数的要求。 if attention_mask is not None: attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) # scaled_dot_product_attention expects attention_mask shape to be # (batch, heads, source_length, target_length) attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) # 对 hidden_states 进行组归一化 if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) # 通过线性变换将 hidden_states 映射到query向量 query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states else: # 分割 encoder_hidden_states 和 ip_hidden_states end_pos = encoder_hidden_states.shape[1] - self.num_tokens encoder_hidden_states, ip_hidden_states = ( encoder_hidden_states[:, :end_pos, :], encoder_hidden_states[:, end_pos:, :], ) if attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) # 将 encoder_hidden_states 映射为多头自注意力计算中的键和值 key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) # 获取每个注意力头的维度 inner_dim = key.shape[-1] head_dim = inner_dim // attn.heads query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # the output of sdp = (batch, num_heads, seq_len, head_dim) # TODO: add support for attn.scale when we move to Torch 2.1 hidden_states = F.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False) # hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask) hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) hidden_states = hidden_states.to(query.dtype) # for ip-adapter # 投影 ip_hidden_states 得到其键和值 ip_key = self.to_k_ip(ip_hidden_states) ip_value = self.to_v_ip(ip_hidden_states) ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # the output of sdp = (batch, num_heads, seq_len, head_dim) # TODO: add support for attn.scale when we move to Torch 2.1 # 注意力计算 得到图像提示的隐藏状态 ip_hidden_states = F.scaled_dot_product_attention(query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False) # ip_hidden_states = xformers.ops.memory_efficient_attention(query, ip_key, ip_value, attn_bias=None) ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) ip_hidden_states = ip_hidden_states.to(query.dtype) # 通过给图像提示隐藏状态加权缩放后与原始隐藏状态相加，实现跨域信息融合 hidden_states = hidden_states + self.scale * ip_hidden_states # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: # 如果输入是 4D 张量（图像数据），则将 hidden_states 转换回原始形状。 hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: # 如果启用了残差连接，则将 residual 添加回 hidden_states hidden_states = hidden_states + residual # 对输出进行缩放 hidden_states = hidden_states / attn.rescale_output_factor return hidden_states ## for controlnet class CNAttnProcessor: r""" Default processor for performing attention-related computations. """ def __init__(self, num_tokens=4): self.num_tokens = num_tokens def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, *args, **kwargs,): residual = hidden_states if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) input_ndim = hidden_states.ndim if input_ndim == 4: batch_size, channel, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states else: end_pos = encoder_hidden_states.shape[1] - self.num_tokens encoder_hidden_states = encoder_hidden_states[:, :end_pos] # only use text if attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) query = attn.head_to_batch_dim(query) key = attn.head_to_batch_dim(key) value = attn.head_to_batch_dim(value) attention_probs = attn.get_attention_scores(query, key, attention_mask) hidden_states = torch.bmm(attention_probs, value) hidden_states = attn.batch_to_head_dim(hidden_states) # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: hidden_states = hidden_states + residual hidden_states = hidden_states / attn.rescale_output_factor return hidden_states class CNAttnProcessor2_0: r""" Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). """ def __init__(self, num_tokens=4): if not hasattr(F, "scaled_dot_product_attention"): raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") self.num_tokens = num_tokens def __call__( self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, *args, **kwargs, ): residual = hidden_states if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) input_ndim = hidden_states.ndim if input_ndim == 4: batch_size, channel, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) if attention_mask is not None: attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) # scaled_dot_product_attention expects attention_mask shape to be # (batch, heads, source_length, target_length) attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states else: end_pos = encoder_hidden_states.shape[1] - self.num_tokens encoder_hidden_states = encoder_hidden_states[:, :end_pos] # only use text if attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) inner_dim = key.shape[-1] head_dim = inner_dim // attn.heads query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # the output of sdp = (batch, num_heads, seq_len, head_dim) # TODO: add support for attn.scale when we move to Torch 2.1 hidden_states = F.scaled_dot_product_attention( query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False ) hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) hidden_states = hidden_states.to(query.dtype) # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: hidden_states = hidden_states + residual hidden_states = hidden_states / attn.rescale_output_factor return hidden_states class IPAttnProcessor2_02(torch.nn.Module): r""" Attention processor for IP-Adapater for PyTorch 2.0. Args: hidden_size (`int`): The hidden size of the attention layer. cross_attention_dim (`int`): The number of channels in the `encoder_hidden_states`. scale (`float`, defaults to 1.0): the weight scale of image prompt. num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16): The context length of the image features. """ def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4): super().__init__() if not hasattr(F, "scaled_dot_product_attention"): raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") self.hidden_size = hidden_size self.cross_attention_dim = cross_attention_dim self.scale = scale self.num_tokens = num_tokens self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False) self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False) def forward( self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, ): residual = hidden_states if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) input_ndim = hidden_states.ndim if input_ndim == 4: batch_size, channel, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) if attention_mask is not None: attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) # scaled_dot_product_attention expects attention_mask shape to be # (batch, heads, source_length, target_length) attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states else: # get encoder_hidden_states, ip_hidden_states end_pos = encoder_hidden_states.shape[1] - self.num_tokens encoder_hidden_states, ip_hidden_states = ( encoder_hidden_states[:, :end_pos, :], encoder_hidden_states[:, end_pos:, :], ) if attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) inner_dim = key.shape[-1] head_dim = inner_dim // attn.heads query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask) # hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False) hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) hidden_states = hidden_states.to(query.dtype) # hidden_states = memory_efficient_attention(query, key, value, attn_mask=attention_mask, dropout_p=0.0) # for ip-adapter ip_key = self.to_k_ip(ip_hidden_states) ip_value = self.to_v_ip(ip_hidden_states) ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # ip_hidden_states = F.scaled_dot_product_attention(query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False) # ip_hidden_states = xformers.ops.memory_efficient_attention(query, ip_key, ip_value, None) ip_hidden_states = self._memory_efficient_attention_xformers(query, ip_key, ip_value, None) with torch.no_grad(): self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1) #print(self.attn_map.shape) ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) ip_hidden_states = ip_hidden_states.to(query.dtype) # region control if len(region_control.prompt_image_conditioning) == 1: region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None) if region_mask is not None: query = query.reshape([-1, query.shape[-2], query.shape[-1]]) h, w = region_mask.shape[:2] ratio = (h * w / query.shape[1]) ** 0.5 mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1]) else: mask = torch.ones_like(ip_hidden_states) ip_hidden_states = ip_hidden_states * mask # ip_hidden_states = memory_efficient_attention(query, ip_key, ip_value, attn_mask=None, dropout_p=0.0) ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * (ip_key.shape[-1] // attn.heads)) hidden_states = hidden_states + self.scale * ip_hidden_states # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: hidden_states = hidden_states + residual hidden_states = hidden_states / attn.rescale_output_factor return hidden_states def _memory_efficient_attention_xformers(self, query, key, value, attention_mask): # TODO attention_mask query = query.contiguous() key = key.contiguous() value = value.contiguous() hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask) # hidden_states = self.reshape_batch_dim_to_heads(hidden_states) return hidden_states class IPAttnProcessor2_00(torch.nn.Module): r""" Attention processor for IP-Adapater for PyTorch 2.0. Args: hidden_size (`int`): The hidden size of the attention layer. cross_attention_dim (`int`): The number of channels in the `encoder_hidden_states`. scale (`float`, defaults to 1.0): the weight scale of image prompt. num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16): The context length of the image features. """ def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4): super().__init__() if not hasattr(F, "scaled_dot_product_attention"): raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") self.hidden_size = hidden_size self.cross_attention_dim = cross_attention_dim self.scale = scale self.num_tokens = num_tokens self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False) self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False) def __call__( self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, ): residual = hidden_states if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) input_ndim = hidden_states.ndim if input_ndim == 4: batch_size, channel, height, width = hidden_states.shape hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) if attention_mask is not None: attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) # scaled_dot_product_attention expects attention_mask shape to be # (batch, heads, source_length, target_length) attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states else: # get encoder_hidden_states, ip_hidden_states end_pos = encoder_hidden_states.shape[1] - self.num_tokens encoder_hidden_states, ip_hidden_states = ( encoder_hidden_states[:, :end_pos, :], encoder_hidden_states[:, end_pos:, :], ) if attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) inner_dim = key.shape[-1] head_dim = inner_dim // attn.heads query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # the output of sdp = (batch, num_heads, seq_len, head_dim) # TODO: add support for attn.scale when we move to Torch 2.1 hidden_states = F.scaled_dot_product_attention( query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False ) hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) hidden_states = hidden_states.to(query.dtype) # for ip-adapter ip_key = self.to_k_ip(ip_hidden_states) ip_value = self.to_v_ip(ip_hidden_states) ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # the output of sdp = (batch, num_heads, seq_len, head_dim) # TODO: add support for attn.scale when we move to Torch 2.1 ip_hidden_states = F.scaled_dot_product_attention( query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False ) ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) ip_hidden_states = ip_hidden_states.to(query.dtype) hidden_states = hidden_states + self.scale * ip_hidden_states # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: hidden_states = hidden_states + residual hidden_states = hidden_states / attn.rescale_output_factor return hidden_states ## for controlnet