Update ip_adapter/attention_processor.py

ip_adapter/attention_processor.py

@@ -1,3 +1,4 @@
+# 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
@@ -6,31 +7,33 @@ try:
     import xformers
     import xformers.ops
     xformers_available = True
-except Exception:
+except Exception as e:
     xformers_available = False
 
-
-# Region Controller (unchanged)
-class RegionControler:
+class RegionControler(object):
     def __init__(self) -> None:
         self.prompt_image_conditioning = []
-
 region_control = RegionControler()
 
-
-# Helper function for weight initialization
-def init_weights(m):
-    if isinstance(m, nn.Linear):
-        nn.init.xavier_uniform_(m.weight)
-
-
-# Base Attention Processor
-class BaseAttnProcessor(nn.Module):
-    def __init__(self):
+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 _process_input(self, attn, hidden_states, encoder_hidden_states, attention_mask, temb):
-        """Handles preprocessing for both AttnProcessor and IPAttnProcessor"""
+    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:
@@ -41,60 +44,286 @@ class BaseAttnProcessor(nn.Module):
         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)
-        else:
-            batch_size, sequence_length, _ = hidden_states.shape
+
+        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)
 
-        return hidden_states, encoder_hidden_states, residual, batch_size, input_ndim, height, width
+        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)
 
-    def _apply_attention(self, attn, query, key, value, attention_mask):
-        """Handles the actual attention operation using either xformers or standard PyTorch"""
         if xformers_available:
-            return xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
         else:
             attention_probs = attn.get_attention_scores(query, key, attention_mask)
-            return torch.bmm(attention_probs, value)
+            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
 
-# Optimized AttnProcessor
-class AttnProcessor(BaseAttnProcessor):
-    def forward(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):
-        hidden_states, encoder_hidden_states, residual, batch_size, input_ndim, height, width = \
-            self._process_input(attn, hidden_states, encoder_hidden_states, attention_mask, temb)
+
+    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)
 
-        query, key, value = map(attn.head_to_batch_dim, (query, key, value))
-        hidden_states = self._apply_attention(attn, query, key, value, attention_mask)
-        hidden_states = attn.batch_to_head_dim(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, -1, height, width)
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
 
         if attn.residual_connection:
             hidden_states = hidden_states + residual
 
-        return hidden_states / attn.rescale_output_factor
+        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.
+    """
 
-# Optimized IPAttnProcessor
-class IPAttnProcessor(BaseAttnProcessor):
     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
@@ -102,49 +331,117 @@ class IPAttnProcessor(BaseAttnProcessor):
 
         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)
-        self.apply(init_weights)
 
-    def forward(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):
-        hidden_states, encoder_hidden_states, residual, batch_size, input_ndim, height, width = \
-            self._process_input(attn, hidden_states, encoder_hidden_states, attention_mask, temb)
+    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
+        )
 
-        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 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)
 
-        query, key, value = map(attn.head_to_batch_dim, (query, key, value))
-        hidden_states = self._apply_attention(attn, query, key, value, attention_mask)
-        hidden_states = attn.batch_to_head_dim(hidden_states)
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
 
-        # Image Prompt Attention
-        ip_key = attn.head_to_batch_dim(self.to_k_ip(ip_hidden_states))
-        ip_value = attn.head_to_batch_dim(self.to_v_ip(ip_hidden_states))
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
 
-        ip_hidden_states = self._apply_attention(attn, query, ip_key, ip_value, None)
-        ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+        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)
 
-        # Region Control
+        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
+        )
+        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)
+            region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None)
             if region_mask is not None:
-                mask = F.interpolate(region_mask[None, None], scale_factor=(ip_hidden_states.shape[1] / region_mask.shape[0]), mode="nearest").reshape([1, -1, 1])
+                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 *= mask
+            ip_hidden_states = ip_hidden_states * mask
 
         hidden_states = hidden_states + self.scale * ip_hidden_states
 
-        # Linear projection and dropout
+        # 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, -1, height, width)
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
 
         if attn.residual_connection:
             hidden_states = hidden_states + residual
 
-        return hidden_states / attn.rescale_output_factor
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states