Add paint by example (huggingface#1533)

* add paint by example

* mkae loading possibel

* up

* Update src/diffusers/models/attention.py

* up

* finalize weight structure

* make example work

* make it work

* up

* up

* fix

* del

* add

* update

* Apply suggestions from code review

* correct transformer 2d

* finish

* up

* up

* up

* up

* fix

* Apply suggestions from code review

Co-authored-by: Pedro Cuenca <[email protected]>

* Apply suggestions from code review

* up

* finish

Co-authored-by: Pedro Cuenca <[email protected]>

Author: patrickvonplaten

__init__.py

@@ -72,6 +72,7 @@
         AltDiffusionPipeline,
         CycleDiffusionPipeline,
         LDMTextToImagePipeline,
+        PaintByExamplePipeline,
         StableDiffusionImageVariationPipeline,
         StableDiffusionImg2ImgPipeline,
         StableDiffusionInpaintPipeline,

models/attention.py

@@ -406,6 +406,9 @@ def __init__(
     ):
         super().__init__()
         self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+
+        # 1. Self-Attn
         self.attn1 = CrossAttention(
             query_dim=dim,
             heads=num_attention_heads,
@@ -415,23 +418,28 @@ def __init__(
             cross_attention_dim=cross_attention_dim if only_cross_attention else None,
         )  # is a self-attention
         self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
-        self.attn2 = CrossAttention(
-            query_dim=dim,
-            cross_attention_dim=cross_attention_dim,
-            heads=num_attention_heads,
-            dim_head=attention_head_dim,
-            dropout=dropout,
-            bias=attention_bias,
-        )  # is self-attn if context is none
 
-        # layer norms
-        self.use_ada_layer_norm = num_embeds_ada_norm is not None
-        if self.use_ada_layer_norm:
-            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
-            self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        # 2. Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = CrossAttention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+            )  # is self-attn if context is none
         else:
-            self.norm1 = nn.LayerNorm(dim)
-            self.norm2 = nn.LayerNorm(dim)
+            self.attn2 = None
+
+        self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+        if cross_attention_dim is not None:
+            self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+        else:
+            self.norm2 = None
+
+        # 3. Feed-forward
         self.norm3 = nn.LayerNorm(dim)
 
         # if xformers is installed try to use memory_efficient_attention by default
@@ -481,11 +489,12 @@ def forward(self, hidden_states, context=None, timestep=None):
         else:
             hidden_states = self.attn1(norm_hidden_states) + hidden_states
 
-        # 2. Cross-Attention
-        norm_hidden_states = (
-            self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
-        )
-        hidden_states = self.attn2(norm_hidden_states, context=context) + hidden_states
+        if self.attn2 is not None:
+            # 2. Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+            )
+            hidden_states = self.attn2(norm_hidden_states, context=context) + hidden_states
 
         # 3. Feed-forward
         hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
@@ -666,14 +675,16 @@ def __init__(
         inner_dim = int(dim * mult)
         dim_out = dim_out if dim_out is not None else dim
 
-        if activation_fn == "geglu":
-            geglu = GEGLU(dim, inner_dim)
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim)
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim)
         elif activation_fn == "geglu-approximate":
-            geglu = ApproximateGELU(dim, inner_dim)
+            act_fn = ApproximateGELU(dim, inner_dim)
 
         self.net = nn.ModuleList([])
         # project in
-        self.net.append(geglu)
+        self.net.append(act_fn)
         # project dropout
         self.net.append(nn.Dropout(dropout))
         # project out
@@ -685,6 +696,27 @@ def forward(self, hidden_states):
         return hidden_states
 
 
+class GELU(nn.Module):
+    r"""
+    GELU activation function
+    """
+
+    def __init__(self, dim_in: int, dim_out: int):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out)
+
+    def gelu(self, gate):
+        if gate.device.type != "mps":
+            return F.gelu(gate)
+        # mps: gelu is not implemented for float16
+        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+
+    def forward(self, hidden_states):
+        hidden_states = self.proj(hidden_states)
+        hidden_states = self.gelu(hidden_states)
+        return hidden_states
+
+
 # feedforward
 class GEGLU(nn.Module):
     r"""

pipelines/__init__.py

@@ -28,6 +28,7 @@
 if is_torch_available() and is_transformers_available():
     from .alt_diffusion import AltDiffusionImg2ImgPipeline, AltDiffusionPipeline
     from .latent_diffusion import LDMTextToImagePipeline
+    from .paint_by_example import PaintByExamplePipeline
     from .stable_diffusion import (
         CycleDiffusionPipeline,
         StableDiffusionImageVariationPipeline,

pipelines/paint_by_example/__init__.py

@@ -0,0 +1,14 @@
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+
+import PIL
+from PIL import Image
+
+from ...utils import is_torch_available, is_transformers_available
+
+
+if is_transformers_available() and is_torch_available():
+    from .image_encoder import PaintByExampleImageEncoder
+    from .pipeline_paint_by_example import PaintByExamplePipeline

pipelines/paint_by_example/image_encoder.py

@@ -0,0 +1,65 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+from torch import nn
+
+from transformers import CLIPPreTrainedModel, CLIPVisionModel
+
+from ...models.attention import BasicTransformerBlock
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class PaintByExampleImageEncoder(CLIPPreTrainedModel):
+    def __init__(self, config, proj_size=768):
+        super().__init__(config)
+        self.proj_size = proj_size
+
+        self.model = CLIPVisionModel(config)
+        self.mapper = PaintByExampleMapper(config)
+        self.final_layer_norm = nn.LayerNorm(config.hidden_size)
+        self.proj_out = nn.Linear(config.hidden_size, self.proj_size)
+
+        # uncondition for scaling
+        self.uncond_vector = nn.Parameter(torch.rand((1, 1, self.proj_size)))
+
+    def forward(self, pixel_values):
+        clip_output = self.model(pixel_values=pixel_values)
+        latent_states = clip_output.pooler_output
+        latent_states = self.mapper(latent_states[:, None])
+        latent_states = self.final_layer_norm(latent_states)
+        latent_states = self.proj_out(latent_states)
+        return latent_states
+
+
+class PaintByExampleMapper(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        num_layers = (config.num_hidden_layers + 1) // 5
+        hid_size = config.hidden_size
+        num_heads = 1
+        self.blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(hid_size, num_heads, hid_size, activation_fn="gelu", attention_bias=True)
+                for _ in range(num_layers)
+            ]
+        )
+
+    def forward(self, hidden_states):
+        for block in self.blocks:
+            hidden_states = block(hidden_states)
+
+        return hidden_states