[Fix Issue #1197 since 2022] Support pre-trained openai/guided-diffusion (ADM) with minimal code change

scripts/convert_adm_to_diffusers.py

@@ -0,0 +1,102 @@
+import argparse
+import os
+
+import torch
+from convert_consistency_to_diffusers import con_pt_to_diffuser
+
+from diffusers import (
+    UNet2DModel,
+)
+
+
+SMALL_256_UNET_CONFIG = {
+    "sample_size": 256,
+    "in_channels": 3,
+    "out_channels": 6,
+    "layers_per_block": 1,
+    "num_class_embeds": None,
+    "block_out_channels": [128, 128, 128 * 2, 128 * 2, 128 * 4, 128 * 4],
+    "attention_head_dim": 64,
+    "down_block_types": [
+        "ResnetDownsampleBlock2D",
+        "ResnetDownsampleBlock2D",
+        "ResnetDownsampleBlock2D",
+        "ResnetDownsampleBlock2D",
+        "AttnDownBlock2D",
+        "ResnetDownsampleBlock2D",
+    ],
+    "up_block_types": [
+        "ResnetUpsampleBlock2D",
+        "AttnUpBlock2D",
+        "ResnetUpsampleBlock2D",
+        "ResnetUpsampleBlock2D",
+        "ResnetUpsampleBlock2D",
+        "ResnetUpsampleBlock2D",
+    ],
+    "resnet_time_scale_shift": "scale_shift",
+    "upsample_type": "resnet",
+    "downsample_type": "resnet",
+    "norm_eps": 1e-06,
+    "norm_num_groups": 32,
+}
+
+
+LARGE_256_UNET_CONFIG = {
+    "sample_size": 256,
+    "in_channels": 3,
+    "out_channels": 6,
+    "layers_per_block": 2,
+    "num_class_embeds": None,
+    "block_out_channels": [256, 256, 256 * 2, 256 * 2, 256 * 4, 256 * 4],
+    "attention_head_dim": 64,
+    "down_block_types": [
+        "ResnetDownsampleBlock2D",
+        "ResnetDownsampleBlock2D",
+        "ResnetDownsampleBlock2D",
+        "AttnDownBlock2D",
+        "AttnDownBlock2D",
+        "AttnDownBlock2D",
+    ],
+    "up_block_types": [
+        "AttnUpBlock2D",
+        "AttnUpBlock2D",
+        "AttnUpBlock2D",
+        "ResnetUpsampleBlock2D",
+        "ResnetUpsampleBlock2D",
+        "ResnetUpsampleBlock2D",
+    ],
+    "resnet_time_scale_shift": "scale_shift",
+    "upsample_type": "resnet",
+    "downsample_type": "resnet",
+    "norm_eps": 1e-06,
+    "norm_num_groups": 32,
+}
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--unet_path", default=None, type=str, required=True, help="Path to the unet.pt to convert.")
+    parser.add_argument(
+        "--dump_path", default=None, type=str, required=True, help="Path to output the converted UNet model."
+    )
+
+    args = parser.parse_args()
+
+    ckpt_name = os.path.basename(args.unet_path)
+    print(f"Checkpoint: {ckpt_name}")
+
+    # Get U-Net config
+    if "ffhq" in ckpt_name:
+        unet_config = SMALL_256_UNET_CONFIG
+    else:
+        unet_config = LARGE_256_UNET_CONFIG
+
+    unet_config["num_class_embeds"] = None
+
+    converted_unet_ckpt = con_pt_to_diffuser(args.unet_path, unet_config)
+
+    image_unet = UNet2DModel(**unet_config)
+    image_unet.load_state_dict(converted_unet_ckpt)
+
+    torch.save(converted_unet_ckpt, args.dump_path)

src/diffusers/models/attention_processor.py

@@ -84,6 +84,8 @@ class Attention(nn.Module):
         processor (`AttnProcessor`, *optional*, defaults to `None`):
             The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and
             `AttnProcessor` otherwise.
+        attention_legacy_order (`bool`, *optional*, defaults to `False`):
+            if attention_legacy_order, split heads before split qkv, see https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/unet.py#L328
     """
 
     def __init__(
@@ -110,6 +112,7 @@ def __init__(
         _from_deprecated_attn_block: bool = False,
         processor: Optional["AttnProcessor"] = None,
         out_dim: int = None,
+        attention_legacy_order: bool = False,
     ):
         super().__init__()
         self.inner_dim = out_dim if out_dim is not None else dim_head * heads
@@ -205,6 +208,7 @@ def __init__(
         # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
         # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
         # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+        self.attention_legacy_order = attention_legacy_order
         if processor is None:
             processor = (
                 AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
@@ -1221,6 +1225,7 @@ def __call__(
             hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
 
         args = () if USE_PEFT_BACKEND else (scale,)
+
         query = attn.to_q(hidden_states, *args)
 
         if encoder_hidden_states is None:
@@ -1234,11 +1239,16 @@ def __call__(
         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)
-
+        if attn.attention_legacy_order:
+            qkv = torch.cat([query, key, value], dim=2).transpose(1, 2)
+            query, key, value = qkv.reshape(batch_size, attn.heads, head_dim * 3, -1).chunk(3, dim=2)
+            query = query.transpose(-1, -2)
+            key = key.transpose(-1, -2)
+            value = value.transpose(-1, -2)
+        else:
+            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(

src/diffusers/models/embeddings.py

@@ -254,6 +254,38 @@ def forward(self, timesteps):
         return t_emb
 
 
+def timestep_embedding_adm(timesteps, dim, max_period=10000):
+    """
+    ADM order embedding from https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/nn.py#L103
+    """
+    half = dim // 2
+    freqs = 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)
+    return embedding
+
+
+class TimestepsADM(nn.Module):
+    """
+    ADM order embedding from https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/nn.py#L103
+    """
+
+    def __init__(self, num_channels: int):
+        super().__init__()
+        self.num_channels = num_channels
+
+    def forward(self, timesteps):
+        t_emb = timestep_embedding_adm(
+            timesteps,
+            self.num_channels,
+        )
+        return t_emb
+
+
 class GaussianFourierProjection(nn.Module):
     """Gaussian Fourier embeddings for noise levels."""
 

src/diffusers/models/unet_2d_blocks.py

@@ -65,6 +65,7 @@ def get_down_block(
     upcast_attention: bool = False,
     resnet_time_scale_shift: str = "default",
     attention_type: str = "default",
+    attention_legacy_order: bool = False,
     resnet_skip_time_act: bool = False,
     resnet_out_scale_factor: float = 1.0,
     cross_attention_norm: Optional[str] = None,
@@ -97,6 +98,7 @@ def get_down_block(
         upcast_attention=upcast_attention,
         resnet_time_scale_shift=resnet_time_scale_shift,
         attention_type=attention_type,
+        attention_legacy_order=attention_legacy_order,
         resnet_skip_time_act=resnet_skip_time_act,
         resnet_out_scale_factor=resnet_out_scale_factor,
         cross_attention_norm=cross_attention_norm,
@@ -202,6 +204,7 @@ def get_up_block(
     upcast_attention: bool = False,
     resnet_time_scale_shift: str = "default",
     attention_type: str = "default",
+    attention_legacy_order: bool = False,
     resnet_skip_time_act: bool = False,
     resnet_out_scale_factor: float = 1.0,
     cross_attention_norm: Optional[str] = None,
@@ -235,6 +238,7 @@ def get_up_block(
         upcast_attention=upcast_attention,
         resnet_time_scale_shift=resnet_time_scale_shift,
         attention_type=attention_type,
+        attention_legacy_order=attention_legacy_order,
         resnet_skip_time_act=resnet_skip_time_act,
         resnet_out_scale_factor=resnet_out_scale_factor,
         cross_attention_norm=cross_attention_norm,

src/diffusers/models/unets/unet_2d.py

@@ -19,7 +19,7 @@
 
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...utils import BaseOutput
-from ..embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
+from ..embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps, TimestepsADM
 from ..modeling_utils import ModelMixin
 from .unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block
 
@@ -72,6 +72,8 @@ class UNet2DModel(ModelMixin, ConfigMixin):
             The upsample type for upsampling layers. Choose between "conv" and "resnet"
         dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
         act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        attention_legacy_order (`bool`, *optional*, defaults to `False`):
+            if attention_legacy_order, split heads before split qkv, see https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/unet.py#L328
         attention_head_dim (`int`, *optional*, defaults to `8`): The attention head dimension.
         norm_num_groups (`int`, *optional*, defaults to `32`): The number of groups for normalization.
         attn_norm_num_groups (`int`, *optional*, defaults to `None`):
@@ -109,6 +111,7 @@ def __init__(
         upsample_type: str = "conv",
         dropout: float = 0.0,
         act_fn: str = "silu",
+        attention_legacy_order: bool = False,
         attention_head_dim: Optional[int] = 8,
         norm_num_groups: int = 32,
         attn_norm_num_groups: Optional[int] = None,
@@ -148,7 +151,9 @@ def __init__(
         elif time_embedding_type == "learned":
             self.time_proj = nn.Embedding(num_train_timesteps, block_out_channels[0])
             timestep_input_dim = block_out_channels[0]
-
+        elif time_embedding_type == "adm":
+            self.time_proj = TimestepsADM(block_out_channels[0])
+            timestep_input_dim = block_out_channels[0]
         self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
 
         # class embedding
@@ -182,6 +187,7 @@ def __init__(
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 resnet_groups=norm_num_groups,
+                attention_legacy_order=attention_legacy_order,
                 attention_head_dim=attention_head_dim if attention_head_dim is not None else output_channel,
                 downsample_padding=downsample_padding,
                 resnet_time_scale_shift=resnet_time_scale_shift,
@@ -191,6 +197,7 @@ def __init__(
             self.down_blocks.append(down_block)
 
         # mid
+        attn_norm_num_groups = norm_num_groups if attention_legacy_order is True else attn_norm_num_groups
         self.mid_block = UNetMidBlock2D(
             in_channels=block_out_channels[-1],
             temb_channels=time_embed_dim,
@@ -203,6 +210,7 @@ def __init__(
             resnet_groups=norm_num_groups,
             attn_groups=attn_norm_num_groups,
             add_attention=add_attention,
+            attention_legacy_order=attention_legacy_order,
         )
 
         # up
@@ -226,6 +234,7 @@ def __init__(
                 resnet_eps=norm_eps,
                 resnet_act_fn=act_fn,
                 resnet_groups=norm_num_groups,
+                attention_legacy_order=attention_legacy_order,
                 attention_head_dim=attention_head_dim if attention_head_dim is not None else output_channel,
                 resnet_time_scale_shift=resnet_time_scale_shift,
                 upsample_type=upsample_type,