genphoto/models/attention.py

# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py

from dataclasses import dataclass
from typing import Optional

import torch
from torch import nn

from ddiffusers.configuration_utils import ConfigMixin, register_to_config
from ddiffusers.models.modeling_utils import ModelMixin
from ddiffusers.utils import BaseOutput
from ddiffusers.models.attention import BasicTransformerBlock
from einops import rearrange, repeat


@dataclass
class Transformer3DModelOutput(BaseOutput):
    sample: torch.FloatTensor


class Transformer3DModel(ModelMixin, ConfigMixin):
    @register_to_config
    def __init__(
            self,
            num_attention_heads: int = 16,
            attention_head_dim: int = 88,
            in_channels: Optional[int] = None,
            num_layers: int = 1,
            dropout: float = 0.0,
            norm_num_groups: int = 32,
            cross_attention_dim: Optional[int] = None,
            attention_bias: bool = False,
            activation_fn: str = "geglu",
            num_embeds_ada_norm: Optional[int] = None,
            use_linear_projection: bool = False,
            only_cross_attention: bool = False,
            upcast_attention: bool = False,
            norm_type: str = "layer_norm",
            norm_elementwise_affine: bool = True,
    ):
        super().__init__()
        self.use_linear_projection = use_linear_projection
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        inner_dim = num_attention_heads * attention_head_dim

        # Define input layers
        self.in_channels = in_channels

        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
        if use_linear_projection:
            self.proj_in = nn.Linear(in_channels, inner_dim)
        else:
            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)

        # Define transformers blocks
        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(
                    inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    dropout=dropout,
                    cross_attention_dim=cross_attention_dim,
                    activation_fn=activation_fn,
                    num_embeds_ada_norm=num_embeds_ada_norm,
                    attention_bias=attention_bias,
                    only_cross_attention=only_cross_attention,
                    upcast_attention=upcast_attention,
                    norm_type=norm_type,
                    norm_elementwise_affine=norm_elementwise_affine,
                )
                for d in range(num_layers)
            ]
        )

        # 4. Define output layers
        if use_linear_projection:
            self.proj_out = nn.Linear(in_channels, inner_dim)
        else:
            self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)

    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True):
        # Input
        assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
        batch_size, _, video_length = hidden_states.shape[:3]
        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")

        if encoder_hidden_states.shape[0] == batch_size:
            encoder_hidden_states = repeat(encoder_hidden_states, 'b n c -> (b f) n c', f=video_length)

        elif encoder_hidden_states.shape[0] == batch_size * video_length:
            pass
        else:
            raise ValueError

        batch, channel, height, weight = hidden_states.shape
        residual = hidden_states

        hidden_states = self.norm(hidden_states)
        if not self.use_linear_projection:
            hidden_states = self.proj_in(hidden_states)
            inner_dim = hidden_states.shape[1]
            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
        else:
            inner_dim = hidden_states.shape[1]
            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
            hidden_states = self.proj_in(hidden_states)

        # Blocks
        for block in self.transformer_blocks:
            hidden_states = block(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                timestep=timestep,
            )

        # Output
        if not self.use_linear_projection:
            hidden_states = (
                hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
            )
            hidden_states = self.proj_out(hidden_states)
        else:
            hidden_states = self.proj_out(hidden_states)
            hidden_states = (
                hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
            )

        output = hidden_states + residual

        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
        if not return_dict:
            return (output,)

        return Transformer3DModelOutput(sample=output)