models/hyvideo/modules/embed_layers.py

import math
import torch
import torch.nn as nn
from einops import rearrange, repeat

from ..utils.helpers import to_2tuple


class PatchEmbed(nn.Module):
    """2D Image to Patch Embedding

    Image to Patch Embedding using Conv2d

    A convolution based approach to patchifying a 2D image w/ embedding projection.

    Based on the impl in https://github.com/google-research/vision_transformer

    Hacked together by / Copyright 2020 Ross Wightman

    Remove the _assert function in forward function to be compatible with multi-resolution images.
    """

    def __init__(
        self,
        patch_size=16,
        in_chans=3,
        embed_dim=768,
        norm_layer=None,
        flatten=True,
        bias=True,
        dtype=None,
        device=None,
    ):
        factory_kwargs = {"dtype": dtype, "device": device}
        super().__init__()
        patch_size = to_2tuple(patch_size)
        self.patch_size = patch_size
        self.flatten = flatten

        self.proj = nn.Conv3d(
            in_chans,
            embed_dim,
            kernel_size=patch_size,
            stride=patch_size,
            bias=bias,
            **factory_kwargs
        )
        nn.init.xavier_uniform_(self.proj.weight.view(self.proj.weight.size(0), -1))
        if bias:
            nn.init.zeros_(self.proj.bias)

        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

    def forward(self, x):
        x = self.proj(x)
        shape = x.shape
        if self.flatten:
            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
        x = self.norm(x)
        return x, shape

class VisionProjection(torch.nn.Module):

    def __init__(self, input_dim, output_dim):
        super().__init__()

        self.proj = torch.nn.Sequential(
            torch.nn.LayerNorm(input_dim), 
            torch.nn.Linear(input_dim, input_dim),
            torch.nn.GELU(), 
            torch.nn.Linear(input_dim, output_dim),
            torch.nn.LayerNorm(output_dim)
        )
        

    def forward(self, vision_embeds):
        return self.proj(vision_embeds)

class TextProjection(nn.Module):
    """
    Projects text embeddings. Also handles dropout for classifier-free guidance.

    Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
    """

    def __init__(self, in_channels, hidden_size, act_layer, dtype=None, device=None):
        factory_kwargs = {"dtype": dtype, "device": device}
        super().__init__()
        self.linear_1 = nn.Linear(
            in_features=in_channels,
            out_features=hidden_size,
            bias=True,
            **factory_kwargs
        )
        self.act_1 = act_layer()
        self.linear_2 = nn.Linear(
            in_features=hidden_size,
            out_features=hidden_size,
            bias=True,
            **factory_kwargs
        )

    def forward(self, caption):
        hidden_states = self.linear_1(caption)
        hidden_states = self.act_1(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states


def timestep_embedding(t, dim, max_period=10000):
    """
    Create sinusoidal timestep embeddings.

    Args:
        t (torch.Tensor): a 1-D Tensor of N indices, one per batch element. These may be fractional.
        dim (int): the dimension of the output.
        max_period (int): controls the minimum frequency of the embeddings.

    Returns:
        embedding (torch.Tensor): An (N, D) Tensor of positional embeddings.

    .. ref_link: https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
    """
    half = dim // 2
    freqs = torch.exp(
        -math.log(max_period)
        * torch.arange(start=0, end=half, dtype=torch.float32)
        / half
    ).to(device=t.device)
    args = t[:, 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 TimestepEmbedder(nn.Module):
    """
    Embeds scalar timesteps into vector representations.
    """

    def __init__(
        self,
        hidden_size,
        act_layer,
        frequency_embedding_size=256,
        max_period=10000,
        out_size=None,
        dtype=None,
        device=None,
    ):
        factory_kwargs = {"dtype": dtype, "device": device}
        super().__init__()
        self.frequency_embedding_size = frequency_embedding_size
        self.max_period = max_period
        if out_size is None:
            out_size = hidden_size

        self.mlp = nn.Sequential(
            nn.Linear(
                frequency_embedding_size, hidden_size, bias=True, **factory_kwargs
            ),
            act_layer(),
            nn.Linear(hidden_size, out_size, bias=True, **factory_kwargs),
        )
        nn.init.normal_(self.mlp[0].weight, std=0.02)
        nn.init.normal_(self.mlp[2].weight, std=0.02)

    def forward(self, t):
        t_freq = timestep_embedding(
            t, self.frequency_embedding_size, self.max_period
        ).type(self.mlp[0].weight.dtype)
        t_emb = self.mlp(t_freq)
        return t_emb