pixart.py

# This code was retrieved on 24/06/2024 from the repository bellow.
# https://github.com/PixArt-alpha/PixArt-alpha/tree/master/diffusion/model/nets

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# GLIDE: https://github.com/openai/glide-text2im
# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py
# --------------------------------------------------------
import math
from typing import Iterable
from itertools import repeat
import torch
import torch.nn as nn
import numpy as np
from einops import rearrange
from timm.models.vision_transformer import PatchEmbed, Mlp, Attention as Attention_
from timm.models.layers import DropPath
import xformers.ops

def _ntuple(n):
    def parse(x):
        if isinstance(x, Iterable) and not isinstance(x, str):
            return x
        return tuple(repeat(x, n))
    return parse

to_1tuple = _ntuple(1)
to_2tuple = _ntuple(2)

def modulate(x, shift, scale):
    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)


def t2i_modulate(x, shift, scale):
    return x * (1 + scale) + shift

import sys
class MultiHeadCrossAttention(nn.Module):
    def __init__(self, d_model, num_heads, attn_drop=0., proj_drop=0., **block_kwargs):
        super(MultiHeadCrossAttention, self).__init__()
        assert d_model % num_heads == 0, "d_model must be divisible by num_heads"

        self.d_model = d_model
        self.num_heads = num_heads
        self.head_dim = d_model // num_heads

        self.q_linear = nn.Linear(d_model, d_model)
        self.kv_linear = nn.Linear(d_model, d_model*2)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(d_model, d_model)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x, cond, mask=None):
        # query: img tokens; key/value: condition; mask: if padding tokens
        B, N, C = x.shape

        q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
        kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
        k, v = kv.unbind(2)
        attn_bias = None
        if mask is not None:
            attn_bias = xformers.ops.fmha.BlockDiagonalMask.from_seqlens([N] * B, mask)
        x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)
        x = x.view(B, -1, C)
        x = self.proj(x)
        x = self.proj_drop(x)

        return x


class WindowAttention(Attention_):
    """Multi-head Attention block with relative position embeddings."""

    def __init__(
        self,
        dim,
        num_heads=8,
        qkv_bias=True,
        use_rel_pos=False,
        rel_pos_zero_init=True,
        input_size=None,
        **block_kwargs,
    ):
        """
        Args:
            dim (int): Number of input channels.
            num_heads (int): Number of attention heads.
            qkv_bias (bool:  If True, add a learnable bias to query, key, value.
            rel_pos (bool): If True, add relative positional embeddings to the attention map.
            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
            input_size (int or None): Input resolution for calculating the relative positional
                parameter size.
        """
        super().__init__(dim, num_heads=num_heads, qkv_bias=qkv_bias, **block_kwargs)

        self.use_rel_pos = use_rel_pos
        if self.use_rel_pos:
            # initialize relative positional embeddings
            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, self.head_dim))
            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, self.head_dim))

            if not rel_pos_zero_init:
                nn.init.trunc_normal_(self.rel_pos_h, std=0.02)
                nn.init.trunc_normal_(self.rel_pos_w, std=0.02)

    def forward(self, x, mask=None):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
        q, k, v = qkv.unbind(2)
        if use_fp32_attention := getattr(self, 'fp32_attention', False):
            q, k, v = q.float(), k.float(), v.float()

        attn_bias = None
        if mask is not None:
            attn_bias = torch.zeros([B * self.num_heads, q.shape[1], k.shape[1]], dtype=q.dtype, device=q.device)
            attn_bias.masked_fill_(mask.squeeze(1).repeat(self.num_heads, 1, 1) == 0, float('-inf'))
        x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)

        x = x.view(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


#################################################################################
#   AMP attention with fp32 softmax to fix loss NaN problem during training     #
#################################################################################
class Attention(Attention_):
    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)  # make torchscript happy (cannot use tensor as tuple)
        use_fp32_attention = getattr(self, 'fp32_attention', False)
        if use_fp32_attention:
            q, k = q.float(), k.float()
        with torch.cuda.amp.autocast(enabled=not use_fp32_attention):
            attn = (q @ k.transpose(-2, -1)) * self.scale
            attn = attn.softmax(dim=-1)

        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class FinalLayer(nn.Module):
    """
    The final layer of PixArt.
    """

    def __init__(self, hidden_size, patch_size, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 2 * hidden_size, bias=True)
        )

    def forward(self, x, c):
        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
        x = modulate(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x


class T2IFinalLayer(nn.Module):
    """
    The final layer of PixArt.
    """

    def __init__(self, hidden_size, patch_size, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
        self.scale_shift_table = nn.Parameter(torch.randn(2, hidden_size) / hidden_size ** 0.5)
        self.out_channels = out_channels

    def forward(self, x, t):
        shift, scale = (self.scale_shift_table[None] + t[:, None]).chunk(2, dim=1)
        x = t2i_modulate(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x


class MaskFinalLayer(nn.Module):
    """
    The final layer of PixArt.
    """

    def __init__(self, final_hidden_size, c_emb_size, patch_size, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(final_hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(final_hidden_size, patch_size * patch_size * out_channels, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(c_emb_size, 2 * final_hidden_size, bias=True)
        )
    def forward(self, x, t):
        shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
        x = modulate(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x


class DecoderLayer(nn.Module):
    """
    The final layer of PixArt.
    """

    def __init__(self, hidden_size, decoder_hidden_size):
        super().__init__()
        self.norm_decoder = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, decoder_hidden_size, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 2 * hidden_size, bias=True)
        )
    def forward(self, x, t):
        shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
        x = modulate(self.norm_decoder(x), shift, scale)
        x = self.linear(x)
        return x

class TimestepEmbedder(nn.Module):
    """
    Embeds scalar timesteps into vector representations.
    """

    def __init__(self, hidden_size, frequency_embedding_size=256):
        super().__init__()
        self.mlp = nn.Sequential(
            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size, bias=True),
        )
        self.frequency_embedding_size = frequency_embedding_size

    @staticmethod
    def timestep_embedding(t, dim, max_period=10000):
        """
        Create sinusoidal timestep embeddings.
        :param t: a 1-D Tensor of N indices, one per batch element.
                          These may be fractional.
        :param dim: the dimension of the output.
        :param max_period: controls the minimum frequency of the embeddings.
        :return: an (N, D) Tensor of positional embeddings.
        """
        # 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, device=t.device) / half)
        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

    def forward(self, t):
        t_freq = self.timestep_embedding(t, self.frequency_embedding_size).to(self.dtype)
        return self.mlp(t_freq)

    @property
    def dtype(self):
        return next(self.parameters()).dtype


class SizeEmbedder(TimestepEmbedder):
    """
    Embeds scalar timesteps into vector representations.
    """

    def __init__(self, hidden_size, frequency_embedding_size=256):
        super().__init__(hidden_size=hidden_size, frequency_embedding_size=frequency_embedding_size)
        self.mlp = nn.Sequential(
            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size, bias=True),
        )
        self.frequency_embedding_size = frequency_embedding_size
        self.outdim = hidden_size

    def forward(self, s, bs):
        if s.ndim == 1:
            s = s[:, None]
        assert s.ndim == 2
        if s.shape[0] != bs:
            s = s.repeat(bs//s.shape[0], 1)
            assert s.shape[0] == bs
        b, dims = s.shape[0], s.shape[1]
        s = rearrange(s, "b d -> (b d)")
        s_freq = self.timestep_embedding(s, self.frequency_embedding_size).to(self.dtype)
        s_emb = self.mlp(s_freq)
        s_emb = rearrange(s_emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim)
        return s_emb

    @property
    def dtype(self):
        return next(self.parameters()).dtype



class PixArtBlock(nn.Module):
    """
    A PixArt block with adaptive layer norm (adaLN-single) conditioning.
    """

    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., window_size=0, input_size=None, use_rel_pos=False, **block_kwargs):
        super().__init__()
        self.hidden_size = hidden_size
        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = WindowAttention(hidden_size, num_heads=num_heads, qkv_bias=True,
                                    input_size=input_size if window_size == 0 else (window_size, window_size),
                                    use_rel_pos=use_rel_pos, **block_kwargs)
        self.cross_attn = MultiHeadCrossAttention(hidden_size, num_heads, **block_kwargs)
        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        # to be compatible with lower version pytorch
        approx_gelu = lambda: nn.GELU(approximate="tanh")
        self.mlp = Mlp(in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.window_size = window_size
        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)

    def forward(self, x, y, t, mask=None, **kwargs):
        B, N, C = x.shape

        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None] + t.reshape(B, 6, -1)).chunk(6, dim=1)
        x = x + self.drop_path(gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)).reshape(B, N, C))
        x = x + self.cross_attn(x, y, mask)
        x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))

        return x


class ImgEmbedder(nn.Module):
    """
    Embeds imgs or latents to vector representations. Also handles label dropout for classifier-free guidance.
    """
    def __init__(self, cond_img_size, cond_channels, hidden_size, uncond_prob, patch_size=4,
                 act_layer=nn.GELU(approximate='tanh'), ):
        super().__init__()
        self.to_patch = PatchEmbed(img_size=cond_img_size, patch_size=patch_size, in_chans=cond_channels,
                                   embed_dim=hidden_size)
        self.y_proj = Mlp(in_features=hidden_size, hidden_features=hidden_size,
                          out_features=hidden_size, act_layer=act_layer, drop=0)
        #self.register_buffer("y_embedding", nn.Parameter(torch.randn(token_num, in_channels) / in_channels ** 0.5))
        self.uncond_prob = uncond_prob

    def token_drop(self, caption, force_drop_ids=None):
        """
        Drops labels to enable classifier-free guidance.
        """
        if force_drop_ids is None:
            drop_ids = torch.rand(caption.shape[0]).cuda() < self.uncond_prob
        else:
            drop_ids = force_drop_ids == 1
        caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
        return caption

    def forward(self, x, train, force_drop_ids=None):
        use_dropout = self.uncond_prob > 0
        # if (train and use_dropout) or (force_drop_ids is not None):
        #     caption = self.token_drop(caption, force_drop_ids)
        patches = self.to_patch(x)[:,None]
        return self.y_proj(patches)

class PixArtImgEmbCrossAttn(nn.Module):
    """
    Diffusion model with a Transformer backbone.
    """

    def __init__(self, input_size=32, patch_size=2, in_channels=4, hidden_size=1152, depth=28, num_heads=16, mlp_ratio=4.0, class_dropout_prob=0.1, pred_sigma=True, drop_path: float = 0., window_size=0, window_block_indexes=None, use_rel_pos=False, cond_channels=4, cond_img_size=32, lewei_scale=1.0, img_emb_patch_size=4, **kwargs):
        if window_block_indexes is None:
            window_block_indexes = []
        super().__init__()
        self.pred_sigma = pred_sigma
        self.in_channels = in_channels
        self.out_channels = in_channels * 2 if pred_sigma else in_channels
        self.patch_size = patch_size
        self.num_heads = num_heads
        self.lewei_scale = lewei_scale,

        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
        self.t_embedder = TimestepEmbedder(hidden_size)
        num_patches = self.x_embedder.num_patches
        self.base_size = input_size // self.patch_size
        # Will use fixed sin-cos embedding:
        self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))

        approx_gelu = lambda: nn.GELU(approximate="tanh")
        self.t_block = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )
        self.y_embedder = ImgEmbedder(cond_channels=cond_channels, cond_img_size=cond_img_size,
                                    hidden_size=hidden_size, uncond_prob=class_dropout_prob,
                                      act_layer=approx_gelu, patch_size=img_emb_patch_size)
        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
        self.blocks = nn.ModuleList([
            PixArtBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
                          input_size=(input_size // patch_size, input_size // patch_size),
                          window_size=window_size if i in window_block_indexes else 0,
                          use_rel_pos=use_rel_pos if i in window_block_indexes else False)
            for i in range(depth)
        ])
        self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)

        self.initialize_weights()

    def forward(self, x, timestep, y, mask=None, data_info=None, **kwargs):
        """
        Forward pass of PixArt.
        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        t: (N,) tensor of diffusion timesteps
        y: (N, C_cond, H_cond, W_cond) tensor of spatial inputs (images or latent representations of images)
        """
        x = x.to(self.dtype)
        timestep = timestep.to(self.dtype)
        y = y.to(self.dtype)
        pos_embed = self.pos_embed.to(self.dtype)
        self.h, self.w = x.shape[-2]//self.patch_size, x.shape[-1]//self.patch_size
        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
        t = self.t_embedder(timestep.to(x.dtype))  # (N, D)
        t0 = self.t_block(t)
        y = self.y_embedder(y, self.training)  # (N, 1, L, D)
        if mask is not None:
            if mask.shape[0] != y.shape[0]:
                mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
            mask = mask.squeeze(1).squeeze(1)
            y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
            y_lens = mask.sum(dim=1).tolist()
        else:
            y_lens = [y.shape[2]] * y.shape[0]
            y = y.squeeze(1).view(1, -1, x.shape[-1])
        for block in self.blocks:
            x = block(x, y, t0, y_lens)  # (N, T, D)
        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
        x = self.unpatchify(x)  # (N, out_channels, H, W)
        return x

    def forward_with_dpmsolver(self, x, timestep, y, mask=None, **kwargs):
        """
        dpm solver donnot need variance prediction
        """
        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
        model_out = self.forward(x, timestep, y, mask)
        return model_out.chunk(2, dim=1)[0]

    def forward_with_cfg(self, x, timestep, y, cfg_scale, mask=None, **kwargs):
        """
        Forward pass of PixArt, but also batches the unconditional forward pass for classifier-free guidance.
        """
        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
        half = x[: len(x) // 2]
        combined = torch.cat([half, half], dim=0)
        model_out = self.forward(combined, timestep, y, mask, kwargs)
        model_out = model_out['x'] if isinstance(model_out, dict) else model_out
        eps, rest = model_out[:, :3], model_out[:, 3:]
        cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
        half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
        eps = torch.cat([half_eps, half_eps], dim=0)
        return torch.cat([eps, rest], dim=1)

    def unpatchify(self, x):
        """
        x: (N, T, patch_size**2 * C)
        imgs: (N, H, W, C)
        """
        c = self.out_channels
        p = self.x_embedder.patch_size[0]
        h = w = int(x.shape[1] ** 0.5)
        assert h * w == x.shape[1]

        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
        x = torch.einsum('nhwpqc->nchpwq', x)
        return x.reshape(shape=(x.shape[0], c, h * p, h * p))

    def initialize_weights(self):
        # Initialize transformer layers:
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)

        self.apply(_basic_init)

        # Initialize (and freeze) pos_embed by sin-cos embedding:
        pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches ** 0.5), lewei_scale=self.lewei_scale, base_size=self.base_size)
        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))

        # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
        w = self.x_embedder.proj.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))

        # Initialize timestep embedding MLP:
        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
        nn.init.normal_(self.t_block[1].weight, std=0.02)

        # Initialize caption embedding MLP:
        nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02)
        nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02)

        # Zero-out adaLN modulation layers in PixArt blocks:
        for block in self.blocks:
            nn.init.constant_(block.cross_attn.proj.weight, 0)
            nn.init.constant_(block.cross_attn.proj.bias, 0)

        # Zero-out output layers:
        nn.init.constant_(self.final_layer.linear.weight, 0)
        nn.init.constant_(self.final_layer.linear.bias, 0)

    @property
    def dtype(self):
        return next(self.parameters()).dtype


def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0, lewei_scale=1.0, base_size=16):
    """
    grid_size: int of the grid height and width
    return:
    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    if isinstance(grid_size, int):
        grid_size = to_2tuple(grid_size)
    grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0]/base_size) / lewei_scale
    grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1]/base_size) / lewei_scale
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)
    grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])

    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if cls_token and extra_tokens > 0:
        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
    return pos_embed


def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
    assert embed_dim % 2 == 0

    # use half of dimensions to encode grid_h
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)

    return np.concatenate([emb_h, emb_w], axis=1)


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position
    pos: a list of positions to be encoded: size (M,)
    out: (M, D)
    """
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=np.float64)
    omega /= embed_dim / 2.
    omega = 1. / 10000 ** omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

    emb_sin = np.sin(out)  # (M, D/2)
    emb_cos = np.cos(out)  # (M, D/2)

    return np.concatenate([emb_sin, emb_cos], axis=1)

class PixArtBlockSA(PixArtBlock):
    """
    A PixArt block with adaptive layer norm (adaLN-single) conditioning.
    And Self-Attention
    """

    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., window_size=0, input_size=None, use_rel_pos=False, **block_kwargs):
        super().__init__(hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path, window_size=window_size, input_size=input_size, use_rel_pos=use_rel_pos, **block_kwargs)

    def forward(self, x, t, mask=None, **kwargs):
        B, N, C = x.shape

        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None] + t.reshape(B, 6, -1)).chunk(6, dim=1)
        x = x + self.drop_path(gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)).reshape(B, N, C))
        x = x + self.cross_attn(x, x, mask)
        x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))

        return x

class PixArtCat(nn.Module):
    """
    Diffusion model with a Transformer backbone.
    """

    def __init__(self, input_size=32, patch_size=2, in_channels=4, out_channels=4, hidden_size=1152, depth=28, num_heads=16, mlp_ratio=4.0, class_dropout_prob=0.1, pred_sigma=True, drop_path: float = 0., window_size=0, window_block_indexes=None, use_rel_pos=False, lewei_scale=1.0, img_emb_patch_size=4, **kwargs):
        if window_block_indexes is None:
            window_block_indexes = []
        super().__init__()
        self.pred_sigma = pred_sigma
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.patch_size = patch_size
        self.num_heads = num_heads
        self.lewei_scale = lewei_scale,

        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
        self.t_embedder = TimestepEmbedder(hidden_size)
        num_patches = self.x_embedder.num_patches
        self.base_size = input_size // self.patch_size
        # Will use fixed sin-cos embedding:
        self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))

        approx_gelu = lambda: nn.GELU(approximate="tanh")
        self.t_block = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )

        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
        self.blocks = nn.ModuleList([
            PixArtBlockSA(hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
                          input_size=(input_size // patch_size, input_size // patch_size),
                          window_size=window_size if i in window_block_indexes else 0,
                          use_rel_pos=use_rel_pos if i in window_block_indexes else False)
            for i in range(depth)
        ])
        self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)

        self.initialize_weights()

    def forward(self, x, timestep, y, mask=None, data_info=None, **kwargs):
        """
        Forward pass of PixArt.
        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        t: (N,) tensor of diffusion timesteps
        y: (N, C_cond, H_cond, W_cond) tensor of spatial inputs (images or latent representations of images)
        """
        x = x.to(self.dtype)
        y = y.to(self.dtype)
        x = torch.cat([x, y], dim=1)

        timestep = timestep.to(self.dtype)
        pos_embed = self.pos_embed.to(self.dtype)
        self.h, self.w = x.shape[-2]//self.patch_size, x.shape[-1]//self.patch_size
        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
        t = self.t_embedder(timestep.to(x.dtype))  # (N, D)
        t0 = self.t_block(t)
        
        for block in self.blocks:
            x = block(x, t0, None)  # (N, T, D)
        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
        x = self.unpatchify(x)  # (N, out_channels, H, W)
        return x
    
    def unpatchify(self, x):
        """
        x: (N, T, patch_size**2 * C)
        imgs: (N, H, W, C)
        """
        c = self.out_channels
        p = self.x_embedder.patch_size[0]
        h = w = int(x.shape[1] ** 0.5)
        assert h * w == x.shape[1]

        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
        x = torch.einsum('nhwpqc->nchpwq', x)
        return x.reshape(shape=(x.shape[0], c, h * p, h * p))

    def initialize_weights(self):
        # Initialize transformer layers:
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)

        self.apply(_basic_init)

        # Initialize (and freeze) pos_embed by sin-cos embedding:
        pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches ** 0.5), lewei_scale=self.lewei_scale, base_size=self.base_size)
        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))

        # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
        w = self.x_embedder.proj.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))

        # Initialize timestep embedding MLP:
        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
        nn.init.normal_(self.t_block[1].weight, std=0.02)

        # Zero-out adaLN modulation layers in PixArt blocks:
        for block in self.blocks:
            nn.init.constant_(block.cross_attn.proj.weight, 0)
            nn.init.constant_(block.cross_attn.proj.bias, 0)

        # Zero-out output layers:
        nn.init.constant_(self.final_layer.linear.weight, 0)
        nn.init.constant_(self.final_layer.linear.bias, 0)

    @property
    def dtype(self):
        return next(self.parameters()).dtype