dit.py

import torch
import torch.nn as nn
import numpy as np
import math
from timm.models.vision_transformer import PatchEmbed, Mlp
from timm.models.vision_transformer import Attention
import torch.nn.functional as F
from einops import repeat, pack, unpack, rearrange


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


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) / 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

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


class IntervalTimestepEmbedder(nn.Module):
    """
    Modified embedder that can handle both single timesteps and intervals [r, t].
    """
    def __init__(self, hidden_size, frequency_embedding_size=256):
        super().__init__()
        
        self.single_embedder = TimestepEmbedder(hidden_size, frequency_embedding_size)

        # Additional embedder for interval information
        self.interval_embedder = nn.Sequential(
            nn.Linear(2 * hidden_size, hidden_size),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size)
        )
        
        # Mode flag
        self.interval_mode = False
        
    def forward(self, r_or_t, t=None):
        """
        Flexible forward that handles both modes:
        - Single timestep mode: forward(t) where r_or_t is the timestep
        - Interval mode: forward(r, t) where r_or_t is r and t is provided
        """
       
        # Interval mode (SplitMeanFlow)
        r_emb = self.single_embedder(r_or_t)
        t_emb = self.single_embedder(t)
        # Combine r and t embeddings
        combined = torch.cat([r_emb, t_emb], dim=-1)
        interval_emb = self.interval_embedder(combined)
        return interval_emb

class LabelEmbedder(nn.Module):
    def __init__(self, num_classes, dim, dropout_prob):
        super().__init__()
        use_cfg_embedding = dropout_prob > 0
        self.embedding = nn.Embedding(num_classes + use_cfg_embedding, dim)
        self.num_classes = num_classes
        self.dropout_prob = dropout_prob

    def token_drop(self, labels, force_drop_ids=None):
        if force_drop_ids is None:
            drop_ids = (
                torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
            )
        else:
            drop_ids = force_drop_ids == 1
        labels = torch.where(drop_ids, self.num_classes, labels)
        return labels

    def forward(self, labels, train, force_drop_ids=None):
        use_dropout = self.dropout_prob > 0
        if (train and use_dropout) or (force_drop_ids is not None):
            labels = self.token_drop(labels, force_drop_ids)
        embeddings = self.embedding(labels)
        return embeddings


class RMSNorm(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.scale = dim**0.5
        self.g = nn.Parameter(torch.ones(1))

    def forward(self, x):
        return F.normalize(x, dim=-1) * self.scale * self.g

class DiTBlock(nn.Module):
    def __init__(self, dim, num_heads, mlp_ratio=4.0):
        super().__init__()
        self.norm1 = RMSNorm(dim)
        self.attn = Attention(dim, num_heads=num_heads, qkv_bias=True, qk_norm=True, norm_layer=RMSNorm)
        self.norm2 = RMSNorm(dim)
        mlp_dim = int(dim * mlp_ratio)
        approx_gelu = lambda: nn.GELU(approximate="tanh")
        self.mlp = Mlp(
            in_features=dim, hidden_features=mlp_dim, act_layer=approx_gelu, drop=0
        )
        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim, 6 * dim))

    def forward(self, x, c):
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
            self.adaLN_modulation(c).chunk(6, dim=-1)
        )
        x = x + gate_msa.unsqueeze(1) * self.attn(
            modulate(self.norm1(x), scale_msa, shift_msa)
        )
        x = x + gate_mlp.unsqueeze(1) * self.mlp(
            modulate(self.norm2(x), scale_mlp, shift_mlp)
        )
        return x


class FinalLayer(nn.Module):
    def __init__(self, dim, patch_size, out_dim):
        super().__init__()
        self.norm_final = RMSNorm(dim)
        self.linear = nn.Linear(dim, patch_size * patch_size * out_dim)
        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim, 2 * dim))

    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 DiT(nn.Module):
    def __init__(
        self,
        input_size=32,
        patch_size=2,
        in_channels=4,
        dim=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        num_register_tokens=4,
        class_dropout_prob=0.1,
        num_classes=1000,
        learn_sigma=True,
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = in_channels
        self.out_channels = in_channels * 2 if learn_sigma else in_channels
        self.patch_size = patch_size
        self.num_heads = num_heads
        self.num_classes = num_classes
        
        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, dim)
        self.t_embedder = TimestepEmbedder(dim)
        
        self.use_cond = num_classes is not None
        self.y_embedder = LabelEmbedder(num_classes, dim, class_dropout_prob) if self.use_cond else None
        
        num_patches = self.x_embedder.num_patches
        # Will use fixed sin-cos embedding:
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, dim), requires_grad=False)
        
        self.blocks = nn.ModuleList([
            DiTBlock(dim, num_heads, mlp_ratio) for _ in range(depth)
        ])
        self.final_layer = FinalLayer(dim, patch_size, self.out_channels)
        
        self.initialize_weights()
    
    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))
        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]))
        nn.init.constant_(self.x_embedder.proj.bias, 0)

        # Initialize label embedding table:
        nn.init.normal_(self.y_embedder.embedding.weight, std=0.02)

        # 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)

        

        # Zero-out adaLN modulation layers in DiT blocks:
        for block in self.blocks:
            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)

        # Zero-out output layers:
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
        nn.init.constant_(self.final_layer.linear.weight, 0)
        nn.init.constant_(self.final_layer.linear.bias, 0)
    
    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)
        imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))
        return imgs

    def forward(self, x, t, y, is_train=False):
        """
        Forward pass of DiT.
        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        t: (N,) tensor of diffusion timesteps
        y: (N,) tensor of class labels
        """
        x = self.x_embedder(x) + self.pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
        t = self.t_embedder(t)                   # (N, D)
        y = self.y_embedder(y, self.training)    # (N, D)
        c = t + y                                # (N, D)
        for block in self.blocks:
            x = block(x, c)
        
        x = self.final_layer(x, c)                # (N, T, patch_size ** 2 * out_channels)
        x = self.unpatchify(x)                   # (N, out_channels, H, W)
        return x

    def forward_with_cfg(self, x, t, y, cfg_scale, is_train_student=False):
        x = x.repeat(2, 1, 1, 1)
        if is_train_student:
            t = t.repeat(2)
        else:
            t = t.repeat(x.shape[0])
        y = y.repeat(2)
        y[len(x) // 2:] = self.y_embedder.num_classes
        print('x.shape', x.shape)
        print('t.shape', t.shape)
        print('y.shape', y.shape)
        model_out = self.forward(x, t, y)
        cond_eps, uncond_eps = model_out.split(len(x) // 2)
        out = uncond_eps + (cond_eps - uncond_eps) * cfg_scale
        return out



class SMDiT(nn.Module):
    def __init__(
        self,
        input_size=32,
        patch_size=2,
        in_channels=4,
        dim=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        num_register_tokens=4,
        class_dropout_prob=0.1,
        num_classes=1000,
        learn_sigma=True,
        interval_mode=True,  # New parameter to enable interval mode
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = in_channels
        self.out_channels = in_channels * 2 if learn_sigma else in_channels
        self.patch_size = patch_size
        self.num_heads = num_heads
        self.num_classes = num_classes
        self.interval_mode = interval_mode
        
        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, dim)
        
        # Use the new interval-aware embedder
        self.t_embedder = IntervalTimestepEmbedder(dim)
        
        self.use_cond = num_classes is not None
        self.y_embedder = LabelEmbedder(num_classes, dim, class_dropout_prob) if self.use_cond else None
        
        num_patches = self.x_embedder.num_patches
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, dim), requires_grad=False)
        
        self.blocks = nn.ModuleList([
            DiTBlock(dim, num_heads, mlp_ratio) for _ in range(depth)
        ])
        self.final_layer = FinalLayer(dim, patch_size, self.out_channels)
        
        self.initialize_weights()
    
    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))
        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]))
        nn.init.constant_(self.x_embedder.proj.bias, 0)

        # Initialize label embedding table:
        if self.y_embedder is not None:
            nn.init.normal_(self.y_embedder.embedding.weight, std=0.02)

        # Initialize interval embedder:
        nn.init.normal_(self.t_embedder.interval_embedder[0].weight, std=0.02)
        nn.init.normal_(self.t_embedder.interval_embedder[2].weight, std=0.02)

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

        # Zero-out adaLN modulation layers in DiT blocks:
        for block in self.blocks:
            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)

        # Zero-out output layers:
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
        nn.init.constant_(self.final_layer.linear.weight, 0)
        nn.init.constant_(self.final_layer.linear.bias, 0)
    
    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)
        imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))
        return imgs

    def forward(self, x, r_or_t, t=None, y=None):
        """
        Forward pass of DiT with flexible timestep handling.
        
        Args:
            x: (N, C, H, W) tensor of spatial inputs
            r_or_t: In single mode, this is t. In interval mode, this is r.
            t: In single mode, this is None. In interval mode, this is t.
            y: (N,) tensor of class labels
        """
        # Patch embedding
        x = self.x_embedder(x) + self.pos_embed
        
        # Time embedding (handles both single and interval modes)
        t_emb = self.t_embedder(r_or_t, t)
        
        # Class embedding
        if self.use_cond and y is not None:
            y_emb = self.y_embedder(y, self.training)
            c = t_emb + y_emb
        else:
            c = t_emb
        
        # Process through transformer blocks
        for block in self.blocks:
            x = block(x, c)
        
        # Final layer
        x = self.final_layer(x, c)
        x = self.unpatchify(x)
        return x

    def forward_with_cfg(self, x, r_or_t, t=None, y=None, cfg_scale=1.0):
        """
        Forward with classifier-free guidance, supporting both modes.
        """
    
        x = x.repeat(2, 1, 1, 1)
        r_or_t = r_or_t.repeat(2)
        t = t.repeat(2)
        y = y.repeat(2) if y is not None else None
        if self.use_cond and y is not None:
            y[len(x) // 2:] = self.y_embedder.num_classes
        
        model_out = self.forward(x, r_or_t, t, y)
        
        cond_out, uncond_out = model_out.split(len(x) // 2)
        out = uncond_out + (cond_out - uncond_out) * cfg_scale
        return out


# Positional embedding from:
# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py

def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
    """
    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)
    """
    grid_h = np.arange(grid_size, dtype=np.float32)
    grid_w = np.arange(grid_size, dtype=np.float32)
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)

    grid = grid.reshape([2, 1, grid_size, grid_size])
    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)

    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
    return emb


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)

    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
    return emb