dit.py

import argparse
import math
from typing import Dict, Any

import numpy as np
import torch
import torch.nn as nn
from timm.models.vision_transformer import Mlp, Attention, PatchEmbed

from routing_module import Router

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


class TimestepEmbedder(nn.Module):
    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):
        # 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 LabelEmbedder(nn.Module):
    def __init__(self, num_classes, hidden_size):
        super().__init__()
        use_cfg_embedding = True
        self.embedding_table = nn.Embedding(
            num_classes + use_cfg_embedding, hidden_size
        )
        self.num_classes = num_classes
        self._init()
        
    def token_drop(self, labels, cond_drop_prob, force_drop_ids=None):
        if force_drop_ids is None:
            drop_ids = (
                torch.rand(labels.shape[0], device=labels.device) < cond_drop_prob
            )
        else:
            drop_ids = force_drop_ids == 1
        labels = torch.where(drop_ids, self.num_classes, labels)
        return labels

    def forward(self, labels, class_drop_prob=0.1, force_drop_ids=None):
        if labels.dim() == 2 and labels.size(1) == self.num_classes:
            labels = labels.argmax(dim=1)
        elif labels.dim() != 1:
            raise ValueError(f"Expected labels to be of shape (batch_size,) or (batch_size, {self.num_classes}), but got {labels.shape}")
        assert labels.max() <= 999
        use_dropout = class_drop_prob > 0
        if use_dropout or (force_drop_ids is not None):
            labels = self.token_drop(labels, class_drop_prob, force_drop_ids)
        embeddings = self.embedding_table(labels)
        return embeddings
    
    def _init(self):
        nn.init.normal_(self.embedding_table.weight, std=0.02)
        
class TransformerBlock(nn.Module):
    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, cond_mode=None, **block_kwargs):
        super().__init__()
        self.cond_mode = cond_mode
        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)
        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        approx_gelu = lambda: nn.GELU(approximate="tanh")
        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)

        if cond_mode == "adaln":
            self.adaLN_modulation = nn.Sequential(
                nn.SiLU(),
                nn.Linear(hidden_size, 6 * hidden_size, bias=True)
            )
            self._init_conditional()
        else:
            self._init_standard()
            
    @torch.compile
    def forward(self, x, c=None):
        if self.cond_mode == "adaln" and c is not None:
            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), shift_msa, scale_msa))
            x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
        else:
            x = x + self.attn(self.norm1(x))
            x = x + self.mlp(self.norm2(x))
        return x

    def _init_standard(self):
        pass

    def _init_conditional(self):
        nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.adaLN_modulation[-1].bias, 0)

class FinalLayer(nn.Module):
    def __init__(self, hidden_size, patch_size, out_channels, cond_mode=None):
        super().__init__()
        self.cond_mode = cond_mode
        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)

        if cond_mode == "adaln":
            self.adaLN_modulation = nn.Sequential(
                nn.SiLU(),
                nn.Linear(hidden_size, 2 * hidden_size, bias=True)
            )
            self._init_conditional()
        else:
            self._init_standard()

    def forward(self, x, c=None):
        if self.cond_mode == "adaln" and c is not None:
            shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
            x = modulate(self.norm_final(x), shift, scale)
        else:
            x = self.norm_final(x)
        x = self.linear(x)
        return x

    def _init_standard(self):
        nn.init.constant_(self.linear.weight, 0)
        nn.init.constant_(self.linear.bias, 0)

    def _init_conditional(self):
        nn.init.constant_(self.linear.weight, 0)
        nn.init.constant_(self.linear.bias, 0)
        nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.adaLN_modulation[-1].bias, 0)


class DiT(nn.Module):
    def __init__(
        self,
        input_size=32,
        patch_size=2,
        in_channels=4,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        num_classes=1000,
        learn_sigma=False,
        cond_mode="adaln",
        enable_routing=False,
        routes=None,
        use_x_T: bool = False,
        use_x_D_last: bool = False,
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = in_channels
        self.input_size = input_size
        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.hidden_size = hidden_size
        self.enable_routing = enable_routing
        self.routes = routes if routes is not None else []

        self.use_x_T = use_x_T
        self.use_x_D_last = use_x_D_last

        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size)
        self.t_embedder = TimestepEmbedder(hidden_size)
        self.y_embedder = LabelEmbedder(num_classes, hidden_size) if num_classes else None

        num_patches = self.x_embedder.num_patches
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size), requires_grad=False)

        self.blocks = nn.ModuleList([
            TransformerBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio, cond_mode=cond_mode)
            for _ in range(depth)
        ])

        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels, cond_mode=cond_mode)
        if enable_routing:
            self.router = Router()
        self.mask_token = None
        self.initialize_weights()

    def initialize_weights(self):
        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)

        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))

        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)

        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)

    def unpatchify(self, x):
        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: torch.Tensor,
                t: torch.Tensor,
                y: torch.Tensor,
                **kwargs
                ) -> Dict[str, torch.Tensor]:
        class_drop_prob = kwargs.get('class_drop_prob', 0.0)
        
        x = self.x_embedder(x) + self.pos_embed
        t = self.t_embedder(t)
        y = self.y_embedder(y, class_drop_prob)

        c = t + y

        if self.enable_routing:
            route_count = 0  
            x_T = x.clone() 
            masks = []
        else:
            route_count = None
            masks = None 
                              
        for idx, block in enumerate(self.blocks):
            if self.training and self.enable_routing and self.routes:
                if idx == self.routes[route_count]['start_layer_idx']:
                    x_D_last = x.clone()
                    mask_info = self.router.get_mask(x, mask_ratio=self.routes[route_count]['selection_ratio'])
                    masks.append(mask_info['mask'].to(torch.int))
                    x = self.router.start_route(x, mask_info)
            
            x = block(x, c)
                
            if self.training and self.enable_routing and self.routes:
                if idx == self.routes[route_count]['end_layer_idx']:
                    x_combined = x_T * self.routes[route_count]['x_T'] + x_D_last * self.routes[route_count]['x_D_last']
                    x = self.router.end_route(x, mask_info, original_x=x_combined)
                    if route_count < len(self.routes) - 1:
                        route_count += 1

        x = self.final_layer(x, c)                
        x = self.unpatchify(x)
        out = {
            'x': x,
            'mask': masks,
        }
        return out

    def forward_with_cfg(self,
                x: torch.Tensor,
                t: torch.Tensor,
                y: torch.Tensor,
                **kwargs
                ) -> Dict[str, torch.Tensor]:
        cfg_scale = kwargs.get('cfg_scale', 0.0)
        cond_logits = self.forward(
            x.clone(), t.clone(), y.clone(), class_drop_prob=0.0, mask_ratio=0.0,
        )['x']
        uncond_logits = self.forward(
            x.clone(), t.clone(), y.clone(), class_drop_prob=1.0, mask_ratio=0.0,
        )['x']

        logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits)
        out = {
            'x': logits,
            'mask': None
        }
        return out


def DiT_XL_2(**kwargs):
    return DiT(depth=28, hidden_size=1152, patch_size=2, num_heads=16, **kwargs)

def DiT_XL_4(**kwargs):
    return DiT(depth=28, hidden_size=1152, patch_size=4, num_heads=16, **kwargs)

def DiT_XL_8(**kwargs):
    return DiT(depth=28, hidden_size=1152, patch_size=8, num_heads=16, **kwargs)

def DiT_L_2(**kwargs):
    return DiT(depth=24, hidden_size=1024, patch_size=2, num_heads=16, **kwargs)

def DiT_L_4(**kwargs):
    return DiT(depth=24, hidden_size=1024, patch_size=4, num_heads=16, **kwargs)

def DiT_L_8(**kwargs):
    return DiT(depth=24, hidden_size=1024, patch_size=8, num_heads=16, **kwargs)

def DiT_B_2(**kwargs):
    return DiT(depth=12, hidden_size=768, patch_size=2, num_heads=12, **kwargs)

def DiT_B_4(**kwargs):
    return DiT(depth=12, hidden_size=768, patch_size=4, num_heads=12, **kwargs)

def DiT_B_8(**kwargs):
    return DiT(depth=12, hidden_size=768, patch_size=8, num_heads=12, **kwargs)

def DiT_S_2(**kwargs):
    return DiT(depth=12, hidden_size=384, patch_size=2, num_heads=6, **kwargs)

def DiT_S_4(**kwargs):
    return DiT(depth=12, hidden_size=384, patch_size=4, num_heads=6, **kwargs)

def DiT_S_8(**kwargs):
    return DiT(depth=12, hidden_size=384, patch_size=8, num_heads=6, **kwargs)


DiT_models = {
    'DiT-XL/2': DiT_XL_2,  'DiT-XL/4': DiT_XL_4,  'DiT-XL/8': DiT_XL_8,
    'DiT-L/2':  DiT_L_2,   'DiT-L/4':  DiT_L_4,   'DiT-L/8':  DiT_L_8,
    'DiT-B/2':  DiT_B_2,   'DiT-B/4':  DiT_B_4,   'DiT-B/8':  DiT_B_8,
    'DiT-S/2':  DiT_S_2,   'DiT-S/4':  DiT_S_4,   'DiT-S/8':  DiT_S_8,
}

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

def get_model(backbone_type: str, params: Dict[str, Any], **kwargs) -> nn.Module:
    """
    Factory function to instantiate a DiT model based on the backbone_type.
    """
    if backbone_type not in DiT_models:
        raise ValueError(
            f"Backbone type '{backbone_type}' is not supported. "
            f"Choose from {list(DiT_models.keys())}"
        )
    
    model_fn = DiT_models[backbone_type]
    
    # Instantiate the model with provided parameters
    model = model_fn(**params, **kwargs)
    
    return model

if __name__ == "__main__":

    parser = argparse.ArgumentParser(description="DiT2D Model Test")
    parser.add_argument("--model", type=str, default="DiT-XL/2", choices=DiT_models.keys(), help="Model type")
    parser.add_argument("--input_size", type=int, default=32, help="Input size of the model")
    parser.add_argument("--in_channels", type=int, default=4, help="Number of input channels")
    parser.add_argument("--num_classes", type=int, default=1000, help="Number of classes")
    args = parser.parse_args()

    # Example: passing the new flags via kwargs
    model_cls = DiT_models[args.model]
    model = model_cls(
        input_size=args.input_size,
        in_channels=args.in_channels,
        num_classes=args.num_classes,
        enable_routing=True,
        routes=[
            {'selection_ratio': 0.5, 'start_layer_idx': 3, 'end_layer_idx': 6},
            {'selection_ratio': 0.3, 'start_layer_idx': 7, 'end_layer_idx': 10}
        ],
        use_x_T=True,         # Set as desired via your config
        use_x_D_last=True     # Set as desired via your config
    )

    x = torch.randn(1, args.in_channels, args.input_size, args.input_size)
    t = torch.randint(0, 1000, (1,))
    y = torch.randint(0, args.num_classes, (1,))

    output = model(x, t, y)
    print("Output shape:", output['x'].shape)