dit.py

from torch import nn
import torch
from torch.nn import functional as F
import numpy as np
import math
from timm.models.vision_transformer import PatchEmbed, Attention, Mlp
from transformer import Beta_DiT
from Unet import UNet
import inspect

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):
        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, dropout_prob):
        super().__init__()
        use_cfg_embedding = dropout_prob > 0
        self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
        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_table(labels)
        return embeddings

class DiTBlock(nn.Module):
    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, **block_kwargs):
        super().__init__()
        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        #yimm.attention uses flash attention if possible
        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)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )

    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), shift_msa, scale_msa))
        x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
        return x

class FinalLayer(nn.Module):
    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 DiT(nn.Module):
    def __init__(self, input_size=32, patch_size=2, in_channels=1, hidden_size=384, depth=12, num_heads=6, mlp_ratio=4.0, class_dropout_prob=0.1, num_classes=10, learn_sigma=False):
        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.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
        self.t_embedder = TimestepEmbedder(hidden_size)
        self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob)
        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([DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(depth)])
        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)
        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.y_embedder.embedding_table.weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)

        for block in self.blocks:
            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)

        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):
        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):
        x = self.x_embedder(x) + self.pos_embed
        t = self.t_embedder(t)
        y = self.y_embedder(y, self.training)
        c = t + y
        for block in self.blocks:
            x = block(x, c)
        x = self.final_layer(x, c)
        x = self.unpatchify(x)
        return x

    def forward_with_cfg(self, x, t, y, cfg_scale):
        half = x[: len(x) // 2]
        combined = torch.cat([half, half], dim=0)
        model_out = self.forward(combined, t, y)
        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 configure_optimizers(self, weight_decay, learning_rate, device):
        #ALl parameters that require grad
        param_dict={pn: p for pn,p in self.named_parameters()}
        param_dict={pn: p for pn,p in param_dict.items() if p.requires_grad}
        #Param wi 2D will be weight decay otherwise no
        decay_params=[p for n,p in param_dict.items() if p.dim() >=2]
        nodecay_params=[p for n,p in param_dict.items() if p.dim() <2]
        optim_groups=[
            {'params': decay_params, 'weight_decay':weight_decay},
            {'params': nodecay_params, 'weight_decay':0.0}
        ]
        num_decay_params=sum(p.numel() for p in decay_params)
        num_nodecay_params=sum(p.numel() for p in nodecay_params)
        #print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
        #print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
        #fused AdamW is a faster only in CUDA
        fused_available='fused' in inspect.signature(torch.optim.AdamW).parameters
        use_fused=fused_available and 'cuda' in device
        #print(f"Using fused AdamW: {use_fused}")
        optimizer=torch.optim.AdamW(optim_groups, lr=learning_rate, betas=(0.9, 0.95), eps=1e-8, fused=use_fused)
        return optimizer

def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
    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)
    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

    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])

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

def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=np.float64)
    omega /= embed_dim / 2.
    omega = 1. / 10000**omega

    pos = pos.reshape(-1)
    out = np.einsum('m,d->md', pos, omega)

    emb_sin = np.sin(out)
    emb_cos = np.cos(out)

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

if __name__ == "__main__":
    in_ch = 1
    img_size = 28
    num_classes = 10

    import time

    start_1=time.time()
    model = DiT(input_size=img_size, patch_size=2, in_channels=in_ch, hidden_size=384, depth=12, num_heads=6, num_classes=num_classes)
    x = torch.randn(10, 1, 28, 28).cuda()
    timesteps = torch.randint(0, 100, (10,)).cuda()
    labels = torch.randint(0, num_classes, (10,)).cuda()
    model = model.cuda()
    print(f"DiT Model parameters: {sum(p.numel() for p in model.parameters()) / 1e6:.2f}M")
    y = model(x, timesteps, labels)
    print(f"Output shape: {y.shape}")
    print(f"Expected shape: {x.shape}")
    end_1=time.time()
    time_1=end_1-start_1

    start_2=time.time()
    model = Beta_DiT(in_ch=in_ch, img_size=img_size)
    x = torch.randn(10, 1, 28, 28).cuda()  # dummy input on CUDA
    timesteps = torch.randint(0, 100, (10,)).cuda()  # dummy timesteps on CUDA
    model = model.cuda()  # move model to CUDA
    print(f"Beta_DiT Model parameters: {sum(p.numel() for p in model.parameters()) / 1e6:.2f}M")
    y = model(x, timesteps)
    print(f"Output shape: {y.shape}")
    print(f"Expected shape: {x.shape}")
    end_2=time.time()
    time_2=end_2-start_2

    start_3=time.time()
    model = UNet()
    x = torch.randn(10, 1, 28, 28).cuda()  # dummy input on CUDA
    timesteps = torch.randint(0, 100, (10,)).cuda()  # dummy timesteps on CUDA
    model = model.cuda()  # move model to CUDA
    print(f"UneT Model parameters: {sum(p.numel() for p in model.parameters()) / 1e6:.2f}M")
    y = model(x, timesteps)
    print(f"Output shape: {y.shape}")
    print(f"Expected shape: {x.shape}")
    end_3=time.time()
    time_3=end_3-start_3

    print(f"Dit: {time_1} Beta_Dit: {time_2} Unet: {time_3}")

    """
    Ideas for better DiT:
    Initialize the weight to 1/sqrt(N)
    Flash attention
    Easier attention to Unet
    """