DiT 代码解析

traing_function:

    # Prepare models for training:
    update_ema(ema, model.module, decay=0)  # Ensure EMA is initialized with synced weights
    model.train()  # important! This enables embedding dropout for classifier-free guidance
    ema.eval()  # EMA model should always be in eval mode

    # Variables for monitoring/logging purposes:
    train_steps = 0
    log_steps = 0
    running_loss = 0
    start_time = time()

    logger.info(f"Training for {args.epochs} epochs...")
    for epoch in range(args.epochs):
        sampler.set_epoch(epoch)
        logger.info(f"Beginning epoch {epoch}...")
        for x, y in loader:
            x = x.to(device)
            y = y.to(device)
            with torch.no_grad():
                # Map input images to latent space + normalize latents:
                x = vae.encode(x).latent_dist.sample().mul_(0.18215)
            t = torch.randint(0, diffusion.num_timesteps, (x.shape[0],), device=device)
            model_kwargs = dict(y=y)
            loss_dict = diffusion.training_losses(model, x, t, model_kwargs)
            loss = loss_dict["loss"].mean()
            opt.zero_grad()
            loss.backward()
            opt.step()
            update_ema(ema, model.module)

class DiTBlock(nn.Module):
    """
    A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
    """
    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)
        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