transfer/code/CCFM/config/config_cascaded.py

from dataclasses import dataclass
import os


@dataclass
class CascadedFlowConfig:
    # === Base (same as scDFM FlowConfig) ===
    model_type: str = "cascaded"
    batch_size: int = 48
    ntoken: int = 512
    d_model: int = 128
    nhead: int = 8
    nlayers: int = 4
    lr: float = 5e-5
    steps: int = 200000
    eta_min: float = 1e-6
    devices: str = "1"
    test_only: bool = False

    data_name: str = "norman"
    perturbation_function: str = "crisper"
    noise_type: str = "Gaussian"
    poisson_alpha: float = 0.8
    poisson_target_sum: int = -1

    print_every: int = 5000
    mode: str = "predict_y"
    result_path: str = "./result_online"
    fusion_method: str = "differential_perceiver"
    infer_top_gene: int = 1000
    n_top_genes: int = 5000
    checkpoint_path: str = ""
    gamma: float = 0.5
    split_method: str = "additive"
    use_mmd_loss: bool = True
    fold: int = 1
    use_negative_edge: bool = True
    topk: int = 30

    # === Cascaded / Latent specific ===
    scgpt_dim: int = 512
    bottleneck_dim: int = 128
    latent_weight: float = 1.0
    choose_latent_p: float = 0.4
    target_std: float = 1.0
    dh_depth: int = 2
    warmup_batches: int = 200  # batches to collect running stats

    # === Cascaded noise (LatentForcing dino_first_cascaded_noised) ===
    noise_beta: float = 0.25  # when training expr flow, t_latent ~ U[1-beta, 1] instead of 1.0

    # === EMA ===
    ema_decay: float = 0.9999

    # === Logit-normal time-step sampling ===
    t_sample_mode: str = "logit_normal"  # "uniform" or "logit_normal"
    t_expr_mean: float = 0.0   # expression flow logit-normal mu
    t_expr_std: float = 1.0    # expression flow logit-normal sigma
    t_latent_mean: float = 0.0  # latent flow logit-normal mu
    t_latent_std: float = 1.0   # latent flow logit-normal sigma

    # === LR warmup ===
    warmup_steps: int = 2000

    # === scGPT paths ===
    scgpt_model_dir: str = "transfer/data/scGPT_pretrained"
    scgpt_max_seq_len: int = 1200
    scgpt_cache_path: str = ""  # Pre-extracted HDF5 path. Empty = online extraction (default)

    # === Inference ===
    latent_steps: int = 20
    expr_steps: int = 20
    ode_method: str = "rk4"  # "euler" or "rk4"

    def __post_init__(self):
        if self.data_name == "norman_umi_go_filtered":
            self.n_top_genes = 5054
        if self.data_name == "norman":
            self.n_top_genes = 5000

    def make_path(self):
        scgpt_mode = "cached" if self.scgpt_cache_path else "online"
        t_mode = "ln" if self.t_sample_mode == "logit_normal" else "uni"
        exp_name = (
            f"ccfm-{self.data_name}-f{self.fold}"
            f"-topk{self.topk}-neg{self.use_negative_edge}"
            f"-d{self.d_model}-lr{self.lr}"
            f"-lw{self.latent_weight}-lp{self.choose_latent_p}"
            f"-ema{self.ema_decay}-{t_mode}-wu{self.warmup_steps}"
            f"-{self.ode_method}-{scgpt_mode}"
        )
        return os.path.join(self.result_path, exp_name)