bnet-ot-fusion / src /otfusion /dropout.py
BioinstLab's picture
Initial commit: BNet-OT multimodal late fusion implementation
3c8ce9d
Raw History Blame Contribute Delete
2.02 kB
"""Simulated modality dropout.
The paper evaluates robustness by randomly omitting one input source at
inference time, at dropout rates of 20% and 50%. A dropped modality is passed
to the fusion layer as missing (mask = 0), which substitutes the uniform
distribution (Eq. 15).
"""
from __future__ import annotations
from typing import Optional
import numpy as np
MODES = ("either", "hist", "risk", "both")
def simulate_dropout(
n: int,
rate: float,
mode: str = "either",
rng: Optional[np.random.Generator] = None,
) -> tuple[np.ndarray, np.ndarray]:
"""Draw modality masks (m_p, m_q) with a given dropout rate.
Parameters
----------
n:
Number of samples.
rate:
Fraction of samples affected by dropout, in [0, 1].
mode:
``"either"`` (default) - each affected sample loses one modality, chosen
uniformly at random (this matches "randomly omitting 1 input source").
``"hist"`` - affected samples lose the histopathology input.
``"risk"`` - affected samples lose the clinical risk input.
``"both"`` - affected samples lose both inputs.
rng:
NumPy random generator, for reproducibility.
Returns
-------
(m_p, m_q):
Boolean presence masks, shape (n,). True means the modality is present.
"""
if not 0.0 <= rate <= 1.0:
raise ValueError("rate must lie in [0, 1]")
if mode not in MODES:
raise ValueError(f"mode must be one of {MODES}")
rng = rng if rng is not None else np.random.default_rng()
affected = rng.random(n) < rate
m_p = np.ones(n, dtype=bool)
m_q = np.ones(n, dtype=bool)
if mode == "hist":
m_p[affected] = False
elif mode == "risk":
m_q[affected] = False
elif mode == "both":
m_p[affected] = False
m_q[affected] = False
else: # either
drop_hist = rng.random(n) < 0.5
m_p[affected & drop_hist] = False
m_q[affected & ~drop_hist] = False
return m_p, m_q