| from __future__ import annotations |
|
|
| import base64 |
| from io import BytesIO |
| from math import erf, sqrt |
|
|
| import numpy as np |
|
|
|
|
| def load_npz_text(path) -> dict[str, np.ndarray]: |
| encoded = open(path, "rt", encoding="ascii").read() |
| with np.load(BytesIO(base64.b64decode(encoded))) as archive: |
| return {name: archive[name] for name in archive.files} |
|
|
|
|
| def linear(x: np.ndarray, weight: np.ndarray, bias: np.ndarray) -> np.ndarray: |
| return x @ weight.T + bias |
|
|
|
|
| def layer_norm( |
| x: np.ndarray, |
| weight: np.ndarray, |
| bias: np.ndarray, |
| eps: float = 1e-5, |
| ) -> np.ndarray: |
| mean = x.mean(axis=-1, keepdims=True) |
| variance = ((x - mean) ** 2).mean(axis=-1, keepdims=True) |
| normalized = (x - mean) / np.sqrt(variance + eps) |
| return normalized * weight + bias |
|
|
|
|
| def gelu(x: np.ndarray) -> np.ndarray: |
| erf_values = np.vectorize(erf, otypes=[np.float32])(x / sqrt(2.0)) |
| return 0.5 * x * (1.0 + erf_values) |
|
|
|
|
| class NumpyStableGateDANN: |
| """Inference-only StableGate-Hybrid DANN disease-classification branch.""" |
|
|
| def __init__(self, weights_path): |
| self.weights = load_npz_text(weights_path) |
|
|
| def predict_logits(self, expression: np.ndarray) -> np.ndarray: |
| weights = self.weights |
| x = np.asarray(expression, dtype=np.float32) * weights["gate_mask"] |
|
|
| x = linear(x, weights["input_projection.0.weight"], weights["input_projection.0.bias"]) |
| x = layer_norm(x, weights["input_projection.1.weight"], weights["input_projection.1.bias"]) |
| x = gelu(x) |
|
|
| block_input = x |
| x = layer_norm(x, weights["encoder.0.block.0.weight"], weights["encoder.0.block.0.bias"]) |
| x = linear(x, weights["encoder.0.block.1.weight"], weights["encoder.0.block.1.bias"]) |
| x = gelu(x) |
| x = linear(x, weights["encoder.0.block.4.weight"], weights["encoder.0.block.4.bias"]) |
| x = block_input + x |
|
|
| x = layer_norm(x, weights["projection_head.0.weight"], weights["projection_head.0.bias"]) |
| x = linear(x, weights["projection_head.1.weight"], weights["projection_head.1.bias"]) |
| x = gelu(x) |
| x = linear(x, weights["projection_head.4.weight"], weights["projection_head.4.bias"]) |
| x = layer_norm(x, weights["projection_head.5.weight"], weights["projection_head.5.bias"]) |
|
|
| x = linear(x, weights["disease_classifier.0.weight"], weights["disease_classifier.0.bias"]) |
| x = gelu(x) |
| return linear(x, weights["disease_classifier.3.weight"], weights["disease_classifier.3.bias"]) |
|
|
