⚡️ Speed up function sigmoid_stable by 236%

Here is a faster and more memory-efficient version of the function. The bottleneck in the original implementation is calling `np.exp(x)` twice for `x < 0` branches. We can compute `exp_x = np.exp(x)` once, then reuse it for both cases.



This version calls `np.exp` only once and chooses the correct formulation based on `x`, which results in a significant speedup and less redundant computation.

Author: codeflash-ai[bot]

codeflash/process/infer.py

@@ -2,7 +2,8 @@
 
 
 def sigmoid_stable(x):
-    return np.where(x >= 0, 1 / (1 + np.exp(-x)), np.exp(x) / (1 + np.exp(x)))
+    exp_x = np.exp(-np.abs(x))
+    return np.where(x >= 0, 1 / (1 + exp_x), exp_x / (1 + exp_x))
 
 
 def postprocess(logits: np.array, max_detections: int = 8):