⚡️ Speed up function sigmoid_stable by 208%

Here is an optimized version of your program. The original code inadvertently computes `np.exp(x)` twice when `x < 0`, incurring redundant computation. By computing it once and caching the result, you eliminate the duplicate work. This reduces runtime, especially on large arrays.



This form reduces repeated computation and will run faster, especially for large NumPy arrays.  
**All comments are preserved unless their code portion changed.**

Author: codeflash-ai[bot]

codeflash/process/infer.py

@@ -0,0 +1,24 @@
+import numpy as np
+
+
+def sigmoid_stable(x):
+    # Avoid redundant computation of 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):
+    batch_size, num_queries, num_classes = logits.shape
+    logits_sigmoid = sigmoid_stable(logits)
+    processed_predictions = []
+    for batch_idx in range(batch_size):
+        logits_flat = logits_sigmoid[batch_idx].reshape(-1)
+
+        sorted_indices = np.argsort(-logits_flat)[:max_detections]
+        processed_predictions.append(sorted_indices)
+    return processed_predictions
+
+
+if __name__ == "__main__":
+    predictions = np.random.normal(size=(8, 1000, 10))
+    result = postprocess(predictions, max_detections=8)