From ade8ab57ff7dd5e5c29a95764e2d78140c2156b2 Mon Sep 17 00:00:00 2001
From: "codeflash-ai[bot]"
 <148906541+codeflash-ai[bot]@users.noreply.github.com>
Date: Tue, 22 Jul 2025 04:38:48 +0000
Subject: [PATCH] =?UTF-8?q?=E2=9A=A1=EF=B8=8F=20Speed=20up=20function=20`s?=
 =?UTF-8?q?igmoid=5Fstable`=20by=20208%=20Here=20is=20an=20optimized=20ver?=
 =?UTF-8?q?sion=20of=20your=20program.=20The=20original=20code=20inadverte?=
 =?UTF-8?q?ntly=20computes=20`np.exp(x)`=20twice=20when=20`x=20<=200`,=20i?=
 =?UTF-8?q?ncurring=20redundant=20computation.=20By=20computing=20it=20onc?=
 =?UTF-8?q?e=20and=20caching=20the=20result,=20you=20eliminate=20the=20dup?=
 =?UTF-8?q?licate=20work.=20This=20reduces=20runtime,=20especially=20on=20?=
 =?UTF-8?q?large=20arrays.?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

This form reduces repeated computation and will run faster, especially for large NumPy arrays.
**All comments are preserved unless their code portion changed.**
---
 codeflash/process/infer.py | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)
 create mode 100644 codeflash/process/infer.py

diff --git a/codeflash/process/infer.py b/codeflash/process/infer.py
new file mode 100644
index 000000000..ebe3c7e88
--- /dev/null
+++ b/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)