From d41f85702d036528342045f4d7e4d0a1a801ec45 Mon Sep 17 00:00:00 2001
From: Nir Ben-Zvi <nzvi@trigosystems.com>
Date: Thu, 22 Nov 2018 15:36:14 +0200
Subject: [PATCH] Fixed NMS bug causing big CPU usage. Note that using
 'cross_class_nms' still takes a huge amount of time and should be fixed
 somehow.

---
 utils/utils.py | 23 +++++++++++++----------
 1 file changed, 13 insertions(+), 10 deletions(-)

diff --git a/utils/utils.py b/utils/utils.py
index 0a4c6f02..8fe216c3 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -285,8 +285,6 @@ def build_targets(pred_boxes, pred_conf, pred_cls, target, anchor_wh, nA, nC, nG
 
 
 def non_max_suppression(prediction, conf_thres=0.5, nms_thres=0.4):
-    prediction = prediction.cpu()
-
     """
     Removes detections with lower object confidence score than 'conf_thres' and performs
     Non-Maximum Suppression to further filter detections.
@@ -302,15 +300,17 @@ def non_max_suppression(prediction, conf_thres=0.5, nms_thres=0.4):
         # cross-class NMS
         cross_class_nms = False
         if cross_class_nms:
-            thresh = 0.85
+            # thresh = 0.85
+            thresh = nms_thres
             a = pred.clone()
-            a = a[np.argsort(-a[:, 4])]  # sort best to worst
+            _, indices = torch.sort(-a[:, 4], 0) # sort best to worst
+            a = a[indices]
             radius = 30  # area to search for cross-class ious
             for i in range(len(a)):
                 if i >= len(a) - 1:
                     break
 
-                close = (np.abs(a[i, 0] - a[i + 1:, 0]) < radius) & (np.abs(a[i, 1] - a[i + 1:, 1]) < radius)
+                close = (torch.abs(a[i, 0] - a[i + 1:, 0]) < radius) & (torch.abs(a[i, 1] - a[i + 1:, 1]) < radius)
                 close = close.nonzero()
 
                 if len(close) > 0:
@@ -324,10 +324,11 @@ def non_max_suppression(prediction, conf_thres=0.5, nms_thres=0.4):
                         a = a[mask]
             pred = a
 
-        x, y, w, h = pred[:, 0].numpy(), pred[:, 1].numpy(), pred[:, 2].numpy(), pred[:, 3].numpy()
+        x, y, w, h = pred[:, 0], pred[:, 1], pred[:, 2], pred[:, 3]
         a = w * h  # area
         ar = w / (h + 1e-16)  # aspect ratio
-        log_w, log_h, log_a, log_ar = np.log(w), np.log(h), np.log(a), np.log(ar)
+
+        log_w, log_h, log_a, log_ar = torch.log(w), torch.log(h), torch.log(a), torch.log(ar)
 
         # n = len(w)
         # shape_likelihood = np.zeros((n, 60), dtype=np.float32)
@@ -338,8 +339,10 @@ def non_max_suppression(prediction, conf_thres=0.5, nms_thres=0.4):
 
         class_prob, class_pred = torch.max(F.softmax(pred[:, 5:], 1), 1)
 
-        v = ((pred[:, 4] > conf_thres) & (class_prob > .3)).numpy()
-        v = v.nonzero()
+        v = ((pred[:, 4] > conf_thres) & (class_prob > .3))
+        v = v.nonzero().squeeze()
+        if len(v.shape) == 0:
+            v = v.unsqueeze(0)
 
         pred = pred[v]
         class_prob = class_prob[v]
@@ -363,7 +366,7 @@ def non_max_suppression(prediction, conf_thres=0.5, nms_thres=0.4):
         # Iterate through all predicted classes
         unique_labels = detections[:, -1].cpu().unique()
         if prediction.is_cuda:
-            unique_labels = unique_labels.cuda()
+            unique_labels = unique_labels.cuda(prediction.device)
 
         nms_style = 'OR'  # 'AND' or 'OR' (classical)
         for c in unique_labels: