From 4120ac3aa685969ea3f729b1b127bbb72a16de7f Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 8 Apr 2020 21:01:58 -0700
Subject: [PATCH] training updates

---
 train.py             | 6 +++---
 utils/torch_utils.py | 2 +-
 utils/utils.py       | 8 ++++----
 3 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/train.py b/train.py
index c4629f36..6c204b7f 100644
--- a/train.py
+++ b/train.py
@@ -209,7 +209,7 @@ def train():
 
     # Start training
     nb = len(dataloader)  # number of batches
-    n_burn = max(3 * nb, 300)  # burn-in iterations, max(3 epochs, 300 iterations)
+    n_burn = max(3 * nb, 500)  # burn-in iterations, max(3 epochs, 300 iterations)
     maps = np.zeros(nc)  # mAP per class
     # torch.autograd.set_detect_anomaly(True)
     results = (0, 0, 0, 0, 0, 0, 0)  # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
@@ -234,8 +234,8 @@ def train():
             targets = targets.to(device)
 
             # Burn-in
-            if ni <= n_burn:
-                model.gr = np.interp(ni, [0, n_burn], [0.0, 1.0])  # giou loss ratio (obj_loss = 1.0 or giou)
+            if ni <= n_burn * 2:
+                model.gr = np.interp(ni, [0, n_burn * 2], [0.0, 1.0])  # giou loss ratio (obj_loss = 1.0 or giou)
                 if ni == n_burn:  # burnin complete
                     print_model_biases(model)
 
diff --git a/utils/torch_utils.py b/utils/torch_utils.py
index 34ebfe65..5819e68e 100644
--- a/utils/torch_utils.py
+++ b/utils/torch_utils.py
@@ -149,7 +149,7 @@ class ModelEMA:
         self.ema = deepcopy(model)
         self.ema.eval()
         self.updates = 0  # number of EMA updates
-        self.decay = lambda x: decay * (1 - math.exp(-x / 1000))  # decay exponential ramp (to help early epochs)
+        self.decay = lambda x: decay * (1 - math.exp(-x / 2000))  # decay exponential ramp (to help early epochs)
         self.device = device  # perform ema on different device from model if set
         if device:
             self.ema.to(device=device)
diff --git a/utils/utils.py b/utils/utils.py
index 51e10da9..4a1a07ab 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -688,11 +688,11 @@ def coco_single_class_labels(path='../coco/labels/train2014/', label_class=43):
             shutil.copyfile(src=img_file, dst='new/images/' + Path(file).name.replace('txt', 'jpg'))  # copy images
 
 
-def kmean_anchors(path='../coco/train2017.txt', n=9, img_size=(608, 608)):
+def kmean_anchors(path='../coco/train2017.txt', n=12, img_size=(320, 1024)):
     # from utils.utils import *; _ = kmean_anchors()
-    # Produces a list of target kmeans suitable for use in *.cfg files
+    # Creaters kmeans anchors for use in *.cfg files
     from utils.datasets import LoadImagesAndLabels
-    thr = 0.20  # IoU threshold
+    thr = 0.225  # IoU threshold
 
     def print_results(k):
         k = k[np.argsort(k.prod(1))]  # sort small to large
@@ -709,7 +709,7 @@ def kmean_anchors(path='../coco/train2017.txt', n=9, img_size=(608, 608)):
     def fitness(k):  # mutation fitness
         iou = wh_iou(wh, torch.Tensor(k))  # iou
         max_iou = iou.max(1)[0]
-        return max_iou.mean()  # product
+        return (max_iou * (max_iou > thr).float()).mean()  # product
 
     # Get label wh
     wh = []