Actual dataset classification is less than labels

[wting861006]

@AlexeyAB Hi,I want to train a model with 4000 categories.But only have 500 classification dataset in this time while I will add more categories in the future.Can you give me some advice now?

I use yolov3-spp.cfg: batch=16 subdivisions=8 width=608 height=608 learning_rate=0.0001 classes=4000 in each of 3 [yolo]-layers filters=12015 in the 3 [convolutional] before each [yolo] layer

[AlexeyAB]

@wting861006 Hi,

Train with

classes=4000 in each of 3 [yolo]-layers
filters=12015 in the 3 [convolutional] before each [yolo] layer

And when you add new objects, then continue training with old_dataset + new_dataset

[wting861006]

@wting861006 Hi, This is my train log,I think the result is not good.

Last accuracy mAP@0.5 = 1.62 %, best = 1.62 % 146880: 121.146484, 128.400024 avg loss, 0.000004 rate, 5.736079 seconds, 4700160 images Loaded: 0.000046 seconds Region 89 Avg IOU: 0.349577, Class: 0.028772, Obj: 0.090707, No Obj: 0.047867, .5R: 0.220779, .75R: 0.045455, count: 154 Region 89 Avg IOU: 0.330239, Class: 0.013179, Obj: 0.100670, No Obj: 0.042956, .5R: 0.230769, .75R: 0.153846, count: 13 Region 89 Avg IOU: 0.310353, Class: 0.021537, Obj: 0.111669, No Obj: 0.046967, .5R: 0.118644, .75R: 0.033898, count: 59 Region 89 Avg IOU: 0.345305, Class: 0.021432, Obj: 0.089220, No Obj: 0.049600, .5R: 0.178808, .75R: 0.039735, count: 151 Region 101 Avg IOU: 0.740141, Class: 0.464832, Obj: 0.515111, No Obj: 0.021187, .5R: 0.948718, .75R: 0.589744, count: 78 Region 101 Avg IOU: 0.726763, Class: 0.366962, Obj: 0.455059, No Obj: 0.021550, .5R: 0.902597, .75R: 0.538961, count: 154 Region 101 Avg IOU: 0.767792, Class: 0.403637, Obj: 0.515740, No Obj: 0.021141, .5R: 0.992537, .75R: 0.597015, count: 134 Region 101 Avg IOU: 0.725869, Class: 0.387593, Obj: 0.499640, No Obj: 0.025871, .5R: 0.897959, .75R: 0.545918, count: 196 Region 113 Avg IOU: 0.741401, Class: 0.100860, Obj: 0.337994, No Obj: 0.002889, .5R: 0.924051, .75R: 0.594937, count: 79 Region 113 Avg IOU: 0.778043, Class: 0.086826, Obj: 0.476780, No Obj: 0.005829, .5R: 0.981707, .75R: 0.682927, count: 164 Region 113 Avg IOU: 0.783053, Class: 0.074814, Obj: 0.530327, No Obj: 0.012129, .5R: 0.995305, .75R: 0.708920, count: 213 Region 113 Avg IOU: 0.775443, Class: 0.074055, Obj: 0.469577, No Obj: 0.006909, .5R: 0.988571, .75R: 0.662857, count: 175 Region 89 Avg IOU: 0.289506, Class: 0.024213, Obj: 0.110561, No Obj: 0.049418, .5R: 0.148352, .75R: 0.010989, count: 182 Region 89 Avg IOU: 0.331488, Class: 0.023914, Obj: 0.071593, No Obj: 0.053193, .5R: 0.175627, .75R: 0.021505, count: 279 Region 89 Avg IOU: 0.298995, Class: 0.020024, Obj: 0.053106, No Obj: 0.045617, .5R: 0.191057, .75R: 0.012195, count: 246 Region 89 Avg IOU: 0.339938, Class: 0.020962, Obj: 0.094843, No Obj: 0.052772, .5R: 0.220339, .75R: 0.016949, count: 59 Region 101 Avg IOU: 0.748451, Class: 0.456646, Obj: 0.534611, No Obj: 0.023076, .5R: 0.953125, .75R: 0.625000, count: 128 Region 101 Avg IOU: 0.772788, Class: 0.411502, Obj: 0.393684, No Obj: 0.018022, .5R: 0.969388, .75R: 0.704082, count: 98 Region 101 Avg IOU: 0.628350, Class: 0.380417, Obj: 0.502259, No Obj: 0.025891, .5R: 0.786047, .75R: 0.306977, count: 215 Region 101 Avg IOU: 0.633628, Class: 0.325094, Obj: 0.583879, No Obj: 0.032071, .5R: 0.756923, .75R: 0.400000, count: 325 Region 113 Avg IOU: 0.690577, Class: 0.084187, Obj: 0.324832, No Obj: 0.004263, .5R: 0.853658, .75R: 0.560976, count: 41 Region 113 Avg IOU: 0.775859, Class: 0.103498, Obj: 0.255589, No Obj: 0.001912, .5R: 0.977778, .75R: 0.733333, count: 45 Region 113 Avg IOU: 0.690974, Class: 0.186987, Obj: 0.171294, No Obj: 0.001211, .5R: 0.888889, .75R: 0.444444, count: 18 Region 113 Avg IOU: 0.734962, Class: 0.072602, Obj: 0.308163, No Obj: 0.004054, .5R: 0.921053, .75R: 0.605263, count: 114 Region 89 Avg IOU: 0.223508, Class: 0.020762, Obj: 0.114924, No Obj: 0.047944, .5R: 0.062500, .75R: 0.000000, count: 32 Region 89 Avg IOU: 0.309066, Class: 0.020677, Obj: 0.091607, No Obj: 0.050668, .5R: 0.155963, .75R: 0.018349, count: 218 Region 89 Avg IOU: 0.304179, Class: 0.019465, Obj: 0.109285, No Obj: 0.053657, .5R: 0.088889, .75R: 0.016667, count: 180 Region 101 Avg IOU: 0.553184, Class: 0.381072, Obj: 0.613604, No Obj: 0.021469, .5R: 0.636986, .75R: 0.246575, count: 146 Region 89 Avg IOU: 0.399214, Class: 0.023076, Obj: 0.108758, No Obj: 0.052571, .5R: 0.317073, .75R: 0.048780, count: 123 Region 101 Avg IOU: 0.777364, Class: 0.344239, Obj: 0.508321, No Obj: 0.022056, .5R: 1.000000, .75R: 0.681818, count: 132 Region 101 Avg IOU: 0.734856, Class: 0.376078, Obj: 0.411875, No Obj: 0.024239, .5R: 0.942308, .75R: 0.519231, count: 104 Region 101 Avg IOU: 0.773506, Class: 0.445470, Obj: 0.570496, No Obj: 0.032258, .5R: 0.984848, .75R: 0.651515, count: 198 Region 113 Avg IOU: 0.734140, Class: 0.061254, Obj: 0.502914, No Obj: 0.009304, .5R: 0.941177, .75R: 0.546219, count: 238 Region 113 Avg IOU: 0.711879, Class: 0.074860, Obj: 0.319056, No Obj: 0.003061, .5R: 0.932203, .75R: 0.457627, count: 59 Region 113 Avg IOU: 0.750571, Class: 0.060877, Obj: 0.388776, No Obj: 0.006074, .5R: 0.958115, .75R: 0.612565, count: 191 Region 113 Avg IOU: 0.764838, Class: 0.073668, Obj: 0.403309, No Obj: 0.006096, .5R: 0.973214, .75R: 0.633929, count: 112 Region 89 Avg IOU: 0.294927, Class: 0.026142, Obj: 0.100348, No Obj: 0.045424, .5R: 0.228070, .75R: 0.000000, count: 57 Region 89 Avg IOU: 0.284458, Class: 0.010538, Obj: 0.128510, No Obj: 0.047431, .5R: 0.142857, .75R: 0.000000, count: 7 Region 101 Avg IOU: 0.801943, Class: 0.490723, Obj: 0.535598, No Obj: 0.019100, .5R: 0.989691, .75R: 0.835052, count: 97 Region 89 Avg IOU: 0.315847, Class: 0.027445, Obj: 0.103298, No Obj: 0.045716, .5R: 0.182540, .75R: 0.007937, count: 126 Region 101 Avg IOU: 0.777160, Class: 0.624860, Obj: 0.471641, No Obj: 0.011672, .5R: 1.000000, .75R: 0.757576, count: 33 Region 89 Avg IOU: 0.224311, Class: 0.016261, Obj: 0.114754, No Obj: 0.042528, .5R: 0.000000, .75R: 0.000000, count: 38 Region 101 Avg IOU: 0.776059, Class: 0.463738, Obj: 0.487168, No Obj: 0.023698, .5R: 0.947368, .75R: 0.714286, count: 133 Region 101 Avg IOU: 0.755165, Class: 0.468994, Obj: 0.724035, No Obj: 0.040469, .5R: 0.989051, .75R: 0.572993, count: 274 Region 113 Avg IOU: 0.769180, Class: 0.077138, Obj: 0.547530, No Obj: 0.009515, .5R: 0.969565, .75R: 0.678261, count: 230 Region 113 Avg IOU: 0.754623, Class: 0.062018, Obj: 0.645483, No Obj: 0.015956, .5R: 0.980687, .75R: 0.564378, count: 466 Region 113 Avg IOU: 0.728218, Class: 0.089087, Obj: 0.450022, No Obj: 0.004633, .5R: 0.932584, .75R: 0.561798, count: 89 Region 113 Avg IOU: 0.740749, Class: 0.065625, Obj: 0.520724, No Obj: 0.008655, .5R: 0.935065, .75R: 0.603896, count: 154

for thresh = 0.25, precision = 0.71, recall = 0.18, F1-score = 0.29 for thresh = 0.25, TP = 458, FP = 184, FN = 2064, average IoU = 54.94 %

IoU threshold = 50 %, used Area-Under-Curve for each unique Recall mean average precision (mAP@0.50) = 0.016151, or 1.62 % Total Detection Time: 20.000000 Seconds

Set -points flag: -points 101 for MS COCO -points 11 for PascalVOC 2007 (uncomment difficult in voc.data) -points 0 (AUC) for ImageNet, PascalVOC 2010-2012, your custom dataset

mean_average_precision (mAP@0.5) = 0.016151 New best mAP! Saving weights to /home/wangting/Chinese/backup/yolov3-spp_best.weights

[AlexeyAB]

It's because you don't use the most of classes.

• How many classes do you actually use?

• How many classes did you set in yolo-layer?

[wting861006]

It's because you don't use the most of classes.

• How many classes do you actually use?
• How many classes did you set in yolo-layer?

@AlexeyAB Hi，Thanks for you reply. I use actually about 100 classes.In the yolo-layer I set 4000 classes.

[AlexeyAB]

So your mAP is calculated as mAP = (AP1 + AP2 + ... AP100 + 0 + 0 + ... + 0) / 4000

So you can think that actually your mAP = 1.62 % * 4000 / 100 = 64.8% it is a good mAP.