yolov4-csp，yolov4-mish，yolov4，yolov4-tiny用自有数据集训练结果大比拼

本文链接：https://blog.csdn.net/jy1023408440/article/details/113184344

4csp: max_batches = 18500 width，height=512

command:./darknet detector train data/9.data yolov4-csp.cfg yolov4-csp.conv.142 -map
模型大小：yolov4x-mish_best.weights 200.67MB

Last accuracy mAP@0.5 = 99.60 %, best = 99.71 %
18500: 2.838509, 4.528679 avg loss, 0.000010 rate, 2.537801 seconds, 1184000 images, 0.071603 hours left
288Total Detection Time: 5 Seconds

calculation mAP (mean average precision)...
Detection layer: 144 - type = 28
Detection layer: 159 - type = 28
Detection layer: 174 - type = 28

detections_count = 6830, unique_truth_count = 548
class_id = 0, name = b6925303773908, ap = 100.00%    (TP = 44, FP = 30)
class_id = 1, name = b6922255451427, ap = 99.91%    (TP = 64, FP = 3)
class_id = 2, name = b6901285991219, ap = 100.00%    (TP = 51, FP = 1)
class_id = 3, name = b6920459905012, ap = 100.00%    (TP = 75, FP = 9)
class_id = 4, name = b6921168509256, ap = 98.03%    (TP = 81, FP = 8)
class_id = 5, name = b6920005772716, ap = 100.00%    (TP = 42, FP = 11)
class_id = 6, name = b6924882485103, ap = 100.00%    (TP = 38, FP = 11)
class_id = 7, name = b6920152471616, ap = 100.00%    (TP = 39, FP = 8)
class_id = 8, name = b6924743915848, ap = 98.38%    (TP = 109, FP = 5)

for conf_thresh = 0.25, precision = 0.86, recall = 0.99, F1-score = 0.92
for conf_thresh = 0.25, TP = 543, FP = 86, FN = 5, average IoU = 81.80 %

IoU threshold = 50 %, used Area-Under-Curve for each unique Recall
mean average precision (mAP@0.50) = 0.995904, or 99.59 %

4mish: max_batches = 18500 width，height=640

command:./darknet detector train data/9.data yolov4x-mish.cfg backup/yolov4x-mish_1000.weights -map -gpus 0,1
模型大小：yolov4x-mish_best.weights 379.06MB

Last accuracy mAP@0.5 = 99.37 %, best = 99.73 %
18500: 5.145534, 5.086271 avg loss, 0.000020 rate, 6.087261 seconds, 2368000 images, 0.166561 hours left

calculation mAP (mean average precision)...
Detection layer: 168 - type = 28
Detection layer: 185 - type = 28
Detection layer: 202 - type = 28
288
detections_count = 9147, unique_truth_count = 548
class_id = 0, name = b6925303773908, ap = 100.00%    (TP = 44, FP = 56)
class_id = 1, name = b6922255451427, ap = 97.11%    (TP = 65, FP = 14)
class_id = 2, name = b6901285991219, ap = 100.00%    (TP = 52, FP = 9)
class_id = 3, name = b6920459905012, ap = 100.00%    (TP = 73, FP = 7)
class_id = 4, name = b6921168509256, ap = 98.05%    (TP = 81, FP = 35)
class_id = 5, name = b6920005772716, ap = 100.00%    (TP = 42, FP = 13)
class_id = 6, name = b6924882485103, ap = 100.00%    (TP = 38, FP = 3)
class_id = 7, name = b6920152471616, ap = 100.00%    (TP = 40, FP = 17)
class_id = 8, name = b6924743915848, ap = 99.17%    (TP = 109, FP = 9)

for conf_thresh = 0.25, precision = 0.77, recall = 0.99, F1-score = 0.87
for conf_thresh = 0.25, TP = 544, FP = 163, FN = 4, average IoU = 73.30 %

IoU threshold = 50 %, used Area-Under-Curve for each unique Recall
mean average precision (mAP@0.50) = 0.993699, or 99.37 %
Total Detection Time: 8 Seconds

4nomal: max_batches = 18500

command:./darknet detector train data/9.data yolov4-custom.cfg yolov4.conv.137 -map

模型大小：yolov4-custom_best.weights 244.32MB

(next mAP calculation at 18658 iterations)
Last accuracy mAP@0.5 = 99.63 %, best = 99.89 %
18500: 0.359358, 0.281362 avg loss, 0.000010 rate, 4.044977 seconds, 1184000 images, 0.162586 hours left
288Total Detection Time: 6 Seconds
Resizing to initial size: 608 x 608 try to allocate additional workspace_size = 52.43 MB
CUDA allocate done!

calculation mAP (mean average precision)...
Detection layer: 139 - type = 28
Detection layer: 150 - type = 28
Detection layer: 161 - type = 28

detections_count = 574, unique_truth_count = 548
class_id = 0, name = b6925303773908, ap = 100.00%    (TP = 44, FP = 0)
class_id = 1, name = b6922255451427, ap = 100.00%    (TP = 64, FP = 3)
class_id = 2, name = b6901285991219, ap = 100.00%    (TP = 52, FP = 0)
class_id = 3, name = b6920459905012, ap = 100.00%    (TP = 75, FP = 0)
class_id = 4, name = b6921168509256, ap = 97.59%       (TP = 81, FP = 2)
class_id = 5, name = b6920005772716, ap = 100.00%    (TP = 42, FP = 0)
class_id = 6, name = b6924882485103, ap = 100.00%    (TP = 38, FP = 0)
class_id = 7, name = b6920152471616, ap = 100.00%    (TP = 40, FP = 0)
class_id = 8, name = b6924743915848, ap = 98.83%       (TP = 109, FP = 3)

for conf_thresh = 0.25, precision = 0.99, recall = 0.99, F1-score = 0.99
for conf_thresh = 0.25, TP = 545, FP = 8, FN = 3, average IoU = 93.66 %

IoU threshold = 50 %, used Area-Under-Curve for each unique Recall
mean average precision (mAP@0.50) = 0.996019, or 99.60 %

4-tiny: max_batches = 18500

command:./darknet detector train data/9.data yolov4-tiny-custom.cfg yolov4-tiny.conv.29 -map

模型大小：yolov4-custom_best.weights 22.50MB

(next mAP calculation at 18658 iterations)
Last accuracy mAP@0.5 = 99.55 %, best = 99.67 %
18500: 0.014559, 0.029347 avg loss, 0.000026 rate, 0.224257 seconds, 1184000 images, 0.009060 hours left

calculation mAP (mean average precision)...
Detection layer: 30 - type = 28
Detection layer: 37 - type = 28
288
detections_count = 600, unique_truth_count = 548
class_id = 0, name = b6925303773908, ap = 100.00%    (TP = 44, FP = 0)
class_id = 1, name = b6922255451427, ap = 100.00%    (TP = 64, FP = 3)
class_id = 2, name = b6901285991219, ap = 100.00%    (TP = 52, FP = 1)
class_id = 3, name = b6920459905012, ap = 100.00%    (TP = 75, FP = 1)
class_id = 4, name = b6921168509256, ap = 97.58%    (TP = 81, FP = 1)
class_id = 5, name = b6920005772716, ap = 100.00%    (TP = 42, FP = 0)
class_id = 6, name = b6924882485103, ap = 100.00%    (TP = 38, FP = 0)
class_id = 7, name = b6920152471616, ap = 100.00%    (TP = 40, FP = 0)
class_id = 8, name = b6924743915848, ap = 98.38%    (TP = 108, FP = 4)

for conf_thresh = 0.25, precision = 0.98, recall = 0.99, F1-score = 0.99
for conf_thresh = 0.25, TP = 544, FP = 10, FN = 4, average IoU = 90.11 %

IoU threshold = 50 %, used Area-Under-Curve for each unique Recall
mean average precision (mAP@0.50) = 0.995511, or 99.55 %
Total Detection Time: 2 Seconds