Training Instructions about freesel HOT 7 CLOSED

Hi thank you for your interest. The object detection code comes from mmdetection. Please clone this repo to your computer, and put "ssd300_voc0712_al.py" to "mmdetection/configs/pascal_voc".

from freesel.

@yichen928 thanks for the response i was able to resolve the issues but currently get output as below
2022-03-05 11:09:15,927 - mmdet - INFO - Saving checkpoint at 30 epochs
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 4952/4952, 60.9 task/s, elapsed: 81s, ETA: 0s
---------------iou_thr: 0.5---------------
2022-03-05 11:10:41,260 - mmdet - INFO -
+-------------+------+-------+--------+-------+
| class | gts | dets | recall | ap |
+-------------+------+-------+--------+-------+
| aeroplane | 285 | 1022 | 0.768 | 0.666 |
| bicycle | 337 | 1034 | 0.804 | 0.719 |
| bird | 459 | 4713 | 0.808 | 0.652 |
| boat | 263 | 2495 | 0.768 | 0.531 |
| bottle | 469 | 4322 | 0.652 | 0.384 |
| bus | 213 | 1114 | 0.873 | 0.673 |
| car | 1201 | 5976 | 0.886 | 0.755 |
| cat | 358 | 1271 | 0.899 | 0.799 |
| chair | 756 | 7239 | 0.716 | 0.434 |
| cow | 244 | 1495 | 0.922 | 0.680 |
| diningtable | 206 | 1305 | 0.782 | 0.590 |
| dog | 489 | 2005 | 0.885 | 0.722 |
| horse | 348 | 1197 | 0.891 | 0.732 |
| motorbike | 325 | 1302 | 0.852 | 0.723 |
| person | 4528 | 20707 | 0.849 | 0.695 |
| pottedplant | 480 | 4687 | 0.610 | 0.367 |
| sheep | 242 | 1483 | 0.785 | 0.608 |
| sofa | 239 | 1125 | 0.808 | 0.640 |
| train | 282 | 1325 | 0.858 | 0.741 |
| tvmonitor | 308 | 2458 | 0.825 | 0.627 |
+-------------+------+-------+--------+-------+
| mAP | | | | 0.637 |
+-------------+------+-------+--------+-------+
2022-03-05 11:10:41,342 - mmdet - INFO - Exp name: ssd300_voc0712_al.py
2022-03-05 11:10:41,342 - mmdet - INFO - Epoch(val) [30][4952] mAP: 0.6370, AP50: 0.6370

any idea y i am getting this less resutls

from freesel.

Can you share the complete log?

from freesel.

20220305_071309_ed.txt
@yichen928 pls find the complete log in the text file

from freesel.

I am not very sure what happens but it may be the following case.

Your loss decreases slowly in the beginning. There seems to be a problem in the training of mmdetection that the training may get stuck in the beginning and the final performance would be poor if the sample number is low. This problem occasionally happens for any sampling strategies including random sampling so it is out of the scope of our active learning research. My suggestion is to try to train SSD again using the same selected samples. It would get improved in most cases.

from freesel.

@yichen928 i did the following

1. i changed the python voc_dense_kmeans.py --selected 4000 and created a new data
2. trained it with updated data and below is the log file
   20220306_142521_edited.txt

i set up mmcv=1.0.5 version is it what causing the issue

from freesel.

I think 67.1 mAP is close to our reported results with 4000 images. I am not sure what causes the problem of mmdetection with 3000 images, but it is less likely to happen as the image number increases.

from freesel.