retrain VOC data got very low IOU
sshuair opened this issue · 1 comments
sshuair commented
I tuning a model from released Model A1(using weight voc_rna-a1_cls21_s8_coco_ep-0001.params ) on VOC data. the training params following:
python issegm/voc.py --gpus 3 --split train --data-root data/VOCdevkit --output output --model voc_rna-a1_cls21 --batch-images 20 --crop-size 224 --origin-size 500 --scale-rate-range 0.7,1.3 --weights models/voc_rna-a1_cls21_s8_coco_ep-0001.params --lr-type fixed --base-lr 0.0016 --to-epoch 50 --prefetch-threads 4 --prefetcher thread --backward-do-mirror
And the Train-fcn_valid is:
Then I use the my traing result voc_rna-a1_cls21_ep-0048.params to check its performance. and the val params following:
python issegm/voc.py --data-root data/VOCdevkit --output output --phase val --weights models/voc_rna-a1_cls21_ep-0048.params --split val --test-scales 500 --test-flipping --gpus 3
but the result IOU is only 56%, not 82.86%(the paper IOU). It's so weird.
2017-03-16 10:14:52,776 Host Done 1448/1449 with speed: 1.09/s
2017-03-16 10:14:52,776 Host pixel acc: 89.40%, mean acc: 67.73%, mean iou: 56.08%
2017-03-16 10:14:52,777 Host
[95.81 79.59 59.37 71.93 60.30 68.02 90.73 77.83 83.10 37.17 48.63 53.65
61.79 68.01 77.06 89.20 40.15 65.97 45.22 69.16 79.67]
2017-03-16 10:14:52,777 Host
[89.65 64.67 38.20 58.13 51.16 54.65 71.36 70.44 68.06 26.75 45.69 47.13
55.05 55.99 66.96 71.79 35.09 54.07 38.45 63.58 50.79]
2017-03-16 10:14:53,781 Host Done 1449/1449 with speed: 1.09/s
2017-03-16 10:14:53,782 Host pixel acc: 89.39%, mean acc: 67.70%, mean iou: 56.07%
2017-03-16 10:14:53,782 Host
[95.81 79.59 59.37 71.93 60.30 67.39 90.73 77.83 83.10 37.17 48.63 53.65
61.79 68.01 77.06 89.20 40.15 65.97 45.22 69.16 79.67]
2017-03-16 10:14:53,783 Host
[89.64 64.67 38.20 58.13 51.07 55.00 71.01 70.44 68.06 26.75 45.69 47.08
55.05 55.99 66.96 71.79 35.09 54.07 38.45 63.58 50.79]
2017-03-16 10:14:53,783 Host Done in 1331.76 s.
It's my training params not correct or other something setting wrong?
I guess the crop-size, origin-size in train and the test-scales in val affect the results. I'm not sure. Does anyone know ?