I meet the problem. about rfbnet HOT 11 CLOSED

What's your python version? We only support python 3+ now.

from rfbnet.

@ruinmessi What's format of img and label？
img: 1.resize, 2.img-mean 3.convert to rgb?
label:?

from rfbnet.

@hewumars I use opencv to process an img, and original vgg-reduced network is ported from caffe framework, so BGR is actually needed. You can check data/data_augment for more details.

from rfbnet.

@ruinmessi Thanks!!

from rfbnet.

@ruinmessi   the conf and loc loss   is  normal?     I trained 6 classes trainset with 30000pcs.    conf loss can not decrease lower.

1 Epoch:113 || epochiter: 710/880|| Totel iter 99270 || L: 0.4974 C: 1.4676||Batch time: 0.7136 sec. ||LR: 0.00000400
2 Epoch:113 || epochiter: 720/880|| Totel iter 99280 || L: 0.4325 C: 1.2438||Batch time: 0.7197 sec. ||LR: 0.00000400
3 Epoch:113 || epochiter: 730/880|| Totel iter 99290 || L: 0.3360 C: 1.0181||Batch time: 0.7640 sec. ||LR: 0.00000400
4 Epoch:113 || epochiter: 740/880|| Totel iter 99300 || L: 0.3455 C: 1.2053||Batch time: 0.7401 sec. ||LR: 0.00000400
5 Epoch:113 || epochiter: 750/880|| Totel iter 99310 || L: 0.3085 C: 1.0816||Batch time: 0.6753 sec. ||LR: 0.00000400
6 Epoch:113 || epochiter: 760/880|| Totel iter 99320 || L: 0.3325 C: 0.9674||Batch time: 0.7058 sec. ||LR: 0.00000400
7 Epoch:113 || epochiter: 770/880|| Totel iter 99330 || L: 0.2641 C: 0.9466||Batch time: 0.7406 sec. ||LR: 0.00000400
8 Epoch:113 || epochiter: 780/880|| Totel iter 99340 || L: 0.3761 C: 1.3036||Batch time: 0.6813 sec. ||LR: 0.00000400
9 Epoch:113 || epochiter: 790/880|| Totel iter 99350 || L: 0.3817 C: 1.1372||Batch time: 0.6823 sec. ||LR: 0.00000400
10 Epoch:113 || epochiter: 800/880|| Totel iter 99360 || L: 0.3597 C: 1.2634||Batch time: 0.7142 sec. ||LR: 0.00000400
11 Epoch:113 || epochiter: 810/880|| Totel iter 99370 || L: 0.3084 C: 1.0017||Batch time: 0.7348 sec. ||LR: 0.00000400
12 Epoch:113 || epochiter: 820/880|| Totel iter 99380 || L: 0.3442 C: 1.0381||Batch time: 0.7488 sec. ||LR: 0.00000400
13 Epoch:113 || epochiter: 830/880|| Totel iter 99390 || L: 0.3259 C: 1.0018||Batch time: 0.8141 sec. ||LR: 0.00000400
14 Epoch:113 || epochiter: 840/880|| Totel iter 99400 || L: 0.4237 C: 1.2106||Batch time: 0.7183 sec. ||LR: 0.00000400
15 Epoch:113 || epochiter: 850/880|| Totel iter 99410 || L: 0.3241 C: 1.1406||Batch time: 0.6899 sec. ||LR: 0.00000400
16 Epoch:113 || epochiter: 860/880|| Totel iter 99420 || L: 0.2067 C: 0.8219||Batch time: 0.7025 sec. ||LR: 0.00000400
17 Epoch:113 || epochiter: 870/880|| Totel iter 99430 || L: 0.3680 C: 1.2261||Batch time: 0.6708 sec. ||LR: 0.00000400

from rfbnet.

@hewumars It seems no problem, you'd better test on your own dataset

from rfbnet.

Can I see your train log ? I use RFB_E_vgg networt to test, the result is good,but I don't know that is it better to be better? Now I want to train RFB_mobile,but It load base weight that report error.

from rfbnet.

@hewumars Sorry, I didn't save my train log actually. And if you wanna train mobilenet version, make sure you install pyinn first (pip install git+https://github.com/szagoruyko/pyinn.git@master) , with a faster implement for Separate Conv.

from rfbnet.

do you plan to pytorch2caffemodel?
pyinn has been installed. I was annotated the code of init weights ,then can train it.
Do you have any comparison between the gpu mem of RFB_vgg\RFB_E_vgg\RFB_mobile?

from rfbnet.

@hewumars I have no plan to do that. There are many tools in github to transform a pytorch model to caffe. When training with 300 input size and 32 batchsize on VOC, RFB_vgg consumes ~9G on a GPU, RFB_E_vgg consumes ~10G, RFB_mobile consumes ~5G. By the way, I don't find any error on init weights, would you tell me more details about your previous error message?

from rfbnet.

@hewumars you can send an email........ [email protected]

from rfbnet.