search

sanghoon / pva-faster-rcnn

Demo code for PVANet
https://arxiv.org/abs/1611.08588
Other
651 stars 241 forks source link

About finetune my own data #27

Open AITTSMD opened 7 years ago

AITTSMD commented 7 years ago

@sanghoon hi,sorry to bother u,but I have some problem when I finetune my own data.My task is to detect the whale in the image.when I use the full/test.model and models/pvanet/example_finetune/solver.prototxt to finetune.I got a bad result(Why?).Then,I use example_train_384/ to finetune.The result is better,but the detection time is long,about 0.18s/img（gpu gtx1070）.The example_train_384/test.prototxt has bn layers.Is it the reason?Can u give some advice about how to finetune my own data? Thank you in advance!

catsdogone commented 7 years ago

@sanghoon I finetune the pvanet/full/test.model using the train.prototxt in pvanet/example_fineturn on KITTI while the result is not good and even worse than the pvanet/full/test.model. Should I use imagenet/full/test.model as pre-train model if I want to train on a dataset with different classes with VOC? Can you give me some advice? Thank you very much. There are related files： log.txt solver_plateau.txt train.txt log txt train_3

The ap @IOU0.7: AP for Car = 0.7321 AP for Van = 0.6796 AP for Truck = 0.8270 AP for Pedestrian = 0.4119 AP for Person_sitting = 0.3100 AP for Tram = 0.6638 AP for Cyclist = 0.4930 Mean AP = 0.5882

AITTSMD commented 7 years ago

@catsdogone HI,I wonder how you draw the pictures

catsdogone commented 7 years ago

@AITTSMD Using the parse.py in caffe-fast-rcnn/tools/extra to parse the log file than plot it.

AITTSMD commented 7 years ago

@catsdogone oh! Thank u,I''ll have a try.By the way,when you have limited data,you can use example_384 to finetune.The result will be better,but the speed will be slower!

catsdogone commented 7 years ago

@AITTSMD Thank you. I will have a try.

theuy commented 7 years ago

@AITTSMD HI, I wonder which model do you use for the initial weights while you use example_384 to fineturne, models/pvanet/full/test.model ?

AITTSMD commented 7 years ago

@catsdogone imagenet/train.model

beihangzxm123 commented 7 years ago

@catsdogone Could you please share the details that how you generate the log file and how to figure the plots of loss-iters?

sanghoon commented 7 years ago

Hi all, Sorry for a late comment. I'm attending a conference right now, I'll check the issue when I get back to my home

One more thing, We've trained a better model with a similar structure. It'll be shared soon (maybe in a couple of weeks)

catsdogone commented 7 years ago

@beihangzxm123 I use the format of script file as py-faster-rcnn/scripts/*.sh. The parse_log.py is from caffe-fast-rcnn/toos/extra. The attachment is the files you may need. kitti_pva.txt--kitti_pva.sh parse_log.txt--parse_log.py plot_loss.txt--plot_loss.py