gurkirt / fpn.pytorch1.0 Goto Github PK

hi
where should i download the pretrained model

emm~ btw i just wonder whether the combo method , fpn+ohem , can achieve the better mAP than pure FPN algorithm on voc2007 or voc2012..?

recently i have trained a model based RESNET101 and achieved 81.2mAP.
now i have some imgs and i wanna test the model's performance (draw the bbox on the imgs) and whether u will upload the inference script in the future? :)

when i run the train.py. i got a hint:
shutil.SameFileError: './weights/voc/cache/FPN600-kmeanssh00-voc-bs16-resnet50-lr00020-bn0/source/kmeans_for_anchors.py' and '/home/cyrus/code/fpn/ohem-fpn/weights/voc/cache/FPN600-kmeanssh00-voc-bs16-resnet50-lr00020-bn0//source/kmeans_for_anchors.py' are the same file

and then

i commented the two lines code to avoid the SameFileError.

# source_dir = args.save_root+'/source/' # where to save the source
# utils.copy_source(source_dir)

(actually i dont even know whether the minor change will arise other ERRORs or not~~)

finally, i got a FileNotFoundError that i didnt have .pth file.

FileNotFoundError: [Errno 2] No such file or directory: '/home/cyrus/code/fpn/ohem-fpn/weights/pretrained/resnet50.pth'

and i found that the download link posted at README.md is doesnt work~ so..where can i have these .pth pretrained files ? could u plz provide a download link again ?

looking forward to ur early reply~

thx~

hi gurkirt.
recently i trained this model on coco dataset .
its weird that when i run evaluate.py on coco the detect time was too long .

Forward Time 0.626
im_detect: 4/4952 time taken 12.002
NMS stuff Time 10.017
Forward Time 0.102
im_detect: 84/4952 time taken 176.985
NMS stuff Time 9.978
Forward Time 0.103
im_detect: 164/4952 time taken 169.440
NMS stuff Time 9.542
Forward Time 0.098
im_detect: 244/4952 time taken 193.901

btw its very normal when i evaluate the voc dataset.

Forward Time 0.133
im_detect: 4/4952 time taken 2.477
NMS stuff Time 1.104
Forward Time 0.082
im_detect: 84/4952 time taken 15.756
NMS stuff Time 0.658
Forward Time 0.083
im_detect: 164/4952 time taken 16.580
NMS stuff Time 0.708
Forward Time 0.083
im_detect: 244/4952 time taken 14.848
NMS stuff Time 0.505
Forward Time 0.087
im_detect: 324/4952 time taken 16.605
NMS stuff Time 0.736
Forward Time 0.084
im_detect: 404/4952 time taken 14.983
NMS stuff Time 0.723
Forward Time 0.083
im_detect: 484/4952 time taken 16.381
NMS stuff Time 0.666
Forward Time 0.085
im_detect: 564/4952 time taken 16.034
NMS stuff Time 0.777

could u do me a favor to solve it ?

hi ,gurkirt:
i benefit greatly from your work! but when i train the net on voc , the first epoch(about the first 1000 iterations）it works well. but when it was trained by the same data for the second time(the second epoch) the location loss gone faulty but the cls loss keep normal. however this cause the validation performance decreases greatly. i have tried many tricks dealing with loss problem and it dosen't works. it is strange that it is the problem happens at exactly the iteration belongs to the second epoch. could you help me? or do you kown where the problem lies. i use the lr0.0005 and default boxes, shared head,mbox loss. looking for your reply,thanks!

line368 in train.py
if loc_loss>1000:
lline = '\n\n\n We got faulty LOCATION loss {} {} \n\n\n'.format(loc_loss, conf_loss)
log_file.write(lline)
print(lline)
loc_loss = 20.0

Why make loc_loss become 20.0 when loc_loss>1000?