yousongzhu / couplenet Goto Github PK

Hi,
I follow the procedure to prepare and I can train py-rfcn. But when I train couple net it occured

F0504 10:19:44.208268 4032 net.cpp:777] Check failed: target_blobs.size() == source_layer.blobs_size() (1 vs. 2) Incompatible number of blobs for layer conv1
*** Check failure stack trace: ***
./experiments/scripts/couplenet_end2end_ohem.sh: line 58: 4032 Aborted (core dumped) ./tools/train_net.py --gpu ${GPU_ID} --solver models/${PT_DIR}/${NET}/couplenet_end2end/solver_ohem.prototxt --weights data/imagenet_models/${NET}-model.caffemodel --imdb ${TRAIN_IMDB} --iters ${ITERS} --cfg experiments/cfgs/couplenet_end2end_ohem.yml ${EXTRA_ARGS}

what's wrong ?

The paper says the global FCN collects the surrounding information by concating the original and 2 times lager proposals together then feding into the later roi-wise subnet. But I dont find that in your code..How you guys do that?

How does concat context information and another feature in goal FCN?
Is it the 0 connection of the concat layer or the 1 connection?

I'd like to train CoupleNet on a custom dataset. Could you tell me how to go about it?

I have no way to download it from baidu pan (do not have a baidu account and it is hard to even register one). Is there any other way I can download the file (dropbox, onedrive?)
Thx

@tshizys HI

In Table4，for Global FCN, it can be regarded as a Faster R-CNN.
But the depth of roI-wise subnetwork is only two, why did it work out so well( 78.5 )?? In our common sense, roi-wise subnetwork should be very deep to be correct.

hi @tshizys,

why do you try to add 1x1 conv to norm them before sum them together?

I'm training CoupleNet on a custom dataset. I have a few questions and I'd greatly appreciate it if you could respond to my queries. First off, How do I run train.py?

Did you use the context prior to the global branch when you get 33.1 mAP result on coco dataset without multi-scale training