search

ZhangHanbo / Visual-Manipulation-Relationship-Network-Pytorch

MIT License
41 stars 13 forks source link

FCGN or MGN training problem #8

Closed xjtAlgo closed 3 years ago

xjtAlgo commented 3 years ago 
Hello,when I use the following commands:
"python main.py --dataset vmrdcompv1 --frame fcgn --net res101 --cuda"
"python main.py --dataset vmrdcompv1 --frame mgn --net res101 --cuda".
I can see the decline in loss, and get a model.
But when I apply this model to the test set, the accuracy rate is equal to zero.
I print my output of the test set, and the predict of grasp box generated by each image is the same grasp box.
The following are the prediction box parameters gained by fcgn test:
# grasps -> array([[5.9643127e+02, 3.6898294e+02, 5.7293231e+02, 4.6866534e+02,
#                  5.0924857e+02, 4.6022067e+02, 5.3274756e+02, 3.6053827e+02,
#                  2.5008336e-01]], dtype=float32))
It should be noted that I have no problem training and predicting on the VMRD dataset for the all_in_one model. 
Why is model FCGN or model MGN training wrong?  I would be very grateful if you can reply me. Thanks!
xjtAlgo commented 3 years ago

You can set "self.use_objdet_branch= True" in MGN. It will run successfully.

YiQi7 commented 10 months ago

I have also encountered the same problem, may I ask how fcgn should solve it