ALFNet
ALFNet copied to clipboard
liuwei16
The MR is not as good as that in the paper.
I have trained the proposed model on my server with P40 GPU. The MR on Cityperson val dataset is 18% with one GPU (16% for two). Where might the problem arise?
i have the similar problem that using alfnet-2s, i spent 2s on per image , so it's hard to acheive 20fps with gtx 1080ti. what's the problem ?
I modified the parameters according to the paper description. The MR decreased to 20.20%.
@Chen94yue Sorry for delayed reply. Did you decrease the learning rate and load the trained weights to continue training?
@rintiunse Sorry for delayed reply. Did you make sure that your tensorflow is running on GPU? I just re-run the exact experiments 4 times. It approximately took 126~132 seconds on 500 images of 1024x2048 pixels on CityPersons validation set. 20fps is achieved on Caltech with images of 480x640 pixels as stated in the paper.
@Chen94yue Sorry for delayed reply. Did you decrease the learning rate and load the trained weights to continue training?
No, I didn't. Let me have a try.
@liuwei16 @VideoObjectSearch Could you update the original code for the following functions? from .utils.cython_bbox import bbox_overlaps from .utils.bbox import box_op The compiled code didn't work because different version of tensorflow.
Also, could you give a detailed description on how to train models? I run many times but also couldn't reproduce as good results as you did.
@VideoObjectSearch im sure that i run the code on GPU(1080TI). But it still spent 2s per image(CityPersons).
@Chen94yue Sorry for delayed reply. Did you decrease the learning rate and load the trained weights to continue training?
I have a problem that what the end total loss value should be?
@Chen94yue Sorry for delayed reply. Did you decrease the learning rate and load the trained weights to continue training?
I have a problem that what the end total loss value should be? I haven't finished training yet. I'm not sure.