Potential serious bug in libs/nets/pyramid_network.py?

[xxxrrr]

Line 252-265: from rois to cropped_rois, the order of rois has been changed, which means the outputs['refined’][‘box’] has different order compared to rois. However, bbox_targets (Line 420) has the same order as rois. So for Line 441 refined_box_loss = bbox_inside_weights * _smooth_l1_dist(boxes, bbox_targets), there is misalignment between boxes and xbox_targets. If this true, this is a serious bug. Or am I misunderstanding something here? Thanks.

[souryuu]

I agree with you. The order is changed and could lead to serious performance drop.

@amirbar: You probably wanna look into this.

[crazylyf]

Is the problem solved?

[souryuu]

No. I have tried to change the order of rois to match with cropped_rois but it still cannot train the refined boxes and classes. The values from refined boxes still close to zero and the network still cannot predict anything else but class 0 and 1(There should be 81 classes predicted).

[realwecan]

@souryuu Do you know why the network cannot predict anything but classes 0 and 1? Thanks!

[souryuu]

I made several changes and got better classification results.

1. Set resnet to be trainable and correct the size of image. @amirbar has already fixed this in his recent commit Fix height and width params, set trainable true #50.
2. As mentioned above, match the order of rois and crop_rois.
3. Add batch normalization. I changed all the _extra_conv_conv_scope in pyramid_network.py to _extra_conv_arg_scope_with_bn.

Changes in 1 and 2 seem to be critical. I did not compare the version with and without the changes in 3 yet but it made the trained network to produce decent (more colorful?) cropped_rois.

[amirbar]

@souryuu can you please come up with a PR? On Jun 24, 2017 5:22 AM, "souryuu" <notifications@github.com> wrote: I made several changes and got better classification results. 1. Set resnet to be trainable and correct the size of image. @amirbar <https://github.com/amirbar> has already fixed this in his recent commit #50 <#50>. 2. As mentioned above, match the order of rois and crop_rois. 3. Add batch normalization. I changed all the _extra_conv_conv_scope in pyramid_network.py to _extra_conv_arg_scope_with_bn. Changes in 1 and 2 seem to be critical. I did not compare the version with and without the changes in 3 yet but it made the trained network to produce decent (more colorful?) cropped_rois. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#68 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGG2FiWecRgXIeOg67zZ9vElJXyncdNiks5sHHL6gaJpZM4Ntt7L> .

[souryuu]

@amirbar #96
Since I changed many parts of the code for visualization, it is quite messy.

[HuangBo-Terraloupe]

@souryuu
Is your code already merged in the master branch?
Because I trained the network for 2 days, the results seems really bad.
And I still do not know how to write a evaluation code. (how to you visualize the output)
can you please help and share your code?

thanks