Ssd.pytorch: Is there a bug in multibox_loss.py?

[kouyichi]

when I train the model, there is a bug in layers/modules/multibox_loss.py
In line 97, loss_c[pos]=0, the shape of mask [32,8732] ar index 0 not match the shape of the indexed tensor [270424,1] at index 0.

[weidi1024]

This problem may be caused by the version of pytorch
here is the solution
1.add code before 'loss_c[pos] = 0' at multibox_loss.py
loss_c = loss_c.view(pos.size()[0], pos.size()[1])
loss_c[pos] = 0 # filter out pos boxes for now
2.at train.py
change
loc_loss += loss_l.data[0]
conf_loss += loss_c.data[0]
to:
loc_loss += loss_l.item()
conf_loss += loss_c.item()
3. and mybe there will be a warning:
UserWarning: size_average and reduce args will be deprecated, please use reduction='sum' instead.
warnings.warn(warning.format(ret))
change
size_average=False
to
reduction='sum'

reference
https://blog.csdn.net/qq_30614451/article/details/100137358

[chaomartin]

modification places are Line 90 and Line 111 in multibox_loss.py
change
loss_l = F.smooth_l1_loss(loc_p, loc_t, size_average=False)
to
loss_l = F.smooth_l1_loss(loc_p, loc_t, reduction='sum')

[Codermanok]

train.py: line 188, should also change
from
print('iter ' + repr(iteration) + ' || Loss: %.4f ||' % (loss.data[0]), end=' ')
to
print('iter ' + repr(iteration) + ' || Loss: %.4f ||' % (loss.item()), end=' ')

[shirui-japina]

This problem may be caused by the version of pytorch
here is the solution
1.add code before 'loss_c[pos] = 0' at multibox_loss.py
loss_c = loss_c.view(pos.size()[0], pos.size()[1])
loss_c[pos] = 0 # filter out pos boxes for now
2.at train.py
change
loc_loss += loss_l.data[0]
conf_loss += loss_c.data[0]
to:
loc_loss += loss_l.item()
conf_loss += loss_c.item()
3. and mybe there will be a warning:
UserWarning: size_average and reduce args will be deprecated, please use reduction='sum' instead.
warnings.warn(warning.format(ret))
change
size_average=False
to
reduction='sum'

reference
https://blog.csdn.net/qq_30614451/article/details/100137358

GREAT CSDN!!!

[TalentWong]

when I train the model, there is a bug in layers/modules/multibox_loss.py
In line 97, loss_c[pos]=0, the shape of mask [32,8732] ar index 0 not match the shape of the indexed tensor [270424,1] at index 0.

Just changing the code order of line 97 and 98, it would work