Unexpected 0 Loss in training a heatmap-predict network

vision
1

During my training process, the loss rapidly decline to 0,so do the output heatmaps.but they are quite diffrent to the target heatmap(ground truth).I wonder how could this be happening. i checked the Loss function,and switched to a simpler model but nothing changed.

report
report1008×600 41.2 KB

Just as the picture shows.
The prediction is a NCW*H tensor ,comparing with the heatmap transformed form the dataset original bounding box’s center points.
the Channels correspoding to the object classes(areoplanes,blablabla)
the targetmap with object in it looks brighter because i set the min value 0.1
i wish the model stop outputing all zeros.Dont care about this trick, it makes no difference.

2

Which loss function are you using and what kind of target distribution are you expecting?
Could you print the number of unique values in your target, where the loss is zero via

print(target.unique(return_counts=True))