why train with your code .Loss too high
Hjy20255 opened this issue · 9 comments
/scripts/learning_curve.py ./scripts/snapshot/log.csv
loss value is too hight. 1e5
in your example log.png .1e6
@wkentaro
Maybe it depends on the dataset you use, and the way to initialize weight.
For example, if you use larger images for training, you have larger loss value, because the loss is not averaged by image size.
Use of pre-training model is also important.
Thanks to your reply, but I am not very understand, the segmentation results is ok， and in your code also used pre-training model ---vgg16 model, why would loss is 1e6
Do you use example code under voc?
Could you please attach log graph? or log.csv posted to Gist.com?
It's similar to the log plot attached to README, isn't it? Where's wrong?
Loss value without normalization in your code ？so, loss is 1e5?
my image is 500×500
for example loss is 4×1e5 /(500*500)=1.6 ?
Yeah.
thank you