Open balabengba opened 2 years ago
This might indicate that an element in the to-be-encoded tensor (like z2_rounded) is out of the predefined range [-511, 511]. Please check the values of elements in such tensors. Are you testing with the pre-trained weights with natural images?
No, I train the model with my own dataset. Why does this phenomenon still occur after input data normalization?
This error occurs when an element in the to-be-encoded tensor falls out of the predefined value range. It may have little to do with input normalization. This can happen when a not-well-trained model takes an "outlier" input, so the model fails to constrain the value range of the to-be-encoded tensor in the middle. You may need to check whether or not the model you trained behaves well enough and whether the input image is an "expected" sample for your training data.
I get it now, thanks!
@huzi96 , I encountered low or high out of range
instead of symbol out of range
. I tried to get the value of low
and high
to check the if expression
in
if ( (low >= high) || ((low & MASK) != low) || ((high & MASK) != high) )
and I got some as below:
value of if expression
is false
, and that is what bothers me. Does low or high out of range
share same cause with symbol out of range
? Thanks in advance~ By the way, I modified the architecture of part of the neural network.
I am not 100% sure but you can try to debug by looking into the actual value of sigma, mu, and the actual symbol to code. Do these values look normal? Is the sigma very small while the symbol value is far from mu?
Appreciate your help. I will try.
During the compressing stage, module_arithmeticcoding report an error:Symbol out of range. Have you encountered this problem? thanks