issues
search
chr5tphr
/
zennit
Zennit is a high-level framework in Python using PyTorch for explaining/exploring neural networks using attribution methods like LRP.
Other
190
stars
32
forks
source link
Fix mod_params breaking gradients from torch 1.10
#61
Closed
chr5tphr
closed
2 years ago
chr5tphr
commented
2 years ago
the data assignment in
mod_params
erroneously overwrote the parameter tensor in-place, which breaks the intended gradient for the autograd update from torch 1.10
this was fixed by overwriting the module's attribute parameter instead
Closes #60
mod_params
erroneously overwrote the parameter tensor in-place, which breaks the intended gradient for the autograd update from torch 1.10Closes #60