- Argmin and Argmax return incorrect result when multiple min or max values are given. #49
- Wrong sqrt gradient. #37
- Small updates
- Wrong gradient computation in the case that a target variable is scalar and its broadcast is involved:
Graph::grad(&[scalar * tensor], &[scalar])
#38
- New
autograd::run
API which is more useful version ofautograd::with
#34
- Fixed some error messages that are not user-friendly (
Tensor::access_elem
,Graph::sparse_softmax_crossentropy
) - Updated dependencies:
num
,rand_distr
- Serious memory bug around gradient computation
- Wrong documentation of
Graph::slice
- Critical memory issue when the number of nodes in a graph exceeded a certain threshold #28
- Misuse of given indexes in
Graph::slice
#30 - Undisposed dependencies
Introduced Graph
and basic API was fixed.