We have released Neural Network Libraries v1.28.0! We have added a new graph converter to support weight pruning!
Spotlight
add a new GraphConverter to support weight pruning
We have implemented a graph converter PruningModifier that prunes the small weight values to 0. The pruning is channel-wise, using the channel-wise L2-normalization to represent the degree of sparsity. If the L2-norm is less than the threshold provided, all values of the target channel will be set to 0.
Currently, pruning is supported for following functions:
- Convolution
- Deconvolution
- DepthwiseConvolution
- DepthwiseDeconvolution
- Affine
Example:
pred = Model(...)
import nnabla.experimental.graph_converters as GC
modifiers = [GC.PruningModifier(pruning_threshold=0.1)]
gc = GC.GraphConverter(modifiers)
pred = gc.convert(pred)
Bugfix
- fix yolov2 load error
- Fix scatter_add for recompute
- fix segmetation load bug
- Fix param dtype translate in python test.
- make varaible’s slice consistent between function call and build-in o…
- Fix: Slow Softmax backward when H and W are 1
