We have released Neural Network Libraries 1.0.18!

We have fixed some of the issues with LSTM/GRU, along with many new models and layers!

Spotlight

[Fix] LSTM/GRU Output Shape, [Fix] PF.lstm, PF.gru weight param mismatch

LSTM/GRU’s output shapes not matching the documentation have been dealt with.

Also, the bug with LSTM/GRU in Parametric Functions , where the memory layout for the weight parameters was different from its cuDNN counterpart, has been fixed.

If you previously trained a model with PF.lstm or PF.gru and saved the parameters, it is still compatible with the new release, although we strongly recommend training again with the updated version to obtain more accurate parameters. We apologize for the inconvenience, and will continue to support and refine our models.

Models

• Add versions v1.0 and v1.1 to SqueezeNet model

• YOLO v2 pretrained model API

Function Layers

• Add 2-D and 3-D nearest neighbor interpolation.

• Add LogSoftmax and disable fp16 Softmax

• Add Fused Batch Normalization

• Convolution and Pooling w/ NHWC layout

• Add CUDNN LogSoftmax and disable fp32 Softmax

• Add CUDNN Fused Batch Normalization and utilize faster CUDNN Batch Normalization

• CUDNN Convolution and Pooling w/ NHWC layout

• Add normalizations

Utilities

• Add more operator support for onnx import, and update doc.

• Add Callback Functionality in CLI

• Add ONNX opset 7 support.

• add dicom support

• Enable CKPT format for tf importer

• Data iterator with npy cache support

• Updating GradDependsOn-argument for trigonometric functions

• Write test for cpp python like functions

• Add 3D linear interpolation support.

• Add 2-D and 3-D nearest neighbor interpolation.

• add dicom support for data source loader

• [python] Changes definition of Variable comparison and hash

• Improve creating cache from CSV dataset

Bug Fix / Documentations

• fix valgrind warning

• [Fix] log_sigmoid computation efficiency

• FIX: Add a missing implemenation to nnp::Monitor

• improve document for tensorflow support

• Fix invalid variable name for tf export

• Fix compile errors in nbla_cli build

• Fix/20190605 fix create cache bug

• Fix AdaBound & AMSBound for learning rate scheduling

• [Fix] Derivatives for Activation Layers