We have Released Neural Network Libraries v1.17.0! First-order motion model, StyleGAN2 distillation, Zooming Slow-mo, and deformable convolution have been added, along with new colab demos!

Spotlight

First-Order Motion Model

We have added first-order motion model, an image animation model where an object in a source image is animated according to the motion of driving video!

We have also added colab demos for this model(), where you can play around with your own image and video! Speaking of Colab demo, our new CenterNet colab() is also available!

StyleGAN2 Distillation

We have added StyleGAN2 distillation model, which distills StyleGAN2 into image-to-image network trained in paired way, enabling gender swap, aging/rejuvenation, style transfer and image morphing, as in the following examples!

Old	Young

More Attractive	Less Attractive

Without Bangs	With Bangs

Zooming Slow-Mo

We have implemented Zooming Slow-Mo, which generates a high-resolution (HR) slow-motion video from a low frame rate (LFR), low-resolution (LR) video, by performing both video frame interpolation (VFI) and video super-resolution (VSR)!

Original	Zooming Slow-Mo	Only Slow-Mo

Deformable Convolution / GPU

Deformable convolution enhances the transformation modeling capacity of CNNs, by augmenting the spatial sampling locations in the modules with additional offsets and learning the offsets from target tasks. Deformable convolution v2 is also included!

Layers

• Change dtype of initializer as float32
• Deformable Convolution/GPU
• Memory optimization for STFT/iSTFT/GPU
• Optimize normalization functions implementation/GPU
• Pad memory optimization and add repeat mode/GPU
• Constant mode optoin for F.random_shift
• Add seed function/GPU
• Uses heuristics to decide conv algorithm as default
• Split convolution workspace size

Build

• Limit numpy version to 1.19.5
• fix link error for extra NBLA_API
• Remove ONNX from NNabla requirement and Update document.
• Update for ARM build
• Sync api_level version from NNabla.

Utilities

• NNP Network Refactor
• add alternative cli callback
• Attach initializer info to Variable.
• Limit numpy version to 1.19.5
• Add cudnn libraries to whl
• check gpu compatibility
• add error message when parameter is illegal
• Negative axis support for nnabla functions

Examples

• Add Zooming-Slo-Mo network
• SAGAN: Replace embedded spectral_norm implementation to F.spectral_norm
• Add First Order Motion Model Colab demo
• Add First Order Motion Model
• add CenterNet Colab demo
• StyleGAN2 distillation
• Change Colab Layout

Bugfix/Documentation

• Fix neu.lms to use GB as an unit.
• Fix a bug in LSTM
• Fix async copy in NNabla::ones and zeros
• Fix a bug in function reference count
• pynvml cannot solve the path to nvml.dll on windows
• Fix generate nnb config file error
• Fix error in using VGG
• Fix segmentation fault caused by use of uninitilized variable
• Add Citation for Arxiv Paper
• add incompatible GPU information for installation doc
• Update nnabla converter document for release nnabla
• enable multi-thread training
• fix relative imports issue
• Fix documentation error in SwapInOutScheduler