We have Released Neural Network Libraries v1.16.0!
We have added out-of-core (OoC) training, which enables execution of model over GPU memory size by automatically assigning parameters and computations to CPU memory during training/inference with GPU. Details of the algorithm are shown in this paper, and this is the publicly released implementation of the method.
OoC functionality can be easily used by adding a few lines to nnabla code.
# Define scheduer scheduler = lms_scheduler(ctx, use_lms=True, gpu_memory_size=3.5e9, window_length=5e9) ... # Select range to apply OoC by with scope with scheduler: loss.forward(clear_no_need_grad=True) solver.zero_grad() loss.backward(clear_buffer=True) solver.update()
We have added an object detection model CenterNet to our examples!
Unlike conventional models like YOLO that rely on anchors for object detection, CenterNet realizes fast and accurate detection by treating an object as a point.
Pre-trained models are also available so that you can immediately run inference with CenterNet. Please give this one a try as well!
- fix backward graph expansion by nn.grad
- update gnumakefile in examples/cpp
- Fix code typo
- support multi-thread training
- Give correct names to parameters
- Fixed multi-node distributed training issue for imagenet classification
- Update nnabla convert doc at 20201228
- Add document and trivial modification for using tensorboard
- fix: solve the nccl error while running more than 2 gpus
- add code to implement asymmetric watch dog
- update GNUmakefile in example/cpp
- Remove unnecessary imports
- fix window build bug
- Fix test shift error
- fix: solve the bug that no found CUDNN_PATH while building in windows
- Fix failed to compile nnabla cuda wheel.
- Fix CI failed for c-runtime build.
- Set colab default accelerator to gpu
- Enable cache option for slice method of DataItertor.
- support the file-like object in nnploader