ZeRO-Offload is a sharded data parallel method for distributed training. It exploits both CPU memory and compute for offloading, while offering a clear path towards efficiently scaling on multiple GPUs by working with [ZeRO-powered data parallelism](https://www.paperswithcode.com/method/zero). The symbiosis allows ZeRO-Offload to maintain a single copy of the optimizer states on the CPU memory regardless of the data parallel degree. Furthermore, it keeps the aggregate communication volume between GPU and CPU, as well as the aggregate CPU computation a constant regardless of data parallelism, allowing ZeRO-Offload to effectively utilize the linear increase in CPU compute with the increase in the data parallelism degree.

**Region of Interest Warping**, or **RoIWarp**, is a form of [RoIPool](https://paperswithcode.com/method/roi-pooling) that is differentiable with respect to the box position. In practice, this takes the form of a RoIWarp layer followed by a standard [Max Pooling](https://paperswithcode.com/method/max-pooling) layer. The RoIWarp layer crops a feature map region and warps it into a target size by interpolation.

RoIWarp

Instance-aware Semantic Segmentation via Multi-task Network Cascades

ZeRO-Offload

ZeRO-Offload: Democratizing Billion-Scale Model Training

**Rectified Linear Units**, or **ReLUs**, are a type of activation function that are linear in the positive dimension, but zero in the negative dimension. The kink in the function is the source of the non-linearity. Linearity in the positive dimension has the attractive property that it prevents non-saturation of gradients (contrast with [sigmoid activations](https://paperswithcode.com/method/sigmoid-activation)), although for half of the real line its gradient is zero.

$$ f\left(x\right) = \max\left(0, x\right) $$

Source	ZeRO-Offload: Democratizing Billion-Scale Model Training
Year	2000
Data Source	CC BY-SA - https://paperswithcode.com