pytorch/vision

Add uint8 support for interpolate and grid-sample in PyTorch

Open

#2,289 建立於 2020年6月5日

在 GitHub 查看
 (4 留言) (6 反應) (1 負責人)Python (15,050 star) (6,858 fork)batch import
help wantedhigh prioritymodule: transforms

描述

🚀 Feature

With the addition of torch tensor support for the transforms (following https://github.com/pytorch/vision/issues/1375), there are two operators that will require extra attention in order for them to be as efficient as the PIL implementations.

Indeed, while we can implement resize via torch.nn.functional.interpolate and rotate via torch.nn.functional.grid_sample, those operators for now only support floating-point types, so for now we need to perform a .float() -> interpolate() -> byte() in order to maintain compatibility, which is wasteful.

It would be great if they could be extended to support uint8 (and maybe other integer types) as well.

A first PR adding support for uint8 to nearest mode interpolate has been sent in https://github.com/pytorch/pytorch/pull/35029

Something to keep in mind: the interpolate function is under optimization in https://github.com/pytorch/pytorch/pull/34864, so this should be kept in mind to avoid conflicts.

貢獻者指南

技術棧
pythonpytorch
領域
machine learning
議題類型
feature
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
4
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
over 1 week
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
blocked
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
clear
前置要求
Basic PyTorch knowledgeFamiliarity with C++ and CUDAUnderstanding of interpolation algorithms
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
15
研究方向
The issue requests adding uint8 support to torch.nn.functional.interpolate and grid sample. A preliminary PR for nearest mode interpolate exists at https://github.com/pytorch/pytorch/pull/35029. Additionally, there is ongoing optimization in PR #34864. Primary files to modify are in `aten/src/ATen/native/` for CPU kernels and `aten/src/ATen/native/cuda/` for GPU. The current code path involves a float conversion, so the fix would add integer type dispatch in the kernel implementations.