From aa417b15332c548de9e3d1bc06b5c67bb52c9e1b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=8C=83=E7=A8=8B?= <568994792@qq.com>
Date: Thu, 10 Nov 2022 07:50:32 +0000
Subject: [PATCH] add
 ACL_PyTorch/contrib/cv/segmentation/Swin-Transformer-Semantic-Segmentation.
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Signed-off-by: 范程 <568994792@qq.com>
---
 .../Swin-Transformer-Semantic-Segmentation          | 13 +++++++++++++
 1 file changed, 13 insertions(+)
 create mode 100644 ACL_PyTorch/contrib/cv/segmentation/Swin-Transformer-Semantic-Segmentation

diff --git a/ACL_PyTorch/contrib/cv/segmentation/Swin-Transformer-Semantic-Segmentation b/ACL_PyTorch/contrib/cv/segmentation/Swin-Transformer-Semantic-Segmentation
new file mode 100644
index 0000000000..096f12667a
--- /dev/null
+++ b/ACL_PyTorch/contrib/cv/segmentation/Swin-Transformer-Semantic-Segmentation
@@ -0,0 +1,13 @@
+1.性能对比
+在 310P 设备上，当 batchsize 为 1 时模型性能为 19.33 fps.
+| batchsize | T4性能    | 310P性能    | 310P/T4   |
+|-----------|---------- |----------- |-----------|
+| 1         | 1.72 fps  | 19.33 fps  | 11.23倍     |
+
+注：当 batchsize 为 4 或更高时，因内存不足导致推理失败，无法获取性能数据。
+
+2.精度对比
+自测了 batchsize 为 1 精度，比开源仓精度的高。
+| Model                                  | batchsize | mIoU(NPU) | mIoU(开源)|
+|----------------------------------------|-----------|-----------|----------|
+| Swin-Transformer-Semantic-Segmentation | 1         | 48.06%    | 47.64%   |
\ No newline at end of file
-- 
Gitee