[Doc] update SAR status (#1789)

* update SAR status and fix torchscript export

* add reminder for SAR

Author: AllentDan

docs/en/04-supported-codebases/mmocr.md

@@ -251,6 +251,9 @@ Besides python API, mmdeploy SDK also provides other FFI (Foreign Function Inter
 
 - ABINet for TensorRT require pytorch1.10+ and TensorRT 8.4+.
 
+- SAR uses `valid_ratio` inside network inference, which causes performance drops. When the `valid_ratio`s between
+  testing image and the image for conversion are quite different, the gap would be enlarged.
+
 - For TensorRT backend, users have to choose the right config. For example, CRNN only accepts 1 channel input. Here is a recommendation table:
 
   | Model    | Config                                                     |

docs/zh_cn/04-supported-codebases/mmocr.md

@@ -255,6 +255,8 @@ print(texts)
 
 - ABINet 在 TensorRT 后端要求使用 pytorch1.10+， TensorRT 8.4+。
 
+- SAR 在网络推广中使用 `valid_ratio`，这会让导出的 ONNX 文件精度下降。当测试图片的 `valid_ratio`s 和转换图片的值差异很大，这种下降就会越多。
+
 - 对于 TensorRT 后端，用户需要使用正确的配置文件。比如 CRNN 只接受单通道输入。下面是一个示例表格:
 
   | Model    | Config                                                     |

mmdeploy/codebase/mmocr/models/text_recognition/sar_decoder.py

@@ -52,7 +52,7 @@ def parallel_sar_decoder__2d_attention(
         attn_mask = torch.zeros(bsz, T, h, w + 1, c).to(attn_weight.device)
         for i, valid_ratio in enumerate(valid_ratios):
             # use torch.ceil to replace original math.ceil and if else in mmocr
-            valid_width = torch.ceil(w * valid_ratio).long()
+            valid_width = torch.tensor(w * valid_ratio).ceil().long()
             # use narrow to replace original [valid_width:] in mmocr
             attn_mask[i].narrow(2, valid_width, w + 1 - valid_width)[:] = 1
         attn_mask = attn_mask[:, :, :, :w, :]
@@ -123,7 +123,7 @@ def sequential_sar_decoder__2d_attention(self,
         attn_mask = torch.zeros(bsz, c, h, w + 1).to(attn_weight.device)
         for i, valid_ratio in enumerate(valid_ratios):
             # use torch.ceil to replace original math.ceil and if else in mmocr
-            valid_width = torch.ceil(w * valid_ratio).long()
+            valid_width = torch.tensor(w * valid_ratio).ceil().long()
             # use narrow to replace original [valid_width:] in mmocr
             attn_mask[i].narrow(2, valid_width, w + 1 - valid_width)[:] = 1
         attn_mask = attn_mask[:, :, :, :w]

mmdeploy/codebase/mmocr/models/text_recognition/sar_encoder.py

@@ -54,7 +54,7 @@ def sar_encoder__forward(
         T = holistic_feat.size(1)
         for i, valid_ratio in enumerate(valid_ratios):
             # use torch.ceil to replace original math.ceil and if else in mmocr
-            valid_step = torch.ceil(T * valid_ratio).long() - 1
+            valid_step = torch.tensor(T * valid_ratio).ceil().long() - 1
             valid_hf.append(holistic_feat[i, valid_step, :])
         valid_hf = torch.stack(valid_hf, dim=0)
     else:

tests/regression/mmocr.yml

@@ -307,7 +307,6 @@ models:
     pipelines:
       - *pipeline_ts_recognition_fp32
       - *pipeline_ort_recognition_dynamic_fp32
-      - *pipeline_trt_recognition_dynamic_fp32_H48_C3
 
   - name: SATRN
     metafile: configs/textrecog/satrn/metafile.yml