[Fix] Dbnet performance of trt8 (#278)

* compatible trt version for dbnet

* judge inside rewrite

.isort.cfg

@@ -1,2 +1,2 @@
 [settings]
-known_third_party = h5py,m2r,mmcls,mmcv,mmdet,mmedit,mmocr,mmseg,ncnn,numpy,onnx,onnxruntime,packaging,pyppeteer,pyppl,pytest,pytorch_sphinx_theme,recommonmark,setuptools,sphinx,tensorrt,torch,torchvision
+known_third_party = h5py,m2r,mmcls,mmcv,mmdet,mmedit,mmocr,mmseg,ncnn,numpy,onnx,onnxruntime,packaging,pyppeteer,pyppl,pytest,pytorch_sphinx_theme,recommonmark,setuptools,sphinx,tensorrt,torch,torchvision,version

docs/benchmark.md

@@ -673,7 +673,7 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](doc
     <td class="tg-0lax">model config file</td>
   </tr>
   <tr>
-    <td class="tg-nrix" rowspan="3">DBNet</td>
+    <td class="tg-nrix" rowspan="3">DBNet*</td>
     <td class="tg-nrix" rowspan="3">TextDetection</td>
     <td class="tg-baqh">recall</td>
     <td class="tg-baqh">0.7310</td>
@@ -823,6 +823,8 @@ Users can directly test the performance through [how_to_evaluate_a_model.md](doc
 
 
 ### Notes
-As some datasets contains images with various resolutions in codebase like MMDet. The speed benchmark is gained through static configs in MMDeploy, while the performance benchmark is gained through dynamic ones.
+- As some datasets contains images with various resolutions in codebase like MMDet. The speed benchmark is gained through static configs in MMDeploy, while the performance benchmark is gained through dynamic ones.
 
-Some int8 performance benchmarks of tensorrt require nvidia cards with tensor core, or the performance would drop heavily.
+- Some int8 performance benchmarks of TensorRT require nvidia cards with tensor core, or the performance would drop heavily.
+
+- DBNet uses the interpolate mode `nearest` in the neck of the model, which TensorRT-7 applies quite different strategy from pytorch. To make the repository compatible with TensorRT-7, we rewrite the neck to use the interpolate mode `bilinear` which improves final detection performance. To get the matched performance with Pytorch, TensorRT-8+ is recommended, which the interpolate methods are all the same as Pytorch.

mmdeploy/codebase/mmocr/models/text_detection/fpn_cat.py

@@ -1,6 +1,7 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 import torch
 import torch.nn.functional as F
+import version
 
 from mmdeploy.core import FUNCTION_REWRITER
 
@@ -11,7 +12,7 @@ def fpnc__forward__tensorrt(ctx, self, inputs, **kwargs):
     """Rewrite `forward` of FPNC for tensorrt backend.
 
     Rewrite this function to replace nearest upsampling with bilinear
-    upsampling. Tensorrt backend applies different nearest sampling strategy
+    upsampling. TensorRT-7 backend applies different nearest sampling strategy
     from pytorch, which heavily influenced the final performance.
 
     Args:
@@ -24,6 +25,10 @@ def fpnc__forward__tensorrt(ctx, self, inputs, **kwargs):
         outs (Tensor): A feature map output from FPNC. The tensor shape
             (N, C, H, W).
     """
+    # TensorRT version 8+ matches the upsampling with pytorch
+    import tensorrt as trt
+    apply_rewrite = version.parse(trt.__version__) < version.parse('8')
+    mode = 'bilinear' if apply_rewrite else 'nearest'
 
     assert len(inputs) == len(self.in_channels)
     # build laterals
@@ -36,7 +41,7 @@ def fpnc__forward__tensorrt(ctx, self, inputs, **kwargs):
     for i in range(used_backbone_levels - 1, 0, -1):
         prev_shape = laterals[i - 1].shape[2:]
         laterals[i - 1] += F.interpolate(
-            laterals[i], size=prev_shape, mode='bilinear')
+            laterals[i], size=prev_shape, mode=mode)
     # build outputs
     # part 1: from original levels
     outs = [
@@ -44,8 +49,7 @@ def fpnc__forward__tensorrt(ctx, self, inputs, **kwargs):
     ]
 
     for i, out in enumerate(outs):
-        outs[i] = F.interpolate(
+        outs[i] = F.interpolate(outs[i], size=outs[0].shape[2:], mode=mode)
-            outs[i], size=outs[0].shape[2:], mode='bilinear')
     out = torch.cat(outs, dim=1)
 
     if self.conv_after_concat: