cherry-pick from commit 197a7ad

mmdeploy/mmcv/cnn/__init__.py

@@ -1,10 +1,8 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-from .conv2d_adaptive_padding import (
+from . import context_block  # noqa: F401,F403
-    AdaptivePadOp, conv2d_adaptive_padding__forward__tensorrt)
+from . import conv2d_adaptive_padding  # noqa: F401,F403
-from .transformer import (MultiHeadAttentionop,
+from . import hsigmoid  # noqa: F401,F403
-                          multiheadattention__forward__ncnn)
+from . import hswish  # noqa: F401,F403
+from .transformer import MultiHeadAttentionop
 
-__all__ = [
+__all__ = ['conv2d_adaptive_padding', 'MultiHeadAttentionop']
-    'multiheadattention__forward__ncnn', 'MultiHeadAttentionop',
-    'conv2d_adaptive_padding__forward__tensorrt', 'AdaptivePadOp'
-]

mmdeploy/mmcv/cnn/conv2d_adaptive_padding.py

@@ -5,7 +5,7 @@ import torch
 import torch.nn.functional as F
 
 from mmdeploy.core import FUNCTION_REWRITER
-from mmdeploy.utils import Backend, is_dynamic_shape
+from mmdeploy.utils import Backend, is_dynamic_batch, is_dynamic_shape
 
 
 def compute_padding(input_size, kernel_size, stride, dilation):
@@ -33,25 +33,22 @@ def compute_padding(input_size, kernel_size, stride, dilation):
 
 
 class AdaptivePadOp(torch.autograd.Function):
-    """AdaptivePadOp."""
+    """Dummy adaptive pad op."""
 
     @staticmethod
-    def forward(ctx, x, kernel, stride, dilation):
+    def forward(ctx, x, padded):
-        padded = compute_padding(x.shape[2:], kernel, stride, dilation)
         if padded is not None:
             x = F.pad(x, padded)
         return x
 
     @staticmethod
-    def symbolic(g, x, kernel, stride, dilation):
+    def symbolic(g, x, padded):
-        padded = compute_padding(x.type().sizes()[2:], kernel, stride,
-                                 dilation)
         if padded is None:
             return g.op('Identity', x)
         padded = g.op(
             'Constant', value_t=torch.tensor(padded, dtype=torch.int64))
         constant_value = g.op(
-            'Constant', value_t=torch.tensor(0, dtype=torch.float32))
+            'Constant', value_t=torch.tensor(0, dtype=torch.int64))
         return g.op(
             'Pad', x, padded, constant_value, mode_s='constant', outputs=1)
 
@@ -76,9 +73,12 @@ def conv2d_adaptive_padding__forward__tensorrt(ctx, self, x):
 
     deploy_cfg = ctx.cfg
     is_dynamic_flag = is_dynamic_shape(deploy_cfg)
-    if not is_dynamic_flag:
+    if (not is_dynamic_flag) or is_dynamic_batch(deploy_cfg):
-        x = AdaptivePadOp.apply(x, self.weight.shape[2:], self.stride,
+        padded = compute_padding(x.shape[2:], self.weight.shape[2:],
-                                self.dilation)
+                                 self.stride, self.dilation)
+        if padded is not None:
+            padded = [int(_) for _ in padded]
+        x = AdaptivePadOp.apply(x, padded)
         return F.conv2d(x, self.weight, self.bias, self.stride, self.padding,
                         self.dilation, self.groups)
     else:

tests/regression/mmcls.yml

@@ -217,6 +217,6 @@ models:
     model_configs:
       - configs/efficientnet/efficientnet-b0_8xb32_in1k.py
     pipelines:
-      - *pipeline_trt_static_fp32
+      - *pipeline_ort_static_fp32
-      - *pipeline_trt_static_fp16
+      - convert_image: *convert_image
-      - *pipeline_trt_static_int8
+        deploy_config: configs/mmcls/classification_tensorrt_dynamic-224x224-224x224.py