[NPU] Optimize Performance on NPU

Author: will-jl944

ppocr/modeling/backbones/custom_devices_layers.py

@@ -0,0 +1,37 @@
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+__all__ = ["AdaptiveAvgPool2D"]
+
+
+class AdaptiveAvgPool2D(nn.AdaptiveAvgPool2D):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        if paddle.device.get_device().startswith("npu"):
+            self.device = "npu"
+        else:
+            self.device = None
+
+        if isinstance(self._output_size, int) and self._output_size == 1:
+            self._gap = True
+        elif isinstance(self._output_size, tuple) and self._output_size[
+                0] == 1 and self._output_size[1] == 1:
+            self._gap = True
+        else:
+            self._gap = False
+
+    def forward(self, x):
+        if self.device == "npu" and self._gap:
+            # Global Average Pooling
+            N, C, _, _ = x.shape
+            x_mean = paddle.mean(x, axis=[2, 3])
+            x_mean = paddle.reshape(x_mean, [N, C, 1, 1])
+            return x_mean
+        else:
+            return F.adaptive_avg_pool2d(
+                x,
+                output_size=self._output_size,
+                data_format=self._data_format,
+                name=self._name, )

ppocr/modeling/backbones/det_pp_lcnet.py

@@ -18,11 +18,13 @@
 import paddle
 import paddle.nn as nn
 from paddle import ParamAttr
-from paddle.nn import AdaptiveAvgPool2D, BatchNorm, Conv2D, Dropout, Linear
+from paddle.nn import BatchNorm, Conv2D, Dropout, Linear
 from paddle.regularizer import L2Decay
 from paddle.nn.initializer import KaimingNormal
 from paddle.utils.download import get_path_from_url
 
+from .custom_devices_layers import AdaptiveAvgPool2D
+
 MODEL_URLS = {
     "PPLCNet_x0.25":
     "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/legendary_models/PPLCNet_x0_25_pretrained.pdparams",
@@ -191,7 +193,7 @@ def __init__(self,
             num_filters=make_divisible(16 * scale),
             stride=2)
 
-        self.blocks2 = nn.Sequential(* [
+        self.blocks2 = nn.Sequential(*[
             DepthwiseSeparable(
                 num_channels=make_divisible(in_c * scale),
                 num_filters=make_divisible(out_c * scale),
@@ -201,7 +203,7 @@ def __init__(self,
             for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks2"])
         ])
 
-        self.blocks3 = nn.Sequential(* [
+        self.blocks3 = nn.Sequential(*[
             DepthwiseSeparable(
                 num_channels=make_divisible(in_c * scale),
                 num_filters=make_divisible(out_c * scale),
@@ -211,7 +213,7 @@ def __init__(self,
             for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks3"])
         ])
 
-        self.blocks4 = nn.Sequential(* [
+        self.blocks4 = nn.Sequential(*[
             DepthwiseSeparable(
                 num_channels=make_divisible(in_c * scale),
                 num_filters=make_divisible(out_c * scale),
@@ -221,7 +223,7 @@ def __init__(self,
             for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks4"])
         ])
 
-        self.blocks5 = nn.Sequential(* [
+        self.blocks5 = nn.Sequential(*[
             DepthwiseSeparable(
                 num_channels=make_divisible(in_c * scale),
                 num_filters=make_divisible(out_c * scale),
@@ -231,7 +233,7 @@ def __init__(self,
             for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks5"])
         ])
 
-        self.blocks6 = nn.Sequential(* [
+        self.blocks6 = nn.Sequential(*[
             DepthwiseSeparable(
                 num_channels=make_divisible(in_c * scale),
                 num_filters=make_divisible(out_c * scale),

ppocr/modeling/backbones/rec_hgnet.py

@@ -16,10 +16,12 @@
 import paddle.nn as nn
 import paddle.nn.functional as F
 from paddle.nn.initializer import KaimingNormal, Constant
-from paddle.nn import Conv2D, BatchNorm2D, ReLU, AdaptiveAvgPool2D, MaxPool2D
+from paddle.nn import Conv2D, BatchNorm2D, ReLU, MaxPool2D
 from paddle.regularizer import L2Decay
 from paddle import ParamAttr
 
+from .custom_devices_layers import AdaptiveAvgPool2D
+
 kaiming_normal_ = KaimingNormal()
 zeros_ = Constant(value=0.)
 ones_ = Constant(value=1.)
@@ -204,13 +206,13 @@ def __init__(
 
         # stem
         stem_channels.insert(0, in_channels)
-        self.stem = nn.Sequential(* [
+        self.stem = nn.Sequential(*[
             ConvBNAct(
                 in_channels=stem_channels[i],
                 out_channels=stem_channels[i + 1],
                 kernel_size=3,
-                stride=2 if i == 0 else 1) for i in range(
-                    len(stem_channels) - 1)
+                stride=2 if i == 0 else 1)
+            for i in range(len(stem_channels) - 1)
         ])
 
         if self.det:

ppocr/modeling/backbones/rec_lcnetv3.py

@@ -21,9 +21,11 @@
 import paddle.nn.functional as F
 from paddle import ParamAttr
 from paddle.nn.initializer import Constant, KaimingNormal
-from paddle.nn import AdaptiveAvgPool2D, BatchNorm2D, Conv2D, Dropout, Hardsigmoid, Hardswish, Identity, Linear, ReLU
+from paddle.nn import BatchNorm2D, Conv2D, Dropout, Hardsigmoid, Hardswish, Identity, Linear, ReLU
 from paddle.regularizer import L2Decay
 
+from .custom_devices_layers import AdaptiveAvgPool2D
+
 NET_CONFIG_det = {
     "blocks2":
     #k, in_c, out_c, s, use_se