From eb7b0ddd7dc6901357c8fb9e92c9ef7bafd0b607 Mon Sep 17 00:00:00 2001
From: Yang Zhang <yangzhang@live.com>
Date: Thu, 27 Feb 2020 19:02:33 +0800
Subject: [PATCH] Force `cudnn` backend for depthwise convs when fp16 is
 enabled (#270)

---
 ppdet/modeling/backbones/blazenet.py  | 14 ++++++++------
 ppdet/modeling/backbones/mobilenet.py |  4 +++-
 2 files changed, 11 insertions(+), 7 deletions(-)

diff --git a/ppdet/modeling/backbones/blazenet.py b/ppdet/modeling/backbones/blazenet.py
index 8b4f63db3..d3987521a 100644
--- a/ppdet/modeling/backbones/blazenet.py
+++ b/ppdet/modeling/backbones/blazenet.py
@@ -19,6 +19,7 @@ from __future__ import print_function
 from paddle import fluid
 from paddle.fluid.param_attr import ParamAttr
 
+from ppdet.experimental import mixed_precision_global_state
 from ppdet.core.workspace import register
 
 __all__ = ['BlazeNet']
@@ -151,6 +152,7 @@ class BlazeNet(object):
         use_pool = not stride == 1
         use_double_block = double_channels is not None
         act = 'relu' if use_double_block else None
+        mixed_precision_enabled = mixed_precision_global_state() is not None
 
         if use_5x5kernel:
             conv_dw = self._conv_norm(
@@ -160,7 +162,7 @@ class BlazeNet(object):
                 stride=stride,
                 padding=2,
                 num_groups=in_channels,
-                use_cudnn=False,
+                use_cudnn=mixed_precision_enabled,
                 name=name + "1_dw")
         else:
             conv_dw_1 = self._conv_norm(
@@ -170,7 +172,7 @@ class BlazeNet(object):
                 stride=1,
                 padding=1,
                 num_groups=in_channels,
-                use_cudnn=False,
+                use_cudnn=mixed_precision_enabled,
                 name=name + "1_dw_1")
             conv_dw = self._conv_norm(
                 input=conv_dw_1,
@@ -179,7 +181,7 @@ class BlazeNet(object):
                 stride=stride,
                 padding=1,
                 num_groups=in_channels,
-                use_cudnn=False,
+                use_cudnn=mixed_precision_enabled,
                 name=name + "1_dw_2")
 
         conv_pw = self._conv_norm(
@@ -199,7 +201,7 @@ class BlazeNet(object):
                     num_filters=out_channels,
                     stride=1,
                     padding=2,
-                    use_cudnn=False,
+                    use_cudnn=mixed_precision_enabled,
                     name=name + "2_dw")
             else:
                 conv_dw_1 = self._conv_norm(
@@ -209,7 +211,7 @@ class BlazeNet(object):
                     stride=1,
                     padding=1,
                     num_groups=out_channels,
-                    use_cudnn=False,
+                    use_cudnn=mixed_precision_enabled,
                     name=name + "2_dw_1")
                 conv_dw = self._conv_norm(
                     input=conv_dw_1,
@@ -218,7 +220,7 @@ class BlazeNet(object):
                     stride=1,
                     padding=1,
                     num_groups=out_channels,
-                    use_cudnn=False,
+                    use_cudnn=mixed_precision_enabled,
                     name=name + "2_dw_2")
 
             conv_pw = self._conv_norm(
diff --git a/ppdet/modeling/backbones/mobilenet.py b/ppdet/modeling/backbones/mobilenet.py
index 56afdf964..e275fe1f1 100644
--- a/ppdet/modeling/backbones/mobilenet.py
+++ b/ppdet/modeling/backbones/mobilenet.py
@@ -20,6 +20,7 @@ from paddle import fluid
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.regularizer import L2Decay
 
+from ppdet.experimental import mixed_precision_global_state
 from ppdet.core.workspace import register
 
 __all__ = ['MobileNet']
@@ -104,6 +105,7 @@ class MobileNet(object):
                             stride,
                             scale,
                             name=None):
+        mixed_precision_enabled = mixed_precision_global_state() is not None
         depthwise_conv = self._conv_norm(
             input=input,
             filter_size=3,
@@ -111,7 +113,7 @@ class MobileNet(object):
             stride=stride,
             padding=1,
             num_groups=int(num_groups * scale),
-            use_cudnn=False,
+            use_cudnn=mixed_precision_enabled,
             name=name + "_dw")
 
         pointwise_conv = self._conv_norm(
-- 
GitLab