From 801159ceeb20e05ca71582f640cba05ee2651ac5 Mon Sep 17 00:00:00 2001
From: hlygit66666 <32728786+hlygit66666@users.noreply.github.com>
Date: Wed, 26 Jan 2022 19:43:42 +0800
Subject: [PATCH] Add FuseBatchNormAddActPass and unittest. (#39178)

* add fuse_relu_depthwise_conv_pass unittest

* fix atol and rtol

* fix according to review

* add FuseBatchNormAddActPass and unittest

* Update test_dist_fuse_bn_add_act_pass.py

* solve conflict
---
 python/paddle/distributed/passes/cpp_pass.py  |  13 +++
 .../test_dist_fuse_bn_add_act_pass.py         | 100 ++++++++++++++++++
 2 files changed, 113 insertions(+)
 create mode 100644 python/paddle/fluid/tests/unittests/distributed_passes/test_dist_fuse_bn_add_act_pass.py

diff --git a/python/paddle/distributed/passes/cpp_pass.py b/python/paddle/distributed/passes/cpp_pass.py
index 6b8ea30f3ba..97243b805cd 100644
--- a/python/paddle/distributed/passes/cpp_pass.py
+++ b/python/paddle/distributed/passes/cpp_pass.py
@@ -39,3 +39,16 @@ class FuseBatchNormActPass(CPPPassWrapper):
 
     def _type(self):
         return PassType.FUSION_OPT
+
+
+@register_pass("fuse_bn_add_act")
+class FuseBatchNormAddActPass(CPPPassWrapper):
+    def __init__(self):
+        super(FuseBatchNormAddActPass, self).__init__()
+
+    @property
+    def cpp_name(self):
+        return "fuse_bn_add_act_pass"
+
+    def _type(self):
+        return PassType.FUSION_OPT
diff --git a/python/paddle/fluid/tests/unittests/distributed_passes/test_dist_fuse_bn_add_act_pass.py b/python/paddle/fluid/tests/unittests/distributed_passes/test_dist_fuse_bn_add_act_pass.py
new file mode 100644
index 00000000000..1b01260eaf2
--- /dev/null
+++ b/python/paddle/fluid/tests/unittests/distributed_passes/test_dist_fuse_bn_add_act_pass.py
@@ -0,0 +1,100 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import paddle
+import paddle.distributed.fleet as fleet
+import numpy as np
+import paddle.nn as nn
+from paddle.distributed.passes import new_pass, PassManager
+import unittest
+from dist_pass_test_base import DistPassTestBase
+
+
+class BatchNormAddActNet(nn.Layer):
+    def __init__(self):
+        super(BatchNormAddActNet, self).__init__()
+
+        self.conv1 = nn.Conv2D(3, 8, (3, 3), data_format="NHWC")
+        self.conv2 = nn.Conv2D(3, 8, (3, 3), data_format="NHWC")
+        self.bn1 = nn.BatchNorm2D(8, data_format="NHWC")
+        self.bn2 = nn.BatchNorm2D(8, data_format="NHWC")
+        self.relu = nn.ReLU()
+
+    def forward(self, x):
+        y = self.conv1(x)
+        y = self.bn1(y)
+        out = self.conv2(x)
+        out = self.bn2(out) + y
+        out = self.relu(out)
+        out = paddle.flatten(out, 1)
+        return out
+
+
+class TestFuseBatchNormAddActPass(DistPassTestBase):
+    def init(self):
+        self.atol = 1e-4
+        self.rtol = 1e-4
+
+    def get_model(self, place, batch_size=32, image_shape=[224, 224, 3]):
+        image = paddle.static.data(
+            shape=[batch_size] + image_shape, dtype='float32', name='image')
+
+        model = BatchNormAddActNet()
+        pred_out = model(image)
+        loss = paddle.mean(pred_out)
+        optimizer = paddle.optimizer.Adam(learning_rate=1e-3)
+
+        dist_strategy = fleet.DistributedStrategy()
+        dist_strategy.fuse_all_reduce_ops = False
+        dist_strategy.without_graph_optimization = True
+        dist_strategy.amp = True
+        dist_strategy.amp_configs = {
+            "init_loss_scaling": 32768,
+            "use_dynamic_loss_scaling": True,
+        }
+        fleet.init(is_collective=True, strategy=dist_strategy)
+        optimizer = fleet.distributed_optimizer(optimizer)
+        optimizer.minimize(loss)
+
+        rank = paddle.distributed.get_rank()
+
+        def reader():
+            seed = int(os.environ.get("SEED", 0))
+            np.random.seed(seed + rank)
+            for _ in range(10):
+                image_np = np.random.random(size=image.shape).astype('float32')
+                yield image_np,
+
+        main_program = paddle.static.default_main_program()
+        startup_program = paddle.static.default_startup_program()
+        return main_program, startup_program, [image], [loss], reader
+
+    def apply_passes(self, main_prog, startup_prog):
+        pass_manager = PassManager([new_pass("fuse_bn_add_act")])
+        pass_manager.apply([main_prog], [startup_prog])
+        print(pass_manager.names)
+
+        op_type = []
+        for op in main_prog.global_block().ops:
+            op_type.append(op.type)
+        self.assertTrue("fused_bn_add_activation" in op_type)
+        self.assertTrue("fused_bn_add_activation_grad" in op_type)
+
+    def test_fuse_bn_add_act(self):
+        self.check_main()
+
+
+if __name__ == "__main__":
+    unittest.main()
-- 
GitLab