Add SyncBatchNorm (#26032)

* add SyncBatchNorm,test=develop

paddle/fluid/pybind/op_function_generator.cc

@@ -57,6 +57,9 @@ std::map<std::string, std::set<std::string>> op_outs_map = {
     {"batch_norm",
      {"Y", "MeanOut", "VarianceOut", "SavedMean", "SavedVariance",
       "ReserveSpace"}},
+    {"sync_batch_norm",
+     {"Y", "MeanOut", "VarianceOut", "SavedMean", "SavedVariance",
+      "ReserveSpace"}},
 };
 
 // NOTE(zhiqiu): Commonly, the outputs in auto-generated OP function are
@@ -76,6 +79,7 @@ std::map<std::string, std::set<std::string>> op_passing_outs_map = {
      {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut"}},
     {"momentum", {"ParamOut", "VelocityOut"}},
     {"batch_norm", {"MeanOut", "VarianceOut"}},
+    {"sync_batch_norm", {"MeanOut", "VarianceOut"}},
     {"accuracy", {"Correct", "Total"}},
     {"fill_constant", {"Out"}},
     {"matmul", {"Out"}},

python/paddle/fluid/dygraph/nn.py

@@ -35,7 +35,7 @@ __all__ = [
     'Conv2D', 'Conv3D', 'Pool2D', 'Linear', 'BatchNorm', 'Dropout', 'Embedding',
     'GRUUnit', 'InstanceNorm', 'LayerNorm', 'NCE', 'PRelu',
     'BilinearTensorProduct', 'Conv2DTranspose', 'Conv3DTranspose', 'GroupNorm',
-    'SpectralNorm', 'TreeConv', 'Flatten'
+    'SpectralNorm', 'TreeConv', 'Flatten', 'SyncBatchNorm'
 ]
 
 
@@ -3202,6 +3202,220 @@ class TreeConv(layers.Layer):
         return self._helper.append_activation(pre_activation, act=self._act)
 
 
+class SyncBatchNorm(layers.Layer):
+    """
+    This interface is used to construct a callable object of the ``SyncBatchNorm`` class.
+    It implements the function of the Cross-GPU Synchronized Batch Normalization Layer, and can 
+    be used as a normalizer function for other operations, such as conv2d and fully connected 
+    operations.
+    The data is normalized by the mean and variance of the channel based on whole mini-batch
+    , which including data in all gpus.
+    Refer to `Batch Normalization: Accelerating Deep Network Training by Reducing
+    Internal Covariate Shift <https://arxiv.org/pdf/1502.03167.pdf>`_
+    for more details.
+
+    When model in training mode, the :math:`\\mu_{\\beta}` 
+    and :math:`\\sigma_{\\beta}^{2}` are the statistics of whole mini-batch data in all gpus.
+    Calculated as follows:
+
+    ..  math::
+
+        \\mu_{\\beta} &\\gets \\frac{1}{m} \\sum_{i=1}^{m} x_i \\qquad &//\\
+        \ mini-batch\ mean \\\\
+        \\sigma_{\\beta}^{2} &\\gets \\frac{1}{m} \\sum_{i=1}^{m}(x_i - \\
+        \\mu_{\\beta})^2 \\qquad &//\ mini-batch\ variance \\\\
+
+    - :math:`x` : whole mini-batch data in all gpus
+    - :math:`m` : the size of the whole mini-batch data
+
+    When model in evaluation mode, the :math:`\\mu_{\\beta}`
+    and :math:`\\sigma_{\\beta}^{2}` are global statistics (moving_mean and moving_variance, 
+    which usually got from the pre-trained model). Global statistics calculated as follows:
+
+    .. math::
+        moving\_mean = moving\_mean * momentum + \mu_{\beta} * (1. - momentum) \quad &// global mean \\
+        moving\_variance = moving\_variance * momentum + \sigma_{\beta}^{2} * (1. - momentum) \quad &// global variance \\
+
+    The formula of normalization is as follows:
+ 
+    ..  math::
+
+        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
+        \\sigma_{\\beta}^{2} + \\eps}} \\qquad &//\ normalize \\\\
+        y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift
+
+    - :math:`\\eps` : add a smaller value to the variance to prevent division by zero
+    - :math:`\\gamma` : trainable scale parameter vector
+    - :math:`\\beta` : trainable shift parameter vector 
+
+    Parameters:
+        num_features(int): Indicate the number of channels of the input ``Tensor``.
+        epsilon(float, optional): The small value added to the variance to prevent division by zero. Default: 1e-5.
+        momentum(float, optional): The value used for the moving_mean and moving_var computation. Default: 0.9.
+        weight_attr(ParamAttr|bool, optional): The parameter attribute for Parameter `scale`
+             of this layer. If it is set to None or one attribute of ParamAttr, this layerr
+             will create ParamAttr as param_attr. If the Initializer of the param_attr
+             is not set, the parameter is initialized with Xavier. If it is set to False, 
+             this layer will not have trainable scale parameter. Default: None.
+        bias_attr(ParamAttr|bool, optional): The parameter attribute for the bias of this layer.
+             If it is set to None or one attribute of ParamAttr, this layer
+             will create ParamAttr as bias_attr. If the Initializer of the bias_attr
+             is not set, the bias is initialized zero. If it is set to False, this layer will not 
+             have trainable bias parameter. Default: None.
+        track_running_stats(bool, optional): Whether to compute global stats, which including running mean and 
+             running variance. Default: True.
+
+    Returns:
+        None
+
+    Examples:
+        .. code-block:: python
+
+          import paddle
+          import paddle.nn as nn
+          import numpy as np
+
+          x = np.array([[[[0.3, 0.4], [0.3, 0.07]], [[0.83, 0.37], [0.18, 0.93]]]]).astype('float32')
+          paddle.disable_static()
+          x = paddle.to_tensor(x)
+          if paddle.fluid.is_compiled_with_cuda():
+              sync_batch_norm = nn.SyncBatchNorm(2)
+              hidden1 = sync_batch_norm(x)
+              print(hidden1.numpy())
+              # [[[[0.26824948, 1.0936325],[0.26824948, -1.6301316]],[[ 0.8095662, -0.665287],[-1.2744656, 1.1301866 ]]]]
+    """
+
+    def __init__(self,
+                 num_features,
+                 epsilon=1e-05,
+                 momentum=0.9,
+                 track_running_stats=True,
+                 weight_attr=None,
+                 bias_attr=None,
+                 data_format='NCHW',
+                 name=None):
+        super(SyncBatchNorm, self).__init__()
+        self._weight_attr = weight_attr
+        self._bias_attr = bias_attr
+        self._num_features = num_features
+        self._data_layout = data_format
+        self._momentum = momentum
+        self._epsilon = epsilon
+        self._track_running_stats = track_running_stats
+
+        if self._track_running_stats == False:
+            logging.warn(
+                "moving mean and moving variance will be calculated whether `track_running_stats` is set to `True` or `False`, we will fix it in the next version."
+            )
+
+        param_shape = [self._num_features]
+
+        # create parameter
+        if weight_attr == False:
+            self.weight = self.create_parameter(
+                attr=None, shape=param_shape, default_initializer=Constant(1.0))
+            self.weight.stop_gradient = True
+        else:
+            self.weight = self.create_parameter(
+                attr=self._weight_attr,
+                shape=param_shape,
+                default_initializer=Constant(1.0))
+            self.weight.stop_gradient = self._weight_attr != None and self._weight_attr.learning_rate == 0.
+
+        if bias_attr == False:
+            self.bias = self.create_parameter(
+                attr=None,
+                shape=param_shape,
+                default_initializer=Constant(0.0),
+                is_bias=True)
+            self.bias.stop_gradient = True
+        else:
+            self.bias = self.create_parameter(
+                attr=self._bias_attr, shape=param_shape, is_bias=True)
+            self.bias.stop_gradient = self._weight_attr != None and self._weight_attr.learning_rate == 0.
+
+        self._mean = self.create_parameter(
+            attr=ParamAttr(
+                name=None,
+                initializer=Constant(0.0),
+                trainable=False,
+                do_model_average=True),
+            shape=param_shape,
+            dtype=self._dtype)
+        self._mean.stop_gradient = True
+
+        self._variance = self.create_parameter(
+            attr=ParamAttr(
+                name=None,
+                initializer=Constant(1.0),
+                trainable=False,
+                do_model_average=True),
+            shape=param_shape,
+            dtype=self._dtype)
+        self._variance.stop_gradient = True
+
+    def forward(self, x):
+        # create output
+        # mean and mean_out share the same memory
+        mean_out = self._mean
+        # variance and variance out share the same memory
+        variance_out = self._variance
+
+        ### train mode: use mini-batch stats, eval mode: use global stats
+        if in_dygraph_mode():
+            attrs = ("momentum", self._momentum, "epsilon", self._epsilon,
+                     "is_test", not self.training, "data_layout",
+                     self._data_layout, "use_mkldnn", False, "fuse_with_relu",
+                     False, "use_global_stats", not self.training,
+                     'trainable_statistics', False)
+            sync_batch_norm_out, _, _, _, _, _ = core.ops.sync_batch_norm(
+                x, self.weight, self.bias, self._mean, self._variance, mean_out,
+                variance_out, *attrs)
+
+            return sync_batch_norm_out
+
+        check_variable_and_dtype(x, 'input', ['float16', 'float32', 'float64'],
+                                 'BatchNorm')
+
+        attrs = {
+            "momentum": self._momentum,
+            "epsilon": self._epsilon,
+            "is_test": not self.training,
+            "data_layout": self._data_layout,
+            "use_mkldnn": False,
+            "fuse_with_relu": False,
+            "use_global_stats": not self.training,
+            "trainable_statistics": False,
+        }
+
+        inputs = {
+            "X": [x],
+            "Scale": [self.weight],
+            "Bias": [self.bias],
+            "Mean": [self._mean],
+            "Variance": [self._variance]
+        }
+
+        saved_mean = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype, stop_gradient=True)
+        saved_variance = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype, stop_gradient=True)
+        sync_batch_norm_out = self._helper.create_variable_for_type_inference(
+            self._dtype)
+
+        outputs = {
+            "Y": [sync_batch_norm_out],
+            "MeanOut": [mean_out],
+            "VarianceOut": [variance_out],
+            "SavedMean": [saved_mean],
+            "SavedVariance": [saved_variance]
+        }
+
+        self._helper.append_op(
+            type="sync_batch_norm", inputs=inputs, outputs=outputs, attrs=attrs)
+        return sync_batch_norm_out
+
+
 class Flatten(layers.Layer):
     """
     :alias_main: paddle.nn.Flatten

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -106,6 +106,7 @@ if (NOT ${WITH_GPU})
     list(REMOVE_ITEM TEST_OPS test_parallel_dygraph_se_resnext)
     LIST(REMOVE_ITEM TEST_OPS test_parallel_dygraph_sparse_embedding)
     LIST(REMOVE_ITEM TEST_OPS test_parallel_dygraph_transformer)
+    LIST(REMOVE_ITEM TEST_OPS test_parallel_dygraph_sync_batch_norm)
     LIST(REMOVE_ITEM TEST_OPS test_imperative_auto_mixed_precision)
 elseif(${CUDNN_VERSION} VERSION_LESS 7100)
     LIST(REMOVE_ITEM TEST_OPS test_conv2d_fusion_op)

python/paddle/fluid/tests/unittests/parallel_dygraph_sync_batch_norm.py

+# Copyright (c) 2020 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.
+
+from __future__ import print_function
+
+import os
+import contextlib
+import unittest
+import numpy as np
+import six
+import pickle
+
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.dygraph as dygraph
+from paddle.fluid import core
+from paddle.fluid.optimizer import SGDOptimizer
+from paddle.nn import Conv2D, Pool2D, Linear, SyncBatchNorm
+from paddle.fluid.dygraph.base import to_variable
+
+from test_dist_base import runtime_main, TestParallelDyGraphRunnerBase
+
+
+class TestLayer(fluid.dygraph.Layer):
+    def __init__(self,
+                 num_channels,
+                 num_filters,
+                 filter_size,
+                 stride=1,
+                 groups=1,
+                 act=None):
+        super(TestLayer, self).__init__()
+
+        self._conv = Conv2D(
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=(filter_size - 1) // 2,
+            groups=groups,
+            act=None,
+            bias_attr=False)
+
+        self._sync_batch_norm = SyncBatchNorm(num_filters)
+
+        self._conv2 = Conv2D(
+            num_channels=num_filters,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=(filter_size - 1) // 2,
+            groups=groups,
+            act=None,
+            bias_attr=False)
+
+        self._sync_batch_norm2 = SyncBatchNorm(
+            num_filters,
+            weight_attr=False,
+            bias_attr=False,
+            track_running_stats=False)
+
+    def forward(self, inputs):
+        y = self._conv(inputs)
+        y = self._sync_batch_norm(y)
+        y = self._conv2(y)
+        y = self._sync_batch_norm2(y)
+
+        return y
+
+
+class TestSyncBatchNorm(TestParallelDyGraphRunnerBase):
+    def get_model(self):
+        model = TestLayer(3, 64, 7)
+        train_reader = paddle.batch(
+            paddle.dataset.flowers.test(use_xmap=False),
+            batch_size=32,
+            drop_last=True)
+        opt = fluid.optimizer.Adam(
+            learning_rate=1e-3, parameter_list=model.parameters())
+        return model, train_reader, opt
+
+    def run_one_loop(self, model, opt, data):
+        batch_size = len(data)
+        dy_x_data = np.array([x[0].reshape(3, 224, 224)
+                              for x in data]).astype('float32')
+        img = to_variable(dy_x_data)
+        img.stop_gradient = False
+
+        out = model(img)
+
+        out = fluid.layers.mean(out)
+
+        return out
+
+
+if __name__ == "__main__":
+    runtime_main(TestSyncBatchNorm)

python/paddle/fluid/tests/unittests/test_layers.py

@@ -283,6 +283,24 @@ class TestLayer(LayerTest):
             with self.assertRaises(ValueError):
                 lm(base.to_variable(inp))
 
+    def test_SyncBatchNorm(self):
+        if core.is_compiled_with_cuda():
+            with self.static_graph():
+                t = layers.data(name='t', shape=[-1, 3, 5, 5], dtype='float32')
+                my_sync_bn = nn.SyncBatchNorm(3)
+                ret = my_sync_bn(t)
+                static_ret = self.get_static_graph_result(
+                    feed={'t': np.ones(
+                        [3, 3, 5, 5], dtype='float32')},
+                    fetch_list=[ret])[0]
+
+            with self.dynamic_graph():
+                t = np.ones([3, 3, 5, 5], dtype='float32')
+                my_syncbn = paddle.nn.SyncBatchNorm(3)
+                dy_ret = my_syncbn(base.to_variable(t))
+                dy_ret_value = dy_ret.numpy()
+            self.assertTrue(np.array_equal(static_ret, static_ret))
+
     def test_relu(self):
         with self.static_graph():
             t = layers.data(name='t', shape=[3, 3], dtype='float32')

python/paddle/fluid/tests/unittests/test_parallel_dygraph_sync_batch_norm.py

+# Copyright (c) 2020 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.
+
+from __future__ import print_function
+import unittest
+from test_dist_base import TestDistBase
+import paddle.fluid as fluid
+
+import os
+flag_name = os.path.splitext(__file__)[0]
+
+
+class TestParallelDygraphMnist(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = False
+        self._nccl2_mode = True
+        self._dygraph = False  #True
+
+    def test_mnist(self):
+        if fluid.core.is_compiled_with_cuda():
+            self.check_with_place(
+                "parallel_dygraph_sync_batch_norm.py",
+                delta=1e-5,
+                check_error_log=True,
+                log_name=flag_name)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_sync_batch_norm_op.py

@@ -25,6 +25,7 @@ import six
 import paddle.fluid.core as core
 import paddle.fluid as fluid
 from paddle.fluid import compiler
+from paddle.fluid import Program, program_guard
 
 from op_test import OpTest, _set_use_system_allocator
 
@@ -202,5 +203,22 @@ class TestFP16SyncBatchNormOpTraining(TestSyncBatchNormOpTraining):
         self.atol = 1e-2
 
 
+class TestDygraphSyncBatchNormAPIError(unittest.TestCase):
+    def test_errors(self):
+        if not core.is_compiled_with_cuda():
+            return
+
+        with program_guard(Program(), Program()):
+            my_sync_batch_norm = fluid.dygraph.SyncBatchNorm(10)
+            x1 = fluid.create_lod_tensor(
+                np.array([-1, 3, 5, 5]), [[1, 1, 1, 1]], fluid.CUDAPlace(0))
+            self.assertRaises(TypeError, my_sync_batch_norm, x1)
+
+            # the input dtype of SyncBatchNorm must be float16 or float32 or float64
+            # float16 only can be set on GPU place
+            x2 = fluid.layers.data(name='x2', shape=[3, 4, 5, 6], dtype="int32")
+            self.assertRaises(TypeError, my_sync_batch_norm, x2)
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/nn/__init__.py

@@ -92,6 +92,7 @@ from .layer.loss import BCELoss  #DEFINE_ALIAS
 from .layer.loss import KLDivLoss  #DEFINE_ALIAS
 from .layer.loss import MarginRankingLoss  #DEFINE_ALIAS
 from .layer.norm import BatchNorm  #DEFINE_ALIAS
+from .layer.norm import SyncBatchNorm  #DEFINE_ALIAS
 from .layer.norm import GroupNorm  #DEFINE_ALIAS
 from .layer.norm import LayerNorm  #DEFINE_ALIAS
 from .layer.norm import SpectralNorm  #DEFINE_ALIAS

python/paddle/nn/layer/__init__.py

@@ -65,6 +65,7 @@ from .loss import BCELoss  #DEFINE_ALIAS
 from .loss import KLDivLoss  #DEFINE_ALIAS
 from .loss import MarginRankingLoss  #DEFINE_ALIAS
 from .norm import BatchNorm  #DEFINE_ALIAS
+from .norm import SyncBatchNorm  #DEFINE_ALIAS
 from .norm import GroupNorm  #DEFINE_ALIAS
 from .norm import LayerNorm  #DEFINE_ALIAS
 from .norm import SpectralNorm  #DEFINE_ALIAS

python/paddle/nn/layer/norm.py

@@ -20,7 +20,9 @@ from ...fluid.dygraph import BatchNorm  #DEFINE_ALIAS
 from ...fluid.dygraph import GroupNorm  #DEFINE_ALIAS
 from ...fluid.dygraph import LayerNorm  #DEFINE_ALIAS
 from ...fluid.dygraph import SpectralNorm  #DEFINE_ALIAS
+from ...fluid.dygraph import SyncBatchNorm  #DEFINE_ALIAS
 
 __all__ = [
-    'BatchNorm', 'GroupNorm', 'LayerNorm', 'SpectralNorm', 'InstanceNorm'
+    'BatchNorm', 'GroupNorm', 'LayerNorm', 'SpectralNorm', 'InstanceNorm',
+    'SyncBatchNorm'
 ]