Merge pull request #7928 from guoshengCS/add-weight-normalization

Add weight normalization wrapper.

paddle/operators/reduce_op.cc

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/reduce_op.h"
-#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace operators {
@@ -38,10 +37,14 @@ class ReduceOp : public framework::OperatorWithKernel {
         dim, x_rank,
         "The dim should be in the range [-rank(input), rank(input)).");
     bool reduce_all = ctx->Attrs().Get<bool>("reduce_all");
+    bool keep_dim = ctx->Attrs().Get<bool>("keep_dim");
     if (reduce_all) {
+      if (keep_dim)
+        ctx->SetOutputDim(
+            "Out", framework::make_ddim(std::vector<int64_t>(x_rank, 1)));
+      else
         ctx->SetOutputDim("Out", {1});
     } else {
-      bool keep_dim = ctx->Attrs().Get<bool>("keep_dim");
       auto dims_vector = vectorize(x_dims);
       if (keep_dim || x_rank == 1) {
         dims_vector[dim] = 1;

python/paddle/v2/fluid/layer_helper.py

@@ -18,7 +18,7 @@ import itertools
 from framework import Variable, Parameter, default_main_program, default_startup_program, \
     unique_name, dtype_is_floating
 from paddle.v2.fluid.initializer import Constant, Xavier
-from param_attr import ParamAttr
+from param_attr import ParamAttr, WeightNormParamAttr
 
 
 class LayerHelper(object):
@@ -104,6 +104,177 @@ class LayerHelper(object):
                                  (dtype, each.dtype))
         return dtype
 
+    def _create_weight_normalize(self, attr, shape, dtype):
+        from .layers import elementwise_mul, elementwise_div, reshape
+
+        # Remove these ops when LayerHelper and layers support indicating
+        # program and block.
+        def __norm_op(x,
+                      out=None,
+                      p=2,
+                      dim=None,
+                      keep_dim=False,
+                      block=self.startup_program.global_block()):
+            if out is None:
+                out = block.create_var(
+                    name=unique_name(".".join([self.name, 'weight_norm_norm'])),
+                    dtype=dtype,
+                    persistable=False)
+            abs_out = block.create_var(
+                name=unique_name(".".join([self.name, 'weight_norm_abs'])),
+                dtype=dtype,
+                persistable=False)
+            block.append_op(
+                type='abs', inputs={'X': x}, outputs={'Out': abs_out})
+            pow_out = block.create_var(
+                name=unique_name(".".join([self.name, 'weight_norm_pow'])),
+                dtype=dtype,
+                persistable=False)
+            block.append_op(
+                type='pow',
+                inputs={'X': abs_out},
+                outputs={'Out': pow_out},
+                attrs={'factor': float(p)})
+            sum_out = block.create_var(
+                name=unique_name(".".join([self.name, 'weight_norm_sum'])),
+                dtype=dtype,
+                persistable=False)
+            block.append_op(
+                type='reduce_sum',
+                inputs={'X': pow_out},
+                outputs={'Out': sum_out},
+                attrs={
+                    'dim': dim,
+                    'keep_dim': keep_dim,
+                    'reduce_all': True if dim is None else False
+                })
+            block.append_op(
+                type='pow',
+                inputs={'X': sum_out},
+                outputs={'Out': out},
+                attrs={'factor': 1. / p})
+            return out
+
+        def __reshape_op(x,
+                         shape,
+                         out=None,
+                         block=self.startup_program.global_block()):
+            if out is None:
+                out = block.create_var(
+                    name=unique_name(".".join(
+                        [self.name, 'weight_norm_reshape'])),
+                    dtype=dtype,
+                    persistable=False)
+            block.append_op(
+                type='reshape',
+                inputs={'X': x},
+                outputs={'Out': out},
+                attrs={'shape': shape})
+            return out
+
+        def __transpose_op(x,
+                           axis,
+                           out=None,
+                           block=self.startup_program.global_block()):
+            if out is None:
+                out = block.create_var(
+                    name=unique_name(".".join(
+                        [self.name, 'weight_norm_transpose'])),
+                    dtype=dtype,
+                    persistable=False)
+            block.append_op(
+                type='transpose',
+                inputs={'X': x},
+                outputs={'Out': out},
+                attrs={'axis': axis})
+            return out
+
+        def __norm_except_dim(x,
+                              out=None,
+                              dim=None,
+                              block=self.startup_program.global_block()):
+            """Computes the norm over all dimensions except dim"""
+            if out is None:
+                out = block.create_var(
+                    name=unique_name(".".join([self.name, 'weight_norm_norm'])),
+                    dtype=dtype,
+                    persistable=False)
+            if dim is None:
+                __norm_op(x, out, dim=dim, block=block)
+            elif dim == 0:
+                out_shape = [x.shape[0]] + [1] * (len(x.shape) - 1)
+                reshape = __reshape_op(x, shape=[x.shape[0], -1], block=block)
+                norm = __norm_op(reshape, dim=1, block=block)
+                __reshape_op(norm, out=out, shape=out_shape, block=block)
+            elif dim == len(x.shape) - 1:
+                out_shape = [1] * (len(x.shape) - 1) + [x.shape[-1]]
+                reshape = __reshape_op(x, shape=[-1, x.shape[-1]], block=block)
+                norm = __norm_op(reshape, dim=0, block=block)
+                __reshape_op(norm, out=out, shape=out_shape, block=block)
+            else:
+                perm = range(len(x.shape))
+                perm[0], perm[dim] = dim, 0
+                transpose = __transpose_op(x, perm, block=block)
+                norm = __norm_op(transpose, dim=0, block=block)
+                __transpose_op(norm, perm, out=out, block=block)
+            return out
+
+        def __weight_normalize(g, v, dim):
+            """Calculations for weight normalization"""
+            norm = __norm_except_dim(
+                v, dim=dim, block=self.main_program.current_block())
+            scale = elementwise_div(
+                x=g, y=norm)  # The shapes of g and norm are the same.
+            # Currently, elementwise_mul only support broadcast when the shape
+            # of y is a subset of the shape of x. Thus, we reshape y to squeeze
+            # to achive the subset.
+            w = elementwise_mul(
+                x=v,
+                y=scale if dim is None else reshape(
+                    x=scale, shape=[v.shape[dim]]),
+                axis=-1 if dim is None else dim)
+            # To serialize the original parameter for inference, maybe a
+            # parameter rather than a variable should be returned.
+            return w
+
+        g_param_attr = copy.deepcopy(attr)
+        g_param_attr.name = attr.name + '_g'
+        g_param_shape = [1] * len(shape)
+        if attr.dim is not None:
+            g_param_shape[attr.dim] = shape[attr.dim]
+        v_param_attr = copy.deepcopy(attr)
+        v_param_attr.name = attr.name + '_v'
+        v_param_shape = shape
+
+        # Add to startup_program to initialize g and v.
+        # Try to reconstruct the initializer of w by initializing g and v.
+        # Set the initializers of g and v as below, then the distribution
+        # of w is the same as initializing w with the given initializer.
+        # For Data-Dependent Initialization, please compute the init-values
+        # of g and v in external and then feed the values to g and v by
+        # executing an extra program.
+        g_param = self.startup_program.global_block().create_parameter(
+            dtype=dtype,
+            shape=g_param_shape,
+            **g_param_attr.to_kwargs(with_initializer=False))
+        v_param = self.startup_program.global_block().create_parameter(
+            dtype=dtype,
+            shape=v_param_shape,
+            **v_param_attr.to_kwargs(with_initializer=True))
+        __norm_except_dim(
+            x=v_param,
+            out=g_param,
+            dim=attr.dim,
+            block=self.startup_program.global_block())
+
+        # Add weight normalization to main_program
+        g_param = self.main_program.global_block().create_parameter(
+            dtype=dtype, shape=g_param_shape, **g_param_attr.to_kwargs())
+        v_param = self.main_program.global_block().create_parameter(
+            dtype=dtype, shape=v_param_shape, **v_param_attr.to_kwargs())
+        w_param = __weight_normalize(g_param, v_param, dim=attr.dim)
+        return w_param
+
     def create_parameter(self,
                          attr,
                          shape,
@@ -114,16 +285,23 @@ class LayerHelper(object):
         attr = copy.deepcopy(attr)
         assert isinstance(attr, ParamAttr)
         suffix = 'b' if is_bias else 'w'
+        if attr.name is None:
+            attr.name = unique_name(".".join([self.name, suffix]))
 
-        if default_initializer is None:
+        if default_initializer is None and attr.initializer is None:
             if is_bias:
                 attr.set_default_bias_initializer()
             else:
                 attr.set_default_param_initializer()
         else:
             attr.set_default_initializer(default_initializer)
-        if attr.name is None:
-            attr.name = unique_name(".".join([self.name, suffix]))
+
+        # If weight normalization is set, insert extra parameters and ops.
+        # Refer to https://arxiv.org/pdf/1602.07868.pdf
+        if isinstance(attr, WeightNormParamAttr):
+            param = self._create_weight_normalize(attr, shape, dtype)
+            WeightNormParamAttr.params_with_weight_norm.append(param)
+            return param
 
         self.startup_program.global_block().create_parameter(
             dtype=dtype, shape=shape, **attr.to_kwargs(with_initializer=True))

python/paddle/v2/fluid/param_attr.py

@@ -15,7 +15,10 @@
 from initializer import Initializer, Xavier, Constant
 from regularizer import WeightDecayRegularizer
 
-__all__ = ['ParamAttr']
+__all__ = [
+    'ParamAttr',
+    'WeightNormParamAttr',
+]
 
 
 class ParamAttr(object):
@@ -82,3 +85,20 @@ class ParamAttr(object):
         if with_initializer:
             kwargs['initializer'] = self.initializer
         return kwargs
+
+
+class WeightNormParamAttr(ParamAttr):
+    """
+    Used for weight normalization. Any field in ParamAttr can also be set here.
+    Besides, an extra field dim can be set to indicate the dimension except 
+    which to normalize.
+    """
+    # List to record the parameters reparameterized by weight normalization.
+    # If these parameters are treated as Variable rather than Parameter,
+    # it can be used to discriminate these parameters and help to serialize
+    # these paramters for inference.
+    params_with_weight_norm = []
+
+    def __init__(self, dim=None, **kwargs):
+        super(WeightNormParamAttr, self).__init__(**kwargs)
+        self.dim = dim

python/paddle/v2/fluid/tests/test_weight_normalization.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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 unittest
+import numpy
+import collections
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.core as core
+from paddle.v2.fluid.initializer import ConstantInitializer
+from paddle.v2.fluid.param_attr import WeightNormParamAttr
+
+
+class TestWeightNormalization(unittest.TestCase):
+    batch_size = 3
+    hidden_size = 5
+    data_desc = (['x', [10], 0], )
+
+    @classmethod
+    def setUpClass(cls):
+        cls.set_program()
+
+    @classmethod
+    def set_program(cls):
+        data = fluid.layers.data(
+            name=cls.data_desc[0][0], shape=cls.data_desc[0][1])
+        out = fluid.layers.fc(input=data,
+                              size=cls.hidden_size,
+                              param_attr=WeightNormParamAttr(
+                                  dim=None,
+                                  name='weight_norm_param',
+                                  initializer=ConstantInitializer(1.0)),
+                              bias_attr=False,
+                              act=None)
+        loss = fluid.layers.reduce_sum(out)
+        fluid.backward.append_backward(loss=loss)
+        cls.fetch_list = [
+            'weight_norm_param_g', 'weight_norm_param_v',
+            'weight_norm_param_g@GRAD'
+        ]
+
+    def run_program(self):
+        outputs = []
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+        for place in places:
+            self.set_inputs(place)
+            exe = fluid.Executor(place)
+            exe.run(fluid.default_startup_program())
+            output = exe.run(fluid.default_main_program(),
+                             feed=self.inputs,
+                             fetch_list=self.fetch_list,
+                             return_numpy=False)
+            outputs.append(output)
+        self.actual_outputs = outputs
+
+    def set_data(self):
+        self.data = collections.OrderedDict()
+        for desc in self.data_desc:
+            data_name = desc[0]
+            data_shape = desc[1]
+            data_lod_level = desc[2]
+            data_lod = []
+            for i in range(data_lod_level):
+                lod_level_i = numpy.random.randint(
+                    low=1,
+                    high=5,
+                    size=self.batch_size if i == 0 else lod_level_i[-1])
+                lod_level_i = [0] + numpy.cumsum(lod_level_i).tolist()
+                data_lod.append(lod_level_i)
+            data_value = numpy.random.random(
+                size=[data_lod[-1][-1] if data_lod else self.batch_size
+                      ] + data_shape).astype('float32')
+            self.data[data_name] = (data_value, data_lod)
+
+    def set_inputs(self, place):
+        self.inputs = {}
+        for desc in self.data_desc:
+            tensor = fluid.Tensor()
+            tensor.set(self.data[desc[0]][0], place)
+            if self.data[desc[0]][1]:
+                tensor.set_lod(self.data[desc[0]][1])
+            self.inputs[desc[0]] = tensor
+
+    def weight_normalize(self):
+        v = numpy.ones((self.data[self.data_desc[0][0]][0].shape[-1],
+                        self.hidden_size))
+        g = numpy.linalg.norm(v, axis=None, keepdims=True)
+        w = g * v / numpy.linalg.norm(v, axis=None, keepdims=True)
+        x = self.data[self.data_desc[0][0]][0]
+        out = numpy.dot(x, w)
+        g_grad = (numpy.dot(x.T, numpy.ones_like(out)) * (v / numpy.linalg.norm(
+            v, axis=None, keepdims=True))).sum(axis=None, keepdims=True)
+        return g, v, g_grad
+
+    def test_weight_normalization(self):
+        self.set_data()
+        self.run_program()
+        expect_output = self.weight_normalize()
+        for actual_output in self.actual_outputs:
+            [
+                self.assertTrue(
+                    numpy.allclose(
+                        numpy.array(actual), expect, atol=0.001))
+                for expect, actual in zip(expect_output, actual_output)
+            ]
+
+
+if __name__ == '__main__':
+    unittest.main()