Add DecoupledWeightDecay (#16427)

* Add DecoupledWeightDecay

Author: chengduo

paddle/fluid/API.spec

@@ -30,6 +30,8 @@
 from .slim import *
 from . import utils
 from .utils import *
+from . import extend_optimizer
+from .extend_optimizer import *
 
 __all__ = []
 __all__ += decoder.__all__
@@ -40,3 +42,4 @@
 __all__ += reader.__all__
 __all__ += slim.__all__
 __all__ += utils.__all__
+__all__ += extend_optimizer.__all__

python/paddle/fluid/contrib/__init__.py

@@ -0,0 +1,20 @@
+#   Copyright (c) 2019 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.
+
+from __future__ import print_function
+from . import extend_optimizer_with_weight_decay
+from .extend_optimizer_with_weight_decay import *
+
+__all__ = []
+__all__ += extend_optimizer_with_weight_decay.__all__

python/paddle/fluid/contrib/extend_optimizer/__init__.py

@@ -0,0 +1,152 @@
+#   Copyright (c) 2019 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 paddle.fluid
+from paddle.fluid import framework as framework
+
+__all__ = ["extend_with_decoupled_weight_decay"]
+
+
+class DecoupledWeightDecay(object):
+    def __init__(self, coeff=0.0, apply_decay_param_fun=None, **kwargs):
+        if not isinstance(coeff, float) and \
+                not isinstance(coeff, framework.Variable):
+            raise TypeError("coeff should be float or Variable.")
+        self._params_name = set()
+        self._apply_decay_param_fun = apply_decay_param_fun
+        self._coeff = coeff
+        super(DecoupledWeightDecay, self).__init__(**kwargs)
+
+    def _scale_parameters(self, params_and_grads):
+        """
+        Adds weight decay ops.
+            scaled_parameter = parameter * coeff
+
+        Args:
+            params_and_grads: A list of (parameters, gradients) pairs,
+                the parameters need to decay.
+        Raises:
+            Exception: The type of coeff and parameter is not consistent.
+        """
+        if isinstance(self._coeff, float) and self._coeff == 0.0:
+            return
+
+        scaled_params = []
+        for param, grad in params_and_grads:
+            # If no gradient then we don't need to do anything
+            if grad is None:
+                continue
+            if self._apply_decay_param_fun is not None \
+                    and not self._apply_decay_param_fun(param.name):
+                continue
+
+            if isinstance(self._coeff, float):
+                assert param.dtype is not paddle.fluid.core.VarDesc.VarType.FP32, \
+                    "the type of coeff(float) and parameter(%s) is not consistent."%(self._coeff.dtype)
+            else:
+                assert self._coeff.dtype == param.dtype, \
+                    "the type of coeff(%s) and parameter(%s) is not consistent."%(self._coeff.dtype, param.dtype)
+
+            with param.block.program._optimized_guard(
+                [param, grad]), framework.name_scope('weight decay'):
+                assert param.name not in self._params_name
+                scaled_params.append((param, grad, param * self._coeff))
+                self._params_name.add(param.name)
+        return scaled_params
+
+    def backward(self, **kargs):
+        return super(DecoupledWeightDecay, self).backward(**kargs)
+
+    def apply_optimize(self, **kargs):
+        return super(DecoupledWeightDecay, self).apply_optimize(**kargs)
+
+    def minimize(self,
+                 loss,
+                 startup_program=None,
+                 parameter_list=None,
+                 no_grad_set=None):
+        params_grads = self.backward(
+            loss=loss,
+            startup_program=startup_program,
+            parameter_list=parameter_list,
+            no_grad_set=no_grad_set)
+        scaled_params = self._scale_parameters(params_grads)
+        for p_grad_sgrad in scaled_params:
+            param, grad, scaled_param = p_grad_sgrad
+            with param.block.program._optimized_guard(
+                [param, grad]), framework.name_scope('weight decay'):
+                updated_param = paddle.fluid.layers.elementwise_sub(
+                    x=param, y=scaled_param)
+                paddle.fluid.layers.assign(input=updated_param, output=param)
+
+        optimize_ops = self.apply_optimize(
+            loss=loss,
+            params_grads=params_grads,
+            startup_program=startup_program)
+        return optimize_ops, params_grads
+
+    def __str__(self):
+        return " ".join(["Weight Decay, params:", ",".join(self._params_name)])
+
+
+def extend_with_decoupled_weight_decay(base_optimizer):
+    """
+    extend_with_decoupled_weight_decay is a decorator function, it returns an
+    optimizer class with decoupled weight decay. The returned optimizer will
+    apply weight decay on the optimized parameters with the parameters before
+    optimization, i.e: new_parameter = optimized_parameter - parameter * coeff.
+    The details of decoupled weight decay yplease refer to this
+    `DECOUPLED WEIGHT DECAY REGULARIZATION <https://arxiv.org/pdf/1711.05101.pdf>`_.
+
+    Args:
+        base_optimizer (Optimizer): The base_optimizer should be a derived class of Optimizer.
+
+    Returns:
+        OptimizerWithDecoupledWeightDecay: the optimizer with decouple weight decay.
+
+    Examples:
+
+      .. code-block:: python
+
+        AdamW = fluid.contrib.extend_with_decoupled_weight_decay(
+            fluid.optimizer.Adam)
+        optimizer = AdamW(learning_rate=0.1,
+                          weight_decay=0.01)
+
+        optimizer.minimize(cost)
+    """
+    if not issubclass(base_optimizer, paddle.fluid.optimizer.Optimizer):
+        raise TypeError(
+            "The input(base_optimizer) should be a derived class of Optimizer.")
+
+    class OptimizerWithDecoupledWeightDecay(DecoupledWeightDecay,
+                                            base_optimizer):
+        """
+        OptimizerWithDecoupledWeightDecay is used to update the optimized parameters
+        with the parameters before optimization. For more information, please refer:
+        https://arxiv.org/pdf/1711.05101.pdf.
+
+        Args:
+            weight_decay (float|Variable): The weight decay coefficient, it can be
+                float or Variable.
+            apply_decay_param_fun (function|None): If it is not None,
+                only variables that makes apply_decay_param_fun(variable)==True
+                will be updated. It only works when we want to specify variables.
+                Default: None.
+        """
+
+        def __init__(self, weight_decay, apply_decay_param_fun=None, **kwargs):
+            super(OptimizerWithDecoupledWeightDecay, self).__init__(
+                weight_decay, apply_decay_param_fun, **kwargs)
+
+    return OptimizerWithDecoupledWeightDecay

python/paddle/fluid/contrib/extend_optimizer/extend_optimizer_with_weight_decay.py

@@ -0,0 +1,151 @@
+#   Copyright (c) 2019 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 functools import partial
+import numpy as np
+import paddle
+import paddle.fluid as fluid
+import contextlib
+
+
+def get_places():
+    places = [fluid.CPUPlace()]
+    if fluid.core.is_compiled_with_cuda():
+        places.append(fluid.CUDAPlace(0))
+    return places
+
+
+@contextlib.contextmanager
+def prog_scope_guard(main_prog, startup_prog):
+    scope = fluid.core.Scope()
+    with fluid.unique_name.guard():
+        with fluid.scope_guard(scope):
+            with fluid.program_guard(main_prog, startup_prog):
+                yield
+
+
+def bow_net(data,
+            label,
+            dict_dim,
+            is_sparse=False,
+            emb_dim=128,
+            hid_dim=128,
+            hid_dim2=96,
+            class_dim=2):
+    """
+    BOW net
+    This model is from https://github.com/PaddlePaddle/models:
+    fluid/PaddleNLP/text_classification/nets.py
+    """
+    emb = fluid.layers.embedding(
+        input=data, is_sparse=is_sparse, size=[dict_dim, emb_dim])
+    bow = fluid.layers.sequence_pool(input=emb, pool_type='sum')
+    bow_tanh = fluid.layers.tanh(bow)
+    fc_1 = fluid.layers.fc(input=bow_tanh, size=hid_dim, act="tanh")
+    fc_2 = fluid.layers.fc(input=fc_1, size=hid_dim2, act="tanh")
+    prediction = fluid.layers.fc(input=[fc_2], size=class_dim, act="softmax")
+    cost = fluid.layers.cross_entropy(input=prediction, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+
+    return avg_cost
+
+
+class TestWeightDecay(unittest.TestCase):
+    def setUp(self):
+        self.word_dict = paddle.dataset.imdb.word_dict()
+        reader = paddle.batch(
+            paddle.dataset.imdb.train(self.word_dict), batch_size=2)()
+        self.train_data = [next(reader) for _ in range(5)]
+        self.learning_rate = .5
+
+    def run_program(self, place, feed_list):
+        exe = fluid.Executor(place)
+        feeder = fluid.DataFeeder(feed_list=feed_list, place=place)
+        exe.run(fluid.default_startup_program())
+
+        main_prog = fluid.default_main_program()
+        param_list = [var.name for var in main_prog.block(0).all_parameters()]
+
+        param_sum = []
+        for data in self.train_data:
+            out = exe.run(main_prog,
+                          feed=feeder.feed(data),
+                          fetch_list=param_list)
+            p_sum = 0
+            for v in out:
+                p_sum += np.sum(np.abs(v))
+            param_sum.append(p_sum)
+        return param_sum
+
+    def check_weight_decay(self, place, model):
+        main_prog = fluid.framework.Program()
+        startup_prog = fluid.framework.Program()
+        startup_prog.random_seed = 1
+        with prog_scope_guard(main_prog=main_prog, startup_prog=startup_prog):
+            data = fluid.layers.data(
+                name="words", shape=[1], dtype="int64", lod_level=1)
+            label = fluid.layers.data(name="label", shape=[1], dtype="int64")
+            avg_cost = model(data, label, len(self.word_dict))
+            AdamW = fluid.contrib.extend_with_decoupled_weight_decay(
+                fluid.optimizer.Adam)
+
+            optimizer = AdamW(
+                learning_rate=self.learning_rate,
+                weight_decay=self.learning_rate)
+
+            optimizer.minimize(avg_cost)
+            param_sum = self.run_program(place, [data, label])
+
+        return param_sum
+
+    def check_weight_decay2(self, place, model):
+        main_prog = fluid.framework.Program()
+        startup_prog = fluid.framework.Program()
+        startup_prog.random_seed = 1
+        with prog_scope_guard(main_prog=main_prog, startup_prog=startup_prog):
+            data = fluid.layers.data(
+                name="words", shape=[1], dtype="int64", lod_level=1)
+            label = fluid.layers.data(name="label", shape=[1], dtype="int64")
+
+            avg_cost = model(data, label, len(self.word_dict))
+
+            param_list = [(var, var * self.learning_rate)
+                          for var in main_prog.block(0).all_parameters()]
+
+            optimizer = fluid.optimizer.Adam(learning_rate=self.learning_rate)
+
+            optimizer.minimize(avg_cost)
+            for params in param_list:
+                updated_p = fluid.layers.elementwise_sub(
+                    x=params[0], y=params[1])
+                fluid.layers.assign(input=updated_p, output=params[0])
+
+            param_sum = self.run_program(place, [data, label])
+        return param_sum
+
+    def test_weight_decay(self):
+        for place in get_places():
+            model = partial(bow_net, is_sparse=False)
+            param_sum1 = self.check_weight_decay(place, model)
+            param_sum2 = self.check_weight_decay2(place, model)
+
+            for i in range(len(param_sum1)):
+                assert np.isclose(a=param_sum1[i], b=param_sum2[i], rtol=5e-5)
+
+
+if __name__ == '__main__':
+    unittest.main()