Fix bugs in piecewise decay

test=develop

python/paddle/fluid/imperative/__init__.py

@@ -26,8 +26,12 @@ from .nn import *
 from . import tracer
 from .tracer import *
 
+from . import learning_rate_scheduler
+from .learning_rate_scheduler import *
+
 __all__ = []
 __all__ += layers.__all__
 __all__ += base.__all__
 __all__ += nn.__all__
 __all__ += tracer.__all__
+__all__ += learning_rate_scheduler.__all__

python/paddle/fluid/imperative/learning_rate_scheduler.py

@@ -14,13 +14,9 @@
 
 from __future__ import print_function
 
-from .. import layers
 from .. import unique_name
 
-__all__ = [
-    'ExponentialDecay', 'NaturalExpDecay', 'InverseTimeDecay',
-    'PolynomialDecay', 'PiecewiseDecay', 'NoamDecay'
-]
+__all__ = ['PiecewiseDecay']
 
 
 class LearningRateDecay(object):
@@ -28,32 +24,35 @@ class LearningRateDecay(object):
     Base class of learning rate decay
     """
 
-    def __init__(self, step, dtype='float32'):
-        self.step = step
+    def __init__(self, begin=0, step=1, dtype='float32'):
+        self.step_num = begin
+        self.step_size = step
         self.dtype = dtype
 
     def __call__(self):
         lr = self.step()
         if isinstance(lr, float):
             lr = self._create_lr_var(lr)
-        self.step += 1
+        self.step_num += self.step_size
         return lr
 
-    def create_lr_var(lr):
+    def create_lr_var(self, lr):
+        from .. import layers
         lr = layers.create_global_var(
             name=unique_name.generate("learning_rate"),
             shape=[1],
             value=float(lr),
             dtype=self.dtype,
             persistable=True)
+        return lr
 
     def step(self):
         raise NotImplementedError()
 
 
-class PiecewiseDecay(object):
-    def __init__(self, boundaries, values, step, dtype='float32'):
-        super(PiecewiseDecay, self).__init__(step, dtype)
+class PiecewiseDecay(LearningRateDecay):
+    def __init__(self, boundaries, values, begin, step=1, dtype='float32'):
+        super(PiecewiseDecay, self).__init__(begin, step, dtype)
         self.boundaries = boundaries
         self.values = values
 
@@ -62,7 +61,7 @@ class PiecewiseDecay(object):
             self.vars.append(self.create_lr_var(value))
 
     def step(self):
-        for i in range(len(boundaries)):
-            if self.step <= boundaries[i]:
+        for i in range(len(self.boundaries)):
+            if self.step_num < self.boundaries[i]:
                 return self.vars[i]
-        return self.vars[len(values) - 1]
+        return self.vars[len(self.values) - 1]

python/paddle/fluid/optimizer.py

@@ -31,6 +31,7 @@ from .layer_helper import LayerHelper
 from .layers import ops
 from .regularizer import append_regularization_ops
 from .imperative import base as imperative_base
+from .imperative.learning_rate_scheduler import LearningRateDecay
 
 __all__ = [
     'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl',
@@ -50,9 +51,19 @@ class Optimizer(object):
     """
 
     def __init__(self, learning_rate, regularization=None, name=None):
+        if framework._in_imperative_mode():
+            if not isinstance(learning_rate, float) and \
+                    not isinstance(learning_rate, LearningRateDecay):
+                raise TypeError(
+                    "learning rate should be float or LearningRateDecay, got %s here"
+                    % type(learning_rate))
+        else:
             if not isinstance(learning_rate, float) and \
                     not isinstance(learning_rate, framework.Variable):
-            raise TypeError("learning rate should be float or Variable")
+                raise TypeError(
+                    "learning rate should be float or Variable, got %s here" %
+                    type(learning_rate))
+
         self._name = name
         self.regularization = regularization
         self._learning_rate = learning_rate
@@ -83,7 +94,7 @@ class Optimizer(object):
                     dtype='float32' if self._dtype is None else self._dtype,
                     persistable=True)
             # get learning rate Variable from LearningRateDecay
-            elif isinstance(self._learning_rate, imperative.LearningRateDecay):
+            elif isinstance(self._learning_rate, LearningRateDecay):
                 self._learning_rate_map[framework.default_main_program(
                 )] = self._learning_rate()
             else:

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

@@ -23,69 +23,129 @@ import paddle
 import paddle.fluid as fluid
 from paddle.fluid import core
 from paddle.fluid.optimizer import SGDOptimizer
-from paddle.fluid.imperative.nn import FC
+from paddle.fluid.imperative.nn import Conv2D, Pool2D, FC
 from paddle.fluid.imperative.base import to_variable
 from test_imperative_base import new_program_scope
 
 
-class MLP(fluid.imperative.Layer):
-    def __init__(self, param_attr=None, bias_attr=None):
-        self._fc1 = FC(10)
-        self._fc2 = FC(10)
+class SimpleImgConvPool(fluid.imperative.Layer):
+    def __init__(self,
+                 name_scope,
+                 num_channels,
+                 num_filters,
+                 filter_size,
+                 pool_size,
+                 pool_stride,
+                 pool_padding=0,
+                 pool_type='max',
+                 global_pooling=False,
+                 conv_stride=1,
+                 conv_padding=0,
+                 conv_dilation=1,
+                 conv_groups=1,
+                 act=None,
+                 use_cudnn=False,
+                 param_attr=None,
+                 bias_attr=None):
+        super(SimpleImgConvPool, self).__init__(name_scope)
+
+        self._conv2d = Conv2D(
+            self.full_name(),
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=conv_stride,
+            padding=conv_padding,
+            dilation=conv_dilation,
+            groups=conv_groups,
+            param_attr=None,
+            bias_attr=None,
+            use_cudnn=use_cudnn)
+
+        self._pool2d = Pool2D(
+            self.full_name(),
+            pool_size=pool_size,
+            pool_type=pool_type,
+            pool_stride=pool_stride,
+            pool_padding=pool_padding,
+            global_pooling=global_pooling,
+            use_cudnn=use_cudnn)
 
     def forward(self, inputs):
-        y = self._fc1(inputs)
-        y = self._fc2(y)
-        return y
+        x = self._conv2d(inputs)
+        x = self._pool2d(x)
+        return x
 
 
-class TestImperativeOptimizerBase(unittest.TestCase):
-    def setUp(self):
-        self.batch_num = 2
+class MNIST(fluid.imperative.Layer):
+    def __init__(self, name_scope):
+        super(MNIST, self).__init__(name_scope)
 
-    def get_optimizer(self):
-        self.optimizer = SGDOptimizer(learning_rate=1e-3)
+        self._simple_img_conv_pool_1 = SimpleImgConvPool(
+            self.full_name(), 1, 20, 5, 2, 2, act="relu")
 
-    def test_optimizer_float32(self):
+        self._simple_img_conv_pool_2 = SimpleImgConvPool(
+            self.full_name(), 20, 50, 5, 2, 2, act="relu")
+
+        pool_2_shape = 50 * 4 * 4
+        SIZE = 10
+        scale = (2.0 / (pool_2_shape**2 * SIZE))**0.5
+        self._fc = FC(self.full_name(),
+                      10,
+                      param_attr=fluid.param_attr.ParamAttr(
+                          initializer=fluid.initializer.NormalInitializer(
+                              loc=0.0, scale=scale)),
+                      act="softmax")
+
+    def forward(self, inputs):
+        x = self._simple_img_conv_pool_1(inputs)
+        x = self._simple_img_conv_pool_2(x)
+        x = self._fc(x)
+        return x
+
+
+class TestImperativeMnist(unittest.TestCase):
+    def test_mnist_float32(self):
         seed = 90
+        epoch_num = 1
         with fluid.imperative.guard():
             fluid.default_startup_program().random_seed = seed
             fluid.default_main_program().random_seed = seed
 
-            mlp = MLP()
-            self.get_optimizer()
+            mnist = MNIST("mnist")
+            sgd = SGDOptimizer(learning_rate=1e-3)
             train_reader = paddle.batch(
                 paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
 
             dy_param_init_value = {}
+            for epoch in range(epoch_num):
                 for batch_id, data in enumerate(train_reader()):
-                if batch_id >= self.batch_num:
-                    break
-
                     dy_x_data = np.array(
-                    [x[0].reshape(1, 28, 28) for x in data]).astype('float32')
-                y_data = np.array([x[1] for x in data]).astype('int64').reshape(
-                    128, 1)
+                        [x[0].reshape(1, 28, 28)
+                         for x in data]).astype('float32')
+                    y_data = np.array(
+                        [x[1] for x in data]).astype('int64').reshape(128, 1)
 
                     img = to_variable(dy_x_data)
                     label = to_variable(y_data)
                     label._stop_gradient = True
 
-                cost = mlp(img)
-                avg_loss = fluid.layers.reduce_mean(cost)
+                    cost = mnist(img)
+                    loss = fluid.layers.cross_entropy(cost, label)
+                    avg_loss = fluid.layers.mean(loss)
+
                     dy_out = avg_loss._numpy()
 
-                if batch_id == 0:
-                    for param in fluid.default_main_program().global_block(
-                    ).all_parameters():
+                    if epoch == 0 and batch_id == 0:
+                        for param in mnist.parameters():
                             dy_param_init_value[param.name] = param._numpy()
 
                     avg_loss._backward()
-                self.optimizer.minimize(avg_loss)
-                mlp.clear_gradients()
+                    sgd.minimize(avg_loss)
+                    mnist.clear_gradients()
+
                     dy_param_value = {}
-                for param in fluid.default_main_program().global_block(
-                ).all_parameters():
+                    for param in mnist.parameters():
                         dy_param_value[param.name] = param._numpy()
 
         with new_program_scope():
@@ -95,8 +155,8 @@ class TestImperativeOptimizerBase(unittest.TestCase):
             exe = fluid.Executor(fluid.CPUPlace(
             ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
 
-            mnist = MNIST()
-            self.get_optimizer()
+            mnist = MNIST("mnist")
+            sgd = SGDOptimizer(learning_rate=1e-3)
             train_reader = paddle.batch(
                 paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
 
@@ -104,8 +164,9 @@ class TestImperativeOptimizerBase(unittest.TestCase):
                 name='pixel', shape=[1, 28, 28], dtype='float32')
             label = fluid.layers.data(name='label', shape=[1], dtype='int64')
             cost = mnist(img)
-            avg_loss = fluid.layers.reduce_mean(cost)
-            self.optimizer.minimize(avg_loss)
+            loss = fluid.layers.cross_entropy(cost, label)
+            avg_loss = fluid.layers.mean(loss)
+            sgd.minimize(avg_loss)
 
             # initialize params and fetch them
             static_param_init_value = {}
@@ -119,18 +180,18 @@ class TestImperativeOptimizerBase(unittest.TestCase):
             for i in range(len(static_param_name_list)):
                 static_param_init_value[static_param_name_list[i]] = out[i]
 
+            for epoch in range(epoch_num):
                 for batch_id, data in enumerate(train_reader()):
-                if batch_id >= self.batch_num:
-                    break
-
                     static_x_data = np.array(
-                    [x[0].reshape(1, 28, 28) for x in data]).astype('float32')
-                y_data = np.array([x[1] for x in data]).astype('int64').reshape(
-                    [128, 1])
+                        [x[0].reshape(1, 28, 28)
+                         for x in data]).astype('float32')
+                    y_data = np.array(
+                        [x[1] for x in data]).astype('int64').reshape([128, 1])
 
                     fetch_list = [avg_loss.name]
                     fetch_list.extend(static_param_name_list)
-                out = exe.run(fluid.default_main_program(),
+                    out = exe.run(
+                        fluid.default_main_program(),
                         feed={"pixel": static_x_data,
                               "label": y_data},
                         fetch_list=fetch_list)
@@ -138,7 +199,10 @@ class TestImperativeOptimizerBase(unittest.TestCase):
                     static_param_value = {}
                     static_out = out[0]
                     for i in range(1, len(out)):
-                    static_param_value[static_param_name_list[i - 1]] = out[i]
+                        static_param_value[static_param_name_list[i - 1]] = out[
+                            i]
+
+        self.assertTrue(np.allclose(dy_x_data.all(), static_x_data.all()))
 
         for key, value in six.iteritems(static_param_init_value):
             self.assertTrue(np.allclose(value, dy_param_init_value[key]))

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

@@ -29,9 +29,11 @@ from test_imperative_base import new_program_scope
 
 
 class MLP(fluid.imperative.Layer):
-    def __init__(self, param_attr=None, bias_attr=None):
-        self._fc1 = FC(10)
-        self._fc2 = FC(10)
+    def __init__(self, name_scope, param_attr=None, bias_attr=None):
+        super(MLP, self).__init__(name_scope)
+
+        self._fc1 = FC(self.full_name(), 10)
+        self._fc2 = FC(self.full_name(), 10)
 
     def forward(self, inputs):
         y = self._fc1(inputs)
@@ -41,10 +43,15 @@ class MLP(fluid.imperative.Layer):
 
 class TestImperativeOptimizerBase(unittest.TestCase):
     def setUp(self):
-        self.batch_num = 2
+        self.batch_num = 10
 
     def get_optimizer(self):
-        self.optimizer = SGDOptimizer(learning_rate=1e-3)
+        bd = [3, 6, 9]
+        self.optimizer = SGDOptimizer(
+            learning_rate=fluid.layers.piecewise_decay(
+                boundaries=bd,
+                values=[0.1 * (0.1**i) for i in range(len(bd) + 1)]))
+        return self.optimizer
 
     def test_optimizer_float32(self):
         seed = 90
@@ -52,8 +59,8 @@ class TestImperativeOptimizerBase(unittest.TestCase):
             fluid.default_startup_program().random_seed = seed
             fluid.default_main_program().random_seed = seed
 
-            mlp = MLP()
-            self.get_optimizer()
+            mlp = MLP('mlp')
+            optimizer = self.get_optimizer()
             train_reader = paddle.batch(
                 paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
 
@@ -81,7 +88,7 @@ class TestImperativeOptimizerBase(unittest.TestCase):
                         dy_param_init_value[param.name] = param._numpy()
 
                 avg_loss._backward()
-                self.optimizer.minimize(avg_loss)
+                optimizer.minimize(avg_loss)
                 mlp.clear_gradients()
                 dy_param_value = {}
                 for param in fluid.default_main_program().global_block(
@@ -95,8 +102,8 @@ class TestImperativeOptimizerBase(unittest.TestCase):
             exe = fluid.Executor(fluid.CPUPlace(
             ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
 
-            mnist = MNIST()
-            self.get_optimizer()
+            mnist = MLP('mlp')
+            optimizer = self.get_optimizer()
             train_reader = paddle.batch(
                 paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
 
@@ -105,7 +112,7 @@ class TestImperativeOptimizerBase(unittest.TestCase):
             label = fluid.layers.data(name='label', shape=[1], dtype='int64')
             cost = mnist(img)
             avg_loss = fluid.layers.reduce_mean(cost)
-            self.optimizer.minimize(avg_loss)
+            optimizer.minimize(avg_loss)
 
             # initialize params and fetch them
             static_param_init_value = {}