add new learing rate strategy to reduce lr when loss reach on plateau (PaddlePaddle#24322) (PaddlePaddle#24979)

zhwesky2010 · web-flow · commit 0b3f6265819d · 2020-06-10T11:27:55.000+08:00
添加loss自适应的学习率衰减策略。
diff --git a/python/paddle/fluid/dygraph/learning_rate_scheduler.py b/python/paddle/fluid/dygraph/learning_rate_scheduler.py
@@ -17,10 +17,13 @@
 import math
 
 from .. import unique_name
+from ..framework import Variable
+from ..data_feeder import check_type
 
 __all__ = [
     'NoamDecay', 'PiecewiseDecay', 'NaturalExpDecay', 'ExponentialDecay',
-    'InverseTimeDecay', 'PolynomialDecay', 'CosineDecay'
+    'InverseTimeDecay', 'PolynomialDecay', 'CosineDecay', 'LinearLrWarmup',
+    'ReduceLROnPlateau'
 ]
 
 
@@ -619,7 +622,7 @@ class LinearLrWarmup(LearningRateDecay):
     
         learning_rate = 0.1 
         warmup_steps = 50
-        start_lr = 1. / 3.
+        start_lr = 0
         end_lr = 0.1
 
         with fluid.dygraph.guard(): 
@@ -660,3 +663,193 @@ def step(self):
             return self.lr_ratio_before_warmup * self.step_num
         else:
             return base_lr
+
+
+class ReduceLROnPlateau(LearningRateDecay):
+    """
+    Reduce learning rate when ``loss`` has stopped descending. Models often benefit from reducing the learning rate 
+    by 2 to 10 times once model performance has no longer improvement.
+
+    The ``loss`` is the one which has been pass into ``step`` , it must be 1-D Tensor with shape [1]. When ``loss`` 
+    stop descending for a ``patience`` number of epochs, the learning rate will be reduced to ``learning_rate * decay_rate`` . 
+    (Specially, ``mode`` can also be set to ``'max`` , in this case, when ``loss`` stop ascending for a ``patience`` number 
+    of epochs, the learning rate will be reduced.)
+
+    In addition, After each reduction, it will wait a ``cooldown`` number of epochs before resuming normal operation.
+
+    Args:
+        learning_rate (Variable|float|int): The initial learning rate. It can be set to python float or int number.
+            If the type is Variable, it should be 1-D Tensor with shape [1], the data type can be 'float32' or 'float64'.
+        mode (str, optional): ``'min'`` or ``'max'`` can be selected. Normally, it is ``'min'`` , which means that the 
+            learning rate will reduce when ``loss`` stops descending. Specially, if it's set to ``'max'`` ,  the learning 
+            rate will reduce when ``loss`` stops ascending. Default: ``'min'`` .
+        decay_rate (float, optional): The Ratio that the learning rate will be reduced. ``new_lr = origin_lr * decay_rate`` . 
+            It should be less than 1.0. Default: 0.1.
+        patience (int, optional): When ``loss`` doesn't improve for this number of epochs, learing rate will be reduced. 
+            Default: 10.
+        verbose (bool, optional): If ``True``, prints a message to stdout for each update. Default: ``False``.
+        threshold (float, optional): ``threshold`` and ``threshold_mode`` will determine the minimum change of ``loss`` . 
+            This make tiny changes of ``loss`` will be ignored. Default: 1e-4.
+        threshold_mode (str, optional): ``'rel'`` or ``'abs'`` can be selected. In ``'rel'`` mode, the minimum change of ``loss``
+            is ``last_loss * threshold`` , where ``last_loss`` is ``loss`` in last epoch. In ``'abs'`` mode, the minimum 
+            change of ``loss`` is ``threshold`` . Default: ``'rel'`` .
+        cooldown (int, optional): The number of epochs to wait before resuming normal operation. Default: 0.
+        min_lr (float, optional): The lower bound of the learning rate after reduction. Default: 0.
+        eps (float, optional): Minimal decay applied to lr. If the difference between new and old lr is smaller than eps, the update is
+            ignored. Default: 1e-8.
+        dtype (str, optional): The data type used to create the learning rate variable. The data type can be set as
+            'float32', 'float64'. Default: 'float32'. 
+    
+    Returns:
+        Reduced learning rate.
+
+    Examples:
+    
+    .. code-block:: python
+
+        import paddle.fluid as fluid
+        import numpy as np
+
+        with fluid.dygraph.guard():
+            x = np.random.uniform(-1, 1, [10, 10]).astype("float32")
+            linear = fluid.dygraph.Linear(10, 10)
+            input = fluid.dygraph.to_variable(x)
+
+            reduce_lr = fluid.dygraph.ReduceLROnPlateau(
+                                    learning_rate = 1.0,
+                                    decay_rate = 0.5,
+                                    patience = 5,
+                                    verbose = True, 
+                                    cooldown = 3)
+            adam = fluid.optimizer.Adam(
+                learning_rate = reduce_lr,
+                parameter_list = linear.parameters())
+
+            for epoch in range(10):
+                total_loss = 0
+                for bath_id in range(5):
+                    out = linear(input)
+                    loss = fluid.layers.reduce_mean(out)
+                    total_loss += loss
+                    adam.minimize(loss)
+                
+                avg_loss = total_loss/5
+
+                # adjust learning rate according to avg_loss
+                reduce_lr.step(avg_loss)
+                lr = adam.current_step_lr()
+                print("current avg_loss is %s, current lr is %s" % (avg_loss.numpy()[0], lr))
+
+    """
+
+    def __init__(self,
+                 learning_rate,
+                 mode='min',
+                 decay_rate=0.1,
+                 patience=10,
+                 verbose=False,
+                 threshold=1e-4,
+                 threshold_mode='rel',
+                 cooldown=0,
+                 min_lr=0,
+                 eps=1e-8,
+                 dtype='float32'):
+        super(ReduceLROnPlateau, self).__init__(dtype=dtype)
+        mode = mode.lower()
+        if mode not in ['min', 'max']:
+            raise ValueError('mode ' + mode + ' is unknown!')
+        self.mode = mode
+
+        if decay_rate >= 1.0:
+            raise ValueError(
+                'new_lr = origin_lr * decay_rate and decay_rate should be < 1.0.'
+            )
+        self.decay_rate = decay_rate
+
+        threshold_mode = threshold_mode.lower()
+        if threshold_mode not in ['rel', 'abs']:
+            raise ValueError('threshold mode ' + threshold_mode +
+                             ' is unknown!')
+        self.threshold_mode = threshold_mode
+
+        check_type(learning_rate, 'learning_rate', (float, int, Variable),
+                   'ReduceLROnPlateau')
+        if isinstance(learning_rate, (float, int)):
+            learning_rate = self.create_lr_var(learning_rate)
+
+        self.learning_rate = learning_rate
+        self.verbose = verbose
+        self.patience = patience
+        self.threshold = threshold
+        self.threshold_mode = threshold_mode
+        self.cooldown = cooldown
+        self.min_lr = self.create_lr_var(min_lr)
+        self.eps = eps
+
+        self.cooldown_counter = 0
+        self.best_loss = None
+        self.num_bad_epochs = 0
+        self.epoch = 0
+
+    def __call__(self):
+        return self.learning_rate
+
+    def step(self, loss):
+        """
+        It should be invoked on each epoch. Update the learning rate in optimizer according to ``loss`` .  
+        The new learning rate will take effect on next call to ``optimizer.minimize`` .
+
+        Args:
+            loss (Variable): A ``Variable`` that will be monitored to determine whether the learning rate will reduce. 
+                If it stop descending for a ``patience`` number of epochs, the learning rate will reduce. It should 
+                be 1-D Tensor with shape [1]. 
+                Specially, if ``mode`` has been set to ``'max'`` ,  the learning rate will reduce when it stops ascending.
+        Returns:
+            None
+        
+        Examples:
+            Please refer to the example of current LearningRateDecay.
+        """
+
+        # loss must be 1-D Tensor with shape [1]
+        check_type(loss, 'loss', Variable, 'ReduceLROnPlateau.step')
+        assert len(loss.shape) == 1 and loss.shape[0] == 1, "the loss.shape " \
+            "should be (1L,), but the current loss.shape is {}. Maybe that "  \
+            "you should call fluid.layers.mean to process it first.".format(loss.shape)
+
+        self.epoch += 1
+        if self.cooldown_counter > 0:
+            self.cooldown_counter -= 1
+        else:
+            if self.best_loss is None or self._is_better(loss, self.best_loss):
+                self.best_loss = loss
+                self.num_bad_epochs = 0
+            else:
+                self.num_bad_epochs += 1
+
+            if self.num_bad_epochs > self.patience:
+                from .. import layers
+                self.cooldown_counter = self.cooldown
+                self.num_bad_epochs = 0
+                new_lr = layers.elementwise_max(self.learning_rate *
+                                                self.decay_rate, self.min_lr)
+                if self.learning_rate - new_lr > self.eps:
+                    if self.verbose:
+                        print('Epoch {}: reducing learning rate from {} to {}.'.
+                              format(self.epoch,
+                                     self.learning_rate.numpy()[0],
+                                     new_lr.numpy()[0]))
+                    self.learning_rate = new_lr
+
+    def _is_better(self, current, best):
+        if self.mode == 'min' and self.threshold_mode == 'rel':
+            return current < best - best * self.threshold
+
+        elif self.mode == 'min' and self.threshold_mode == 'abs':
+            return current < best - self.threshold
+
+        elif self.mode == 'max' and self.threshold_mode == 'rel':
+            return current > best + best * self.threshold
+
+        else:
+            return current > best + self.threshold
diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py
@@ -708,7 +708,7 @@ def apply_gradients(self, params_grads):
         params_grads, table_param_and_grad, table_optimize_op = \
             self._process_distribute_lookuptable(params_grads)
 
-        # 'minimize(grad_clip)' or 'set_gradient_clip'
+        # 'optimizer(grad_clip)' or 'set_gradient_clip'
         if self._grad_clip is not None:
             params_grads = self._grad_clip(params_grads)
         else:
@@ -1460,7 +1460,7 @@ def apply_gradients(self, params_grads):
             else:
                 dgc_params_grads.append((param, grad))
 
-        # 'minimize(grad_clip)' or 'set_gradient_clip'
+        # 'optimizer(grad_clip)' or 'set_gradient_clip'
         if self._grad_clip is not None:
             not_dgc_params_grads = self._grad_clip(not_dgc_params_grads)
         else:
diff --git a/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py b/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py
@@ -199,7 +199,7 @@ def test_decay(self):
         ]
 
         for py_decay_fn, fluid_decay_fn, kwargs in decay_fns:
-            print("class=" + self.__class__.__name__ + "decay_fn=" +
+            print("class=" + self.__class__.__name__ + " decay_fn=" +
                   py_decay_fn.__name__ + " kwargs=" + str(kwargs))
             main_program = framework.Program()
             startup_program = framework.Program()
@@ -335,5 +335,111 @@ def test_dygraph_mode(self):
                     end_lr=1.0)
 
 
+def reduce_lr_on_plateau(decay_rate, threshold, cooldown, patience, m, n, loss,
+                         var_list):
+    def is_better(current, best, m, n):
+        if m == 'min' and n == 'rel':
+            return current < best - best * threshold
+        elif m == 'min' and n == 'abs':
+            return current < best - threshold
+        elif m == 'max' and n == 'rel':
+            return current > best + best * threshold
+        else:  # mode == 'max' and epsilon_mode == 'abs':
+            return current > best + threshold
+
+    if var_list[2] > 0:
+        var_list[2] -= 1
+        return var_list[1]
+
+    if is_better(loss, var_list[0], m, n):
+        var_list[0] = loss
+        var_list[3] = 0
+    else:
+        var_list[3] += 1
+        if var_list[3] > patience:
+            var_list[2] = cooldown
+            var_list[3] = 0
+            new_lr = var_list[1] * decay_rate
+            var_list[1] = new_lr if var_list[1] - new_lr > 1e-8 else var_list[1]
+
+    return var_list[1]
+
+
+class TestReduceLROnPlateauDecay(unittest.TestCase):
+    def test_dygraph_mode(self):
+        with fluid.dygraph.guard():
+            # the decay rate must be less than 1.0
+            with self.assertRaises(ValueError):
+                fluid.dygraph.ReduceLROnPlateau(
+                    learning_rate=1.0, decay_rate=2.0)
+            # the mode must be "min" or "max"
+            with self.assertRaises(ValueError):
+                fluid.dygraph.ReduceLROnPlateau(learning_rate=1.0, mode="test")
+            # the threshold_mode must be "rel" or "abs"
+            with self.assertRaises(ValueError):
+                fluid.dygraph.ReduceLROnPlateau(
+                    learning_rate=1.0, threshold_mode="test")
+
+            base_lr = 1.0
+            patience = 3
+            cooldown = 1
+            decay_rate = 0.5
+            threshold = 1e-4
+            linear = fluid.dygraph.Linear(10, 10)
+
+            for m, n in zip(['min', 'max', 'min', 'max'],
+                            ['rel', 'rel', 'abs', 'abs']):
+                kwargs = {
+                    'learning_rate': base_lr,
+                    'decay_rate': decay_rate,
+                    'threshold': threshold,
+                    'verbose': True,
+                    'patience': patience,
+                    'cooldown': cooldown,
+                    'mode': m,
+                    'threshold_mode': n,
+                    'eps': 1e-6
+                }
+                print("class=" + fluid.dygraph.ReduceLROnPlateau.__name__ +
+                      " kwargs=" + str(kwargs))
+                lr = fluid.dygraph.ReduceLROnPlateau(**kwargs)
+                sgd = fluid.optimizer.SGD(learning_rate=lr,
+                                          parameter_list=linear.parameters())
+
+                best = float("-10000") if m == "max" else float("10000")
+                expected_lr = 1.0
+                cooldown_counter = 0
+                num_bad_epochs = 0
+                var_list = [best, expected_lr, cooldown_counter, num_bad_epochs]
+                step_num = 0
+                epoch_num = 0
+                for epoch in range(30):
+                    total_loss = 0
+
+                    for batch_id in range(2):
+                        step_num += 1
+                        x = fluid.dygraph.to_variable(
+                            np.array([step_num]).astype('float32'))
+                        loss = layers.sin(x)
+                        sgd.minimize(loss)
+                        total_loss += loss
+
+                    epoch_num += 1
+                    # get expected lr from fluid
+                    avg_loss = total_loss / 1
+                    lr.step(avg_loss)
+                    actual_lr = lr().numpy()[0]
+
+                    # get expected lr form python
+                    expected_lr = reduce_lr_on_plateau(decay_rate, threshold,
+                                                       cooldown, patience, m, n,
+                                                       avg_loss, var_list)
+                    self.assertEqual(
+                        expected_lr,
+                        actual_lr,
+                        msg='Failed reduce lr scheduler in epoch {0}, Python result is {1}, Fluid result is {2}'.
+                        format(epoch_num, expected_lr, actual_lr))
+
+
 if __name__ == '__main__':
     unittest.main()
