# Copyright (c) 2016 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 copy import math import numpy as np import unittest import paddle import paddle.fluid as fluid import paddle.fluid.layers as layers import paddle.fluid.framework as framework import paddle.fluid.core as core 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 TestReduceOnPlateauDecay(object): def test_ReduceLR(self): # the decay rate must be less than 1.0 with self.assertRaises(ValueError): paddle.optimizer.lr.ReduceOnPlateau(learning_rate=1.0, factor=2.0) # the mode must be "min" or "max" with self.assertRaises(ValueError): paddle.optimizer.lr.ReduceOnPlateau(learning_rate=1.0, mode="test") # the threshold_mode must be "rel" or "abs" with self.assertRaises(ValueError): paddle.optimizer.lr.ReduceOnPlateau(learning_rate=1.0, threshold_mode="test") with self.assertRaises(TypeError): paddle.optimizer.lr.ReduceOnPlateau(learning_rate="test") with self.assertRaises(TypeError): paddle.optimizer.lr.ReduceOnPlateau(learning_rate=0.5).step("test") places = [paddle.CPUPlace()] if core.is_compiled_with_cuda(): places.append(paddle.CUDAPlace(0)) for place in places: for m, n in zip(['min', 'max', 'min', 'max'], ['rel', 'rel', 'abs', 'abs']): kwargs = { 'learning_rate': 1.0, 'mode': m, 'factor': 0.5, 'patience': 3, 'threshold': 1e-4, 'threshold_mode': n, 'cooldown': 1, 'min_lr': 0, 'epsilon': 1e-8, 'verbose': False, } paddle.enable_static() self._test_static(place, kwargs) paddle.disable_static(place) self._test_dygraph(place, kwargs) paddle.enable_static() def _test_static(self, place, kwargs): paddle.enable_static() best = float("-10000") if kwargs['mode'] == "max" else float("10000") current_lr = 1.0 cooldown_counter = 0 num_bad_epochs = 0 var_list = [best, current_lr, cooldown_counter, num_bad_epochs] main_prog = paddle.static.Program() start_prog = paddle.static.Program() with paddle.static.program_guard(main_prog, start_prog): x = fluid.layers.create_global_var([1], 1, 'float32', persistable=True) paddle.increment(x) loss = paddle.sin(x) scheduler = paddle.optimizer.lr.ReduceOnPlateau(**kwargs) adam = paddle.optimizer.Adam(learning_rate=scheduler) adam.minimize(loss) lr_var = adam._global_learning_rate() test_prog = main_prog.clone() exe = paddle.static.Executor(place) exe.run(start_prog) for epoch in range(20): for batch_id in range(1): out, actual_lr = exe.run(main_prog, fetch_list=[loss.name, lr_var.name]) expected_lr = reduce_lr_on_plateau( kwargs['factor'], kwargs['threshold'], kwargs['cooldown'], kwargs['patience'], kwargs['mode'], kwargs['threshold_mode'], out[0], var_list) scheduler.step(out[0]) actual_lr = scheduler() self.assertEqual(actual_lr, np.array(expected_lr)) for epoch in range(10): for batch_id in range(1): out, actual_lr = exe.run(test_prog, fetch_list=[loss.name, lr_var.name]) expected_lr = reduce_lr_on_plateau( kwargs['factor'], kwargs['threshold'], kwargs['cooldown'], kwargs['patience'], kwargs['mode'], kwargs['threshold_mode'], out[0], var_list) scheduler.step(out[0]) actual_lr = scheduler() self.assertEqual(actual_lr, np.array(expected_lr)) def _test_dygraph(self, place, kwargs): paddle.disable_static(place) best = float("-10000") if kwargs['mode'] == "max" else float("10000") current_lr = 1.0 cooldown_counter = 0 num_bad_epochs = 0 var_list = [best, current_lr, cooldown_counter, num_bad_epochs] linear = paddle.nn.Linear(10, 10) scheduler = paddle.optimizer.lr.ReduceOnPlateau(**kwargs) adam = paddle.optimizer.Adam(learning_rate=scheduler, parameters=linear.parameters()) for epoch in range(20): for batch_id in range(1): x = paddle.to_tensor(epoch).astype('float32') loss = paddle.sin(x) loss.backward() adam.step() adam.clear_grad() scheduler.step(loss) # get lr from paddle current_lr = adam.get_lr() # get lr form python expected_lr = reduce_lr_on_plateau( kwargs['factor'], kwargs['threshold'], kwargs['cooldown'], kwargs['patience'], kwargs['mode'], kwargs['threshold_mode'], loss, var_list) self.assertEqual(current_lr, expected_lr) state_dict = adam.state_dict() scheduler1 = paddle.optimizer.lr.ReduceOnPlateau(**kwargs) adam1 = paddle.optimizer.Adam(learning_rate=scheduler1, parameters=linear.parameters()) adam1.set_state_dict(state_dict) self.assertEqual(scheduler.cooldown_counter, scheduler1.cooldown_counter) self.assertEqual(scheduler.best.numpy()[0], scheduler1.best) self.assertEqual(scheduler.num_bad_epochs, scheduler1.num_bad_epochs) self.assertEqual(scheduler.last_epoch, scheduler1.last_epoch) self.assertEqual(scheduler.last_lr, scheduler1.last_lr) def noam_lr(epoch_num, d_model, warmup_steps, learning_rate=1.0, verbose=False): if epoch_num == 0: a = 1 else: a = math.pow(epoch_num, -0.5) b = math.pow(warmup_steps, -1.5) * epoch_num return learning_rate * math.pow(d_model, -0.5) * min(a, b) def lambda_lr(epoch_num, learning_rate, lr_lambda, verbose=False): return learning_rate * lr_lambda(epoch_num) def multiplicative_lr(epoch_num, learning_rate, lr_lambda, verbose=False): latest_lr = learning_rate for i in range(epoch_num): latest_lr = latest_lr * lr_lambda(i + 1) return latest_lr def piecewise_lr(epoch_num, boundaries, values, verbose=False): assert len(boundaries) + 1 == len(values) for i in range(len(boundaries)): if epoch_num < boundaries[i]: return values[i] return values[len(values) - 1] def exponential_lr(epoch_num, learning_rate, gamma, verbose=False): return learning_rate * gamma**epoch_num def natural_exp_lr(epoch_num, learning_rate, gamma, verbose=False): return learning_rate * math.exp(-1 * gamma * epoch_num) def inverse_time_lr(epoch_num, learning_rate, gamma, verbose=False): return learning_rate / (1 + gamma * epoch_num) def polynomial_lr(epoch_num, learning_rate, decay_steps, end_lr=0.0001, power=1.0, cycle=False, verbose=False): if cycle: div = math.ceil(epoch_num / float(decay_steps)) if epoch_num == 0: div = 1 decay_steps = decay_steps * div else: epoch_num = min(epoch_num, decay_steps) return (learning_rate - end_lr) * ( (1 - float(epoch_num) / float(decay_steps))**power) + end_lr def get_lr(self): if self.last_epoch == 0: return self.base_lr elif (self.last_epoch - 1 - self.T_max) % (2 * self.T_max) == 0: return self.last_lr + (self.base_lr - self.eta_min) * ( 1 - math.cos(math.pi / self.T_max)) / 2 return (1 + math.cos(math.pi * self.last_epoch / self.T_max)) / ( 1 + math.cos(math.pi * (self.last_epoch - 1) / self.T_max)) * ( self.last_lr - self.eta_min) + self.eta_min cosine_annealing_lr_current = None def cosine_annealing_lr(epoch_num, learning_rate, T_max, eta_min=0, verbose=False): global cosine_annealing_lr_current if epoch_num == 0: cosine_annealing_lr_current = learning_rate elif (epoch_num - 1 - T_max) % (2 * T_max) == 0: cosine_annealing_lr_current = cosine_annealing_lr_current + ( learning_rate - eta_min) * (1 - math.cos(math.pi / float(T_max))) / 2 else: cosine_annealing_lr_current = ( 1 + math.cos(math.pi * epoch_num / float(T_max))) / ( 1 + math.cos(math.pi * (epoch_num - 1) / float(T_max))) * ( cosine_annealing_lr_current - eta_min) + eta_min return cosine_annealing_lr_current def linear_warmup_lr(epoch_num, learning_rate, warmup_steps, start_lr, end_lr, verbose=False): tmp = epoch_num - warmup_steps if tmp < 0: return start_lr + (end_lr - start_lr) * (float(epoch_num) / float(warmup_steps)) elif paddle.in_dynamic_mode(): if tmp < 3: return 0.5 elif tmp < 6: return 0.2 else: return 0.1 else: return 0.5 def multi_step_lr(epoch_num, learning_rate, milestones, gamma=0.1, verbose=False): for i in range(len(milestones)): if epoch_num < milestones[i]: return learning_rate * (gamma**i) return learning_rate * (gamma**len(milestones)) def step_lr(epoch_num, learning_rate, step_size, gamma=0.1, verbose=False): return learning_rate * math.pow(gamma, epoch_num // step_size) def one_cycle_lr(epoch_num, max_learning_rate, total_steps, divide_factor=25, end_learning_rate=0.0001, phase_pct=0.3, anneal_strategy='cos', three_phase=False, verbose=False): initial_lr = max_learning_rate / divide_factor if three_phase: _end_steps = [ float(phase_pct * total_steps) - 1, float(2 * phase_pct * total_steps) - 2, total_steps - 1 ] _schedule_phases = [ { 'start_lr': initial_lr, 'end_lr': max_learning_rate, }, { 'start_lr': max_learning_rate, 'end_lr': initial_lr, }, { 'start_lr': initial_lr, 'end_lr': end_learning_rate, }, ] else: _end_steps = [float(phase_pct * total_steps) - 1, total_steps - 1] _schedule_phases = [ { 'start_lr': initial_lr, 'end_lr': max_learning_rate, }, { 'start_lr': max_learning_rate, 'end_lr': end_learning_rate, }, ] if anneal_strategy == 'cos': def anneal_func(start, end, pct): cos_out = math.cos(math.pi * pct) + 1 return end + (start - end) / 2.0 * cos_out else: def anneal_func(start, end, pct): return (end - start) * pct + start start_step = 0 for i, phase in enumerate(_schedule_phases): end_step = _end_steps[i] if epoch_num <= end_step or i == len(_schedule_phases) - 1: pct = (epoch_num - start_step) / (end_step - start_step) computed_lr = anneal_func(phase['start_lr'], phase['end_lr'], pct) break start_step = end_step return computed_lr class TestLRScheduler(unittest.TestCase): def _test_static(self, python_func, paddle_api, kwarg, place): scheduler = paddle_api(**kwarg) adam = paddle.optimizer.Adam(learning_rate=scheduler) main_prog = paddle.static.Program() start_prog = paddle.static.Program() with paddle.static.program_guard(main_prog, start_prog): x = paddle.static.data(name='x', shape=[3, 4, 5]) loss = paddle.mean(x) adam.minimize(loss) lr_var = adam._global_learning_rate() test_prog = main_prog.clone() num = 0 exe = paddle.static.Executor(place) exe.run(start_prog) for epoch in range(5): for batch_id in range(2): out = exe.run( main_prog, feed={'x': np.random.randn(3, 4, 5).astype('float32')}, fetch_list=lr_var.name) self.assertEqual(out, np.array(python_func(num, **kwarg))) scheduler.step() num += 1 for epoch in range(5): for batch_id in range(2): out = exe.run( test_prog, feed={'x': np.random.randn(3, 4, 5).astype('float32')}, fetch_list=lr_var.name) self.assertEqual(out, np.array(python_func(num, **kwarg))) scheduler.step() num += 1 if isinstance(place, paddle.CPUPlace): compiled_train_prog = paddle.static.CompiledProgram( main_prog).with_data_parallel(loss_name=loss.name, places=fluid.cpu_places(4)) for epoch in range(5): python_result = python_func(num, **kwarg) for batch_id in range(2): _ = exe.run( compiled_train_prog, feed={'x': np.random.randn(12, 4, 5).astype('float32')}, fetch_list=lr_var.name) scopes = compiled_train_prog._executor.local_scopes() out = np.array(scopes[0].var(lr_var.name).get_tensor()) self.assertEqual(out, np.array(python_result)) out = np.array(scopes[1].var(lr_var.name).get_tensor()) self.assertEqual(out, np.array(python_result)) out = np.array(scopes[2].var(lr_var.name).get_tensor()) self.assertEqual(out, np.array(python_result)) out = np.array(scopes[3].var(lr_var.name).get_tensor()) self.assertEqual(out, np.array(python_result)) scheduler.step() num += 1 compiled_test_prog = paddle.static.CompiledProgram( test_prog).with_data_parallel( loss_name=loss.name, share_vars_from=compiled_train_prog, places=fluid.cpu_places(4)) for epoch in range(5): python_result = python_func(num, **kwarg) for batch_id in range(2): _ = exe.run( compiled_test_prog, feed={'x': np.random.randn(12, 4, 5).astype('float32')}, fetch_list=lr_var.name) scopes = compiled_test_prog._executor.local_scopes() out = np.array(scopes[0].var(lr_var.name).get_tensor()) self.assertEqual(out, np.array(python_result)) out = np.array(scopes[1].var(lr_var.name).get_tensor()) self.assertEqual(out, np.array(python_result)) out = np.array(scopes[2].var(lr_var.name).get_tensor()) self.assertEqual(out, np.array(python_result)) out = np.array(scopes[3].var(lr_var.name).get_tensor()) self.assertEqual(out, np.array(python_result)) scheduler.step() num += 1 def _test_dygraph(self, python_func, paddle_api, kwarg, place): paddle.disable_static(place) x = np.random.uniform(-1, 1, [10, 10]).astype("float32") linear = paddle.nn.Linear(10, 10) if paddle_api.__name__ == "LinearWarmup": kwarg['learning_rate'] = paddle.optimizer.lr.PiecewiseDecay( [3, 6], [0.5, 0.2, 0.1]) scheduler = paddle_api(**kwarg) adam = paddle.optimizer.Adam(learning_rate=scheduler, parameters=linear.parameters()) for epoch in range(20): for batch_id in range(2): x = paddle.to_tensor(x) out = linear(x) loss = paddle.mean(out) loss.backward() adam.step() adam.clear_grad() current_lr = adam.get_lr() expected_lr = python_func(epoch, **kwarg) if paddle_api.__name__ == "CosineAnnealingDecay": self.assertAlmostEqual(current_lr, expected_lr) scheduler.step(epoch + 1) elif paddle_api.__name__ == "LinearWarmup": self.assertAlmostEqual(current_lr, expected_lr) state_dict = adam.state_dict() scheduler1 = paddle.optimizer.lr.LinearWarmup(**kwarg) adam1 = paddle.optimizer.Adam(learning_rate=scheduler1, parameters=linear.parameters()) adam1.set_state_dict(state_dict) self.assertEqual(scheduler.last_epoch, scheduler1.last_epoch) self.assertEqual(scheduler.last_lr, scheduler1.last_lr) self.assertEqual(scheduler.learning_rate.last_lr, scheduler1.learning_rate.last_lr) self.assertEqual(scheduler.learning_rate.last_epoch, scheduler1.learning_rate.last_epoch) scheduler.step() else: self.assertEqual(current_lr, expected_lr) scheduler.step() def test_scheduler(self): with self.assertRaises(NotImplementedError): paddle.optimizer.lr.LRScheduler().step() with self.assertRaises(TypeError): paddle.optimizer.lr.MultiStepDecay(learning_rate="test", milestones=[1, 2, 3]) with self.assertRaises(TypeError): paddle.optimizer.lr.MultiStepDecay(learning_rate=0.5, milestones='test') with self.assertRaises(ValueError): paddle.optimizer.lr.MultiStepDecay(learning_rate=0.5, milestones=[3, 2, 1]) with self.assertRaises(ValueError): paddle.optimizer.lr.MultiStepDecay(learning_rate=0.5, milestones=[1, 2, 3], gamma=2) with self.assertRaises(TypeError): paddle.optimizer.lr.OneCycleLR(max_learning_rate='test', total_steps=20) with self.assertRaises(ValueError): paddle.optimizer.lr.OneCycleLR(max_learning_rate=-1.5, total_steps=20) with self.assertRaises(TypeError): paddle.optimizer.lr.OneCycleLR(max_learning_rate=0.1, total_steps=20, end_learning_rate='test') with self.assertRaises(ValueError): paddle.optimizer.lr.OneCycleLR(max_learning_rate=0.1, total_steps=20, end_learning_rate=-1) with self.assertRaises(TypeError): paddle.optimizer.lr.OneCycleLR(max_learning_rate=0.1, total_steps='test') with self.assertRaises(ValueError): paddle.optimizer.lr.OneCycleLR(max_learning_rate=0.1, total_steps=-10) with self.assertRaises(ValueError): paddle.optimizer.lr.OneCycleLR(max_learning_rate=0.1, total_steps=20, anneal_strategy='test') with self.assertRaises(ValueError): paddle.optimizer.lr.OneCycleLR(max_learning_rate=0.1, total_steps=20, phase_pct=0.6, three_phase=True) func_api_kwargs = [ (noam_lr, paddle.optimizer.lr.NoamDecay, { "d_model": 0.01, "warmup_steps": 100, "verbose": False }), (piecewise_lr, paddle.optimizer.lr.PiecewiseDecay, { "boundaries": [3, 6, 9, 15, 20], "values": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6], "verbose": False }), (natural_exp_lr, paddle.optimizer.lr.NaturalExpDecay, { "learning_rate": 0.5, "gamma": 0.1, "verbose": True }), (inverse_time_lr, paddle.optimizer.lr.InverseTimeDecay, { "learning_rate": 0.5, "gamma": 0.1, "verbose": False }), (polynomial_lr, paddle.optimizer.lr.PolynomialDecay, { "learning_rate": 0.5, "decay_steps": 20, "end_lr": 0, "power": 1.0, "cycle": False }), (polynomial_lr, paddle.optimizer.lr.PolynomialDecay, { "learning_rate": 0.5, "decay_steps": 20, "end_lr": 0, "power": 1.0, "cycle": True, "verbose": False }), (linear_warmup_lr, paddle.optimizer.lr.LinearWarmup, { 'learning_rate': 0.5, 'warmup_steps': 10, 'start_lr': 0, 'end_lr': 0.5 }), (exponential_lr, paddle.optimizer.lr.ExponentialDecay, { "learning_rate": 0.5, "gamma": 0.9, "verbose": False }), (multi_step_lr, paddle.optimizer.lr.MultiStepDecay, { "learning_rate": 0.5, "milestones": [3, 6, 9, 15, 20], "gamma": 0.8 }), (step_lr, paddle.optimizer.lr.StepDecay, { "learning_rate": 0.5, "step_size": 2, "gamma": 0.8, "verbose": False }), (lambda_lr, paddle.optimizer.lr.LambdaDecay, { "learning_rate": 0.5, "lr_lambda": lambda x: 0.95**x, "verbose": True }), (multiplicative_lr, paddle.optimizer.lr.MultiplicativeDecay, { "learning_rate": 0.5, "lr_lambda": lambda x: 0.95, "verbose": True }), (cosine_annealing_lr, paddle.optimizer.lr.CosineAnnealingDecay, { "learning_rate": 0.5, "T_max": 10, "verbose": False }), (one_cycle_lr, paddle.optimizer.lr.OneCycleLR, { "max_learning_rate": 0.1, "total_steps": 20, "divide_factor": 5, "end_learning_rate": 0.0001, "anneal_strategy": 'cos', "phase_pct": 0.3, "three_phase": False, }), (one_cycle_lr, paddle.optimizer.lr.OneCycleLR, { "max_learning_rate": 0.5, "total_steps": 20, "divide_factor": 10, "end_learning_rate": 0.001, "anneal_strategy": 'linear', "phase_pct": 0.4, "three_phase": False, }), (one_cycle_lr, paddle.optimizer.lr.OneCycleLR, { "max_learning_rate": 1.0, "total_steps": 20, "divide_factor": 9, "end_learning_rate": 0.0001, "anneal_strategy": 'cos', "phase_pct": 0.3, "three_phase": True, }), (one_cycle_lr, paddle.optimizer.lr.OneCycleLR, { "max_learning_rate": 0.3, "total_steps": 20, "divide_factor": 25, "end_learning_rate": 0.0005, "anneal_strategy": 'linear', "phase_pct": 0.2, "three_phase": True, }) ] for python_func, paddle_api, kwarg in func_api_kwargs: places = [paddle.CPUPlace()] if core.is_compiled_with_cuda(): places.append(paddle.CUDAPlace(0)) for place in places: paddle.enable_static() self._test_static(python_func, paddle_api, kwarg, place) paddle.disable_static(place) self._test_dygraph(python_func, paddle_api, kwarg, place) paddle.enable_static() def test_linear_warmp(self): natural_lr = paddle.optimizer.lr.NaturalExpDecay(learning_rate=0.5, gamma=0.1) natural_lr_warmup = paddle.optimizer.lr.LinearWarmup( learning_rate=natural_lr, warmup_steps=10, start_lr=0.0, end_lr=0.1) for idx in range(30): if idx >= 10: self.assertEqual(natural_lr_warmup.get_lr(), natural_lr.get_lr()) natural_lr.step() natural_lr_warmup.step() if __name__ == '__main__': paddle.enable_static() unittest.main()