def test_multi_scheduler_without_state(self):
start_value = 100.0
decay_steps = 100
stop_value = 0
lin_sche = sche.LinearScheduler(start_value, decay_steps, stop_value)
constant = 50.0
const_sche = sche.ConstantScheduler(constant)
m_sches = sche.MultiScheduler([lin_sche, const_sche], op='min')
self.assertAlmostEqual(m_sches(0), constant)
self.assertAlmostEqual(m_sches(51), 49.0)
def test_multi_scheduler_custom_op(self):
start_value = 100.0
decay_steps = 100
stop_value = 0
lin_sche = sche.LinearScheduler(start_value, decay_steps, stop_value)
constant = 50.0
const_sche = sche.ConstantScheduler(constant)
custom_op = lambda x: np.sum(x, axis=0)
m_sches = sche.MultiScheduler([lin_sche, const_sche], op=custom_op)
self.assertAlmostEqual(m_sches(0), constant + start_value)
self.assertAlmostEqual(m_sches(100), constant + stop_value)
def test_multi_scheduler_dump_load(self):
state = create_empty_state()
start_value = 100.0
decay_steps = 100
stop_value = 0
lin_sche = sche.LinearScheduler(start_value,
decay_steps,
stop_value,
unit='progress')
constant = 0.5
const_sche = sche.ConstantScheduler(constant,
unit='epoch',
state=state)
m_sches = sche.MultiScheduler([lin_sche, const_sche], op='mean')
d = json.loads(json.dumps(m_sches))
self.assertEqual(set(d.keys()),
set(['type', 'unit', 'schedulers', 'op']))
self.assertEqual(d['type'], 'MultiScheduler')
self.assertEqual(d['op'], 'mean')
self.assertEqual(len(d['schedulers']), 2)
d0 = d['schedulers'][0]
self.assertEqual(
set(d0.keys()),
set(['type', 'unit', 'start_value', 'decay_steps', 'stop_value']))
self.assertEqual(d0['type'], 'LinearScheduler')
self.assertEqual(d0['unit'], 'progress')
self.assertEqual(d0['start_value'], start_value)
self.assertEqual(d0['decay_steps'], decay_steps)
self.assertEqual(d0['stop_value'], stop_value)
d1 = d['schedulers'][1]
self.assertEqual(set(d1.keys()), set(['type', 'unit', 'value']))
self.assertEqual(d1['type'], 'ConstantScheduler')
self.assertEqual(d1['unit'], 'epoch')
self.assertEqual(d1['value'], constant)
new_sche = sche.get_scheduler(d)
self.assertTrue(isinstance(new_sche, sche.MultiScheduler))
self.assertEqual(new_sche.op, 'mean')
self.assertEqual(len(new_sche.schedulers), 2)
sche0 = new_sche.schedulers[0]
self.assertTrue(isinstance(sche0, sche.LinearScheduler))
self.assertEqual(sche0.unit, sche.Unit.progress)
self.assertEqual(sche0.start_value, start_value)
self.assertEqual(sche0.decay_steps, decay_steps)
self.assertEqual(sche0.stop_value, stop_value)
sche1 = new_sche.schedulers[1]
self.assertTrue(isinstance(sche1, sche.ConstantScheduler))
self.assertEqual(sche1.unit, sche.Unit.epoch)
self.assertEqual(sche1.value, constant)
def test_multi_scheduler_custom_op_dump_load(self):
start_value = 100.0
decay_steps = 100
stop_value = 0
lin_sche = sche.LinearScheduler(start_value, decay_steps, stop_value)
constant = 50.0
const_sche = sche.ConstantScheduler(constant)
custom_op = lambda x: np.sum(x, axis=0)
m_sches = sche.MultiScheduler([lin_sche, const_sche], op=custom_op)
d = json.loads(json.dumps(m_sches))
self.assertEqual(set(d.keys()),
set(['type', 'unit', 'schedulers', 'op']))
self.assertEqual(d['type'], 'MultiScheduler')
self.assertEqual(d['op'], None)
self.assertEqual(len(d['schedulers']), 2)
d0 = d['schedulers'][0]
self.assertEqual(
set(d0.keys()),
set(['type', 'unit', 'start_value', 'decay_steps', 'stop_value']))
self.assertEqual(d0['type'], 'LinearScheduler')
self.assertEqual(d0['start_value'], start_value)
self.assertEqual(d0['decay_steps'], decay_steps)
self.assertEqual(d0['stop_value'], stop_value)
d1 = d['schedulers'][1]
self.assertEqual(set(d1.keys()), set(['type', 'unit', 'value']))
self.assertEqual(d1['type'], 'ConstantScheduler')
self.assertEqual(d1['value'], constant)
# custom op not specified, raise ValueError
with self.assertRaises(ValueError):
new_sche = sche.get_scheduler(d)
new_sche = sche.get_scheduler(d, op=custom_op)
self.assertTrue(isinstance(new_sche, sche.MultiScheduler))
self.assertEqual(new_sche.op, None)
self.assertEqual(len(new_sche.schedulers), 2)
sche0 = new_sche.schedulers[0]
self.assertTrue(isinstance(sche0, sche.LinearScheduler))
self.assertEqual(sche0.start_value, start_value)
self.assertEqual(sche0.decay_steps, decay_steps)
self.assertEqual(sche0.stop_value, stop_value)
sche1 = new_sche.schedulers[1]
self.assertTrue(isinstance(sche1, sche.ConstantScheduler))
self.assertEqual(sche1.value, constant)
self.assertAlmostEqual(new_sche(0), constant + start_value)
self.assertAlmostEqual(new_sche(100), constant + stop_value)
def test_multi_scheduler_with_state(self):
state = create_empty_state()
start_value = 100.0
decay_steps = 100
stop_value = 0
lin_sche = sche.LinearScheduler(start_value,
decay_steps,
stop_value,
unit='timestep')
constant = 50.0
const_sche = sche.ConstantScheduler(constant, unit='timestep')
m_sches = sche.MultiScheduler([lin_sche, const_sche], op='min')
m_sches.bind(state)
self.assertTrue(lin_sche.state is state)
self.assertTrue(const_sche.state is state)
state.num_timesteps = 0
self.assertAlmostEqual(m_sches(), constant)
state.num_timesteps = 51
self.assertAlmostEqual(m_sches(), 49.0)
def test_constant_scheduler_with_state(self):
state = create_empty_state()
constant = 1.0
const_sche = sche.ConstantScheduler(constant, unit='timestep')
self.assertEqual(const_sche(100), constant)
def test_constant_scheduler_without_state(self):
constant = 1.0
const_sche = sche.ConstantScheduler(constant)
self.assertEqual(const_sche(100), constant)