def test_minibatch_score_trainer():
X = np.random.random((100, 10))
Z = np.random.random((100, 2))
X, Z = theano_floatx(X, Z)
data = {
'train': (X, Z),
'val': (X, Z),
'test': (X, Z)
}
cut_size = 10
class MyModel(mlp.Mlp):
def score(self, X, Z):
assert X.shape[0] <= cut_size
return super(MyModel, self).score(X, Z)
m = MyModel(10, [100], 2, ['tanh'], 'identity', 'squared',
max_iter=10)
score = MinibatchScore(cut_size, [0, 0])
trainer = Trainer(
m, data, score=score, pause=lambda info: True, stop=lambda info: False)
trainer.val_key = 'val'
for _ in trainer.iter_fit(X, Z):
break
def test_training_continuation():
# Make model and data for the test.
X = np.random.random((10, 2))
X, = theano_floatx(X)
optimizer = 'gd'
m = autoencoder.AutoEncoder(2, [2], ['tanh'],
'identity',
'squared',
tied_weights=True,
max_iter=10,
optimizer=optimizer)
# Train the mdoel with a trainer for 2 epochs.
stopper = climin.stops.OnSignal()
print stopper.sig
stops = climin.stops.Any([stopper, climin.stops.AfterNIterations(5)])
t = Trainer(m, stop=stops, pause=climin.stops.always)
t.val_key = 'val'
t.eval_data = {'val': (X, )}
killed = False
for info in t.iter_fit(X):
os.kill(os.getpid(), stopper.sig)
assert info['n_iter'] == 1
def test_training_continuation():
if sys.platform == 'win32':
raise SkipTest()
# Make model and data for the test.
X = np.random.random((100, 10))
Z = np.random.random((100, 2))
X, Z = theano_floatx(X, Z)
data = {
'train': (X, Z),
'val': (X, Z),
'test': (X, Z)
}
optimizer = 'rmsprop', {'step_rate': 0.0001}
m = mlp.Mlp(10, [2], 2, ['tanh'], 'identity', 'squared', max_iter=10,
optimizer=optimizer)
# Train the mdoel with a trainer for 2 epochs.
stopper = climin.stops.OnSignal()
stops = climin.stops.Any([stopper, climin.stops.AfterNIterations(5)])
t = Trainer(
m,
data,
stop=stops,
pause=climin.stops.always)
t.val_key = 'val'
for info in t.iter_fit(X, Z):
os.kill(os.getpid(), stopper.sig)
assert info['n_iter'] == 1
def test_minibatch_score_trainer():
X = np.random.random((100, 10))
X, = theano_floatx(X)
cut_size = 10
class MyAutoEncoder(autoencoder.AutoEncoder):
def score(self, X):
assert X.shape[0] <= cut_size
return super(MyAutoEncoder, self).score(X)
m = MyAutoEncoder(10, [100], ['tanh'],
'identity',
'squared',
tied_weights=True,
max_iter=10)
score = MinibatchScore(cut_size, [0])
trainer = Trainer(m,
score=score,
pause=lambda info: True,
stop=lambda info: False)
trainer.eval_data = {'val': (X, )}
trainer.val_key = 'val'
for _ in trainer.iter_fit(X):
break
def test_training_continuation():
# Make model and data for the test.
X = np.random.random((10, 2))
X, = theano_floatx(X)
optimizer = 'gd'
m = autoencoder.AutoEncoder(2, [2], ['tanh'], 'identity', 'squared',
tied_weights=True, max_iter=10,
optimizer=optimizer)
# Train the mdoel with a trainer for 2 epochs.
stopper = climin.stops.OnSignal()
print stopper.sig
stops = climin.stops.Any([stopper, climin.stops.AfterNIterations(5)])
t = Trainer(
m,
stop=stops,
pause=climin.stops.always)
t.val_key = 'val'
t.eval_data = {'val': (X,)}
killed = False
for info in t.iter_fit(X):
os.kill(os.getpid(), stopper.sig)
assert info['n_iter'] == 1
def test_minibatch_score_trainer():
X = np.random.random((100, 10))
Z = np.random.random((100, 2))
X, Z = theano_floatx(X, Z)
data = {'train': (X, Z), 'val': (X, Z), 'test': (X, Z)}
cut_size = 10
class MyModel(mlp.Mlp):
def score(self, X, Z):
assert X.shape[0] <= cut_size
return super(MyModel, self).score(X, Z)
m = MyModel(10, [100], 2, ['tanh'], 'identity', 'squared', max_iter=10)
score = MinibatchScore(cut_size, [0, 0])
trainer = Trainer(m,
data,
score=score,
pause=lambda info: True,
stop=lambda info: False)
trainer.val_key = 'val'
for _ in trainer.iter_fit(X, Z):
break
def test_training_continuation():
if sys.platform == 'win32':
raise SkipTest()
# Make model and data for the test.
X = np.random.random((100, 10))
Z = np.random.random((100, 2))
X, Z = theano_floatx(X, Z)
data = {'train': (X, Z), 'val': (X, Z), 'test': (X, Z)}
optimizer = 'rmsprop', {'step_rate': 0.0001}
m = mlp.Mlp(10, [2],
2, ['tanh'],
'identity',
'squared',
max_iter=10,
optimizer=optimizer)
# Train the mdoel with a trainer for 2 epochs.
stopper = climin.stops.OnSignal()
stops = climin.stops.Any([stopper, climin.stops.AfterNIterations(5)])
t = Trainer(m, data, stop=stops, pause=climin.stops.always)
t.val_key = 'val'
for info in t.iter_fit(X, Z):
os.kill(os.getpid(), stopper.sig)
assert info['n_iter'] == 1
def new_trainer(pars, data):
X, Z = data
m = Mlp(2, [pars['n_hidden']], 1,
hidden_transfers=[pars['hidden_transfer']], out_transfer='sigmoid',
loss='bern_ces',
optimizer=pars['optimizer'])
climin.initialize.randomize_normal(m.parameters.data, 0, pars['par_std'])
n_report = 100
interrupt = climin.stops.OnSignal()
print dir(climin.stops)
stop = climin.stops.Any([
climin.stops.AfterNIterations(10000),
climin.stops.OnSignal(signal.SIGTERM),
climin.stops.BetterThan('train_loss', 1e-4),
])
pause = climin.stops.ModuloNIterations(n_report)
reporter = KeyPrinter(['n_iter', 'train_loss'])
t = Trainer(
m,
stop=stop, pause=pause, report=reporter,
interrupt=interrupt)
t.val_key = 'train'
t.eval_data['train'] = (X, Z)
return t
def test_checkpoint_trainer():
# Make model and data for the test.
X = np.random.random((10, 2))
X, = theano_floatx(X)
optimizer = 'rmsprop', {'step_rate': 0.0001}
m = autoencoder.AutoEncoder(2, [2], ['tanh'],
'identity',
'squared',
tied_weights=True,
max_iter=10,
optimizer=optimizer)
# Train the mdoel with a trainer for 2 epochs.
t = Trainer(m,
stop=climin.stops.AfterNIterations(2),
pause=climin.stops.always)
t.val_key = 'val'
t.eval_data = {'val': (X, )}
t.fit(X)
# Make a copy of the trainer.
t2 = copy.deepcopy(t)
intermediate_pars = t2.model.parameters.data.copy()
intermediate_info = t2.current_info.copy()
# Train original for 2 more epochs.
t.stop = climin.stops.AfterNIterations(4)
t.fit(X)
# Check that the snapshot has not changed
assert np.all(t2.model.parameters.data == intermediate_pars)
final_pars = t.model.parameters.data.copy()
final_info = t.current_info.copy()
check_infos(intermediate_info, t2.current_info)
t2.stop = climin.stops.AfterNIterations(4)
t2.fit(X)
check_infos(final_info, t2.current_info)
assert np.allclose(final_pars, t2.model.parameters.data)
t_pickled = cPickle.dumps(t2)
t_unpickled = cPickle.loads(t_pickled)
t.stop = climin.stops.AfterNIterations(4)
t_unpickled.fit(X)
assert np.allclose(final_pars,
t_unpickled.model.parameters.data,
atol=5.e-3)
def new_trainer(self, pars, data):
modules = ['theano', 'breze', 'climin', 'alchemie']
git_log(modules)
copy_theanorc()
m = Mlp(2, [pars['n_hidden']],
1,
hidden_transfers=[pars['hidden_transfer']],
out_transfer='sigmoid',
loss='bern_ces',
optimizer=pars['optimizer'])
climin.initialize.randomize_normal(m.parameters.data, 0,
pars['par_std'])
n_report = 100
t = Trainer(
model=m,
data=data,
stop=climin.stops.Any([
climin.stops.AfterNIterations(10000),
climin.stops.NotBetterThanAfter(1e-1, 5000, key='val_loss')
]),
pause=climin.stops.ModuloNIterations(n_report),
report=OneLinePrinter(
keys=['n_iter', 'runtime', 'train_loss', 'val_loss'],
spaces=[6, '10.2f', '15.8f', '15.8f']),
interrupt=climin.stops.OnSignal(),
)
return t
def new_trainer(pars, data):
m = Mlp(2, [pars['n_hidden']],
1,
hidden_transfers=[pars['hidden_transfer']],
out_transfer='sigmoid',
loss='bern_ces',
optimizer=pars['optimizer'])
climin.initialize.randomize_normal(m.parameters.data, 0, pars['par_std'])
n_report = 100
interrupt = climin.stops.OnSignal()
print dir(climin.stops)
stop = climin.stops.Any([
climin.stops.AfterNIterations(10000),
climin.stops.OnSignal(signal.SIGTERM),
climin.stops.NotBetterThanAfter(1e-1, 500, key='train_loss'),
])
pause = climin.stops.ModuloNIterations(n_report)
reporter = KeyPrinter(['n_iter', 'train_loss'])
t = Trainer(m,
stop=stop,
pause=pause,
report=reporter,
interrupt=interrupt)
make_data_dict(t, data)
return t
def test_checkpoint_trainer():
# Make model and data for the test.
X = np.random.random((100, 10))
Z = np.random.random((100, 2))
X, Z = theano_floatx(X, Z)
data = {
'train': (X, Z),
'val': (X, Z),
'test': (X, Z)
}
optimizer = 'rmsprop', {'step_rate': 0.0001}
m = mlp.Mlp(10, [2], 2, ['tanh'], 'identity', 'squared', max_iter=10,
optimizer=optimizer)
# Train the mdoel with a trainer for 2 epochs.
t = Trainer(
m,
data,
stop=climin.stops.AfterNIterations(2),
pause=climin.stops.always)
t.val_key = 'val'
t.fit()
# Make a copy of the trainer.
t2 = copy.deepcopy(t)
t2.data = t.data
intermediate_pars = t2.model.parameters.data.copy()
intermediate_info = t2.current_info.copy()
# Train original for 2 more epochs.
t.stop = climin.stops.AfterNIterations(4)
t.fit()
# Check that the snapshot has not changed
assert np.all(t2.model.parameters.data == intermediate_pars)
final_pars = t.model.parameters.data.copy()
final_info = t.current_info.copy()
check_infos(intermediate_info, t2.current_info)
t2.stop = climin.stops.AfterNIterations(4)
t2.fit()
check_infos(final_info, t2.current_info)
assert np.allclose(final_pars, t2.model.parameters.data)
t_pickled = cPickle.dumps(t2)
t_unpickled = cPickle.loads(t_pickled)
t_unpickled.data = data
t.stop = climin.stops.AfterNIterations(4)
t_unpickled.fit()
assert np.allclose(
final_pars, t_unpickled.model.parameters.data, atol=5.e-3)
def test_minibatch_score_trainer():
X = np.random.random((100, 10))
X, = theano_floatx(X)
cut_size = 10
class MyAutoEncoder(autoencoder.AutoEncoder):
def score(self, X):
assert X.shape[0] <= cut_size
return super(MyAutoEncoder, self).score(X)
m = MyAutoEncoder(10, [100], ['tanh'], 'identity', 'squared',
tied_weights=True, max_iter=10)
score = MinibatchScore(cut_size, [0])
trainer = Trainer(
m, score=score, pause=lambda info: True, stop=lambda info: False)
trainer.eval_data = {'val': (X,)}
trainer.val_key = 'val'
for _ in trainer.iter_fit(X):
break