def entrenar(checkpoint, entrRuedas, entrOperaciones, input_dim, num_output_classes, testRuedas, testOperaciones):
minibatch_size = 100;
epocs=900;
minibatchIteraciones = int(len(entrOperaciones) / minibatch_size);
# Input variables denoting the features and label data
feature = input((input_dim), np.float32)
label = input((num_output_classes), np.float32)
netout = crearRed(input_dim, num_output_classes, feature);
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
lr_per_minibatch=learning_rate_schedule(0.25, UnitType.minibatch)
# Instantiate the trainer object to drive the model training
learner = sgd(netout.parameters, lr=lr_per_minibatch)
progress_printer = ProgressPrinter(log_to_file=checkpoint+".log", num_epochs=epocs);
trainer = Trainer(netout, (ce, pe), learner, progress_printer)
if os.path.isfile(checkpoint):
trainer.restore_from_checkpoint(checkpoint);
npentrRuedas = np.array(entrRuedas).astype(np.float32);
npentrOperaciones = np.array(entrOperaciones).astype(np.float32);
#iteramos una vez por cada "epoc"
for i in range(0, epocs):
p = np.random.permutation(len(entrRuedas));
npentrOperaciones = npentrOperaciones[p];
npentrRuedas = npentrRuedas[p];
#ahora partimos los datos en "minibatches" y entrenamos
for j in range(0, minibatchIteraciones):
features = npentrRuedas[j*minibatch_size:(j+1)*minibatch_size];
labels = npentrOperaciones[j*minibatch_size:(j+1)*minibatch_size];
trainer.train_minibatch({feature: features, label: labels});
trainer.summarize_training_progress()
trainer.save_checkpoint(checkpoint);
minibatchIteraciones = int(len(testOperaciones) / minibatch_size);
avg_error = 0;
for j in range(0, minibatchIteraciones):
test_features = np.array(testRuedas[j*minibatch_size:(j+1)*minibatch_size]).astype(np.float32);
test_labels = np.array(testOperaciones[j*minibatch_size:(j+1)*minibatch_size]).astype(np.float32);
#test_features = np.array( entrRuedas[0:minibatch_size]).astype(np.float32);
#test_labels = np.array(entrOperaciones[0:minibatch_size]).astype(np.float32);
avg_error = avg_error + ( trainer.test_minibatch(
{feature: test_features, label: test_labels}) / minibatchIteraciones)
return avg_error
def test_sgd_with_noise():
# Runs a network where the number of parameters is odd
# in some layers. This tests that cuRand library will not
# complain about generating an odd number of random values
np.random.seed(98052)
learner = lambda params: sgd(params, lr=learning_rate_schedule(0.125, UnitType.minibatch), gaussian_noise_injection_std_dev=0.01)
ffnet(learner)
# We just verify that we did not crash
assert(True)
def test_sgd_with_noise():
# Runs a network where the number of parameters is odd
# in some layers. This tests that cuRand library will not
# complain about generating an odd number of random values
np.random.seed(98052)
learner = lambda params: sgd(params, lr=learning_rate_schedule(0.125, UnitType.minibatch), gaussian_noise_injection_std_dev=0.01)
ffnet(learner)
# We just verify that we did not crash
assert(True)
def test_universal():
np.random.seed(98052)
builtin_sgd = lambda params: sgd(params, lr=learning_rate_schedule(0.125, UnitType.minibatch))
builtin_last_avg_error, builtin_avg_error, _ = ffnet(builtin_sgd)
np.random.seed(98052)
my_sgd = lambda ps, gs: C.combine([C.assign(p, p - 0.125/25 * g) for p, g in zip(ps, gs)])
universal_sgd = lambda params: universal(my_sgd, params)
my_last_avg_error, my_avg_error, _ = ffnet(universal_sgd)
assert np.all(np.less_equal(my_last_avg_error, builtin_last_avg_error))
assert np.all(np.less_equal(my_avg_error, builtin_avg_error))
def test_universal():
np.random.seed(98052)
builtin_sgd = lambda params: sgd(params, lr=learning_rate_schedule(0.125, UnitType.minibatch))
builtin_last_avg_error, builtin_avg_error = ffnet(builtin_sgd)
np.random.seed(98052)
my_sgd = lambda p, g: C.assign(p, p - 0.125/25 * g)
universal_sgd = lambda params: universal(my_sgd, params)
my_last_avg_error, my_avg_error = ffnet(universal_sgd)
assert np.allclose(my_last_avg_error, builtin_last_avg_error)
assert np.allclose(my_avg_error, builtin_avg_error)
def test_universal():
np.random.seed(98052)
builtin_sgd = lambda params: sgd(params, lr=learning_rate_schedule(0.125, UnitType.minibatch))
builtin_last_avg_error, builtin_avg_error, _ = ffnet(builtin_sgd)
np.random.seed(98052)
my_sgd = lambda ps, gs: C.combine([C.assign(p, p - 0.125/25 * g) for p, g in zip(ps, gs)])
universal_sgd = lambda params: universal(my_sgd, params)
my_last_avg_error, my_avg_error, _ = ffnet(universal_sgd)
assert np.all(np.less_equal(my_last_avg_error, builtin_last_avg_error))
assert np.all(np.less_equal(my_avg_error, builtin_avg_error))
def train_sequence_classifier(debug_output=False):
input_dim = 2000
cell_dim = 25
hidden_dim = 25
embedding_dim = 50
num_output_classes = 5
# Input variables denoting the features and label data
features = sequence.input(shape=input_dim, is_sparse=True)
label = input(num_output_classes)
# Instantiate the sequence classification model
classifier_output = LSTM_sequence_classifer_net(features,
num_output_classes,
embedding_dim, hidden_dim,
cell_dim)
ce = cross_entropy_with_softmax(classifier_output, label)
pe = classification_error(classifier_output, label)
rel_path = r"../../../../Tests/EndToEndTests/Text/SequenceClassification/Data/Train.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
reader = create_reader(path, True, input_dim, num_output_classes)
input_map = {
features: reader.streams.features,
label: reader.streams.labels
}
lr_per_sample = learning_rate_schedule(0.0005, UnitType.sample)
# Instantiate the trainer object to drive the model training
trainer = Trainer(classifier_output, (ce, pe),
sgd(classifier_output.parameters, lr=lr_per_sample))
# Get minibatches of sequences to train with and perform model training
minibatch_size = 200
training_progress_output_freq = 10
if debug_output:
training_progress_output_freq = training_progress_output_freq / 3
for i in range(251):
mb = reader.next_minibatch(minibatch_size, input_map=input_map)
trainer.train_minibatch(mb)
print_training_progress(trainer, i, training_progress_output_freq)
import copy
evaluation_average = copy.copy(
trainer.previous_minibatch_evaluation_average)
loss_average = copy.copy(trainer.previous_minibatch_loss_average)
return evaluation_average, loss_average
def __init__(self,
name,
observation_space_shape,
num_actions,
pretrained_policy=None,
*args,
**kwargs):
self.name = name
self.observation_space_shape = observation_space_shape
self.num_actions = num_actions
self._build_network(pretrained_policy)
self.trainer = Trainer(self.q, self.loss,
[sgd(self.q.parameters, lr=5e-4)])
def ffnet():
inputs = 3
outputs = 3
layers = 2
hidden_dimension = 3
# input variables denoting the features and label data
features = C.input((inputs), np.float32)
label = C.input((outputs), np.float32)
# Instantiate the feedforward classification model
my_model = Sequential(
[Dense(hidden_dimension, activation=C.sigmoid),
Dense(outputs)])
z = my_model(features)
ce = C.cross_entropy_with_softmax(z, label)
pe = C.classification_error(z, label)
# Instantiate the trainer object to drive the model training
lr_per_minibatch = learning_rate_schedule(0.125, UnitType.minibatch)
progress_printer = ProgressPrinter(0)
trainer = C.Trainer(z, (ce, pe), [
sgd(z.parameters,
lr=lr_per_minibatch,
gaussian_noise_injection_std_dev=0.01)
], [progress_printer])
# Get minibatches of training data and perform model training
minibatch_size = 25
num_minibatches_to_train = 100
aggregate_loss = 0.0
for i in range(num_minibatches_to_train):
train_features, labels = generate_random_data(minibatch_size, inputs,
outputs)
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
trainer.train_minibatch({features: train_features, label: labels})
sample_count = trainer.previous_minibatch_sample_count
aggregate_loss += trainer.previous_minibatch_loss_average * sample_count
last_avg_error = aggregate_loss / trainer.total_number_of_samples_seen
test_features, test_labels = generate_random_data(minibatch_size, inputs,
outputs)
avg_error = trainer.test_minibatch({
features: test_features,
label: test_labels
})
print(' error rate on an unseen minibatch: {}'.format(avg_error))
return last_avg_error, avg_error
def run_model(create_model_fn, **params):
NUM_TRAIN_SAMPLES = params.get('NUM_TRAIN_SAMPLES', 60000)
NUM_TEST_SAMPLES = params.get('NUM_TEST_SAMPLES', 10000)
INPUT_DIM_MODEL = params.get('INPUT_DIM_MODEL', 28 * 28)
INPUT_DIM = params.get('INPUT_DIM', 28 * 28)
NUM_OUTPUT_CLASSES = params.get('NUM_OUTPUT_CLASSES', 10)
LEARNING_RATE = params.get('LEARNING_RATE', 0.2)
MINIBATCH_SIZE = params.get('MINIBATCH_SIZE', 64)
NUM_SAMPLES_PER_SWEEP = params.get('NUM_SAMPLES_PER_SWEEP', 60000)
NUM_SWEEP_TO_TRAIN = params.get('NUM_SWEEP_TO_TRAIN', 10)
train_file, test_file = load_and_save(NUM_TRAIN_SAMPLES, NUM_TEST_SAMPLES)
input = C.input_variable(INPUT_DIM_MODEL)
label = C.input_variable(NUM_OUTPUT_CLASSES)
z = create_model_fn(input / 255.0, NUM_OUTPUT_CLASSES, **params)
loss = C.cross_entropy_with_softmax(z, label)
label_error = C.classification_error(z, label)
lr_schedule = learning_rate_schedule(LEARNING_RATE, UnitType.minibatch)
learner = sgd(z.parameters, lr_schedule)
trainer = C.Trainer(z, (loss, label_error), [learner])
# Create the reader to training data set
reader_train = create_reader(train_file, True, INPUT_DIM,
NUM_OUTPUT_CLASSES)
# Map the data streams to the input and labels.
input_map = {
label: reader_train.streams.labels,
input: reader_train.streams.features
}
tr = Trainer(MINIBATCH_SIZE, NUM_SAMPLES_PER_SWEEP, NUM_SWEEP_TO_TRAIN,
trainer, reader_train)
plotdata = tr.train(input_map)
plot_learning(plotdata)
# Read the training data
reader_test = create_reader(test_file, False, INPUT_DIM,
NUM_OUTPUT_CLASSES)
test_input_map = {
label: reader_test.streams.labels,
input: reader_test.streams.features,
}
test_model(test_input_map, reader_test, trainer, NUM_TEST_SAMPLES, 512)
return z
def ffnet(optimizer, num_minibatches_to_train, learning_rate_func, lr_args,
learner_kwargs):
inputs = 2
outputs = 2
hidden_dimension = 50
# input variables denoting the features and label data
features = C.input_variable((inputs), np.float32)
label = C.input_variable((outputs), np.float32)
# Instantiate the feedforward classification model
my_model = Sequential([
Dense(hidden_dimension,
activation=C.sigmoid,
init=C.glorot_uniform(seed=SEED)),
Dense(outputs, init=C.glorot_uniform(seed=SEED))
])
z = my_model(features)
ce = C.cross_entropy_with_softmax(z, label)
pe = C.classification_error(z, label)
# Instantiate the trainer object to drive the model training
lr = learning_rate_func(0.125, *lr_args)
progress_printer = ProgressPrinter(0)
learner = optimizer(z.parameters, lr) if optimizer != sgd else sgd(
z.parameters, lr, **learner_kwargs)
trainer = C.Trainer(z, (ce, pe), [learner], progress_printer)
# Get minibatches of training data and perform model training
minibatch_size = 25
for i in range(num_minibatches_to_train):
train_features, labels = generate_random_data(minibatch_size, inputs,
outputs)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
trainer.train_minibatch({features: train_features, label: labels})
test_features, test_labels = generate_random_data(minibatch_size, inputs,
outputs)
avg_error = trainer.test_minibatch({
features: test_features,
label: test_labels
})
print(' error rate on an unseen minibatch: {}'.format(avg_error))
return z.parameters
def train_sequence_classifier():
input_dim = 2000
cell_dim = 25
hidden_dim = 25
embedding_dim = 50
num_output_classes = 5
# Input variables denoting the features and label data
features = sequence.input_variable(shape=input_dim, is_sparse=True)
label = input_variable(num_output_classes)
# Instantiate the sequence classification model
classifier_output = LSTM_sequence_classifier_net(features,
num_output_classes,
embedding_dim, hidden_dim,
cell_dim)
ce = cross_entropy_with_softmax(classifier_output, label)
pe = classification_error(classifier_output, label)
rel_path = ("../../../Tests/EndToEndTests/Text/" +
"SequenceClassification/Data/Train.ctf")
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
reader = create_reader(path, True, input_dim, num_output_classes)
input_map = {
features: reader.streams.features,
label: reader.streams.labels
}
lr_per_sample = learning_parameter_schedule_per_sample(0.0005)
# Instantiate the trainer object to drive the model training
progress_printer = ProgressPrinter(0)
trainer = Trainer(classifier_output, (ce, pe),
sgd(classifier_output.parameters, lr=lr_per_sample),
progress_printer)
# Get minibatches of sequences to train with and perform model training
minibatch_size = 200
for i in range(255):
mb = reader.next_minibatch(minibatch_size, input_map=input_map)
trainer.train_minibatch(mb)
evaluation_average = float(trainer.previous_minibatch_evaluation_average)
loss_average = float(trainer.previous_minibatch_loss_average)
return evaluation_average, loss_average
def test_learner_update():
i = C.input_variable(shape=(1,), needs_gradient=True, name='a')
w_init = 1
w = parameter(shape=(1,), init=w_init)
res = i * w
learner = sgd(res.parameters, lr=C.learning_parameter_schedule([0.1]*50 + [0.2]*50, minibatch_size = 1, epoch_size=1))
assert learner.learning_rate() == 0.1
x = learner.update({w: np.asarray([[2.]], dtype=np.float32)}, 100)
assert learner.learning_rate() == 0.2
assert w.value < w_init
learner.reset_learning_rate(learning_parameter_schedule([0.3]*50 + [0.4]*50, minibatch_size = 1, epoch_size=1));
assert learner.learning_rate() == 0.3
x = learner.update({w: np.asarray([[2.]], dtype=np.float32)}, 100)
assert learner.learning_rate() == 0.4
def test_learner_update():
i = C.input_variable(shape=(1,), needs_gradient=True, name='a')
w_init = 1
w = parameter(shape=(1,), init=w_init)
res = i * w
learner = sgd(res.parameters, lr=learning_rate_schedule([0.1]*50 + [0.2]*50, UnitType.sample, 1))
assert learner.learning_rate() == 0.1
x = learner.update({w: np.asarray([[2.]], dtype=np.float32)}, 100)
assert learner.learning_rate() == 0.2
assert w.value < w_init
learner.reset_learning_rate(learning_rate_schedule([0.3]*50 + [0.4]*50, UnitType.sample, 1));
assert learner.learning_rate() == 0.3
x = learner.update({w: np.asarray([[2.]], dtype=np.float32)}, 100)
assert learner.learning_rate() == 0.4
def train_sequence_classifier():
input_dim = 2000
cell_dim = 25
hidden_dim = 25
embedding_dim = 50
num_output_classes = 5
# Input variables denoting the features and label data
features = sequence.input_variable(shape=input_dim, is_sparse=True)
label = input_variable(num_output_classes)
# Instantiate the sequence classification model
classifier_output = LSTM_sequence_classifier_net(
features, num_output_classes, embedding_dim, hidden_dim, cell_dim)
ce = cross_entropy_with_softmax(classifier_output, label)
pe = classification_error(classifier_output, label)
rel_path = ("../../../Tests/EndToEndTests/Text/" +
"SequenceClassification/Data/Train.ctf")
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
reader = create_reader(path, True, input_dim, num_output_classes)
input_map = {
features: reader.streams.features,
label: reader.streams.labels
}
lr_per_sample = learning_parameter_schedule_per_sample(0.0005)
# Instantiate the trainer object to drive the model training
progress_printer = ProgressPrinter(0)
trainer = Trainer(classifier_output, (ce, pe),
sgd(classifier_output.parameters, lr=lr_per_sample),
progress_printer)
# Get minibatches of sequences to train with and perform model training
minibatch_size = 200
for i in range(255):
mb = reader.next_minibatch(minibatch_size, input_map=input_map)
trainer.train_minibatch(mb)
evaluation_average = float(trainer.previous_minibatch_evaluation_average)
loss_average = float(trainer.previous_minibatch_loss_average)
return evaluation_average, loss_average
def __init__(self,
name,
num_frames_to_stack,
observation_space_shape,
num_actions,
pretrained_policy=None,
*args,
**kwargs):
self.name = name
self.num_frames_to_stack = num_frames_to_stack
self.observation_space_shape = observation_space_shape
self.frame_stacker = FrameStacker(stack_size=num_frames_to_stack,
frame_shape=observation_space_shape)
self.num_actions = num_actions
self._build_network(pretrained_policy)
self.trainer = Trainer(self.q, self.loss,
[sgd(self.q.parameters, lr=0.000001)])
def ffnet():
input_dim = 2
num_output_classes = 2
num_hidden_layers = 2
hidden_layers_dim = 50
# Input variables denoting the features and label data
feature = input_variable((input_dim), np.float32)
label = input_variable((num_output_classes), np.float32)
netout = Sequential([
For(range(num_hidden_layers),
lambda i: Dense(hidden_layers_dim, activation=sigmoid)),
Dense(num_output_classes)
])(feature)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
lr_per_minibatch = learning_parameter_schedule(0.5)
# Instantiate the trainer object to drive the model training
learner = sgd(netout.parameters, lr=lr_per_minibatch)
progress_printer = ProgressPrinter(128)
trainer = Trainer(netout, (ce, pe), learner, progress_printer)
# Get minibatches of training data and perform model training
minibatch_size = 25
for i in range(1024):
features, labels = generate_random_data(minibatch_size, input_dim,
num_output_classes)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
trainer.train_minibatch({feature: features, label: labels})
trainer.summarize_training_progress()
test_features, test_labels = generate_random_data(minibatch_size,
input_dim,
num_output_classes)
avg_error = trainer.test_minibatch({
feature: test_features,
label: test_labels
})
return avg_error
def ffnet():
input_dim = 2
num_output_classes = 2
num_hidden_layers = 2
hidden_layers_dim = 50
# Input variables denoting the features and label data
input = input_variable((input_dim), np.float32)
label = input_variable((num_output_classes), np.float32)
# Instantiate the feedforward classification model
netout = fully_connected_classifier_net(input, num_output_classes,
hidden_layers_dim,
num_hidden_layers, sigmoid)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
lr_per_minibatch = learning_rate_schedule(0.5, UnitType.minibatch)
# Instantiate the trainer object to drive the model training
learner = sgd(netout.parameters, lr=lr_per_minibatch)
progress_printer = ProgressPrinter(128)
trainer = Trainer(netout, (ce, pe), learner, progress_printer)
# Get minibatches of training data and perform model training
minibatch_size = 25
for i in range(1024):
features, labels = generate_random_data(minibatch_size, input_dim,
num_output_classes)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
trainer.train_minibatch({input: features, label: labels})
trainer.summarize_training_progress()
test_features, test_labels = generate_random_data(minibatch_size,
input_dim,
num_output_classes)
avg_error = trainer.test_minibatch({
input: test_features,
label: test_labels
})
return avg_error
def ffnet(optimizer, num_minibatches_to_train, learning_rate_func, lr_args, learner_kwargs):
inputs = 2
outputs = 2
hidden_dimension = 50
# input variables denoting the features and label data
features = C.input_variable((inputs), np.float32)
label = C.input_variable((outputs), np.float32)
# Instantiate the feedforward classification model
my_model = Sequential([
Dense(hidden_dimension, activation=C.sigmoid,
init=C.glorot_uniform(seed=SEED)),
Dense(outputs, init=C.glorot_uniform(seed=SEED))])
z = my_model(features)
ce = C.cross_entropy_with_softmax(z, label)
pe = C.classification_error(z, label)
# Instantiate the trainer object to drive the model training
lr= learning_rate_func(0.125, *lr_args)
progress_printer = ProgressPrinter(0)
learner = optimizer(z.parameters, lr) if optimizer != sgd else sgd(z.parameters, lr, **learner_kwargs)
trainer = C.Trainer(z, (ce, pe), [learner], progress_printer)
# Get minibatches of training data and perform model training
minibatch_size = 25
for i in range(num_minibatches_to_train):
train_features, labels = generate_random_data(
minibatch_size, inputs, outputs)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
trainer.train_minibatch({features: train_features, label: labels})
test_features, test_labels = generate_random_data(
minibatch_size, inputs, outputs)
avg_error = trainer.test_minibatch(
{features: test_features, label: test_labels})
print(' error rate on an unseen minibatch: {}'.format(avg_error))
return z.parameters
def ffnet():
inputs = 2
outputs = 2
layers = 2
hidden_dimension = 50
# input variables denoting the features and label data
features = C.input_variable((inputs), np.float32)
label = C.input_variable((outputs), np.float32)
# Instantiate the feedforward classification model
my_model = Sequential ([
Dense(hidden_dimension, activation=C.sigmoid),
Dense(outputs)])
z = my_model(features)
ce = C.cross_entropy_with_softmax(z, label)
pe = C.classification_error(z, label)
# Instantiate the trainer object to drive the model training
lr_per_minibatch = C.learning_parameter_schedule(0.125)
progress_printer = ProgressPrinter(0)
trainer = C.Trainer(z, (ce, pe), [sgd(z.parameters, lr=lr_per_minibatch)], [progress_printer])
# Get minibatches of training data and perform model training
minibatch_size = 25
num_minibatches_to_train = 1024
aggregate_loss = 0.0
for i in range(num_minibatches_to_train):
train_features, labels = generate_random_data(minibatch_size, inputs, outputs)
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
trainer.train_minibatch({features : train_features, label : labels})
sample_count = trainer.previous_minibatch_sample_count
aggregate_loss += trainer.previous_minibatch_loss_average * sample_count
last_avg_error = aggregate_loss / trainer.total_number_of_samples_seen
test_features, test_labels = generate_random_data(minibatch_size, inputs, outputs)
avg_error = trainer.test_minibatch({features : test_features, label : test_labels})
print(' error rate on an unseen minibatch: {}'.format(avg_error))
return last_avg_error, avg_error
def ffnet(data, labels):
input_dim = 800
num_output_classes = 3
num_hidden_layers = 2
hidden_layers_dim = 50
# Input variables denoting the features and label data
feature = input((input_dim), np.float32)
label = input((num_output_classes), np.float32)
netout = Sequential([For(range(num_hidden_layers), lambda i: Dense(hidden_layers_dim, activation=sigmoid)),
Dense(num_output_classes)])(feature)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
lr_per_minibatch=learning_rate_schedule(0.5, UnitType.minibatch)
# Instantiate the trainer object to drive the model training
learner = sgd(netout.parameters, lr=lr_per_minibatch)
progress_printer = ProgressPrinter(128)
trainer = Trainer(netout, (ce, pe), learner, progress_printer)
# Get minibatches of training data and perform model training
minibatch_size = 25
features, labels = generate_stock_data(minibatch_size);
for i in range(1024):
# features, labels = generate_random_data(
# minibatch_size, input_dim, num_output_classes)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
trainer.train_minibatch({feature: features, label: labels})
trainer.summarize_training_progress()
test_features, test_labels = generate_random_data(
minibatch_size, input_dim, num_output_classes)
avg_error = trainer.test_minibatch(
{feature: test_features, label: test_labels})
return avg_error
def cargarRedDesdeArchivo(archivo):
input_dim = 800;
num_output_classes = 3;
feature = input((input_dim), np.float32);
label = input((num_output_classes), np.float32)
netout = crearRed(input_dim, 3, feature);
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
lr_per_minibatch=learning_rate_schedule(0.5, UnitType.minibatch)
# Instantiate the trainer object to drive the model training
learner = sgd(netout.parameters, lr=lr_per_minibatch)
progress_printer = ProgressPrinter(1)
trainer = Trainer(netout, (ce, pe), learner, progress_printer)
trainer.restore_from_checkpoint(archivo);
return netout;
def simple_mnist(tensorboard_logdir=None):
input_dim = 784
num_output_classes = 10
num_hidden_layers = 2
hidden_layers_dim = 200
# Input variables denoting the features and label data
feature = C.input_variable(input_dim, np.float32)
label = C.input_variable(num_output_classes, np.float32)
# Instantiate the feedforward classification model
scaled_input = element_times(constant(0.00390625), feature)
z = Sequential([
For(range(num_hidden_layers),
lambda i: Dense(hidden_layers_dim, activation=relu)),
Dense(num_output_classes)
])(scaled_input)
ce = cross_entropy_with_softmax(z, label)
pe = classification_error(z, label)
data_dir = os.path.dirname(os.path.abspath(__file__))
path = os.path.join(data_dir, 'Train-28x28_cntk_text.txt')
reader_train = create_reader(path, True, input_dim, num_output_classes)
input_map = {
feature: reader_train.streams.features,
label: reader_train.streams.labels
}
# Training config
minibatch_size = 64
num_samples_per_sweep = 60000
num_sweeps_to_train_with = 10
# Instantiate progress writers.
# training_progress_output_freq = 100
progress_writers = [
ProgressPrinter(
# freq=training_progress_output_freq,
tag='Training',
num_epochs=num_sweeps_to_train_with)
]
if tensorboard_logdir is not None:
progress_writers.append(
TensorBoardProgressWriter(freq=10,
log_dir=tensorboard_logdir,
model=z))
# Instantiate the trainer object to drive the model training
lr = 0.001
trainer = Trainer(z, (ce, pe), sgd(z.parameters, lr), progress_writers)
training_session(trainer=trainer,
mb_source=reader_train,
mb_size=minibatch_size,
model_inputs_to_streams=input_map,
max_samples=num_samples_per_sweep *
num_sweeps_to_train_with,
progress_frequency=num_samples_per_sweep).train()
# Load test data
path = os.path.normpath(os.path.join(data_dir, "Test-28x28_cntk_text.txt"))
check_path(path)
reader_test = create_reader(path, False, input_dim, num_output_classes)
input_map = {
feature: reader_test.streams.features,
label: reader_test.streams.labels
}
# Test data for trained model
C.debugging.start_profiler()
C.debugging.enable_profiler()
C.debugging.set_node_timing(True)
test_minibatch_size = 1024
num_samples = 10000
num_minibatches_to_test = num_samples / test_minibatch_size
test_result = 0.0
for i in range(0, int(num_minibatches_to_test)):
mb = reader_test.next_minibatch(test_minibatch_size,
input_map=input_map)
eval_error = trainer.test_minibatch(mb)
test_result = test_result + eval_error
C.debugging.stop_profiler()
trainer.print_node_timing()
# Average of evaluation errors of all test minibatches
return test_result * 100 / num_minibatches_to_test
def test_sweep_based_schedule(tmpdir, device_id):
from cntk.io import MinibatchSource, CTFDeserializer, StreamDef, StreamDefs
from cntk import cross_entropy_with_softmax, classification_error, plus, reduce_sum, sequence
from cntk import Trainer
input_dim = 69
ctf_data = '''\
0 |S0 3:1 |S1 3:1 |# <s>
0 |S0 4:1 |# A |S1 32:1 |# ~AH
0 |S0 5:1 |# B |S1 36:1 |# ~B
0 |S0 4:1 |# A |S1 31:1 |# ~AE
0 |S0 7:1 |# D |S1 38:1 |# ~D
0 |S0 12:1 |# I |S1 47:1 |# ~IY
0 |S0 1:1 |# </s> |S1 1:1 |# </s>
2 |S0 60:1 |# <s> |S1 3:1 |# <s>
2 |S0 61:1 |# A |S1 32:1 |# ~AH
'''
ctf_file = str(tmpdir/'2seqtest.txt')
with open(ctf_file, 'w') as f:
f.write(ctf_data)
mbs = MinibatchSource(CTFDeserializer(ctf_file, StreamDefs(
features = StreamDef(field='S0', shape=input_dim, is_sparse=True),
labels = StreamDef(field='S1', shape=input_dim, is_sparse=True)
)), randomize=False)
in1 = sequence.input_variable(shape=(input_dim,))
labels = sequence.input_variable(shape=(input_dim,))
p = parameter(shape=(input_dim,), init=10)
z = plus(in1, reduce_sum(p), name='z')
ce = cross_entropy_with_softmax(z, labels)
errs = classification_error(z, labels)
lr_per_sample = learning_rate_schedule([0.3, 0.2, 0.1, 0.0], UnitType.sample)
learner = sgd(z.parameters, lr_per_sample)
trainer = Trainer(z, (ce, errs), [learner])
input_map = {
in1 : mbs.streams.features,
labels : mbs.streams.labels
}
# fetch minibatch (first sequence)
data = mbs.next_minibatch(1, input_map=input_map)
trainer.train_minibatch(data)
assert learner.learning_rate() == 0.3
# fetch minibatch (second sequence, sweep ends at this point)
data = mbs.next_minibatch(1, input_map=input_map)
trainer.train_minibatch(data)
assert learner.learning_rate() == 0.2
# fetch minibatch (both sequences -- entire sweep in one go)
data = mbs.next_minibatch(9, input_map=input_map)
trainer.train_minibatch(data)
assert learner.learning_rate() == 0.1
# fetch minibatch (multiple sweeps)
data = mbs.next_minibatch(30, input_map=input_map)
trainer.train_minibatch(data, outputs=[z.output])
assert learner.learning_rate() == 0.0
label : training_reader.streams.labels,
feature : training_reader.streams.features
}
"""
Set loss and evaluation functions
"""
loss = squared_error(netout, label)
evaluation = squared_error(netout, label)
lr_per_minibatch=learning_rate_schedule(learning_rate, UnitType.minibatch)
"""
Instantiate the trainer object to drive the model training
See: https://www.cntk.ai/pythondocs/cntk.learners.html
"""
learner = sgd(netout.parameters, lr=lr_per_minibatch)
# Other learners to try
#learner = momentum_sgd(netout.parameters, lr=lr_per_minibatch, momentum = momentum_schedule(0.9))
#learner = adagrad(netout.parameters, lr=lr_per_minibatch)
progress_printer = ProgressPrinter(minibatch_size)
"""
Instantiate the trainer
See: https://www.cntk.ai/pythondocs/cntk.train.html#module-cntk.train.trainer
"""
trainer = Trainer(netout, (loss, evaluation), learner, progress_printer)
#%%
"""
def simple_mnist(tensorboard_logdir=None):
input_dim = 784
num_output_classes = 10
num_hidden_layers = 1
hidden_layers_dim = 200
# Input variables denoting the features and label data
input = input_variable(input_dim, np.float32)
label = input_variable(num_output_classes, np.float32)
# Instantiate the feedforward classification model
scaled_input = element_times(constant(0.00390625), input)
z = fully_connected_classifier_net(
scaled_input, num_output_classes, hidden_layers_dim, num_hidden_layers, relu)
ce = cross_entropy_with_softmax(z, label)
pe = classification_error(z, label)
data_dir = os.path.join(abs_path, "..", "..", "..", "DataSets", "MNIST")
path = os.path.normpath(os.path.join(data_dir, "Train-28x28_cntk_text.txt"))
check_path(path)
reader_train = create_reader(path, True, input_dim, num_output_classes)
input_map = {
input : reader_train.streams.features,
label : reader_train.streams.labels
}
# Training config
minibatch_size = 64
num_samples_per_sweep = 60000
num_sweeps_to_train_with = 10
# Instantiate progress writers.
#training_progress_output_freq = 100
progress_writers = [ProgressPrinter(
#freq=training_progress_output_freq,
tag='Training',
num_epochs=num_sweeps_to_train_with)]
if tensorboard_logdir is not None:
progress_writers.append(TensorBoardProgressWriter(freq=10, log_dir=tensorboard_logdir, model=z))
# Instantiate the trainer object to drive the model training
lr_per_minibatch = learning_rate_schedule(0.2, UnitType.minibatch)
trainer = Trainer(z, (ce, pe), sgd(z.parameters, lr=lr_per_minibatch), progress_writers)
training_session(
trainer=trainer,
mb_source = reader_train,
mb_size = minibatch_size,
var_to_stream = input_map,
max_samples = num_samples_per_sweep * num_sweeps_to_train_with,
progress_frequency=num_samples_per_sweep
).train()
# Load test data
path = os.path.normpath(os.path.join(data_dir, "Test-28x28_cntk_text.txt"))
check_path(path)
reader_test = create_reader(path, False, input_dim, num_output_classes)
input_map = {
input : reader_test.streams.features,
label : reader_test.streams.labels
}
# Test data for trained model
test_minibatch_size = 1024
num_samples = 10000
num_minibatches_to_test = num_samples / test_minibatch_size
test_result = 0.0
for i in range(0, int(num_minibatches_to_test)):
mb = reader_test.next_minibatch(test_minibatch_size, input_map=input_map)
eval_error = trainer.test_minibatch(mb)
test_result = test_result + eval_error
# Average of evaluation errors of all test minibatches
return test_result / num_minibatches_to_test
##################
###### Loss ######
##################
# Defining loss function and evaluation metric
loss = cntk.squared_error(pred, label)
eval_fun = cntk.squared_error(pred, label)
######################
###### Training ######
######################
# Instantiate the trainer object to drive the model training
learning_rate = learning_rate_schedule(args.initial_learning_rate,
UnitType.minibatch)
optimizer_op = sgd(pred.parameters, learning_rate)
train_op = Trainer(pred, (loss, eval_fun), [optimizer_op])
for step in range(0, args.num_iterations):
for batch_num in range(0, num_minibatches_to_train):
batch_features = features[(batch_num * args.batch_size):(
batch_num * args.batch_size + args.batch_size), :]
batch_labels = predictions[(batch_num * args.batch_size):(
batch_num * args.batch_size + args.batch_size), :]
train_op.train_minibatch({input: batch_features, label: batch_labels})
training_loss = train_op.previous_minibatch_loss_average
eval_value = train_op.previous_minibatch_evaluation_average
print("Minibatch: {0}, Loss: {1:.2f}".format(batch_num, training_loss))
##############################
###### Model Evaluation ######
def test_learner_init_legacy():
i = C.input_variable(shape=(1,), needs_gradient=True, name='a')
w = parameter(shape=(1,))
res = i * w
# for backcompatibility test
# this will be deprecated in future version
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.sample))
assert learner._learning_rate_schedule.minibatch_size == 1 # the deprecated per sample schedule should not use compatible mode
assert learner.learning_rate() == 0.1
# for backcompatibility test
# this will be deprecated in future version
# The UnitType will provide per minibatch instruction for the learner
# this will be deprecated in future version
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.minibatch))
assert learner.is_compatible_mode() == False
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == 0
# for backcompatibility test, in reset learning rate, the learner won't receive the reference minibatch size from the schedule
# user will need to specify the reference minibatch size explicitly
# this will be deprecated in future version
learner = sgd(res.parameters, lr=0.1)
learner.reset_learning_rate(learning_rate_schedule([1, 2, 3], UnitType.minibatch))
assert learner.learning_rate() == 1.0
learner.minibatch_size = C.learners.IGNORE # reset to be per minibatch
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.is_compatible_mode() == True
# for backcompatibility test
# this will be deprecated in future version
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.sample), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE # the learner's reference minibatch size is still 0
# this will be deprecated in future version: This is logical invalid combination but it was the only way to use mean gradient and set learning rate in the past.
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.sample), use_mean_gradient=True)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
#test the override in the new version
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.minibatch_size == C.learners.IGNORE # the learner's reference minibatch size is still 0
# for backcompatibility test
# this will be deprecated in future version
# The UnitType will provide per minibatch instruction for the learner
# this will be deprecated in future version
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.minibatch), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
# for backcompatibility test, in reset learning rate, the learner won't receive the reference minibatch size from the schedule
# user will need to specify the reference minibatch size explicitly
# this will be deprecated in future version
learner = sgd(res.parameters, lr=0.1)
learner.reset_learning_rate(learning_rate_schedule([1, 2, 3], UnitType.minibatch))
assert learner.learning_rate() == 1.0
learner.minibatch_size = C.learners.IGNORE # reset to be per minibatch
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.is_compatible_mode() == True
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
# back compatible API test
momentum_time_constant = C.momentum_as_time_constant_schedule(1100)
lr_per_sample = learning_parameter_schedule(0.1, minibatch_size=1)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant, unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_rate_schedule([(3, 0.1), (2, 0.2), (1, 0.3)], unit=UnitType.sample)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant)
C.fsadagrad(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant, unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.5, 1.2, 0.7, 10, 1e-8]
lr_per_sample = learning_rate_schedule([0.1, 0.2], unit=UnitType.sample, epoch_size=100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.adadelta(res.parameters, lr_per_sample, use_mean_gradient=True)
def test_learner_init():
i = C.input_variable(shape=(1,), needs_gradient=True, name='a')
w = parameter(shape=(1,))
res = i * w
#test new API: learning_parameter_schedule
#explicitly specify reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=0.1, minibatch_size = 25)
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == 25 #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 25
assert learner.learning_rate() == 0.1
#no explicitly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1))
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1, 20), minibatch_size = 25)
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == 25 #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1, 20))
assert learner.is_compatible_mode() == False
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
#no explicitly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1))
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
#no explicitly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1, 20), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
#no explicitly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
mysgd = C.sgd(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert mysgd.minibatch_size == 32
assert mysgd._learning_rate_schedule.minibatch_size == 32
assert mysgd.learning_rate() == 0.4
mymomentum = C.momentum_sgd(parameters=res.parameters, lr=0.4, momentum=0.9, minibatch_size=32)
assert mymomentum.minibatch_size == 32
assert mymomentum._learning_rate_schedule.minibatch_size == 32
assert mymomentum.learning_rate() == 0.4
myadadelta = C.adadelta(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert myadadelta.minibatch_size == 32
assert myadadelta._learning_rate_schedule.minibatch_size == 32
assert myadadelta.learning_rate() == 0.4
myadam = C.adam(parameters=res.parameters, lr=0.4, momentum=0.9, variance_momentum=0.9, minibatch_size=32)
assert myadam.minibatch_size == 32
assert myadam._learning_rate_schedule.minibatch_size == 32
assert myadam.learning_rate() == 0.4
myadagrad = C.adagrad(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad.learning_rate() == 0.4
myfsadagrad = C.fsadagrad(parameters=res.parameters, lr=0.4, momentum=0.9, variance_momentum=0.9,
minibatch_size=32)
assert myfsadagrad.minibatch_size == 32
assert myfsadagrad._learning_rate_schedule.minibatch_size == 32
assert myfsadagrad.learning_rate() == 0.4
mynesterov = C.nesterov(parameters=res.parameters, lr=0.4, momentum=0.9, minibatch_size=32)
assert mynesterov.minibatch_size == 32
assert mynesterov._learning_rate_schedule.minibatch_size == 32
assert mynesterov.learning_rate() == 0.4
myrmsrop = C.rmsprop(parameters=res.parameters, lr=0.4, gamma=0.5, inc=1.2, dec=0.7, max=10, min=1e-8,
minibatch_size=32)
assert myrmsrop.minibatch_size == 32
assert myrmsrop._learning_rate_schedule.minibatch_size == 32
assert myrmsrop.learning_rate() == 0.4
mysgd = C.sgd(parameters=res.parameters, lr=[0.4, 0.1, 0.001], minibatch_size=32, epoch_size=512)
assert mysgd.minibatch_size == 32
assert mysgd._learning_rate_schedule.minibatch_size == 32
assert mysgd._learning_rate_schedule[0] == 0.4
assert mysgd._learning_rate_schedule[512] == 0.1
assert mysgd._learning_rate_schedule[512 * 2] == 0.001
mymomentum = C.momentum_sgd(parameters=res.parameters, lr=[0.4, 0.1, 0.001], momentum=[0.9],
minibatch_size=32, epoch_size=512)
assert mymomentum.minibatch_size == 32
assert mymomentum._learning_rate_schedule.minibatch_size == 32
assert mymomentum._learning_rate_schedule[0] == 0.4
assert mymomentum._learning_rate_schedule[512] == 0.1
assert mymomentum._learning_rate_schedule[512 * 2] == 0.001
myadadelta = C.adadelta(parameters=res.parameters, lr=[0.4, 0.1, 0.001],
minibatch_size=32, epoch_size=512)
assert myadadelta.minibatch_size == 32
assert myadadelta._learning_rate_schedule.minibatch_size == 32
assert myadadelta._learning_rate_schedule[0] == 0.4
assert myadadelta._learning_rate_schedule[512] == 0.1
assert myadadelta._learning_rate_schedule[512 * 2] == 0.001
myadam = C.adam(parameters=res.parameters, lr=[0.4, 0.1, 0.001], momentum=[0.9, 0.1, 0.001], variance_momentum=[0.9],
minibatch_size=32, epoch_size=512)
assert myadam.minibatch_size == 32
assert myadam._learning_rate_schedule.minibatch_size == 32
assert myadam._learning_rate_schedule[0] == 0.4
assert myadam._learning_rate_schedule[512] == 0.1
assert myadam._learning_rate_schedule[512 * 2] == 0.001
myadagrad = C.adagrad(parameters=res.parameters, lr=[0.4, 0.1, 0.001], minibatch_size=32, epoch_size=512)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad._learning_rate_schedule[0] == 0.4
assert myadagrad._learning_rate_schedule[512] == 0.1
assert myadagrad._learning_rate_schedule[512 * 2] == 0.001
myfsadagrad = C.fsadagrad(parameters=res.parameters, lr=[0.4, 0.1, 0.001], momentum=[0.9],
variance_momentum=[0.9],
minibatch_size=32, epoch_size=512)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad._learning_rate_schedule[0] == 0.4
assert myadagrad._learning_rate_schedule[512] == 0.1
assert myadagrad._learning_rate_schedule[512 * 2] == 0.001
mynesterov = C.nesterov(parameters=res.parameters, lr=[0.4, 0.1, 0.001], momentum=[0.9],
minibatch_size=32, epoch_size=512)
assert mynesterov.minibatch_size == 32
assert mynesterov._learning_rate_schedule.minibatch_size == 32
assert mynesterov._learning_rate_schedule[0] == 0.4
assert mynesterov._learning_rate_schedule[512] == 0.1
assert mynesterov._learning_rate_schedule[512 * 2] == 0.001
myrmsrop = C.rmsprop(parameters=res.parameters, lr=[0.4, 0.1, 0.001], gamma=0.5, inc=1.2, dec=0.7, max=10,
min=1e-8,
minibatch_size=32, epoch_size=512)
assert myrmsrop.minibatch_size == 32
assert myrmsrop._learning_rate_schedule.minibatch_size == 32
assert myrmsrop._learning_rate_schedule[0] == 0.4
assert myrmsrop._learning_rate_schedule[512] == 0.1
assert myrmsrop._learning_rate_schedule[512 * 2] == 0.001
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
momentum = C.momentum_schedule(0.999, minibatch_size=1)
lr_per_sample = learning_parameter_schedule(0.1, minibatch_size = 1)
C.momentum_sgd(res.parameters, lr_per_sample, momentum)
C.momentum_sgd(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.momentum_sgd(res.parameters, lr_per_sample, momentum, unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
lr_per_sample = learning_parameter_schedule([0.1, 0.2], minibatch_size = 1)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum)
C.nesterov(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum, unit_gain=unit_gain_value)
lr_per_sample = learning_parameter_schedule([0.1]*3 +[0.2]*2 +[0.3], minibatch_size=1)
C.adagrad(res.parameters, lr=lr_per_sample, need_ave_multiplier=True)
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_parameter_schedule([(3,0.1), (2, 0.2), (1, 0.3)], minibatch_size=1)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum)
C.fsadagrad(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum, unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.5, 1.2, 0.7, 10, 1e-8]
lr_per_sample = learning_parameter_schedule([0.1, 0.2], minibatch_size = 1, epoch_size = 100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.adadelta(res.parameters, lr_per_sample)
def test_learner_init():
i = C.input_variable(shape=(1, ), needs_gradient=True, name='a')
w = parameter(shape=(1, ))
res = i * w
#test new API: learning_parameter_schedule
#explictly specify reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=0.1, minibatch_size=25)
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == 25 #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 25
assert learner.learning_rate() == 0.1
#no explictly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1))
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters,
lr=learning_parameter_schedule(0.1, 20),
minibatch_size=25)
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == 25 #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1, 20))
assert learner.is_compatible_mode() == False
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
#no explictly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1))
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
#no explictly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters,
lr=learning_parameter_schedule(0.1),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters,
lr=learning_parameter_schedule(0.1, 20),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
#no explictly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters,
lr=learning_parameter_schedule(0.1),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
mysgd = C.sgd(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert mysgd.minibatch_size == 32
assert mysgd._learning_rate_schedule.minibatch_size == 32
assert mysgd.learning_rate() == 0.4
mymomentum = C.momentum_sgd(parameters=res.parameters,
lr=0.4,
momentum=0.9,
minibatch_size=32)
assert mymomentum.minibatch_size == 32
assert mymomentum._learning_rate_schedule.minibatch_size == 32
assert mymomentum.learning_rate() == 0.4
myadadelta = C.adadelta(parameters=res.parameters,
lr=0.4,
minibatch_size=32)
assert myadadelta.minibatch_size == 32
assert myadadelta._learning_rate_schedule.minibatch_size == 32
assert myadadelta.learning_rate() == 0.4
myadam = C.adam(parameters=res.parameters,
lr=0.4,
momentum=0.9,
variance_momentum=0.9,
minibatch_size=32)
assert myadam.minibatch_size == 32
assert myadam._learning_rate_schedule.minibatch_size == 32
assert myadam.learning_rate() == 0.4
myadagrad = C.adagrad(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad.learning_rate() == 0.4
myfsadagrad = C.fsadagrad(parameters=res.parameters,
lr=0.4,
momentum=0.9,
variance_momentum=0.9,
minibatch_size=32)
assert myfsadagrad.minibatch_size == 32
assert myfsadagrad._learning_rate_schedule.minibatch_size == 32
assert myfsadagrad.learning_rate() == 0.4
mynesterov = C.nesterov(parameters=res.parameters,
lr=0.4,
momentum=0.9,
minibatch_size=32)
assert mynesterov.minibatch_size == 32
assert mynesterov._learning_rate_schedule.minibatch_size == 32
assert mynesterov.learning_rate() == 0.4
myrmsrop = C.rmsprop(parameters=res.parameters,
lr=0.4,
gamma=0.5,
inc=1.2,
dec=0.7,
max=10,
min=1e-8,
minibatch_size=32)
assert myrmsrop.minibatch_size == 32
assert myrmsrop._learning_rate_schedule.minibatch_size == 32
assert myrmsrop.learning_rate() == 0.4
mysgd = C.sgd(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
minibatch_size=32,
epoch_size=512)
assert mysgd.minibatch_size == 32
assert mysgd._learning_rate_schedule.minibatch_size == 32
assert mysgd._learning_rate_schedule[0] == 0.4
assert mysgd._learning_rate_schedule[512] == 0.1
assert mysgd._learning_rate_schedule[512 * 2] == 0.001
mymomentum = C.momentum_sgd(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
momentum=[0.9],
minibatch_size=32,
epoch_size=512)
assert mymomentum.minibatch_size == 32
assert mymomentum._learning_rate_schedule.minibatch_size == 32
assert mymomentum._learning_rate_schedule[0] == 0.4
assert mymomentum._learning_rate_schedule[512] == 0.1
assert mymomentum._learning_rate_schedule[512 * 2] == 0.001
myadadelta = C.adadelta(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
minibatch_size=32,
epoch_size=512)
assert myadadelta.minibatch_size == 32
assert myadadelta._learning_rate_schedule.minibatch_size == 32
assert myadadelta._learning_rate_schedule[0] == 0.4
assert myadadelta._learning_rate_schedule[512] == 0.1
assert myadadelta._learning_rate_schedule[512 * 2] == 0.001
myadam = C.adam(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
momentum=[0.9, 0.1, 0.001],
variance_momentum=[0.9],
minibatch_size=32,
epoch_size=512)
assert myadam.minibatch_size == 32
assert myadam._learning_rate_schedule.minibatch_size == 32
assert myadam._learning_rate_schedule[0] == 0.4
assert myadam._learning_rate_schedule[512] == 0.1
assert myadam._learning_rate_schedule[512 * 2] == 0.001
myadagrad = C.adagrad(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
minibatch_size=32,
epoch_size=512)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad._learning_rate_schedule[0] == 0.4
assert myadagrad._learning_rate_schedule[512] == 0.1
assert myadagrad._learning_rate_schedule[512 * 2] == 0.001
myfsadagrad = C.fsadagrad(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
momentum=[0.9],
variance_momentum=[0.9],
minibatch_size=32,
epoch_size=512)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad._learning_rate_schedule[0] == 0.4
assert myadagrad._learning_rate_schedule[512] == 0.1
assert myadagrad._learning_rate_schedule[512 * 2] == 0.001
mynesterov = C.nesterov(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
momentum=[0.9],
minibatch_size=32,
epoch_size=512)
assert mynesterov.minibatch_size == 32
assert mynesterov._learning_rate_schedule.minibatch_size == 32
assert mynesterov._learning_rate_schedule[0] == 0.4
assert mynesterov._learning_rate_schedule[512] == 0.1
assert mynesterov._learning_rate_schedule[512 * 2] == 0.001
myrmsrop = C.rmsprop(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
gamma=0.5,
inc=1.2,
dec=0.7,
max=10,
min=1e-8,
minibatch_size=32,
epoch_size=512)
assert myrmsrop.minibatch_size == 32
assert myrmsrop._learning_rate_schedule.minibatch_size == 32
assert myrmsrop._learning_rate_schedule[0] == 0.4
assert myrmsrop._learning_rate_schedule[512] == 0.1
assert myrmsrop._learning_rate_schedule[512 * 2] == 0.001
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
momentum = C.momentum_schedule(0.999, minibatch_size=1)
lr_per_sample = learning_parameter_schedule(0.1, minibatch_size=1)
C.momentum_sgd(res.parameters, lr_per_sample, momentum)
C.momentum_sgd(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.momentum_sgd(res.parameters,
lr_per_sample,
momentum,
unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
lr_per_sample = learning_parameter_schedule([0.1, 0.2], minibatch_size=1)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum)
C.nesterov(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.nesterov(res.parameters,
lr=lr_per_sample,
momentum=momentum,
unit_gain=unit_gain_value)
lr_per_sample = learning_parameter_schedule([0.1] * 3 + [0.2] * 2 + [0.3],
minibatch_size=1)
C.adagrad(res.parameters, lr=lr_per_sample, need_ave_multiplier=True)
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_parameter_schedule([(3, 0.1), (2, 0.2), (1, 0.3)],
minibatch_size=1)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum)
C.fsadagrad(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.fsadagrad(res.parameters,
lr=lr_per_sample,
momentum=momentum,
unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.5, 1.2, 0.7, 10, 1e-8]
lr_per_sample = learning_parameter_schedule([0.1, 0.2],
minibatch_size=1,
epoch_size=100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.adadelta(res.parameters, lr_per_sample)
init = C.initializer.normal(0.01)
with C.layers.default_options(init=init):
z = C.layers.Sequential(
[C.layers.Dense(12, activation=C.relu),
C.layers.Dense(3)])
y = C.cross_entropy_with_softmax(z(x), t)
acc = C.classification_error(z(x), t)
batch_size = 20
from cntk.learners import sgd, learning_parameter_schedule
lr = learning_parameter_schedule([.5 * (.1**i) for i in range(10000)],
minibatch_size=batch_size,
epoch_size=1000 * batch_size)
learner = sgd(z.parameters, lr)
trainer = C.Trainer(z(x), (y, acc), [learner])
for i in range(min(dataset_size, 100000) // batch_size):
sample = X[batch_size * i:batch_size * (i + 1)]
target = labels[batch_size * i:batch_size * (i + 1)]
trainer.train_minibatch({x: sample, t: target})
loss = trainer.previous_minibatch_loss_average
acc = trainer.previous_minibatch_evaluation_average
print("cost {} - classification error {} - learning rate {}".format(
loss, acc, learner.learning_rate()))
y = C.argmax(z(x))
accuracy = 0
for i in range(1000):
sample = X[batch_size * i:batch_size * (i + 1)]
def test_learner_init():
i = C.input_variable(shape=(1,), needs_gradient=True, name='a')
w = parameter(shape=(1,))
res = i * w
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.sample))
assert learner.learning_rate() == 0.1
learner.reset_learning_rate(learning_rate_schedule([1,2,3], UnitType.minibatch));
assert learner.learning_rate() == 1.0
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
momentum_time_constant = C.momentum_as_time_constant_schedule(1100)
lr_per_sample = learning_rate_schedule(0.1, UnitType.sample)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant, unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
lr_per_sample = learning_rate_schedule([0.1, 0.2], UnitType.sample)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant)
C.nesterov(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant, unit_gain=unit_gain_value)
lr_per_sample = learning_rate_schedule([0.1]*3 +[0.2]*2 +[0.3], UnitType.sample)
C.adagrad(res.parameters, lr=lr_per_sample, need_ave_multiplier=True)
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_rate_schedule([(3,0.1), (2, 0.2), (1, 0.3)], UnitType.sample)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant)
C.fsadagrad(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant, unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.1]*5
lr_per_sample = learning_rate_schedule([0.1, 0.2], UnitType.sample, 100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.set_default_use_mean_gradient_value(False)
use_mean_gradient_value = C.default_use_mean_gradient_value()
assert not use_mean_gradient_value
C.adadelta(res.parameters, lr_per_sample)
C.set_default_use_mean_gradient_value(True)
use_mean_gradient_value = C.default_use_mean_gradient_value()
assert use_mean_gradient_value
C.adadelta(res.parameters, lr_per_sample)
def test_sweep_based_schedule(tmpdir, device_id):
from cntk.io import MinibatchSource, CTFDeserializer, StreamDef, StreamDefs
from cntk import cross_entropy_with_softmax, classification_error, plus, reduce_sum, sequence
from cntk import Trainer
input_dim = 69
ctf_data = '''\
0 |S0 3:1 |S1 3:1 |# <s>
0 |S0 4:1 |# A |S1 32:1 |# ~AH
0 |S0 5:1 |# B |S1 36:1 |# ~B
0 |S0 4:1 |# A |S1 31:1 |# ~AE
0 |S0 7:1 |# D |S1 38:1 |# ~D
0 |S0 12:1 |# I |S1 47:1 |# ~IY
0 |S0 1:1 |# </s> |S1 1:1 |# </s>
2 |S0 60:1 |# <s> |S1 3:1 |# <s>
2 |S0 61:1 |# A |S1 32:1 |# ~AH
'''
ctf_file = str(tmpdir/'2seqtest.txt')
with open(ctf_file, 'w') as f:
f.write(ctf_data)
mbs = MinibatchSource(CTFDeserializer(ctf_file, StreamDefs(
features = StreamDef(field='S0', shape=input_dim, is_sparse=True),
labels = StreamDef(field='S1', shape=input_dim, is_sparse=True)
)), randomize=False)
in1 = sequence.input_variable(shape=(input_dim,))
labels = sequence.input_variable(shape=(input_dim,))
p = parameter(shape=(input_dim,), init=10)
z = plus(in1, reduce_sum(p), name='z')
ce = cross_entropy_with_softmax(z, labels)
errs = classification_error(z, labels)
lr_per_sample = learning_rate_schedule([0.3, 0.2, 0.1, 0.0], UnitType.sample)
learner = sgd(z.parameters, lr_per_sample)
trainer = Trainer(z, (ce, errs), [learner])
input_map = {
in1 : mbs.streams.features,
labels : mbs.streams.labels
}
# fetch minibatch (first sequence)
data = mbs.next_minibatch(1, input_map=input_map)
trainer.train_minibatch(data)
assert learner.learning_rate() == 0.3
# fetch minibatch (second sequence, sweep ends at this point)
data = mbs.next_minibatch(1, input_map=input_map)
trainer.train_minibatch(data)
assert learner.learning_rate() == 0.2
# fetch minibatch (both sequences -- entire sweep in one go)
data = mbs.next_minibatch(9, input_map=input_map)
trainer.train_minibatch(data)
assert learner.learning_rate() == 0.1
# fetch minibatch (multiple sweeps)
data = mbs.next_minibatch(30, input_map=input_map)
trainer.train_minibatch(data, outputs=[z.output])
assert learner.learning_rate() == 0.0
feature : training_reader.streams.features,
#xy : training_reader.streams.xy
}
"""
Set loss and evaluation functions
"""
loss = squared_error(netout, label)
evaluation = squared_error(netout, label)
lr_per_minibatch=learning_rate_schedule(learning_rate, UnitType.minibatch)
lr_fine_tuning=learning_rate_schedule(learning_rate_fine_tuning, UnitType.minibatch)
"""
Instantiate the trainer object to drive the model training
See: https://www.cntk.ai/pythondocs/cntk.learners.html
"""
learner = sgd(netout.parameters, lr=lr_per_minibatch)
learner_fine_tuning = sgd(netout.parameters, lr=lr_fine_tuning)
# Other learners to try
#learner = momentum_sgd(netout.parameters, lr=lr_per_minibatch, momentum = momentum_schedule(0.9))
#learner = adagrad(netout.parameters, lr=lr_per_minibatch)
progress_printer = ProgressPrinter(100)
"""
Instantiate the trainer
See: https://www.cntk.ai/pythondocs/cntk.train.html#module-cntk.train.trainer
"""
trainer = Trainer(netout, (loss, evaluation), learner, progress_printer)
trainer_fine_tune = Trainer(netout, (loss, evaluation), learner_fine_tuning, progress_printer)
#%%
def test_learner_init_legacy():
i = C.input_variable(shape=(1, ), needs_gradient=True, name='a')
w = parameter(shape=(1, ))
res = i * w
# for backcompatibility test
# this will be deprecated in future version
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.sample))
assert learner._learning_rate_schedule.minibatch_size == 1 # the deprecated per sample schedule should not use compatible mode
assert learner.learning_rate() == 0.1
# for backcompatibility test
# this will be deprecated in future version
# The UnitType will provide per minibatch instruction for the learner
# this will be deprecated in future version
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.minibatch))
assert learner.is_compatible_mode() == False
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == 0
# for backcompatibility test, in reset learning rate, the learner won't receive the reference minibatch size from the schedule
# user will need to specify the reference minibatch size explicitly
# this will be deprecated in future version
learner = sgd(res.parameters, lr=0.1)
learner.reset_learning_rate(
learning_rate_schedule([1, 2, 3], UnitType.minibatch))
assert learner.learning_rate() == 1.0
learner.minibatch_size = C.learners.IGNORE # reset to be per minibatch
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.is_compatible_mode() == True
# for backcompatibility test
# this will be deprecated in future version
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.sample),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE # the learner's reference minibatch size is still 0
# this will be deprecated in future version: This is logical invalid combination but it was the only way to use mean gradient and set learning rate in the past.
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.sample),
use_mean_gradient=True)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
#test the override in the new version
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.minibatch_size == C.learners.IGNORE # the learner's reference minibatch size is still 0
# for backcompatibility test
# this will be deprecated in future version
# The UnitType will provide per minibatch instruction for the learner
# this will be deprecated in future version
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.minibatch),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
# for backcompatibility test, in reset learning rate, the learner won't receive the reference minibatch size from the schedule
# user will need to specify the reference minibatch size explicitly
# this will be deprecated in future version
learner = sgd(res.parameters, lr=0.1)
learner.reset_learning_rate(
learning_rate_schedule([1, 2, 3], UnitType.minibatch))
assert learner.learning_rate() == 1.0
learner.minibatch_size = C.learners.IGNORE # reset to be per minibatch
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.is_compatible_mode() == True
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
# back compatible API test
momentum_time_constant = C.momentum_as_time_constant_schedule(1100)
lr_per_sample = learning_parameter_schedule(0.1, minibatch_size=1)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant,
unit_gain_value)
C.momentum_sgd(res.parameters,
lr_per_sample,
momentum_time_constant,
unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_rate_schedule([(3, 0.1), (2, 0.2), (1, 0.3)],
unit=UnitType.sample)
C.fsadagrad(res.parameters,
lr=lr_per_sample,
momentum=momentum_time_constant)
C.fsadagrad(res.parameters, lr_per_sample, momentum_time_constant,
unit_gain_value)
C.fsadagrad(res.parameters,
lr=lr_per_sample,
momentum=momentum_time_constant,
unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.5, 1.2, 0.7, 10, 1e-8]
lr_per_sample = learning_rate_schedule([0.1, 0.2],
unit=UnitType.sample,
epoch_size=100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.adadelta(res.parameters, lr_per_sample, use_mean_gradient=True)
label = C.input_variable((n_output))
# ネットワークを構築する。
model = Sequential([
Dense(n_hidden, activation=C.relu),
Dense(n_hidden, activation=C.relu),
Dense(n_hidden, activation=C.relu),
Dense(n_output)
])(features)
ce = C.cross_entropy_with_softmax(model, label)
pe = C.classification_error(model, label)
minibatch = C.learning_parameter_schedule(0.125)
progress_printer = ProgressPrinter(0)
trainer = C.Trainer(model, (ce, pe), [sgd(model.parameters, lr=minibatch)],
[progress_printer])
# 学習する。
n_epoch = 30
n_batchsize = 16
for epoch in range(n_epoch):
order = np.random.permutation(range(len(x_train)))
aggregate_loss = 0.0
for i in range(0, len(order), n_batchsize):
index = order[i:i + n_batchsize]
x_train_batch = x_train[index, :]
_t_train_batch = t_train[index]
