def train(reader, model, max_epochs):
# Input variables denoting the features and label data
query = Input(input_dim, is_sparse=False)
slot_labels = Input(num_labels, is_sparse=True) # TODO: make sparse once it works
# apply model to input
z = model(query)
# loss and metric
ce = cross_entropy_with_softmax(z, slot_labels)
pe = classification_error (z, slot_labels)
# training config
epoch_size = 36000
minibatch_size = 70
num_mbs_to_show_result = 100
momentum_time_constant = momentum_as_time_constant_schedule(minibatch_size / -math.log(0.9)) # TODO: Change to round number. This is 664.39. 700?
lr_schedule = [0.003]*2+[0.0015]*12+[0.0003] # LR schedule over epochs (we don't run that many epochs, but if we did, these are good values)
# trainer object
lr_per_sample = learning_rate_schedule(lr_schedule, UnitType.sample, epoch_size)
learner = adam_sgd(z.parameters,
lr=lr_per_sample, momentum=momentum_time_constant,
unit_gain=True,
low_memory=True,
gradient_clipping_threshold_per_sample=15, gradient_clipping_with_truncation=True)
trainer = Trainer(z, (ce, pe), [learner])
# define mapping from reader streams to network inputs
input_map = {
query : reader.streams.query,
slot_labels : reader.streams.slot_labels
}
# process minibatches and perform model training
log_number_of_parameters(z) ; print()
# more detailed logging
progress_printer = ProgressPrinter(freq=100, first=10, tag='Training', tensorboard_log_dir='atis_log', model=z)
#progress_printer = ProgressPrinter(tag='Training')
t = 0
for epoch in range(max_epochs): # loop over epochs
epoch_end = (epoch+1) * epoch_size
while t < epoch_end: # loop over minibatches on the epoch
# BUGBUG? The change of minibatch_size parameter vv has no effect.
data = reader.next_minibatch(min(minibatch_size, epoch_end-t), input_map=input_map) # fetch minibatch
trainer.train_minibatch(data) # update model with it
t += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
#def trace_node(name):
# nl = [n for n in z.parameters if n.name() == name]
# if len(nl) > 0:
# print (name, np.asarray(nl[0].value))
#trace_node('W')
#trace_node('stabilizer_param')
loss, metric, actual_samples = progress_printer.epoch_summary(with_metric=True)
return loss, metric
def create_resnet_network(network_name):
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# create model, and configure learning parameters
if network_name == 'resnet20':
z = create_cifar10_model(input_var, 3, num_classes)
elif network_name == 'resnet110':
z = create_cifar10_model(input_var, 18, num_classes)
else:
return RuntimeError("Unknown model name!")
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
return {
'name' : network_name,
'feature': input_var,
'label': label_var,
'ce' : ce,
'pe' : pe,
'output': z
}
def criterion(input, labels):
# criterion function must drop the <s> from the labels
postprocessed_labels = sequence.slice(labels, 1, 0) # <s> A B C </s> --> A B C </s>
z = model(input, postprocessed_labels)
ce = cross_entropy_with_softmax(z, postprocessed_labels)
errs = classification_error (z, postprocessed_labels)
return (ce, errs)
def train_lm(training_file, epochs, max_num_minibatches):
# load the data and vocab
data, char_to_ix, ix_to_char, data_size, vocab_dim = load_data_and_vocab(training_file)
# Model the source and target inputs to the model
input_sequence, label_sequence = create_inputs(vocab_dim)
# create the model
model = create_model(vocab_dim)
# and apply it to the input sequence
z = model(input_sequence)
# setup the criterions (loss and metric)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# Instantiate the trainer object to drive the model training
lr_per_sample = learning_rate_schedule(0.001, UnitType.sample)
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 5.0
gradient_clipping_with_truncation = True
learner = momentum_sgd(z.parameters, lr_per_sample, momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
progress_printer = ProgressPrinter(freq=100, tag='Training')
trainer = Trainer(z, (ce, errs), learner, progress_printer)
sample_freq = 1000
minibatches_per_epoch = min(data_size // minibatch_size, max_num_minibatches // epochs)
# print out some useful training information
log_number_of_parameters(z)
print ("Running %d epochs with %d minibatches per epoch" % (epochs, minibatches_per_epoch))
print()
for e in range(0, epochs):
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
# If it's the start of the data, we specify that we are looking at a new sequence (True)
mask = [True]
for b in range(0, minibatches_per_epoch):
# get the data
features, labels = get_data(b, minibatch_size, data, char_to_ix, vocab_dim)
arguments = ({input_sequence : features, label_sequence : labels}, mask)
mask = [False]
trainer.train_minibatch(arguments)
global_minibatch = e*minibatches_per_epoch + b
if global_minibatch % sample_freq == 0:
print(sample(z, ix_to_char, vocab_dim, char_to_ix))
model_filename = "models/shakespeare_epoch%d.dnn" % (e+1)
z.save_model(model_filename)
print("Saved model to '%s'" % model_filename)
def train_fast_rcnn(debug_output=False):
if debug_output:
print("Storing graphs and intermediate models to %s." % os.path.join(abs_path, "Output"))
# Create the minibatch source
minibatch_source = create_mb_source(image_height, image_width, num_channels,
num_classes, num_rois, base_path, "train")
# Input variables denoting features, rois and label data
image_input = input_variable((num_channels, image_height, image_width))
roi_input = input_variable((num_rois, 4))
label_input = input_variable((num_rois, num_classes))
# define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source[features_stream_name],
roi_input: minibatch_source[roi_stream_name],
label_input: minibatch_source[label_stream_name]
}
# Instantiate the Fast R-CNN prediction model and loss function
frcn_output = frcn_predictor(image_input, roi_input, num_classes)
ce = cross_entropy_with_softmax(frcn_output, label_input, axis=1)
pe = classification_error(frcn_output, label_input, axis=1)
if debug_output:
plot(frcn_output, os.path.join(abs_path, "Output", "graph_frcn.png"))
# Set learning parameters
l2_reg_weight = 0.0005
lr_per_sample = [0.00001] * 10 + [0.000001] * 5 + [0.0000001]
lr_schedule = learning_rate_schedule(lr_per_sample, unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# Instantiate the trainer object
learner = momentum_sgd(frcn_output.parameters, lr_schedule, mm_schedule, l2_regularization_weight=l2_reg_weight)
trainer = Trainer(frcn_output, (ce, pe), learner)
# Get minibatches of images and perform model training
print("Training Fast R-CNN model for %s epochs." % max_epochs)
log_number_of_parameters(frcn_output)
progress_printer = ProgressPrinter(tag='Training', num_epochs=max_epochs)
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(mb_size, epoch_size-sample_count), input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
if debug_output:
frcn_output.save(os.path.join(abs_path, "Output", "frcn_py_%s.model" % (epoch+1)))
return frcn_output
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 = input_variable(shape=input_dim, is_sparse=True)
label = input_variable(num_output_classes, dynamic_axes=[
Axis.default_batch_axis()])
# 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 simple_mnist():
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)
netout = fully_connected_classifier_net(scaled_input, num_output_classes, hidden_layers_dim, num_hidden_layers, sigmoid)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
rel_path = os.path.join(*"../../../../Examples/Image/MNIST/Data/Train-28x28_cntk_text.txt".split("/"))
path = os.path.normpath(os.path.join(abs_path, rel_path))
if not os.path.exists(path):
readme_file = os.path.normpath(os.path.join(os.path.dirname(path), "..", "README.md"))
raise RuntimeError("File '%s' does not exist. Please follow the instructions at %s to download and prepare it."%(path, readme_file))
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration( feature_stream_name, input_dim ),
StreamConfiguration( labels_stream_name, num_output_classes) ])
features_si = mb_source.stream_info(feature_stream_name)
labels_si = mb_source.stream_info(labels_stream_name)
# Instantiate the trainer object to drive the model training
lr = learning_rates_per_sample(0.003125)
trainer = Trainer(netout, ce, pe, [sgd_learner(netout.owner.parameters(), lr)])
# Get minibatches of images to train with and perform model training
minibatch_size = 32
num_samples_per_sweep = 60000
num_sweeps_to_train_with = 1
num_minibatches_to_train = (num_samples_per_sweep * num_sweeps_to_train_with) / minibatch_size
training_progress_output_freq = 20
for i in range(0, int(num_minibatches_to_train)):
mb = mb_source.get_next_minibatch(minibatch_size)
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
arguments = {input : mb[features_si].m_data, label : mb[labels_si].m_data}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
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 = input_variable(shape=input_dim, is_sparse=True)
label = input_variable(num_output_classes, dynamic_axes = [Axis.default_batch_axis()])
# 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)
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration( feature_stream_name, input_dim, True, 'x' ),
StreamConfiguration( labels_stream_name, num_output_classes, False, 'y')], 0)
features_si = mb_source.stream_info(features)
labels_si = mb_source.stream_info(label)
# Instantiate the trainer object to drive the model training
lr = lr = learning_rates_per_sample(0.0005)
trainer = Trainer(classifier_output, ce, pe, [sgd_learner(classifier_output.owner.parameters(), lr)])
# Get minibatches of sequences to train with and perform model training
minibatch_size = 200
training_progress_output_freq = 10
i = 0;
while True:
mb = mb_source.get_next_minibatch(minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
arguments = {features : mb[features_si].m_data, label : mb[labels_si].m_data}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
i += 1
def train_model(base_model_file, feature_node_name, last_hidden_node_name,
image_width, image_height, num_channels, num_classes, train_map_file,
num_epochs, max_images=-1, freeze=False):
epoch_size = sum(1 for line in open(train_map_file))
if max_images > 0:
epoch_size = min(epoch_size, max_images)
# Create the minibatch source and input variables
minibatch_source = create_mb_source(train_map_file, image_width, image_height, num_channels, num_classes)
image_input = input_variable((num_channels, image_height, image_width))
label_input = input_variable(num_classes)
# Define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source[features_stream_name],
label_input: minibatch_source[label_stream_name]
}
# Instantiate the transfer learning model and loss function
tl_model = create_model(base_model_file, feature_node_name, last_hidden_node_name, num_classes, image_input, freeze)
ce = cross_entropy_with_softmax(tl_model, label_input)
pe = classification_error(tl_model, label_input)
# Instantiate the trainer object
lr_schedule = learning_rate_schedule(lr_per_mb, unit=UnitType.minibatch)
mm_schedule = momentum_schedule(momentum_per_mb)
learner = momentum_sgd(tl_model.parameters, lr_schedule, mm_schedule, l2_regularization_weight=l2_reg_weight)
trainer = Trainer(tl_model, (ce, pe), learner)
# Get minibatches of images and perform model training
print("Training transfer learning model for {0} epochs (epoch_size = {1}).".format(num_epochs, epoch_size))
log_number_of_parameters(tl_model)
progress_printer = ProgressPrinter(tag='Training', num_epochs=num_epochs)
for epoch in range(num_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(mb_size, epoch_size-sample_count), input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
if sample_count % (100 * mb_size) == 0:
print ("Processed {0} samples".format(sample_count))
progress_printer.epoch_summary(with_metric=True)
return tl_model
def ffnet(debug_output=False):
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)
# Instantiate the trainer object to drive the model training
trainer = Trainer(netout, ce, pe, [sgd(netout.parameters(), lr=0.02)])
# Get minibatches of training data and perform model training
minibatch_size = 25
num_samples_per_sweep = 10000
num_sweeps_to_train_with = 2
num_minibatches_to_train = (
num_samples_per_sweep * num_sweeps_to_train_with) / minibatch_size
training_progress_output_freq = 60
if debug_output:
training_progress_output_freq = training_progress_output_freq/3
for i in range(0, int(num_minibatches_to_train)):
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})
print_training_progress(trainer, i, training_progress_output_freq)
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 create_recurrent_network():
# Input variables denoting the features and label data
features = input_variable(((2*context+1)*feature_dim))
labels = input_variable((num_classes))
# create network
model = Sequential([For(range(3), lambda : Recurrence(LSTM(256))),
Dense(num_classes)])
z = model(features)
ce = cross_entropy_with_softmax(z, labels)
errs = classification_error (z, labels)
return {
'feature': features,
'label': labels,
'ce' : ce,
'errs' : errs,
'output': z
}
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():
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
trainer = Trainer(netout, (ce, pe), sgd(netout.parameters, lr=lr_per_minibatch))
# Get minibatches of training data and perform model training
minibatch_size = 25
pp = ProgressPrinter(128)
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})
pp.update_with_trainer(trainer)
pp.epoch_summary()
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 cifar_resnet():
image_height = 32
image_width = 32
num_channels = 3
num_classes = 10
feats_stream_name = 'features'
labels_stream_name = 'labels'
minibatch_source = create_mb_source(feats_stream_name, labels_stream_name,
image_height, image_width, num_channels, num_classes)
features_si = minibatch_source.stream_info(feats_stream_name)
labels_si = minibatch_source.stream_info(labels_stream_name)
# Input variables denoting the features and label data
image_input = input_variable((num_channels, image_height, image_width), features_si.m_element_type)
label_var = input_variable((num_classes), features_si.m_element_type)
# Instantiate the resnet classification model
classifier_output = resnet_classifer(image_input, num_classes)
ce = cross_entropy_with_softmax(classifier_output, label_var)
pe = classification_error(classifier_output, label_var)
# Instantiate the trainer object to drive the model training
lr = learning_rates_per_sample(0.0078125)
trainer = Trainer(classifier_output, ce, pe, [sgd_learner(classifier_output.owner.parameters(), lr)])
# Get minibatches of images to train with and perform model training
mb_size = 32
training_progress_output_freq = 20
num_mbs = 1000
for i in range(0, num_mbs):
mb=minibatch_source.get_next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
arguments = {image_input : mb[features_si].m_data, label_var : mb[labels_si].m_data}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
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)
# Instantiate the trainer object to drive the model training
lr = learning_rates_per_sample(0.02)
trainer = Trainer(netout, ce, pe,
[sgd_learner(netout.owner.parameters(), lr)])
# Get minibatches of training data and perform model training
minibatch_size = 25
num_samples_per_sweep = 10000
num_sweeps_to_train_with = 2
num_minibatches_to_train = (num_samples_per_sweep *
num_sweeps_to_train_with) / minibatch_size
training_progress_output_freq = 20
for i in range(0, int(num_minibatches_to_train)):
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})
print_training_progress(trainer, i, training_progress_output_freq)
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 = input_variable(shape=input_dim, is_sparse=True)
label = input_variable(num_output_classes,
dynamic_axes=[Axis.default_batch_axis()])
# 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)
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration(feature_stream_name, input_dim, True, 'x'),
StreamConfiguration(labels_stream_name, num_output_classes, False, 'y')
], 0)
features_si = mb_source.stream_info(features)
labels_si = mb_source.stream_info(label)
# Instantiate the trainer object to drive the model training
trainer = Trainer(classifier_output, ce, pe,
[sgd(classifier_output.parameters(), lr=0.0005)])
# Get minibatches of sequences to train with and perform model training
minibatch_size = 200
training_progress_output_freq = 10
i = 0
while True:
mb = mb_source.get_next_minibatch(minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {
features: mb[features_si].m_data,
label: mb[labels_si].m_data
}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
i += 1
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 simple_mnist():
input_dim = 784
num_output_classes = 10
num_hidden_layers = 1
hidden_layers_dim = 200
# Input variables denoting the features and label data
features = 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), features)
netout = fully_connected_classifier_net(
scaled_input, num_output_classes, hidden_layers_dim, num_hidden_layers, relu)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
try:
rel_path = os.path.join(os.environ['CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY'],
*"Image/MNIST/v0/Train-28x28_cntk_text.txt".split("/"))
except KeyError:
rel_path = os.path.join(*"../Image/DataSets/MNIST/Train-28x28_cntk_text.txt".split("/"))
path = os.path.normpath(os.path.join(abs_path, rel_path))
check_path(path)
reader_train = create_reader(path, True, input_dim, num_output_classes)
input_map = {
features: reader_train.streams.features,
label: reader_train.streams.labels
}
# Instantiate progress writers.
logdir = os.path.join(os.path.dirname(os.path.abspath(__file__)), "mnist_log")
tensorboard_writer = TensorBoardProgressWriter(freq=1, log_dir=logdir, model=netout)
progress_printer = ProgressPrinter(freq=10, tag='Training')
# Instantiate the trainer object to drive the model training
lr_per_minibatch = learning_rate_schedule(0.2, UnitType.minibatch)
learner = sgd(netout.parameters, lr=lr_per_minibatch)
trainer = Trainer(netout, (ce, pe), learner, [tensorboard_writer, progress_printer])
# Get minibatches of images to train with and perform model training
minibatch_size = 64
num_samples_per_sweep = 6000
num_sweeps_to_train_with = 2
num_minibatches_to_train = (num_samples_per_sweep * num_sweeps_to_train_with) / minibatch_size
for minibatch_idx in range(0, int(num_minibatches_to_train)):
trainer.train_minibatch(reader_train.next_minibatch(minibatch_size, input_map=input_map))
# Log max/min/mean of each parameter tensor, so that we can confirm that the parameters change indeed.
# Don't want to do that very often though, otherwise will spend too much time computing min/max/mean.
if minibatch_idx % 10 == 9:
for p in netout.parameters:
tensorboard_writer.write_value(p.uid + "/max", reduce_max(p).eval(), minibatch_idx)
tensorboard_writer.write_value(p.uid + "/min", reduce_min(p).eval(), minibatch_idx)
tensorboard_writer.write_value(p.uid + "/mean", reduce_mean(p).eval(), minibatch_idx)
trainer.summarize_training_progress()
# Load test data
try:
rel_path = os.path.join(os.environ['CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY'],
*"Image/MNIST/v0/Test-28x28_cntk_text.txt".split("/"))
except KeyError:
rel_path = os.path.join(*"../Image/DataSets/MNIST/Test-28x28_cntk_text.txt".split("/"))
path = os.path.normpath(os.path.join(abs_path, rel_path))
check_path(path)
reader_test = create_reader(path, False, input_dim, num_output_classes)
input_map = {
features: 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)
test_result += trainer.test_minibatch(mb)
# Average of evaluation errors of all test minibatches
trainer.summarize_test_progress()
return test_result / num_minibatches_to_test
def sequence_to_sequence_translator(debug_output=False, run_test=False):
input_vocab_dim = 69
label_vocab_dim = 69
# network complexity; initially low for faster testing
hidden_dim = 256
num_layers = 1
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(shape=(input_vocab_dim),
dynamic_axes=input_dynamic_axes,
name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(shape=(label_vocab_dim),
dynamic_axes=label_dynamic_axes,
name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = sequence.slice(raw_labels, 1,
0) # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
is_first_label = sequence.is_first(label_sequence) # <s> 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(label_sentence_start,
is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH,
encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value,
future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(thought_vectorH,
label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(thought_vectorC,
label_sequence)
# Decoder
decoder_history_hook = alias(
label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label,
label_sentence_start_scattered,
past_value(decoder_history_hook))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH,
encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH,
recurrence_hookC)
decoder_output = decoder_outputH
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
# Criterion nodes
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# network output for decoder history
net_output = hardmax(z)
# make a clone of the graph where the ground truth is replaced by the network output
ng = z.clone(CloneMethod.share,
{decoder_history_hook.output: net_output.output})
# Instantiate the trainer object to drive the model training
lr_per_minibatch = learning_rate_schedule(0.5, UnitType.minibatch)
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
learner = momentum_sgd(
z.parameters,
lr_per_minibatch,
momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
trainer = Trainer(z, ce, errs, learner)
# setup data
train_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..",
"Data", "cmudict-0.7b.train-dev-20-21.ctf")
valid_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..",
"Data", "tiny.ctf")
# readers
randomize_data = True
if run_test:
randomize_data = False # because we want to get an exact error
train_reader = create_reader(train_path, randomize_data, input_vocab_dim,
label_vocab_dim)
train_bind = {
raw_input: train_reader.streams.features,
raw_labels: train_reader.streams.labels
}
# get the vocab for printing output sequences in plaintext
vocab_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..",
"Data", "cmudict-0.7b.mapping")
vocab = [w.strip() for w in open(vocab_path).readlines()]
i2w = {i: ch for i, ch in enumerate(vocab)}
# Get minibatches of sequences to train with and perform model training
i = 0
mbs = 0
minibatch_size = 72
epoch_size = 908241
max_epochs = 10
training_progress_output_freq = 500
# make things more basic for running a quicker test
if run_test:
epoch_size = 5000
max_epochs = 1
training_progress_output_freq = 30
valid_reader = create_reader(valid_path, False, input_vocab_dim,
label_vocab_dim)
valid_bind = {
find_arg_by_name('raw_input', ng): valid_reader.streams.features,
find_arg_by_name('raw_labels', ng): valid_reader.streams.labels
}
for epoch in range(max_epochs):
loss_numer = 0
metric_numer = 0
denom = 0
while i < (epoch + 1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size,
input_map=train_bind)
trainer.train_minibatch(mb_train)
# collect epoch-wide stats
samples = trainer.previous_minibatch_sample_count
loss_numer += trainer.previous_minibatch_loss_average * samples
metric_numer += trainer.previous_minibatch_evaluation_average * samples
denom += samples
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % training_progress_output_freq == 0:
mb_valid = valid_reader.next_minibatch(minibatch_size,
input_map=valid_bind)
e = ng.eval(mb_valid)
print_sequences(e, i2w)
print_training_progress(trainer, mbs,
training_progress_output_freq)
i += mb_train[raw_labels].num_samples
mbs += 1
print("--- EPOCH %d DONE: loss = %f, errs = %f ---" %
(epoch, loss_numer / denom, 100.0 * (metric_numer / denom)))
error1 = translator_test_error(z, trainer, input_vocab_dim,
label_vocab_dim)
save_model(z, "seq2seq.dnn")
z = load_model("seq2seq.dnn")
label_seq_axis = Axis('labelAxis')
label_sequence = sequence.slice(find_arg_by_name('raw_labels', z), 1, 0)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
trainer = Trainer(z, ce, errs, [
momentum_sgd(z.parameters, lr_per_minibatch, momentum_time_constant,
clipping_threshold_per_sample,
gradient_clipping_with_truncation)
])
error2 = translator_test_error(z, trainer, input_vocab_dim,
label_vocab_dim)
assert error1 == error2
return error1
def convnet_cifar10_dataaug(reader_train,
reader_test,
distributed_trainer,
max_epochs=80):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# apply model to input
scaled_input = element_times(constant(0.00390625), input_var)
with default_options(activation=relu, pad=True):
z = Sequential([
LayerStack(
2, lambda: [
Convolution((3, 3), 64),
Convolution((3, 3), 64),
MaxPooling((3, 3), (2, 2))
]),
LayerStack(2, lambda i: [Dense([256, 128][i]),
Dropout(0.5)]),
Dense(num_classes, activation=None)
])(scaled_input)
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# training config
epoch_size = 50000 # for now we manually specify epoch size
minibatch_size = 64
# Set learning parameters
lr_per_sample = [0.0015625] * 20 + [0.00046875] * 20 + [
0.00015625
] * 20 + [0.000046875] * 10 + [0.000015625]
lr_schedule = learning_rate_schedule(lr_per_sample,
unit=UnitType.sample,
epoch_size=epoch_size)
momentum_time_constant = [0] * 20 + [600] * 20 + [1200]
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant,
epoch_size=epoch_size)
l2_reg_weight = 0.002
# trainer object
learner = momentum_sgd(z.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(z, ce, pe, learner, distributed_trainer)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# perform model training
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(
min(minibatch_size, epoch_size - sample_count),
input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
if distributed_trainer.communicator().current_worker(
).global_rank == 0:
persist.save_model(
z,
os.path.join(model_path,
"ConvNet_CIFAR10_DataAug_{}.dnn".format(epoch)))
### Evaluation action
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch,
input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += trainer.previous_minibatch_sample_count
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
return metric_numer / metric_denom
def convnet_cifar10_dataaug(reader_train, reader_test, max_epochs = 80):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# apply model to input
scaled_input = element_times(constant(0.00390625), input_var)
with default_options (activation=relu, pad=True):
z = Sequential([
LayerStack(2, lambda : [
Convolution((3,3), 64),
Convolution((3,3), 64),
MaxPooling((3,3), (2,2))
]),
LayerStack(2, lambda i: [
Dense([256,128][i]),
Dropout(0.5)
]),
Dense(num_classes, activation=None)
])(scaled_input)
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# training config
epoch_size = 50000 # for now we manually specify epoch size
minibatch_size = 64
# Set learning parameters
lr_per_sample = [0.0015625]*20+[0.00046875]*20+[0.00015625]*20+[0.000046875]*10+[0.000015625]
lr_schedule = learning_rate_schedule(lr_per_sample, unit=UnitType.sample, epoch_size=epoch_size)
mm_time_constant = [0]*20+[600]*20+[1200]
mm_schedule = momentum_as_time_constant_schedule(mm_time_constant, epoch_size=epoch_size)
l2_reg_weight = 0.002
# trainer object
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule,
l2_regularization_weight = l2_reg_weight)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z) ; print()
progress_printer = ProgressPrinter(tag='Training')
# perform model training
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(min(minibatch_size, epoch_size-sample_count), input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
persist.save_model(z, os.path.join(model_path, "ConvNet_CIFAR10_DataAug_{}.dnn".format(epoch)))
### Evaluation action
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch, input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(minibatch_index+1, (metric_numer*100.0)/metric_denom, metric_denom))
print("")
return metric_numer/metric_denom
def simple_mnist(debug_output=False):
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)
try:
rel_path = os.path.join(os.environ['CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY'],
*"Image/MNIST/v0/Train-28x28_cntk_text.txt".split("/"))
except KeyError:
rel_path = os.path.join(abs_path, "..", "..", "..", "..", "..", "Examples", "Image", "DataSets", "MNIST", "Train-28x28_cntk_text.txt")
path = os.path.normpath(os.path.join(abs_path, rel_path))
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
}
lr_per_minibatch=learning_rate_schedule(0.2, UnitType.minibatch)
# Instantiate the trainer object to drive the model training
trainer = Trainer(z, ce, pe, sgd(z.parameters, lr=lr_per_minibatch))
# Get minibatches of images to train with and perform model training
minibatch_size = 64
num_samples_per_sweep = 60000
num_sweeps_to_train_with = 10
num_minibatches_to_train = (num_samples_per_sweep * num_sweeps_to_train_with) / minibatch_size
training_progress_output_freq = 500
if debug_output:
training_progress_output_freq = training_progress_output_freq/4
for i in range(0, int(num_minibatches_to_train)):
mb = reader_train.next_minibatch(minibatch_size, input_map=input_map)
trainer.train_minibatch(mb)
print_training_progress(trainer, i, training_progress_output_freq)
# Load test data
try:
rel_path = os.path.join(os.environ['CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY'],
*"Image/MNIST/v0/Test-28x28_cntk_text.txt".split("/"))
except KeyError:
rel_path = os.path.join(abs_path, "..", "..", "..", "..", "..", "Examples", "Image", "DataSets", "MNIST", "Test-28x28_cntk_text.txt")
path = os.path.normpath(os.path.join(abs_path, rel_path))
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
def cifar_resnet_distributed(data_path, run_test, num_epochs, communicator=None, save_model_filename=None, load_model_filename=None, debug_output=False):
image_height = 32
image_width = 32
num_channels = 3
num_classes = 10
feats_stream_name = 'features'
labels_stream_name = 'labels'
minibatch_source = create_reader(os.path.join(data_path, 'train_map.txt'), os.path.join(data_path, 'CIFAR-10_mean.xml'), True,
distributed_communicator = communicator)
features_si = minibatch_source[feats_stream_name]
labels_si = minibatch_source[labels_stream_name]
# Instantiate the resnet classification model, or load from file
if load_model_filename:
print("Loading model:", load_model_filename)
classifier_output = persist.load_model(load_model_filename)
image_input = classifier_output.arguments[0]
else:
image_input = input_variable(
(num_channels, image_height, image_width), features_si.m_element_type)
classifier_output = create_resnet_model(image_input, num_classes)
# Input variables denoting the features and label data
label_var = input_variable((num_classes), features_si.m_element_type)
ce = cross_entropy_with_softmax(classifier_output, label_var)
pe = classification_error(classifier_output, label_var)
# Instantiate the trainer object to drive the model training
mb_size = 128
num_mb_per_epoch = 100
num_mbs = num_mb_per_epoch * num_epochs
lr_per_sample = [1/mb_size]*80+[0.1/mb_size]*40+[0.01/mb_size]
lr_schedule = learning_rate_schedule(lr_per_sample, units = mb_size * num_mb_per_epoch)
momentum_time_constant = -mb_size/np.log(0.9)
# create data parallel distributed trainer if needed
dist_trainer = distributed.data_parallel_distributed_trainer(communicator, False) if communicator else None
# Instantiate the trainer object to drive the model training
trainer = Trainer(classifier_output, ce, pe,
[momentum_sgd(classifier_output.parameters, lr_schedule, momentum_time_constant, l2_regularization_weight=0.0001)],
distributed_trainer = dist_trainer)
# Get minibatches of images to train with and perform model training
training_progress_output_freq = 100 if communicator else 20
if debug_output:
training_progress_output_freq = training_progress_output_freq/4
for i in range(0, num_mbs):
# NOTE: depends on network, the mb_size can be changed dynamically here
mb = minibatch_source.next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {
image_input: mb[features_si],
label_var: mb[labels_si]
}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
if save_model_filename:
print("Saving model:", save_model_filename)
persist.save_model(classifier_output, save_model_filename)
if run_test:
test_minibatch_source = create_reader(os.path.join(data_path, 'test_map.txt'), os.path.join(data_path, 'CIFAR-10_mean.xml'), False)
features_si = test_minibatch_source[feats_stream_name]
labels_si = test_minibatch_source[labels_stream_name]
mb_size = 128
num_mbs = 100
total_error = 0.0
for i in range(0, num_mbs):
mb = test_minibatch_source.next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {
image_input: mb[features_si],
label_var: mb[labels_si]
}
error = trainer.test_minibatch(arguments)
total_error += error
return total_error / num_mbs
else:
return 0
def train_sequence_to_sequence_translator():
input_vocab_dim = 69
label_vocab_dim = 69
hidden_dim = 512
num_layers = 2
# Source and target inputs to the model
input_dynamic_axes = [ Axis('inputAxis'), Axis.default_batch_axis() ]
raw_input = input_variable(shape=(input_vocab_dim), dynamic_axes = input_dynamic_axes)
label_dynamic_axes = [ Axis('labelAxis'), Axis.default_batch_axis() ]
raw_labels = input_variable(shape=(label_vocab_dim), dynamic_axes = label_dynamic_axes)
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = slice(raw_labels, label_dynamic_axes[0], 1, 0)
label_sentence_start = sequence.first(raw_labels)
is_first_label = sequence.is_first(label_sequence)
label_sentence_start_scattered = sequence.scatter(label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(encoder_outputH, hidden_dim, hidden_dim, future_value, future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(thought_vectorC, label_sequence)
# Decoder
decoder_history_from_ground_truth = label_sequence
decoder_input = element_select(is_first_label, label_sentence_start_scattered, past_value(decoder_history_from_ground_truth))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i == 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(decoder_outputH, hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC)
decoder_output = decoder_outputH
decoder_dim = hidden_dim
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.train-dev-20-21.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration( feature_stream_name, input_vocab_dim, True, 'S0' ),
StreamConfiguration( labels_stream_name, label_vocab_dim, True, 'S1') ], 10000)
features_si = mb_source.stream_info(feature_stream_name)
labels_si = mb_source.stream_info(labels_stream_name)
# Instantiate the trainer object to drive the model training
lr = learning_rates_per_sample(0.007)
momentum_time_constant = 1100
momentum_per_sample = momentums_per_sample(math.exp(-1.0 / momentum_time_constant))
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
trainer = Trainer(z, ce, errs, [momentum_sgd_learner(z.owner.parameters(), lr, momentum_per_sample, clipping_threshold_per_sample, gradient_clipping_with_truncation)])
# Get minibatches of sequences to train with and perform model training
minibatch_size = 72
training_progress_output_freq = 10
while True:
mb = mb_source.get_next_minibatch(minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
arguments = {raw_input : mb[features_si].m_data, raw_labels : mb[labels_si].m_data}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
i += 1
def train(reader, model, max_epochs):
# Input variables denoting the features and label data
query = Input(input_dim, is_sparse=False) # TODO: make sparse once it works
slot_labels = Input(num_labels, is_sparse=True)
# apply model to input
z = model(query)
# loss and metric
ce = cross_entropy_with_softmax(z, slot_labels)
pe = classification_error (z, slot_labels)
# training config
epoch_size = 36000
minibatch_size = 70
num_mbs_to_show_result = 100
lr_per_sample = [0.003]*2+[0.0015]*12+[0.0003]
momentum = 0.9**(1/minibatch_size) # TODO: change to time constant
# trainer object
lr_schedule = learning_rates_per_sample(lr_per_sample, units=epoch_size)
learner = fsadagrad(z.parameters(), lr_schedule, momentum,
targetAdagradAvDenom=1, clipping_threshold_per_sample=15, gradient_clipping_with_truncation=True)
trainer = Trainer(z, ce, pe, [learner])
#_extend_Trainer(trainer) # TODO: should be just baked in
# define mapping from reader streams to network inputs
input_map = {
query : reader.streams.query,
slot_labels : reader.streams.slot_labels
}
# process minibatches and perform model training
t = 0
mbs = 0
for epoch in range(max_epochs):
loss_numer = 0 # TODO: find a nicer way of tracking, this is clumsy
loss_denom = 0
metric_numer = 0
metric_denom = 0
epoch_end = (epoch+1) * epoch_size
while t < epoch_end:
# BUGBUG: RuntimeError: GetNextMinibatch: Changing minibatch sizes across calls is currently unsupported
#data, num_samples = next_minibatch(reader, min(minibatch_size, epoch_size-t), input_map)
data, num_samples = next_minibatch(reader, minibatch_size, input_map)
if data is None:
break
trainer.train_minibatch(data)
loss_numer += trainer.previous_minibatch_loss_average() * trainer.previous_minibatch_sample_count() # too much code for something this simple
loss_denom += trainer.previous_minibatch_sample_count()
metric_numer += trainer.previous_minibatch_evaluation_average() * trainer.previous_minibatch_sample_count()
metric_denom += trainer.previous_minibatch_sample_count()
print_training_progress(trainer, mbs if mbs > 10 else 0, num_mbs_to_show_result)
t += num_samples[slot_labels]
#print (num_samples[slot_labels], t)
mbs += 1
print("--- EPOCH {} DONE: loss = {:0.6f} * {}, metric = {:0.1f}% * {} ---".format(epoch+1, loss_numer/loss_denom, loss_denom, metric_numer/metric_denom*100.0, metric_denom))
return loss_numer/loss_denom, metric_numer/metric_denom
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
def create_network(input_vocab_dim, label_vocab_dim):
# network complexity; initially low for faster testing
hidden_dim = 256
num_layers = 1
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(
shape=(input_vocab_dim), dynamic_axes=input_dynamic_axes, name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(
shape=(label_vocab_dim), dynamic_axes=label_dynamic_axes, name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = sequence.slice(raw_labels, 1, 0) # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
is_first_label = sequence.is_first(label_sequence) # <s> 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(
label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value, future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(
thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(
thought_vectorC, label_sequence)
# Decoder
decoder_history_hook = alias(label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label, label_sentence_start_scattered, past_value(
decoder_history_hook))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC)
decoder_output = decoder_outputH
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
# Criterion nodes
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# network output for decoder history
net_output = hardmax(z)
# make a clone of the graph where the ground truth is replaced by the network output
ng = z.clone(CloneMethod.share, {decoder_history_hook.output : net_output.output})
return {
'raw_input' : raw_input,
'raw_labels' : raw_labels,
'ce' : ce,
'pe' : errs,
'ng' : ng,
'output': z
}
def train_and_evaluate(reader_train, reader_test, network_name, epoch_size, max_epochs, profiler_dir=None,
model_dir=None, tensorboard_logdir=None):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# create model, and configure learning parameters
if network_name == 'resnet20':
z = create_cifar10_model(input_var, 3, num_classes)
lr_per_mb = [1.0]*80+[0.1]*40+[0.01]
elif network_name == 'resnet110':
z = create_cifar10_model(input_var, 18, num_classes)
lr_per_mb = [0.1]*1+[1.0]*80+[0.1]*40+[0.01]
else:
return RuntimeError("Unknown model name!")
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# shared training parameters
minibatch_size = 128
momentum_time_constant = -minibatch_size/np.log(0.9)
l2_reg_weight = 0.0001
# Set learning parameters
lr_per_sample = [lr/minibatch_size for lr in lr_per_mb]
lr_schedule = learning_rate_schedule(lr_per_sample, epoch_size=epoch_size, unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# progress writers
progress_writers = [ProgressPrinter(tag='Training', num_epochs=max_epochs)]
tensorboard_writer = None
if tensorboard_logdir is not None:
tensorboard_writer = TensorBoardProgressWriter(freq=10, log_dir=tensorboard_logdir, model=z)
progress_writers.append(tensorboard_writer)
# trainer object
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule,
l2_regularization_weight = l2_reg_weight)
trainer = Trainer(z, (ce, pe), learner, progress_writers)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z) ; print()
# perform model training
if profiler_dir:
start_profiler(profiler_dir, True)
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(min(minibatch_size, epoch_size-sample_count), input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
trainer.summarize_training_progress()
# Log mean of each parameter tensor, so that we can confirm that the parameters change indeed.
if tensorboard_writer:
for parameter in z.parameters:
tensorboard_writer.write_value(parameter.uid + "/mean", reduce_mean(parameter).eval(), epoch)
if model_dir:
z.save(os.path.join(model_dir, network_name + "_{}.dnn".format(epoch)))
enable_profiler() # begin to collect profiler data after first epoch
if profiler_dir:
stop_profiler()
# Evaluation parameters
test_epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
while sample_count < test_epoch_size:
current_minibatch = min(minibatch_size, test_epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch, input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
print("")
trainer.summarize_test_progress()
print("")
return metric_numer/metric_denom
def train_sequence_classifier(device):
input_dim = 2000
cell_dim = 25
hidden_dim = 25
embedding_dim = 50
num_output_classes = 5
features = variable(shape=input_dim, is_sparse=True, name="features")
classifier_output = LSTM_sequence_classifer_net(features,
num_output_classes,
embedding_dim, hidden_dim,
cell_dim, device)
label = variable(num_output_classes,
dynamic_axes=[Axis.default_batch_axis()],
name="labels")
ce = cross_entropy_with_softmax(classifier_output, label)
pe = classification_error(classifier_output, label)
#TODO: add save and load module code
lstm_net = combine([ce, pe, classifier_output], "classifier_model")
rel_path = r"../../../../Tests/EndToEndTests/Text/SequenceClassification/Data/Train.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
cm = create_text_mb_source(path, input_dim, num_output_classes, 0, True,
False, "x", "y")
stream_infos = cm.stream_infos()
for si in stream_infos:
if si.m_name == 'features':
features_si = si
elif si.m_name == 'labels':
labels_si = si
minibatch_size = 200
lr = lr = learning_rates_per_sample(0.0005)
trainer = Trainer(classifier_output, ce, pe,
[sgdlearner(classifier_output.owner.parameters(), lr)])
freq = 1
i = 0
cntk_dev = cntk_device(device)
while True:
mb = cm.get_next_minibatch(minibatch_size, cntk_dev)
if len(mb) == 0:
break
arguments = dict()
arguments[features] = mb[features_si].m_data
arguments[label] = mb[labels_si].m_data
trainer.train_minibatch(arguments, cntk_dev)
if i % freq == 0:
training_loss = get_train_loss(trainer)
eval_crit = get_train_eval_criterion(trainer)
print(
"Minibatch: {}, Train Loss: {}, Train Evaluation Criterion: {}"
.format(i, training_loss, eval_crit))
i += 1
def simple_mnist():
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)
netout = fully_connected_classifier_net(scaled_input, num_output_classes,
hidden_layers_dim,
num_hidden_layers, sigmoid)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
rel_path = os.path.join(
*"../../../../Examples/Image/MNIST/Data/Train-28x28_cntk_text.txt".
split("/"))
path = os.path.normpath(os.path.join(abs_path, rel_path))
if not os.path.exists(path):
readme_file = os.path.normpath(
os.path.join(os.path.dirname(path), "..", "README.md"))
raise RuntimeError(
"File '%s' does not exist. Please follow the instructions at %s to download and prepare it."
% (path, readme_file))
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration(feature_stream_name, input_dim),
StreamConfiguration(labels_stream_name, num_output_classes)
])
features_si = mb_source.stream_info(feature_stream_name)
labels_si = mb_source.stream_info(labels_stream_name)
# Instantiate the trainer object to drive the model training
trainer = Trainer(netout, ce, pe,
[sgd_learner(netout.owner.parameters(), lr=0.003125)])
# Get minibatches of images to train with and perform model training
minibatch_size = 32
num_samples_per_sweep = 60000
num_sweeps_to_train_with = 1
num_minibatches_to_train = (num_samples_per_sweep *
num_sweeps_to_train_with) / minibatch_size
training_progress_output_freq = 20
for i in range(0, int(num_minibatches_to_train)):
mb = mb_source.get_next_minibatch(minibatch_size)
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
arguments = {
input: mb[features_si].m_data,
label: mb[labels_si].m_data
}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
def simple_mnist():
input_dim = 784
num_output_classes = 10
num_hidden_layers = 1
hidden_layers_dim = 200
epoch_size = sys.maxsize
minibatch_size = 32
num_samples_per_sweep = 60000
num_sweeps_to_train_with = 1
num_minibatches_to_train = (num_samples_per_sweep *
num_sweeps_to_train_with) / minibatch_size
lr = learning_rates_per_sample(0.003125)
input = variable(input_dim,
np.float32,
needs_gradient=False,
name="features")
scaled_input = element_times(constant((), 0.00390625), input)
label = variable(num_output_classes,
np.float32,
needs_gradient=False,
name="labels")
dev = -1
cntk_dev = cntk_device(dev)
netout = fully_connected_classifier_net(scaled_input, num_output_classes,
hidden_layers_dim,
num_hidden_layers, dev, sigmoid)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
#TODO: add save and load module code
ffnet = combine([ce, pe, netout], "classifier_model")
rel_path = r"../../../../Examples/Image/MNIST/Data/Train-28x28_cntk_text.txt"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
cm = create_text_mb_source(path, input_dim, num_output_classes, epoch_size)
stream_infos = cm.stream_infos()
for si in stream_infos:
if si.m_name == 'features':
features_si = si
elif si.m_name == 'labels':
labels_si = si
trainer = Trainer(netout, ce, pe,
[sgdlearner(netout.owner.parameters(), lr)])
for i in range(0, int(num_minibatches_to_train)):
mb = cm.get_next_minibatch(minibatch_size, cntk_dev)
arguments = dict()
arguments[input] = mb[features_si].m_data
arguments[label] = mb[labels_si].m_data
trainer.train_minibatch(arguments, cntk_dev)
freq = 20
if i % freq == 0:
training_loss = get_train_loss(trainer)
eval_crit = get_train_eval_criterion(trainer)
print(
"Minibatch: {}, Train Loss: {}, Train Evaluation Criterion: {}"
.format(i, training_loss, eval_crit))
def conv3d_ucf11(train_reader, test_reader, max_epochs=30):
# Replace 0 with 1 to get detailed log.
set_computation_network_trace_level(0)
# These values must match for both train and test reader.
image_height = train_reader.height
image_width = train_reader.width
num_channels = train_reader.channel_count
sequence_length = train_reader.sequence_length
num_output_classes = train_reader.label_count
# Input variables denoting the features and label data
input_var = input_variable((num_channels, sequence_length, image_height, image_width), np.float32)
label_var = input_variable(num_output_classes, np.float32)
# Instantiate simple 3D Convolution network inspired by VGG network
# and http://vlg.cs.dartmouth.edu/c3d/c3d_video.pdf
with default_options (activation=relu):
z = Sequential([
Convolution3D((3,3,3), 64, pad=True),
MaxPooling((1,2,2), (1,2,2)),
For(range(3), lambda i: [
Convolution3D((3,3,3), [96, 128, 128][i], pad=True),
Convolution3D((3,3,3), [96, 128, 128][i], pad=True),
MaxPooling((2,2,2), (2,2,2))
]),
For(range(2), lambda : [
Dense(1024),
Dropout(0.5)
]),
Dense(num_output_classes, activation=None)
])(input_var)
# loss and classification error.
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# training config
epoch_size = 1322 # for now we manually specify epoch size
minibatch_size = 4
# Set learning parameters
lr_per_sample = [0.01]*10+[0.001]*10+[0.0001]
lr_schedule = learning_rate_schedule(lr_per_sample, epoch_size=epoch_size, unit=UnitType.sample)
momentum_time_constant = 4096
mm_schedule = momentum_as_time_constant_schedule([momentum_time_constant], epoch_size=epoch_size)
# Instantiate the trainer object to drive the model training
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule, True)
progress_printer = ProgressPrinter(tag='Training', num_epochs=max_epochs)
trainer = Trainer(z, (ce, pe), learner, progress_printer)
log_number_of_parameters(z) ; print()
# Get minibatches of images to train with and perform model training
for epoch in range(max_epochs): # loop over epochs
train_reader.reset()
while train_reader.has_more():
videos, labels, current_minibatch = train_reader.next_minibatch(minibatch_size)
trainer.train_minibatch({input_var : videos, label_var : labels})
trainer.summarize_training_progress()
# Test data for trained model
epoch_size = 332
minibatch_size = 2
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
minibatch_index = 0
test_reader.reset()
while test_reader.has_more():
videos, labels, current_minibatch = test_reader.next_minibatch(minibatch_size)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch({input_var : videos, label_var : labels}) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(minibatch_index+1, (metric_numer*100.0)/metric_denom, metric_denom))
print("")
return metric_numer/metric_denom
def train_and_evaluate(reader_train, reader_test, network_name, max_epochs):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# create model, and configure learning parameters
if network_name == 'resnet20':
z = create_cifar10_model(input_var, 3, num_classes)
lr_per_mb = [1.0] * 80 + [0.1] * 40 + [0.01]
elif network_name == 'resnet110':
z = create_cifar10_model(input_var, 18, num_classes)
lr_per_mb = [0.1] * 1 + [1.0] * 80 + [0.1] * 40 + [0.01]
else:
return RuntimeError("Unknown model name!")
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# shared training parameters
epoch_size = 50000 # for now we manually specify epoch size
minibatch_size = 128
momentum_time_constant = -minibatch_size / np.log(0.9)
l2_reg_weight = 0.0001
# Set learning parameters
lr_per_sample = [lr / minibatch_size for lr in lr_per_mb]
lr_schedule = learning_rate_schedule(lr_per_sample,
epoch_size=epoch_size,
unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# trainer object
learner = momentum_sgd(z.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# perform model training
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(
min(minibatch_size, epoch_size - sample_count),
input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
persist.save_model(
z, os.path.join(model_path,
network_name + "_{}.dnn".format(epoch)))
# Evaluation parameters
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch,
input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
return metric_numer / metric_denom
def simple_mnist(debug_output=False):
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)
netout = fully_connected_classifier_net(scaled_input, num_output_classes,
hidden_layers_dim,
num_hidden_layers, sigmoid)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
try:
rel_path = os.path.join(
os.environ['CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY'],
*"Image/MNIST/v0/Train-28x28_cntk_text.txt".split("/"))
except KeyError:
rel_path = os.path.join(
*"../../../../Examples/Image/MNIST/Data/Train-28x28_cntk_text.txt".
split("/"))
path = os.path.normpath(os.path.join(abs_path, rel_path))
check_path(path)
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration(feature_stream_name, input_dim),
StreamConfiguration(labels_stream_name, num_output_classes)
])
features_si = mb_source.stream_info(feature_stream_name)
labels_si = mb_source.stream_info(labels_stream_name)
# Instantiate the trainer object to drive the model training
trainer = Trainer(netout, ce, pe, [sgd(netout.parameters(), lr=0.003125)])
# Get minibatches of images to train with and perform model training
minibatch_size = 32
num_samples_per_sweep = 60000
num_sweeps_to_train_with = 1
num_minibatches_to_train = (num_samples_per_sweep *
num_sweeps_to_train_with) / minibatch_size
training_progress_output_freq = 20
for i in range(0, int(num_minibatches_to_train)):
mb = mb_source.get_next_minibatch(minibatch_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {
input: mb[features_si].m_data,
label: mb[labels_si].m_data
}
trainer.train_minibatch(arguments)
if debug_output:
print_training_progress(trainer, i, training_progress_output_freq)
# Load test data
try:
rel_path = os.path.join(
os.environ['CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY'],
*"Image/MNIST/v0/Test-28x28_cntk_text.txt".split("/"))
except KeyError:
rel_path = os.path.join(
*"../../../../Examples/Image/MNIST/Data/Test-28x28_cntk_text.txt".
split("/"))
path = os.path.normpath(os.path.join(abs_path, rel_path))
check_path(path)
test_mb_source = text_format_minibatch_source(path, [
StreamConfiguration(feature_stream_name, input_dim),
StreamConfiguration(labels_stream_name, num_output_classes)
])
features_si = test_mb_source.stream_info(feature_stream_name)
labels_si = test_mb_source.stream_info(labels_stream_name)
# Test data for trained model
test_minibatch_size = 512
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 = test_mb_source.get_next_minibatch(test_minibatch_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be tested with
arguments = {
input: mb[features_si].m_data,
label: mb[labels_si].m_data
}
eval_error = trainer.test_minibatch(arguments)
test_result = test_result + eval_error
# Average of evaluation errors of all test minibatches
return test_result / num_minibatches_to_test
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 = input_variable(shape=input_dim, is_sparse=True)
label = input_variable(num_output_classes, dynamic_axes=[
Axis.default_batch_axis()])
# 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)
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration(feature_stream_name, input_dim, True, 'x'),
StreamConfiguration(labels_stream_name, num_output_classes, False, 'y')], 0)
features_si = mb_source[features]
labels_si = mb_source[label]
# Instantiate the trainer object to drive the model training
trainer = Trainer(classifier_output, ce, pe,
[sgd(classifier_output.parameters(), lr=0.0005)])
# Get minibatches of sequences to train with and perform model training
minibatch_size = 200
training_progress_output_freq = 10
i = 0
if debug_output:
training_progress_output_freq = training_progress_output_freq/3
while True:
mb = mb_source.get_next_minibatch(minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {features: mb[features_si],
label: mb[labels_si]}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
i += 1
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 sequence_to_sequence_translator(debug_output=False, run_test=False):
input_vocab_dim = 69
label_vocab_dim = 69
# network complexity; initially low for faster testing
hidden_dim = 256
num_layers = 1
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(shape=(input_vocab_dim),
dynamic_axes=input_dynamic_axes,
name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(shape=(label_vocab_dim),
dynamic_axes=label_dynamic_axes,
name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = slice(raw_labels, label_seq_axis, 1,
0) # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
is_first_label = sequence.is_first(label_sequence) # <s> 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(label_sentence_start,
is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH,
encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value,
future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(thought_vectorH,
label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(thought_vectorC,
label_sequence)
# Decoder
decoder_history_hook = alias(
label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label,
label_sentence_start_scattered,
past_value(decoder_history_hook))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH,
encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH,
recurrence_hookC)
decoder_output = decoder_outputH
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
# Criterion nodes
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# network output for decoder history
net_output = hardmax(z)
# make a clone of the graph where the ground truth is replaced by the network output
ng = z.clone(CloneMethod.share,
{decoder_history_hook.output: net_output.output})
# Instantiate the trainer object to drive the model training
lr = 0.007
minibatch_size = 72
momentum_time_constant = 1100
m_schedule = momentum_schedule(momentum_time_constant)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
learner = momentum_sgd(z.parameters, lr, m_schedule,
clipping_threshold_per_sample,
gradient_clipping_with_truncation)
trainer = Trainer(z, ce, errs, learner)
# setup data
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.train-dev-20-21.ctf"
train_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
rel_path)
valid_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
"tiny.ctf")
feature_stream_name = 'features'
labels_stream_name = 'labels'
# readers
randomize_data = True
if run_test:
randomize_data = False # because we want to get an exact error
train_reader = text_format_minibatch_source(train_path, [
StreamConfiguration(feature_stream_name, input_vocab_dim, True, 'S0'),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, 'S1')
],
randomize=randomize_data)
features_si_tr = train_reader.stream_info(feature_stream_name)
labels_si_tr = train_reader.stream_info(labels_stream_name)
valid_reader = text_format_minibatch_source(valid_path, [
StreamConfiguration(feature_stream_name, input_vocab_dim, True, 'S0'),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, 'S1')
],
randomize=False)
features_si_va = valid_reader.stream_info(feature_stream_name)
labels_si_va = valid_reader.stream_info(labels_stream_name)
# get the vocab for printing output sequences in plaintext
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.mapping"
vocab_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
rel_path)
vocab = [w.strip() for w in open(vocab_path).readlines()]
i2w = {i: ch for i, ch in enumerate(vocab)}
# Get minibatches of sequences to train with and perform model training
i = 0
mbs = 0
epoch_size = 908241
max_epochs = 10
training_progress_output_freq = 500
# make things more basic for running a quicker test
if run_test:
epoch_size = 5000
max_epochs = 1
training_progress_output_freq = 30
for epoch in range(max_epochs):
loss_numer = 0
metric_numer = 0
denom = 0
while i < (epoch + 1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size)
train_args = {
'raw_input': mb_train[features_si_tr],
'raw_labels': mb_train[labels_si_tr]
}
trainer.train_minibatch(train_args)
# collect epoch-wide stats
samples = trainer.previous_minibatch_sample_count
loss_numer += trainer.previous_minibatch_loss_average * samples
metric_numer += trainer.previous_minibatch_evaluation_average * samples
denom += samples
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % training_progress_output_freq == 0:
mb_valid = valid_reader.next_minibatch(minibatch_size)
valid_args = {
'raw_input': mb_valid[features_si_va],
'raw_labels': mb_valid[labels_si_va]
}
e = ng.eval(valid_args)
print_sequences(e, i2w)
print_training_progress(trainer, mbs,
training_progress_output_freq)
i += mb_train[labels_si_tr].num_samples
mbs += 1
print("--- EPOCH %d DONE: loss = %f, errs = %f ---" %
(epoch, loss_numer / denom, 100.0 * (metric_numer / denom)))
# now setup a test run
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.test.ctf"
test_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
rel_path)
test_reader = text_format_minibatch_source(test_path, [
StreamConfiguration(feature_stream_name, input_vocab_dim, True, 'S0'),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, 'S1')
],
10000,
randomize=False)
features_si_te = test_reader.stream_info(feature_stream_name)
labels_si_te = test_reader.stream_info(labels_stream_name)
test_minibatch_size = 1024
# Get minibatches of sequences to test and perform testing
i = 0
total_error = 0.0
while True:
mb = test_reader.next_minibatch(test_minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be tested with
arguments = {
raw_input: mb[features_si_te],
raw_labels: mb[labels_si_te]
}
mb_error = trainer.test_minibatch(arguments)
total_error += mb_error
if debug_output:
print("Minibatch {}, Error {} ".format(i, mb_error))
i += 1
# Average of evaluation errors of all test minibatches
return total_error / i
def sequence_to_sequence_translator(debug_output=False, run_test=False):
input_vocab_dim = 69
label_vocab_dim = 69
# network complexity; initially low for faster testing
hidden_dim = 256
num_layers = 1
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(
shape=(input_vocab_dim), dynamic_axes=input_dynamic_axes, name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(
shape=(label_vocab_dim), dynamic_axes=label_dynamic_axes, name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = sequence.slice(raw_labels, 1, 0) # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
is_first_label = sequence.is_first(label_sequence) # <s> 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(
label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value, future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(
thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(
thought_vectorC, label_sequence)
# Decoder
decoder_history_hook = alias(label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label, label_sentence_start_scattered, past_value(
decoder_history_hook))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC)
decoder_output = decoder_outputH
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
# Criterion nodes
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# network output for decoder history
net_output = hardmax(z)
# make a clone of the graph where the ground truth is replaced by the network output
ng = z.clone(CloneMethod.share, {decoder_history_hook.output : net_output.output})
# Instantiate the trainer object to drive the model training
lr_per_minibatch = learning_rate_schedule(0.5, UnitType.minibatch)
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
learner = momentum_sgd(z.parameters,
lr_per_minibatch, momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
trainer = Trainer(z, ce, errs, learner)
# setup data
train_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "Data", "cmudict-0.7b.train-dev-20-21.ctf")
valid_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "Data", "tiny.ctf")
# readers
randomize_data = True
if run_test:
randomize_data = False # because we want to get an exact error
train_reader = create_reader(train_path, randomize_data, input_vocab_dim, label_vocab_dim)
train_bind = {
raw_input : train_reader.streams.features,
raw_labels : train_reader.streams.labels
}
# get the vocab for printing output sequences in plaintext
vocab_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "Data", "cmudict-0.7b.mapping")
vocab = [w.strip() for w in open(vocab_path).readlines()]
i2w = { i:ch for i,ch in enumerate(vocab) }
# Get minibatches of sequences to train with and perform model training
i = 0
mbs = 0
minibatch_size = 72
epoch_size = 908241
max_epochs = 10
training_progress_output_freq = 500
# make things more basic for running a quicker test
if run_test:
epoch_size = 5000
max_epochs = 1
training_progress_output_freq = 30
valid_reader = create_reader(valid_path, False, input_vocab_dim, label_vocab_dim)
valid_bind = {
find_arg_by_name('raw_input',ng) : valid_reader.streams.features,
find_arg_by_name('raw_labels',ng) : valid_reader.streams.labels
}
for epoch in range(max_epochs):
loss_numer = 0
metric_numer = 0
denom = 0
while i < (epoch+1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size, input_map=train_bind)
trainer.train_minibatch(mb_train)
# collect epoch-wide stats
samples = trainer.previous_minibatch_sample_count
loss_numer += trainer.previous_minibatch_loss_average * samples
metric_numer += trainer.previous_minibatch_evaluation_average * samples
denom += samples
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % training_progress_output_freq == 0:
mb_valid = valid_reader.next_minibatch(minibatch_size, input_map=valid_bind)
e = ng.eval(mb_valid)
print_sequences(e, i2w)
print_training_progress(trainer, mbs, training_progress_output_freq)
i += mb_train[raw_labels].num_samples
mbs += 1
print("--- EPOCH %d DONE: loss = %f, errs = %f ---" % (epoch, loss_numer/denom, 100.0*(metric_numer/denom)))
error1 = translator_test_error(z, trainer, input_vocab_dim, label_vocab_dim)
z.save_model("seq2seq.dnn")
z.restore_model("seq2seq.dnn")
label_seq_axis = Axis('labelAxis')
label_sequence = sequence.slice(find_arg_by_name('raw_labels',z), 1, 0)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
trainer = Trainer(z, ce, errs, [momentum_sgd(
z.parameters, lr_per_minibatch, momentum_time_constant, clipping_threshold_per_sample, gradient_clipping_with_truncation)])
error2 = translator_test_error(z, trainer, input_vocab_dim, label_vocab_dim)
assert error1 == error2
return error1
def simple_mnist():
input_dim = 784
num_output_classes = 10
num_hidden_layers = 1
hidden_layers_dim = 200
# Input variables denoting the features and label data
features = 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), features)
netout = fully_connected_classifier_net(scaled_input, num_output_classes,
hidden_layers_dim,
num_hidden_layers, relu)
ce = cross_entropy_with_softmax(netout, label)
pe = classification_error(netout, label)
try:
rel_path = os.path.join(
os.environ['CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY'],
*"Image/MNIST/v0/Train-28x28_cntk_text.txt".split("/"))
except KeyError:
rel_path = os.path.join(
*"../Image/DataSets/MNIST/Train-28x28_cntk_text.txt".split("/"))
path = os.path.normpath(os.path.join(abs_path, rel_path))
check_path(path)
reader_train = create_reader(path, True, input_dim, num_output_classes)
input_map = {
features: reader_train.streams.features,
label: reader_train.streams.labels
}
# Instantiate progress writers.
logdir = os.path.join(os.path.dirname(os.path.abspath(__file__)),
"mnist_log")
tensorboard_writer = TensorBoardProgressWriter(freq=1,
log_dir=logdir,
model=netout)
progress_printer = ProgressPrinter(freq=10, tag='Training')
# Instantiate the trainer object to drive the model training
lr_per_minibatch = learning_rate_schedule(0.2, UnitType.minibatch)
learner = sgd(netout.parameters, lr=lr_per_minibatch)
trainer = Trainer(netout, (ce, pe), learner,
[tensorboard_writer, progress_printer])
# Get minibatches of images to train with and perform model training
minibatch_size = 64
num_samples_per_sweep = 6000
num_sweeps_to_train_with = 2
num_minibatches_to_train = (num_samples_per_sweep *
num_sweeps_to_train_with) / minibatch_size
for minibatch_idx in range(0, int(num_minibatches_to_train)):
trainer.train_minibatch(
reader_train.next_minibatch(minibatch_size, input_map=input_map))
# Log max/min/mean of each parameter tensor, so that we can confirm that the parameters change indeed.
# Don't want to do that very often though, otherwise will spend too much time computing min/max/mean.
if minibatch_idx % 10 == 9:
for p in netout.parameters:
tensorboard_writer.write_value(p.uid + "/max",
reduce_max(p).eval(),
minibatch_idx)
tensorboard_writer.write_value(p.uid + "/min",
reduce_min(p).eval(),
minibatch_idx)
tensorboard_writer.write_value(p.uid + "/mean",
reduce_mean(p).eval(),
minibatch_idx)
trainer.summarize_training_progress()
# Load test data
try:
rel_path = os.path.join(
os.environ['CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY'],
*"Image/MNIST/v0/Test-28x28_cntk_text.txt".split("/"))
except KeyError:
rel_path = os.path.join(
*"../Image/DataSets/MNIST/Test-28x28_cntk_text.txt".split("/"))
path = os.path.normpath(os.path.join(abs_path, rel_path))
check_path(path)
reader_test = create_reader(path, False, input_dim, num_output_classes)
input_map = {
features: 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)
test_result += trainer.test_minibatch(mb)
# Average of evaluation errors of all test minibatches
trainer.summarize_test_progress()
return test_result / num_minibatches_to_test
def train_fast_rcnn(debug_output=False):
if debug_output:
print("Storing graphs and intermediate models to %s." %
os.path.join(abs_path, "Output"))
# Create the minibatch source
minibatch_source = create_mb_source(image_height, image_width,
num_channels, num_classes, num_rois,
base_path, "train")
# Input variables denoting features, rois and label data
image_input = input((num_channels, image_height, image_width))
roi_input = input((num_rois, 4))
label_input = input((num_rois, num_classes))
# define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source[features_stream_name],
roi_input: minibatch_source[roi_stream_name],
label_input: minibatch_source[label_stream_name]
}
# Instantiate the Fast R-CNN prediction model and loss function
frcn_output = frcn_predictor(image_input, roi_input, num_classes)
ce = cross_entropy_with_softmax(frcn_output, label_input, axis=1)
pe = classification_error(frcn_output, label_input, axis=1)
if debug_output:
plot(frcn_output, os.path.join(abs_path, "Output", "graph_frcn.png"))
# Set learning parameters
l2_reg_weight = 0.0005
lr_per_sample = [0.00001] * 10 + [0.000001] * 5 + [0.0000001]
lr_schedule = learning_rate_schedule(lr_per_sample, unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# Instantiate the trainer object
learner = momentum_sgd(frcn_output.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
progress_printer = ProgressPrinter(tag='Training', num_epochs=max_epochs)
trainer = Trainer(frcn_output, (ce, pe), learner, progress_printer)
# Get minibatches of images and perform model training
print("Training Fast R-CNN model for %s epochs." % max_epochs)
log_number_of_parameters(frcn_output)
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(
mb_size, epoch_size - sample_count),
input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
trainer.summarize_training_progress()
if debug_output:
frcn_output.save(
os.path.join(abs_path, "Output",
"frcn_py_%s.model" % (epoch + 1)))
return frcn_output
def simple_mnist(debug_output=False):
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
}
lr_per_minibatch=learning_rate_schedule(0.2, UnitType.minibatch)
# Instantiate the trainer object to drive the model training
trainer = Trainer(z, ce, pe, sgd(z.parameters, lr=lr_per_minibatch))
# Get minibatches of images to train with and perform model training
minibatch_size = 64
num_samples_per_sweep = 60000
num_sweeps_to_train_with = 10
num_minibatches_to_train = (num_samples_per_sweep * num_sweeps_to_train_with) / minibatch_size
training_progress_output_freq = 500
if debug_output:
training_progress_output_freq = training_progress_output_freq/4
for i in range(0, int(num_minibatches_to_train)):
mb = reader_train.next_minibatch(minibatch_size, input_map=input_map)
trainer.train_minibatch(mb)
print_training_progress(trainer, i, training_progress_output_freq)
# 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
def cifar_resnet(base_path, debug_output=False):
image_height = 32
image_width = 32
num_channels = 3
num_classes = 10
feats_stream_name = 'features'
labels_stream_name = 'labels'
minibatch_source = create_mb_source(feats_stream_name, labels_stream_name,
image_height, image_width, num_channels, num_classes, base_path)
features_si = minibatch_source[feats_stream_name]
labels_si = minibatch_source[labels_stream_name]
# Input variables denoting the features and label data
image_input = input_variable(
(num_channels, image_height, image_width), features_si.m_element_type)
label_var = input_variable((num_classes), features_si.m_element_type)
# Instantiate the resnet classification model
classifier_output = resnet_classifer(image_input, num_classes)
ce = cross_entropy_with_softmax(classifier_output, label_var)
pe = classification_error(classifier_output, label_var)
# Instantiate the trainer object to drive the model training
trainer = Trainer(classifier_output, ce, pe,
[sgd(classifier_output.parameters(), lr=0.0078125)])
# Get minibatches of images to train with and perform model training
mb_size = 32
training_progress_output_freq = 60
num_mbs = 1000
if debug_output:
training_progress_output_freq = training_progress_output_freq/3
for i in range(0, num_mbs):
mb = minibatch_source.get_next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {
image_input: mb[features_si],
label_var: mb[labels_si]
}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
test_minibatch_source = create_test_mb_source(feats_stream_name, labels_stream_name,
image_height, image_width, num_channels, num_classes, base_path)
features_si = test_minibatch_source[feats_stream_name]
labels_si = test_minibatch_source[labels_stream_name]
mb_size = 64
num_mbs = 300
total_error = 0.0
for i in range(0, num_mbs):
mb = test_minibatch_source.get_next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {
image_input: mb[features_si],
label_var: mb[labels_si]
}
error = trainer.test_minibatch(arguments)
total_error += error
return total_error / num_mbs
def train_and_evaluate(reader_train, reader_test, max_epochs):
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# Normalize the input
feature_scale = 1.0 / 256.0
input_var_norm = element_times(feature_scale, input_var)
# apply model to input
z = create_vgg9_model(input_var_norm, 10)
#
# Training action
#
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# training config
epoch_size = 50000
minibatch_size = 64
# Set learning parameters
lr_per_minibatch = learning_rate_schedule(
[0.01] * 10 + [0.003] * 10 + [0.001], epoch_size, UnitType.minibatch)
momentum_time_constant = momentum_as_time_constant_schedule(
-minibatch_size / np.log(0.9))
l2_reg_weight = 0.0001
# trainer object
learner = momentum_sgd(z.parameters,
lr=lr_per_minibatch,
momentum=momentum_time_constant,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# perform model training
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(
min(minibatch_size, epoch_size - sample_count),
input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += data[
label_var].num_samples # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
#
# Evaluation action
#
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
#progress_printer = ProgressPrinter(freq=100, first=10, tag='Eval')
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch,
input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.1f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
# return evaluation error.
return metric_numer / metric_denom
recurrence_hook_h, recurrence_hook_c)
# Linear output layer
W = parameter(shape=(decoder_output_h.shape[0], label_vocab_dim),
init=glorot_uniform())
B = parameter(shape=(label_vocab_dim), init=0)
z = plus(B, times(stabilize(decoder_output_h), W))
return z
model = create_model()
label_sequence = model.find_by_name('label_sequence')
# Criterion nodes
ce = cross_entropy_with_softmax(model, label_sequence)
errs = classification_error(model, label_sequence)
# let's show the required arguments for this model
print([x.name for x in model.arguments])
lr_per_sample = learning_rate_schedule(0.007, UnitType.sample)
minibatch_size = 72
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
learner = momentum_sgd(
model.parameters,
lr_per_sample,
momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
def train(reader, model, max_epochs):
# Input variables denoting the features and label data
query = Input(input_dim, is_sparse=False)
slot_labels = Input(num_labels,
is_sparse=True) # TODO: make sparse once it works
# apply model to input
z = model(query)
# loss and metric
ce = cross_entropy_with_softmax(z, slot_labels)
pe = classification_error(z, slot_labels)
# training config
epoch_size = 36000
minibatch_size = 70
num_mbs_to_show_result = 100
momentum_time_constant = momentum_as_time_constant_schedule(
minibatch_size /
-math.log(0.9)) # TODO: Change to round number. This is 664.39. 700?
lr_schedule = [0.003] * 2 + [0.0015] * 12 + [
0.0003
] # LR schedule over epochs (we don't run that many epochs, but if we did, these are good values)
# trainer object
lr_per_sample = learning_rate_schedule(lr_schedule, UnitType.sample,
epoch_size)
learner = adam_sgd(z.parameters,
lr=lr_per_sample,
momentum=momentum_time_constant,
low_memory=True,
gradient_clipping_threshold_per_sample=15,
gradient_clipping_with_truncation=True)
trainer = Trainer(z, ce, pe, [learner])
# define mapping from reader streams to network inputs
input_map = {
query: reader.streams.query,
slot_labels: reader.streams.slot_labels
}
# process minibatches and perform model training
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(freq=100, first=10,
tag='Training') # more detailed logging
#progress_printer = ProgressPrinter(tag='Training')
t = 0
for epoch in range(max_epochs): # loop over epochs
epoch_end = (epoch + 1) * epoch_size
while t < epoch_end: # loop over minibatches on the epoch
# BUGBUG? The change of minibatch_size parameter vv has no effect.
data = reader.next_minibatch(
min(minibatch_size,
epoch_end - t), input_map=input_map) # fetch minibatch
trainer.train_minibatch(data) # update model with it
t += data[
slot_labels].num_samples # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
#def trace_node(name):
# nl = [n for n in z.parameters if n.name() == name]
# if len(nl) > 0:
# print (name, np.asarray(nl[0].value))
#trace_node('W')
#trace_node('stabilizer_param')
loss, metric, actual_samples = progress_printer.epoch_summary(
with_metric=True)
return loss, metric
def train(train_reader, valid_reader, vocab, i2w, model, max_epochs):
# do some hooks that we won't need in the future
label_sequence = model.find_by_name('label_sequence')
decoder_history_hook = model.find_by_name('decoder_history_hook')
# Criterion nodes
ce = cross_entropy_with_softmax(model, label_sequence)
errs = classification_error(model, label_sequence)
def clone_and_hook():
# network output for decoder history
net_output = hardmax(model)
# make a clone of the graph where the ground truth is replaced by the network output
return model.clone(CloneMethod.share,
{decoder_history_hook.output: net_output.output})
# get a new model that uses the past network output as input to the decoder
new_model = clone_and_hook()
# Instantiate the trainer object to drive the model training
lr_per_sample = learning_rate_schedule(0.007, UnitType.sample)
minibatch_size = 72
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
learner = momentum_sgd(
model.parameters,
lr_per_sample,
momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
trainer = Trainer(model, ce, errs, learner)
# Get minibatches of sequences to train with and perform model training
i = 0
mbs = 0
# Set epoch size to a larger number of lower training error
epoch_size = 5000 if isFast else 908241
training_progress_output_freq = 100
# bind inputs to data from readers
train_bind = {
find_arg_by_name('raw_input', model): train_reader.streams.features,
find_arg_by_name('raw_labels', model): train_reader.streams.labels
}
valid_bind = {
find_arg_by_name('raw_input', new_model):
valid_reader.streams.features,
find_arg_by_name('raw_labels', new_model): valid_reader.streams.labels
}
for epoch in range(max_epochs):
loss_numer = 0
metric_numer = 0
denom = 0
while i < (epoch + 1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size,
input_map=train_bind)
trainer.train_minibatch(mb_train)
# collect epoch-wide stats
samples = trainer.previous_minibatch_sample_count
loss_numer += trainer.previous_minibatch_loss_average * samples
metric_numer += trainer.previous_minibatch_evaluation_average * samples
denom += samples
# every N MBs evaluate on a test sequence to visually show how we're doing; also print training stats
if mbs % training_progress_output_freq == 0:
print(
"Minibatch: {0}, Train Loss: {1:2.3f}, Train Evaluation Criterion: {2:2.3f}"
.format(mbs, get_train_loss(trainer),
get_train_eval_criterion(trainer)))
mb_valid = valid_reader.next_minibatch(minibatch_size,
input_map=valid_bind)
e = new_model.eval(mb_valid)
print_sequences(e, i2w)
i += mb_train[find_arg_by_name('raw_labels', model)].num_samples
mbs += 1
print("--- EPOCH %d DONE: loss = %f, errs = %f ---" %
(epoch, loss_numer / denom, 100.0 * (metric_numer / denom)))
return 100.0 * (metric_numer / denom)
def train_model(base_model_file,
feature_node_name,
last_hidden_node_name,
image_width,
image_height,
num_channels,
num_classes,
train_map_file,
num_epochs,
max_images=-1,
freeze=False):
epoch_size = sum(1 for line in open(train_map_file))
if max_images > 0:
epoch_size = min(epoch_size, max_images)
# Create the minibatch source and input variables
minibatch_source = create_mb_source(train_map_file, image_width,
image_height, num_channels,
num_classes)
image_input = input_variable((num_channels, image_height, image_width))
label_input = input_variable(num_classes)
# Define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source[features_stream_name],
label_input: minibatch_source[label_stream_name]
}
# Instantiate the transfer learning model and loss function
tl_model = create_model(base_model_file, feature_node_name,
last_hidden_node_name, num_classes, image_input,
freeze)
ce = cross_entropy_with_softmax(tl_model, label_input)
pe = classification_error(tl_model, label_input)
# Instantiate the trainer object
lr_schedule = learning_rate_schedule(lr_per_mb, unit=UnitType.minibatch)
mm_schedule = momentum_schedule(momentum_per_mb)
learner = momentum_sgd(tl_model.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(tl_model, (ce, pe), learner)
# Get minibatches of images and perform model training
print(
"Training transfer learning model for {0} epochs (epoch_size = {1}).".
format(num_epochs, epoch_size))
log_number_of_parameters(tl_model)
progress_printer = ProgressPrinter(tag='Training', num_epochs=num_epochs)
for epoch in range(num_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(
mb_size, epoch_size - sample_count),
input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
if sample_count % (100 * mb_size) == 0:
print("Processed {0} samples".format(sample_count))
progress_printer.epoch_summary(with_metric=True)
return tl_model
def convnet_mnist(debug_output=False):
image_height = 28
image_width = 28
num_channels = 1
input_dim = image_height * image_width * num_channels
num_output_classes = 10
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width), np.float32)
label_var = input_variable(num_output_classes, np.float32)
# Instantiate the feedforward classification model
scaled_input = element_times(constant(0.00390625), input_var)
with default_options (activation=relu, pad=False):
conv1 = Convolution((5,5), 32, pad=True)(scaled_input)
pool1 = MaxPooling((3,3), (2,2))(conv1)
conv2 = Convolution((3,3), 48)(pool1)
pool2 = MaxPooling((3,3), (2,2))(conv2)
conv3 = Convolution((3,3), 64)(pool2)
f4 = Dense(96)(conv3)
drop4 = Dropout(0.5)(f4)
z = Dense(num_output_classes, activation=None)(drop4)
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
reader_train = create_reader(os.path.join(data_path, 'Train-28x28_cntk_text.txt'), True, input_dim, num_output_classes)
# training config
epoch_size = 60000 # for now we manually specify epoch size
minibatch_size = 128
# Set learning parameters
lr_per_sample = [0.001]*10+[0.0005]*10+[0.0001]
lr_schedule = learning_rate_schedule(lr_per_sample, epoch_size)
momentum_time_constant = [0]*5+[1024]
# Instantiate the trainer object to drive the model training
learner = momentum_sgd(z.parameters, lr_schedule, momentum_time_constant)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var : reader_train.streams.features,
label_var : reader_train.streams.labels
}
log_number_of_parameters(z) ; print()
progress_printer = ProgressPrinter(tag='Training')
# Get minibatches of images to train with and perform model training
max_epochs = 40
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(min(minibatch_size, epoch_size - sample_count), input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += data[label_var].num_samples # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
persist.save_model(z, os.path.join(model_path, "ConvNet_MNIST_{}.dnn".format(epoch)))
# Load test data
reader_test = create_reader(os.path.join(data_path, 'Test-28x28_cntk_text.txt'), False, input_dim, num_output_classes)
input_map = {
input_var : reader_test.streams.features,
label_var : reader_test.streams.labels
}
# Test data for trained model
epoch_size = 10000
minibatch_size = 1024
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch, input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(minibatch_index+1, (metric_numer*100.0)/metric_denom, metric_denom))
print("")
return metric_numer/metric_denom
def convnet_cifar10(debug_output=False):
set_computation_network_trace_level(0)
image_height = 32
image_width = 32
num_channels = 3
input_dim = image_height * image_width * num_channels
num_output_classes = 10
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width),
np.float32)
label_var = input_variable(num_output_classes, np.float32)
# Instantiate the feedforward classification model
input_removemean = minus(input_var, constant(128))
scaled_input = element_times(constant(0.00390625), input_removemean)
with default_options(activation=relu, pad=True):
z = Sequential([
LayerStack(
2, lambda: [
Convolution((3, 3), 64),
Convolution((3, 3), 64),
MaxPooling((3, 3), (2, 2))
]),
LayerStack(2, lambda i: [Dense([256, 128][i]),
Dropout(0.5)]),
Dense(num_output_classes, activation=None)
])(scaled_input)
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
reader_train = create_reader(
os.path.join(data_path, 'Train_cntk_text.txt'), True, input_dim,
num_output_classes)
# training config
epoch_size = 50000 # for now we manually specify epoch size
minibatch_size = 64
# Set learning parameters
lr_per_sample = [0.0015625] * 10 + [0.00046875] * 10 + [0.00015625]
lr_schedule = learning_rate_schedule(lr_per_sample,
epoch_size=epoch_size,
unit=UnitType.sample)
momentum_time_constant = [0] * 20 + [-minibatch_size / np.log(0.9)]
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant,
epoch_size=epoch_size)
l2_reg_weight = 0.002
# Instantiate the trainer object to drive the model training
learner = momentum_sgd(z.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# Get minibatches of images to train with and perform model training
max_epochs = 30
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(
min(minibatch_size, epoch_size - sample_count),
input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
persist.save_model(
z, os.path.join(model_path,
"ConvNet_CIFAR10_{}.dnn".format(epoch)))
# Load test data
reader_test = create_reader(os.path.join(data_path, 'Test_cntk_text.txt'),
False, input_dim, num_output_classes)
input_map = {
input_var: reader_test.streams.features,
label_var: reader_test.streams.labels
}
# Test data for trained model
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch,
input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
return metric_numer / metric_denom
def convnet_cifar10(debug_output=False):
set_computation_network_trace_level(0)
image_height = 32
image_width = 32
num_channels = 3
input_dim = image_height * image_width * num_channels
num_output_classes = 10
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width), np.float32)
label_var = input_variable(num_output_classes, np.float32)
# Instantiate the feedforward classification model
input_removemean = minus(input_var, constant(128))
scaled_input = element_times(constant(0.00390625), input_removemean)
with default_options (activation=relu, pad=True):
z = Sequential([
LayerStack(2, lambda : [
Convolution((3,3), 64),
Convolution((3,3), 64),
MaxPooling((3,3), (2,2))
]),
LayerStack(2, lambda i: [
Dense([256,128][i]),
Dropout(0.5)
]),
Dense(num_output_classes, activation=None)
])(scaled_input)
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
reader_train = create_reader(os.path.join(data_path, 'Train_cntk_text.txt'), True, input_dim, num_output_classes)
# training config
epoch_size = 50000 # for now we manually specify epoch size
minibatch_size = 64
# Set learning parameters
lr_per_sample = [0.0015625]*10+[0.00046875]*10+[0.00015625]
lr_schedule = learning_rate_schedule(lr_per_sample, epoch_size=epoch_size)
momentum_time_constant = [0]*20+[-minibatch_size/np.log(0.9)]
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant, epoch_size=epoch_size)
l2_reg_weight = 0.002
# Instantiate the trainer object to drive the model training
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule, l2_regularization_weight = l2_reg_weight)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var : reader_train.streams.features,
label_var : reader_train.streams.labels
}
log_number_of_parameters(z) ; print()
progress_printer = ProgressPrinter(tag='Training')
# Get minibatches of images to train with and perform model training
max_epochs = 30
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(min(minibatch_size, epoch_size - sample_count), input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += data[label_var].num_samples # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
persist.save_model(z, os.path.join(model_path, "ConvNet_CIFAR10_{}.dnn".format(epoch)))
# Load test data
reader_test = create_reader(os.path.join(data_path, 'Test_cntk_text.txt'), False, input_dim, num_output_classes)
input_map = {
input_var : reader_test.streams.features,
label_var : reader_test.streams.labels
}
# Test data for trained model
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch, input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(minibatch_index+1, (metric_numer*100.0)/metric_denom, metric_denom))
print("")
return metric_numer/metric_denom
def train_lm(training_file, max_num_minibatches):
# load the data and vocab
data, char_to_ix, ix_to_char, data_size, vocab_dim = load_data_and_vocab(
training_file)
# Model the source and target inputs to the model
input_sequence, label_sequence = create_inputs(vocab_dim)
# create the model
model = create_model(vocab_dim)
# and apply it to the input sequence
z = model(input_sequence)
# setup the criterions (loss and metric)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# Instantiate the trainer object to drive the model training
lr_per_sample = learning_rate_schedule(0.001, UnitType.sample)
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 5.0
gradient_clipping_with_truncation = True
learner = momentum_sgd(
z.parameters,
lr_per_sample,
momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
trainer = Trainer(z, (ce, errs), learner)
sample_freq = 1000
epochs = 50
minibatches_per_epoch = int((data_size / minibatch_size))
minibatches = min(epochs * minibatches_per_epoch, max_num_minibatches)
# print out some useful training information
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(freq=100, tag='Training')
e = 0
p = 0
for i in range(0, minibatches):
if p + minibatch_size + 1 >= data_size:
p = 0
e += 1
model_filename = "models/shakespeare_epoch%d.dnn" % e
z.save(model_filename)
print("Saved model to '%s'" % model_filename)
# get the data
features, labels = get_data(p, minibatch_size, data, char_to_ix,
vocab_dim)
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
# If it's the start of the data, we specify that we are looking at a new sequence (True)
mask = [False]
if p == 0:
mask = [True]
arguments = ({input_sequence: features, label_sequence: labels}, mask)
trainer.train_minibatch(arguments)
progress_printer.update_with_trainer(trainer,
with_metric=True) # log progress
if i % sample_freq == 0:
print(sample(z, ix_to_char, vocab_dim, char_to_ix))
p += minibatch_size
# Do a final save of the model
model_filename = "models/shakespeare_epoch%d.dnn" % e
z.save(model_filename)
def cifar_resnet(base_path, debug_output=False):
image_height = 32
image_width = 32
num_channels = 3
num_classes = 10
feats_stream_name = 'features'
labels_stream_name = 'labels'
minibatch_source = create_mb_source(feats_stream_name, labels_stream_name,
image_height, image_width,
num_channels, num_classes, base_path)
features_si = minibatch_source[feats_stream_name]
labels_si = minibatch_source[labels_stream_name]
# Input variables denoting the features and label data
image_input = input_variable((num_channels, image_height, image_width),
features_si.m_element_type)
label_var = input_variable((num_classes), features_si.m_element_type)
# Instantiate the resnet classification model
classifier_output = resnet_classifer(image_input, num_classes)
ce = cross_entropy_with_softmax(classifier_output, label_var)
pe = classification_error(classifier_output, label_var)
# Instantiate the trainer object to drive the model training
trainer = Trainer(classifier_output, ce, pe,
[sgd(classifier_output.parameters(), lr=0.0078125)])
# Get minibatches of images to train with and perform model training
mb_size = 32
training_progress_output_freq = 60
num_mbs = 1000
if debug_output:
training_progress_output_freq = training_progress_output_freq / 3
for i in range(0, num_mbs):
mb = minibatch_source.get_next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {image_input: mb[features_si], label_var: mb[labels_si]}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
test_minibatch_source = create_test_mb_source(feats_stream_name,
labels_stream_name,
image_height, image_width,
num_channels, num_classes,
base_path)
features_si = test_minibatch_source[feats_stream_name]
labels_si = test_minibatch_source[labels_stream_name]
mb_size = 64
num_mbs = 300
total_error = 0.0
for i in range(0, num_mbs):
mb = test_minibatch_source.get_next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {image_input: mb[features_si], label_var: mb[labels_si]}
error = trainer.test_minibatch(arguments)
total_error += error
return total_error / num_mbs
def train_and_evaluate(reader_train, reader_test, max_epochs):
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# Normalize the input
feature_scale = 1.0 / 256.0
input_var_norm = element_times(feature_scale, input_var)
# apply model to input
z = create_vgg9_model(input_var_norm, 10)
#
# Training action
#
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# training config
epoch_size = 50000
minibatch_size = 64
# Set learning parameters
lr_per_minibatch = learning_rate_schedule([0.01]*10 + [0.003]*10 + [0.001], epoch_size, UnitType.minibatch)
momentum_time_constant = momentum_as_time_constant_schedule(-minibatch_size/np.log(0.9))
l2_reg_weight = 0.0001
# trainer object
learner = momentum_sgd(z.parameters,
lr = lr_per_minibatch, momentum = momentum_time_constant,
l2_regularization_weight = l2_reg_weight)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z) ; print()
progress_printer = ProgressPrinter(tag='Training')
# perform model training
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(min(minibatch_size, epoch_size - sample_count), input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += data[label_var].num_samples # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
#
# Evaluation action
#
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
#progress_printer = ProgressPrinter(freq=100, first=10, tag='Eval')
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch, input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.1f}% * {}".format(minibatch_index+1, (metric_numer*100.0)/metric_denom, metric_denom))
print("")
# return evaluation error.
return metric_numer/metric_denom
def sequence_to_sequence_translator(debug_output=False):
input_vocab_dim = 69
label_vocab_dim = 69
hidden_dim = 512
num_layers = 2
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis("inputAxis")
label_seq_axis = Axis("labelAxis")
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(shape=(input_vocab_dim), dynamic_axes=input_dynamic_axes)
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(shape=(label_vocab_dim), dynamic_axes=label_dynamic_axes)
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = slice(raw_labels, label_seq_axis, 1, 0)
label_sentence_start = sequence.first(raw_labels)
is_first_label = sequence.is_first(label_sequence)
label_sentence_start_scattered = sequence.scatter(label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output(), hidden_dim, hidden_dim, future_value, future_value
)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(thought_vectorC, label_sequence)
# Decoder
decoder_history_from_ground_truth = label_sequence
decoder_input = element_select(
is_first_label, label_sentence_start_scattered, past_value(decoder_history_from_ground_truth)
)
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if i > 0:
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output(), hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC
)
decoder_output = decoder_outputH
decoder_dim = hidden_dim
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# Instantiate the trainer object to drive the model training
lr = 0.007
momentum_time_constant = 1100
momentum_per_sample = momentums_per_sample(math.exp(-1.0 / momentum_time_constant))
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
trainer = Trainer(
z,
ce,
errs,
[
momentum_sgd(
z.parameters(),
lr,
momentum_per_sample,
clipping_threshold_per_sample,
gradient_clipping_with_truncation,
)
],
)
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.train-dev-20-21.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
feature_stream_name = "features"
labels_stream_name = "labels"
mb_source = text_format_minibatch_source(
path,
[
StreamConfiguration(feature_stream_name, input_vocab_dim, True, "S0"),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, "S1"),
],
10000,
)
features_si = mb_source[feature_stream_name]
labels_si = mb_source[labels_stream_name]
# Get minibatches of sequences to train with and perform model training
minibatch_size = 72
training_progress_output_freq = 30
if debug_output:
training_progress_output_freq = training_progress_output_freq / 3
while True:
mb = mb_source.get_next_minibatch(minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {raw_input: mb[features_si], raw_labels: mb[labels_si]}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
i += 1
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.test.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
test_mb_source = text_format_minibatch_source(
path,
[
StreamConfiguration(feature_stream_name, input_vocab_dim, True, "S0"),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, "S1"),
],
10000,
False,
)
features_si = test_mb_source[feature_stream_name]
labels_si = test_mb_source[labels_stream_name]
# choose this to be big enough for the longest sentence
train_minibatch_size = 1024
# Get minibatches of sequences to test and perform testing
i = 0
total_error = 0.0
while True:
mb = test_mb_source.get_next_minibatch(train_minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be tested with
arguments = {raw_input: mb[features_si], raw_labels: mb[labels_si]}
mb_error = trainer.test_minibatch(arguments)
total_error += mb_error
if debug_output:
print("Minibatch {}, Error {} ".format(i, mb_error))
i += 1
# Average of evaluation errors of all test minibatches
return total_error / i
def train_and_evaluate(reader_train,
reader_test,
network_name,
max_epochs,
distributed_trainer,
scale_up=False):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# create model, and configure learning parameters
if network_name == 'resnet20':
z = create_cifar10_model(input_var, 3, num_classes)
lr_per_mb = [1.0] * 80 + [0.1] * 40 + [0.01]
elif network_name == 'resnet110':
z = create_cifar10_model(input_var, 18, num_classes)
lr_per_mb = [0.1] * 1 + [1.0] * 80 + [0.1] * 40 + [0.01]
else:
return RuntimeError("Unknown model name!")
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# shared training parameters
epoch_size = 50000 # for now we manually specify epoch size
# NOTE: scaling up minibatch_size increases sample throughput. In 8-GPU machine,
# ResNet110 samples-per-second is ~7x of single GPU, comparing to ~3x without scaling
# up. However, bigger minimatch size on the same number of samples means less updates,
# thus leads to higher training error. This is a trade-off of speed and accuracy
minibatch_size = 128 * (len(distributed_trainer.communicator().workers())
if scale_up else 1)
momentum_time_constant = -minibatch_size / np.log(0.9)
l2_reg_weight = 0.0001
# Set learning parameters
lr_per_sample = [lr / minibatch_size for lr in lr_per_mb]
lr_schedule = learning_rate_schedule(lr_per_sample,
epoch_size=epoch_size,
unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# trainer object
learner = momentum_sgd(z.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(z, ce, pe, learner, distributed_trainer)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# perform model training
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(
min(minibatch_size, epoch_size - sample_count),
input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
# save model only in worker0, otherwise there will be file write conflicts for multi GPU on the same machine
if distributed_trainer.communicator().current_worker(
).global_rank == 0:
z.save_model(
os.path.join(model_path,
network_name + "_{}.dnn".format(epoch)))
# Evaluation parameters
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch,
input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
return metric_numer / metric_denom
def train_and_evaluate(create_train_reader,
test_reader,
network_name,
max_epochs,
create_dist_learner,
scale_up=False):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# create model, and configure learning parameters
if network_name == 'resnet20':
z = create_cifar10_model(input_var, 3, num_classes)
lr_per_mb = [1.0] * 80 + [0.1] * 40 + [0.01]
elif network_name == 'resnet110':
z = create_cifar10_model(input_var, 18, num_classes)
lr_per_mb = [0.1] * 1 + [1.0] * 80 + [0.1] * 40 + [0.01]
else:
return RuntimeError("Unknown model name!")
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# shared training parameters
epoch_size = 50000 # for now we manually specify epoch size
# NOTE: scaling up minibatch_size increases sample throughput. In 8-GPU machine,
# ResNet110 samples-per-second is ~7x of single GPU, comparing to ~3x without scaling
# up. However, bigger minimatch size on the same number of samples means less updates,
# thus leads to higher training error. This is a trade-off of speed and accuracy
minibatch_size = 128 * (distributed.Communicator.num_workers()
if scale_up else 1)
momentum_time_constant = -minibatch_size / np.log(0.9)
l2_reg_weight = 0.0001
# Set learning parameters
lr_per_sample = [lr / minibatch_size for lr in lr_per_mb]
lr_schedule = learning_rate_schedule(lr_per_sample,
epoch_size=epoch_size,
unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# trainer object
learner = create_dist_learner(
momentum_sgd(z.parameters,
lr_schedule,
mm_schedule,
unit_gain=True,
l2_regularization_weight=l2_reg_weight))
trainer = Trainer(z, ce, pe, learner)
total_number_of_samples = max_epochs * epoch_size
train_reader = create_train_reader(total_number_of_samples)
# define mapping from reader streams to network inputs
input_map = {
input_var: train_reader.streams.features,
label_var: train_reader.streams.labels
}
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# perform model training
current_epoch = 0
updated = True
while updated:
data = train_reader.next_minibatch(
minibatch_size, input_map=input_map) # fetch minibatch.
updated = trainer.train_minibatch(data) # update model with it
progress_printer.update_with_trainer(trainer,
with_metric=True) # log progress
epoch_index = int(trainer.total_number_of_samples_seen / epoch_size)
if current_epoch != epoch_index: # new epoch reached
progress_printer.epoch_summary(with_metric=True)
current_epoch = epoch_index
trainer.save_checkpoint(
os.path.join(model_path,
network_name + "_{}.dnn".format(current_epoch)))
# Evaluation parameters
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while True:
data = test_reader.next_minibatch(minibatch_size, input_map=input_map)
if not data: break
local_mb_samples = data[label_var].num_samples
metric_numer += trainer.test_minibatch(data) * local_mb_samples
metric_denom += local_mb_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
return metric_numer / metric_denom
def cifar_resnet_distributed(data_path,
run_test,
num_epochs,
communicator=None,
save_model_filename=None,
load_model_filename=None,
debug_output=False):
image_height = 32
image_width = 32
num_channels = 3
num_classes = 10
feats_stream_name = 'features'
labels_stream_name = 'labels'
minibatch_source = create_reader(os.path.join(data_path, 'train_map.txt'),
os.path.join(data_path,
'CIFAR-10_mean.xml'),
True,
distributed_communicator=communicator)
features_si = minibatch_source[feats_stream_name]
labels_si = minibatch_source[labels_stream_name]
# Instantiate the resnet classification model, or load from file
if load_model_filename:
print("Loading model:", load_model_filename)
classifier_output = persist.load_model(load_model_filename)
image_input = classifier_output.arguments[0]
else:
image_input = input_variable((num_channels, image_height, image_width),
features_si.m_element_type)
classifier_output = create_resnet_model(image_input, num_classes)
# Input variables denoting the features and label data
label_var = input_variable((num_classes), features_si.m_element_type)
ce = cross_entropy_with_softmax(classifier_output, label_var)
pe = classification_error(classifier_output, label_var)
# Instantiate the trainer object to drive the model training
mb_size = 128
num_mb_per_epoch = 100
num_mbs = num_mb_per_epoch * num_epochs
lr_schedule = [1.0 / mb_size] * 80 + [0.1 / mb_size] * 40 + [
0.01 / mb_size
]
lr_per_minibatch = learning_rate_schedule(lr_schedule, UnitType.minibatch,
mb_size * num_mb_per_epoch)
momentum_time_constant = momentum_as_time_constant_schedule(-mb_size /
np.log(0.9))
# create data parallel distributed trainer if needed
dist_trainer = distributed.data_parallel_distributed_trainer(
communicator, False) if communicator else None
# Instantiate the trainer object to drive the model training
trainer = Trainer(classifier_output,
ce,
pe, [
momentum_sgd(classifier_output.parameters,
lr=lr_per_minibatch,
momentum=momentum_time_constant,
l2_regularization_weight=0.0001)
],
distributed_trainer=dist_trainer)
# Get minibatches of images to train with and perform model training
training_progress_output_freq = 100 if communicator else 20
if debug_output:
training_progress_output_freq = training_progress_output_freq / 4
for i in range(0, num_mbs):
# NOTE: depends on network, the mb_size can be changed dynamically here
mb = minibatch_source.next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {image_input: mb[features_si], label_var: mb[labels_si]}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
if save_model_filename:
print("Saving model:", save_model_filename)
persist.save_model(classifier_output, save_model_filename)
if run_test:
test_minibatch_source = create_reader(
os.path.join(data_path, 'test_map.txt'),
os.path.join(data_path, 'CIFAR-10_mean.xml'), False)
features_si = test_minibatch_source[feats_stream_name]
labels_si = test_minibatch_source[labels_stream_name]
mb_size = 128
num_mbs = 100
total_error = 0.0
for i in range(0, num_mbs):
mb = test_minibatch_source.next_minibatch(mb_size)
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {
image_input: mb[features_si],
label_var: mb[labels_si]
}
error = trainer.test_minibatch(arguments)
total_error += error
return total_error / num_mbs
else:
return 0
def conv3d_ucf11(train_reader, test_reader, max_epochs=30):
# Replace 0 with 1 to get detailed log.
set_computation_network_trace_level(0)
# These values must match for both train and test reader.
image_height = train_reader.height
image_width = train_reader.width
num_channels = train_reader.channel_count
sequence_length = train_reader.sequence_length
num_output_classes = train_reader.label_count
# Input variables denoting the features and label data
input_var = input_variable(
(num_channels, sequence_length, image_height, image_width), np.float32)
label_var = input_variable(num_output_classes, np.float32)
# Instantiate simple 3D Convolution network inspired by VGG network
# and http://vlg.cs.dartmouth.edu/c3d/c3d_video.pdf
with default_options(activation=relu):
z = Sequential([
Convolution((3, 3, 3), 64, pad=True),
MaxPooling((1, 2, 2), (1, 2, 2)),
LayerStack(
3, lambda i: [
Convolution((3, 3, 3), [96, 128, 128][i], pad=True),
Convolution((3, 3, 3), [96, 128, 128][i], pad=True),
MaxPooling((2, 2, 2), (2, 2, 2))
]),
LayerStack(2, lambda: [Dense(1024), Dropout(0.5)]),
Dense(num_output_classes, activation=None)
])(input_var)
# loss and classification error.
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# training config
epoch_size = 1322 # for now we manually specify epoch size
minibatch_size = 4
# Set learning parameters
lr_per_sample = [0.01] * 10 + [0.001] * 10 + [0.0001]
lr_schedule = learning_rate_schedule(lr_per_sample,
epoch_size=epoch_size,
unit=UnitType.sample)
momentum_time_constant = 4096
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant,
epoch_size=epoch_size)
# Instantiate the trainer object to drive the model training
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule, True)
trainer = Trainer(z, ce, pe, learner)
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# Get minibatches of images to train with and perform model training
for epoch in range(max_epochs): # loop over epochs
train_reader.reset()
while train_reader.has_more():
videos, labels, current_minibatch = train_reader.next_minibatch(
minibatch_size)
trainer.train_minibatch({input_var: videos, label_var: labels})
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
# Test data for trained model
epoch_size = 332
minibatch_size = 2
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
minibatch_index = 0
test_reader.reset()
while test_reader.has_more():
videos, labels, current_minibatch = test_reader.next_minibatch(
minibatch_size)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch({
input_var: videos,
label_var: labels
}) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
return metric_numer / metric_denom
def convnet_mnist(debug_output=False):
image_height = 28
image_width = 28
num_channels = 1
input_dim = image_height * image_width * num_channels
num_output_classes = 10
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width),
np.float32)
label_var = input_variable(num_output_classes, np.float32)
# Instantiate the feedforward classification model
scaled_input = element_times(constant(0.00390625), input_var)
with default_options(activation=relu, pad=False):
conv1 = Convolution((5, 5), 32, pad=True)(scaled_input)
pool1 = MaxPooling((3, 3), (2, 2))(conv1)
conv2 = Convolution((3, 3), 48)(pool1)
pool2 = MaxPooling((3, 3), (2, 2))(conv2)
conv3 = Convolution((3, 3), 64)(pool2)
f4 = Dense(96)(conv3)
drop4 = Dropout(0.5)(f4)
z = Dense(num_output_classes, activation=None)(drop4)
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
reader_train = create_reader(
os.path.join(data_path, 'Train-28x28_cntk_text.txt'), True, input_dim,
num_output_classes)
# training config
epoch_size = 60000 # for now we manually specify epoch size
minibatch_size = 128
# Set learning parameters
lr_per_sample = [0.001] * 10 + [0.0005] * 10 + [0.0001]
lr_schedule = learning_rate_schedule(lr_per_sample, UnitType.sample,
epoch_size)
mm_time_constant = [0] * 5 + [1024]
mm_schedule = momentum_as_time_constant_schedule(mm_time_constant,
epoch_size)
# Instantiate the trainer object to drive the model training
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# Get minibatches of images to train with and perform model training
max_epochs = 40
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(
min(minibatch_size, epoch_size - sample_count),
input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += data[
label_var].num_samples # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
z.save_model(
os.path.join(model_path, "ConvNet_MNIST_{}.dnn".format(epoch)))
# Load test data
reader_test = create_reader(
os.path.join(data_path, 'Test-28x28_cntk_text.txt'), False, input_dim,
num_output_classes)
input_map = {
input_var: reader_test.streams.features,
label_var: reader_test.streams.labels
}
# Test data for trained model
epoch_size = 10000
minibatch_size = 1024
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch,
input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += trainer.previous_minibatch_sample_count
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
return metric_numer / metric_denom
def train_and_evaluate(reader_train, reader_test, network_name):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width))
label_var = input_variable((num_classes))
# create model, and configure learning parameters
if network_name == 'resnet20':
z = create_cifar10_model(input_var, 3, num_classes)
lr_per_mb = [1.0]*80+[0.1]*40+[0.01]
elif network_name == 'resnet110':
z = create_cifar10_model(input_var, 18, num_classes)
lr_per_mb = [0.1]*1+[1.0]*80+[0.1]*40+[0.01]
else:
return RuntimeError("Unknown model name!")
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
# shared training parameters
epoch_size = 50000 # for now we manually specify epoch size
minibatch_size = 128
max_epochs = 160
momentum_time_constant = -minibatch_size/np.log(0.9)
l2_reg_weight = 0.0001
# Set learning parameters
lr_per_sample = [lr/minibatch_size for lr in lr_per_mb]
lr_schedule = learning_rate_schedule(lr_per_sample, epoch_size=epoch_size)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# trainer object
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule,
l2_regularization_weight = l2_reg_weight)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z) ; print()
progress_printer = ProgressPrinter(tag='Training')
# perform model training
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(min(minibatch_size, epoch_size-sample_count), input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += data[label_var].num_samples # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
persist.save_model(z, os.path.join(model_path, network_name + "_{}.dnn".format(epoch)))
# Evaluation parameters
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch, input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(minibatch_index+1, (metric_numer*100.0)/metric_denom, metric_denom))
print("")
return metric_numer/metric_denom
