def create_trainer(network, minibatch_size, epoch_size, num_quantization_bits, block_size, warm_up, progress_printer):
if network['name'] == 'resnet20':
lr_per_mb = [1.0]*80+[0.1]*40+[0.01]
elif network['name'] == 'resnet110':
lr_per_mb = [0.1]*1+[1.0]*80+[0.1]*40+[0.01]
else:
return RuntimeError("Unknown model name!")
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)
# learner object
if block_size != None and num_quantization_bits != 32:
raise RuntimeError("Block momentum cannot be used with quantization, please remove quantized_bits option.")
local_learner = momentum_sgd(network['output'].parameters, lr_schedule, mm_schedule,
l2_regularization_weight = l2_reg_weight)
if block_size != None:
learner = block_momentum_distributed_learner(local_learner, block_size=block_size)
else:
learner = data_parallel_distributed_learner(local_learner, num_quantization_bits=num_quantization_bits, distributed_after=warm_up)
return Trainer(network['output'], (network['ce'], network['pe']), learner, progress_printer)
def create_trainer(network, minibatch_size, epoch_size, num_quantization_bits, block_size, warm_up):
if network['name'] == 'resnet20':
lr_per_mb = [1.0]*80+[0.1]*40+[0.01]
elif network['name'] == 'resnet110':
lr_per_mb = [0.1]*1+[1.0]*80+[0.1]*40+[0.01]
else:
return RuntimeError("Unknown model name!")
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)
# learner object
if block_size != None and num_quantization_bits != 32:
raise RuntimeError("Block momentum cannot be used with quantization, please remove quantized_bits option.")
local_learner = momentum_sgd(network['output'].parameters, lr_schedule, mm_schedule,
l2_regularization_weight = l2_reg_weight)
if block_size != None:
learner = block_momentum_distributed_learner(local_learner, block_size=block_size)
else:
learner = data_parallel_distributed_learner(local_learner, num_quantization_bits=num_quantization_bits, distributed_after=warm_up)
return Trainer(network['output'], (network['ce'], network['pe']), learner)
def create_trainer(network, minibatch_size, epoch_size, num_quantization_bits):
if network['name'] == 'resnet20':
lr_per_mb = [1.0] * 80 + [0.1] * 40 + [0.01]
elif network['name'] == 'resnet110':
lr_per_mb = [0.1] * 1 + [1.0] * 80 + [0.1] * 40 + [0.01]
else:
return RuntimeError("Unknown model name!")
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)
# learner object
local_learner = momentum_sgd(network['output'].parameters,
lr_schedule,
mm_schedule,
unit_gain=True,
l2_regularization_weight=l2_reg_weight)
learner = data_parallel_distributed_learner(
learner=local_learner,
num_quantization_bits=num_quantization_bits,
distributed_after=0)
return Trainer(network['output'], network['ce'], network['pe'], learner)
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 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 += 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 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_lm():
data = DataReader(token_to_id_path, segment_sepparator)
# Create model nodes for the source and target inputs
input_sequence, label_sequence = create_inputs(data.vocab_dim)
# Create the model. It has three output nodes
# z: the input to softmax that provides the latent representation of the next token
# cross_entropy: this is used training criterion
# error: this a binary indicator if the model predicts the correct token
z, cross_entropy, error = create_model(input_sequence, label_sequence, data.vocab_dim, hidden_dim)
# For measurement we use the (build in) full softmax.
full_ce = C.cross_entropy_with_softmax(z, label_sequence)
# print out some useful training information
log_number_of_parameters(z) ; print()
# Run the training loop
num_trained_samples = 0
num_trained_samples_since_last_report = 0
# Instantiate the trainer object to drive the model training
lr_schedule = learning_rate_schedule(learning_rate, UnitType.sample)
momentum_schedule = momentum_as_time_constant_schedule(momentum_as_time_constant)
gradient_clipping_with_truncation = True
learner = momentum_sgd(z.parameters, lr_schedule, momentum_schedule,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
trainer = Trainer(z, (cross_entropy, error), learner)
for epoch_count in range(num_epochs):
for features, labels, token_count in data.minibatch_generator(train_file_path, sequence_length, sequences_per_batch):
arguments = ({input_sequence : features, label_sequence : labels})
t_start = timeit.default_timer()
trainer.train_minibatch(arguments)
t_end = timeit.default_timer()
samples_per_second = token_count / (t_end - t_start)
# Print progress report every num_samples_between_progress_report samples
if num_trained_samples_since_last_report >= num_samples_between_progress_report or num_trained_samples == 0:
av_ce = average_cross_entropy(full_ce, input_sequence, label_sequence, data)
print_progress(samples_per_second, av_ce, num_trained_samples, t_start)
num_trained_samples_since_last_report = 0
num_trained_samples += token_count
num_trained_samples_since_last_report += token_count
# after each epoch save the model
model_filename = "models/lm_epoch%d.dnn" % epoch_count
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(reader, model, max_epochs):
# declare the model's input dimension, so that the saved model is usable
model.update_signature(Sequence[SparseTensor[vocab_size]])
#model.declare_args(vocab_size)
# criterion: (model args, labels) -> (loss, metric)
# here (query, slot_labels) -> (ce, errs)
criterion = create_criterion_function(model)
labels = reader.streams.slot_labels
#labels = reader.streams.intent_labels # for intent classification
#from cntk.graph import plot
#plot(criterion, filename=data_dir + "/model.pdf")
# iteration parameters --needed here because learner schedule needs it
epoch_size = 36000
minibatch_size = 70
#epoch_size = 1000 ; max_epochs = 1 # uncomment for faster testing
# SGD parameters
learner = adam_sgd(criterion.parameters,
lr = learning_rate_schedule([0.003]*2+[0.0015]*12+[0.0003], UnitType.sample, epoch_size),
momentum = momentum_as_time_constant_schedule(minibatch_size / -math.log(0.9)),
low_memory = True,
gradient_clipping_threshold_per_sample = 15,
gradient_clipping_with_truncation = True)
# trainer
trainer = Trainer(None, criterion, learner)
# process minibatches and perform model training
log_number_of_parameters(model) ; 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
peek(model, epoch) # log some interesting info
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.
# TODO: change all examples to this pattern; then remove this comment
data = reader.next_minibatch(min(minibatch_size, epoch_end-t)) # fetch minibatch
#trainer.train_minibatch(data[reader.streams.query], data[labels]) # update model with it
trainer.train_minibatch({criterion.arguments[0]: data[reader.streams.query], criterion.arguments[1]: data[labels]}) # update model with it
t += data[labels].num_samples # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
loss, metric, actual_samples = progress_printer.epoch_summary(with_metric=True)
return loss, metric # return values from last epoch
def Evaluator(model, criterion):
from cntk import Trainer
from cntk.learner import momentum_sgd, learning_rate_schedule, UnitType, momentum_as_time_constant_schedule
loss, metric = Trainer._get_loss_metric(criterion)
parameters = set(loss.parameters)
if model:
parameters |= set(model.parameters)
if metric:
parameters |= set(metric.parameters)
dummy_learner = momentum_sgd(tuple(parameters),
lr = learning_rate_schedule(1, UnitType.minibatch),
momentum = momentum_as_time_constant_schedule(0))
return Trainer(model, (loss, metric), dummy_learner)
def Evaluator(model, criterion):
from cntk import Trainer
from cntk.learner import momentum_sgd, learning_rate_schedule, UnitType, momentum_as_time_constant_schedule
loss, metric = Trainer._get_loss_metric(criterion)
parameters = set(loss.parameters)
if model:
parameters |= set(model.parameters)
if metric:
parameters |= set(metric.parameters)
dummy_learner = momentum_sgd(
tuple(parameters),
lr=learning_rate_schedule(1, UnitType.minibatch),
momentum=momentum_as_time_constant_schedule(0))
return Trainer(model, (loss, metric), dummy_learner)
def create_trainer():
# Will take the model and the batch generator to create a Trainer
# Will return the input variables, trainer variable, model and the embedding layer
##################################################
################### Inputs #######################
##################################################
word_one_hot = C.input_variable((G.embedding_vocab_size),
np.float32,
is_sparse=True,
name='word_input')
context_one_hots = [
C.input_variable((G.embedding_vocab_size),
np.float32,
is_sparse=True,
name='context_input{}'.format(i))
for i in range(context_size)
]
negative_one_hots = [
C.input_variable((G.embedding_vocab_size),
np.float32,
is_sparse=True,
name='negative_input{}'.format(i))
for i in range(G.negative)
]
# The target labels should have first as 1 and rest as 0
target = C.input_variable((G.negative + 1), np.float32)
word_negative_context_product, embedding_layer = create_word2vec_cbow_model(
word_one_hot, context_one_hots, negative_one_hots)
loss = C.binary_cross_entropy(word_negative_context_product, target)
eval_loss = C.binary_cross_entropy(word_negative_context_product, target)
lr_schedule = learning_rate_schedule(G.learning_rate, UnitType.minibatch)
learner = adam_sgd(word_negative_context_product.parameters,
lr=lr_schedule,
momentum=momentum_as_time_constant_schedule(700))
trainer = Trainer(word_negative_context_product, (loss, eval_loss),
learner)
return word_one_hot, context_one_hots, negative_one_hots, target, trainer, word_negative_context_product, embedding_layer
def create_trainer(network, epoch_size, num_quantization_bits, block_size, warm_up):
# 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
# Create learner
if block_size is not None and num_quantization_bits != default_quantization_bits:
raise RuntimeError("Block momentum cannot be used with quantization, please remove quantized_bits option.")
local_learner = momentum_sgd(network['output'].parameters,
lr_per_minibatch, momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
if block_size != None:
learner = block_momentum_distributed_learner(local_learner, block_size=block_size)
else:
learner = data_parallel_distributed_learner(local_learner, num_quantization_bits=num_quantization_bits, distributed_after=warm_up)
return Trainer(network['output'], (network['ce'], network['pe']), learner)
def train(train_reader, valid_reader, vocab, i2w, s2smodel, max_epochs,
epoch_size):
# Note: We would like to set the signature of 's2smodel' (s2smodel.update_signature()), but that will cause
# an error since the training criterion uses a reduced sequence axis for the labels.
# This is because it removes the initial <s> symbol. Hence, we must leave the model
# with unspecified input shapes and axes.
# create the training wrapper for the s2smodel, as well as the criterion function
model_train = create_model_train(s2smodel)
criterion = create_criterion_function(model_train)
# also wire in a greedy decoder so that we can properly log progress on a validation example
# This is not used for the actual training process.
model_greedy = create_model_greedy(s2smodel)
# This does not need to be done in training generally though
# Instantiate the trainer object to drive the model training
minibatch_size = 72
lr = 0.001 if use_attention else 0.005 # TODO: can we use the same value for both?
learner = adam_sgd(
model_train.parameters,
lr=learning_rate_schedule([lr] * 2 + [lr / 2] * 3 + [lr / 4],
UnitType.sample, epoch_size),
momentum=momentum_as_time_constant_schedule(1100),
gradient_clipping_threshold_per_sample=2.3,
gradient_clipping_with_truncation=True)
trainer = Trainer(None, criterion, learner)
# Get minibatches of sequences to train with and perform model training
total_samples = 0
mbs = 0
eval_freq = 100
# print out some useful training information
log_number_of_parameters(model_train)
print()
progress_printer = ProgressPrinter(freq=30, tag='Training')
#progress_printer = ProgressPrinter(freq=30, tag='Training', log_to_file=model_path_stem + ".log") # use this to log to file
sparse_to_dense = create_sparse_to_dense(input_vocab_dim)
for epoch in range(max_epochs):
print("Saving model to '%s'" % model_path(epoch))
s2smodel.save(model_path(epoch))
while total_samples < (epoch + 1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size)
#trainer.train_minibatch(mb_train[train_reader.streams.features], mb_train[train_reader.streams.labels])
trainer.train_minibatch({
criterion.arguments[0]:
mb_train[train_reader.streams.features],
criterion.arguments[1]:
mb_train[train_reader.streams.labels]
})
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % eval_freq == 0:
mb_valid = valid_reader.next_minibatch(1)
# run an eval on the decoder output model (i.e. don't use the groundtruth)
e = model_greedy(mb_valid[valid_reader.streams.features])
print(
format_sequences(
sparse_to_dense(
mb_valid[valid_reader.streams.features]), i2w))
print("->")
print(format_sequences(e, i2w))
# debugging attention
if use_attention:
debug_attention(model_greedy,
mb_valid[valid_reader.streams.features])
total_samples += mb_train[train_reader.streams.labels].num_samples
mbs += 1
# log a summary of the stats for the epoch
progress_printer.epoch_summary(with_metric=True)
# done: save the final model
print("Saving final model to '%s'" % model_path(max_epochs))
s2smodel.save(model_path(max_epochs))
print("%d epochs complete." % max_epochs)
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,
reader_test,
model,
epoch_size=50000,
max_epochs=5):
# declare the model's input dimension
# Training does not require this, but it is needed for deployment.
model.update_signature((num_channels, image_height, image_width))
# criterion function. This is what is being trained trained.
# Model gets "sandwiched" between normalization (not part of model proper) and criterion.
criterion = create_criterion_function(model, normalize=lambda x: x / 256)
#debughelpers.dump_function(criterion, 'criterion')
#from cntk.graph import plot
#plot(criterion, filename=os.path.join(model_path, "ConvNet_CIFAR10_DataAug.pdf"))
# iteration parameters
minibatch_size = 64
#epoch_size = 1000 ; max_epochs = 1 # for faster testing
# learning parameters
learner = momentum_sgd(
model.parameters,
lr=learning_rate_schedule([0.0015625] * 20 + [0.00046875] * 20 +
[0.00015625] * 20 + [0.000046875] * 10 +
[0.000015625],
unit=UnitType.sample,
epoch_size=epoch_size),
momentum=momentum_as_time_constant_schedule([0] * 20 + [600] * 20 +
[1200],
epoch_size=epoch_size),
l2_regularization_weight=0.002)
# trainer object
trainer = Trainer(None, criterion, learner)
# perform model training
log_number_of_parameters(model)
print()
progress_printer = ProgressPrinter(tag='Training')
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
mb = reader.next_minibatch(
min(minibatch_size,
epoch_size - sample_count)) # fetch minibatch.
#trainer.train_minibatch(mb[reader.streams.features], mb[reader.streams.labels])
trainer.train_minibatch({
criterion.arguments[0]:
mb[reader.streams.features],
criterion.arguments[1]:
mb[reader.streams.labels]
})
sample_count += mb[
reader.streams.
labels].num_samples # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
loss, metric, actual_samples = progress_printer.epoch_summary(
with_metric=True)
model.save(
os.path.join(model_path,
"ConvNet_CIFAR10_DataAug_{}.dnn".format(epoch)))
progress_printer.epoch_summary(with_metric=True)
# TODO: we should be done here
#return metric_numer/metric_denom
### Evaluation action
# evaluate with current Trainer instance; just to make sure we save and load the model correctly and BN works now --TODO: delete once confirmed
epoch_size = 10000
minibatch_size = 16
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
mbsize = min(minibatch_size, epoch_size - sample_count)
mb = reader_test.next_minibatch(mbsize)
metric_numer += mbsize * trainer.test_minibatch(
{
criterion.arguments[0]: mb[reader_test.streams.features],
criterion.arguments[1]: mb[reader_test.streams.labels]
})
metric_denom += mbsize
sample_count += mb[reader_test.streams.labels].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 evaluation error.
return loss, metric # return values from last epoch
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 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 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 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 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 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(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)
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)
trainer = Trainer(model, ce, errs, learner)
# helper function to find variables by name
def find_arg_by_name(name, expression):
vars = [i for i in expression.arguments if i.name == name]
assert len(vars) == 1
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)
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 += 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_lm():
data = DataReader(token_to_id_path, segment_sepparator)
# Create model nodes for the source and target inputs
input_sequence, label_sequence = create_inputs(data.vocab_dim)
# Create the model. It has three output nodes
# z: the input to softmax that provides the latent representation of the next token
# cross_entropy: this is used training criterion
# error: this a binary indicator if the model predicts the correct token
z, cross_entropy, error = create_model(input_sequence, label_sequence,
data.vocab_dim, hidden_dim)
# For measurement we use the (build in) full softmax.
full_ce = C.cross_entropy_with_softmax(z, label_sequence)
# print out some useful training information
log_number_of_parameters(z)
print()
# Run the training loop
num_trained_samples = 0
num_trained_samples_since_last_report = 0
# Instantiate the trainer object to drive the model training
lr_schedule = learning_rate_schedule(learning_rate, UnitType.sample)
momentum_schedule = momentum_as_time_constant_schedule(
momentum_as_time_constant)
gradient_clipping_with_truncation = True
learner = momentum_sgd(
z.parameters,
lr_schedule,
momentum_schedule,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
trainer = Trainer(z, (cross_entropy, error), learner)
for epoch_count in range(num_epochs):
for features, labels, token_count in data.minibatch_generator(
train_file_path, sequence_length, sequences_per_batch):
arguments = ({input_sequence: features, label_sequence: labels})
t_start = timeit.default_timer()
trainer.train_minibatch(arguments)
t_end = timeit.default_timer()
samples_per_second = token_count / (t_end - t_start)
# Print progress report every num_samples_between_progress_report samples
if num_trained_samples_since_last_report >= num_samples_between_progress_report or num_trained_samples == 0:
av_ce = average_cross_entropy(full_ce, input_sequence,
label_sequence, data)
print_progress(samples_per_second, av_ce, num_trained_samples,
t_start)
num_trained_samples_since_last_report = 0
num_trained_samples += token_count
num_trained_samples_since_last_report += token_count
# after each epoch save the model
model_filename = "models/lm_epoch%d.dnn" % epoch_count
z.save_model(model_filename)
print("Saved model to '%s'" % model_filename)
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 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_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 train(reader, model, max_epochs):
# declare the model's input dimension, so that the saved model is usable
model.update_signature(Sequence[SparseTensor[vocab_size]])
#model.declare_args(vocab_size)
# criterion: (model args, labels) -> (loss, metric)
# here (query, slot_labels) -> (ce, errs)
criterion = create_criterion_function(model)
labels = reader.streams.slot_labels
#labels = reader.streams.intent_labels # for intent classification
#from cntk.graph import plot
#plot(criterion, filename=data_dir + "/model.pdf")
# iteration parameters --needed here because learner schedule needs it
epoch_size = 36000
minibatch_size = 70
#epoch_size = 1000 ; max_epochs = 1 # uncomment for faster testing
# SGD parameters
learner = adam_sgd(
criterion.parameters,
lr=learning_rate_schedule([0.003] * 2 + [0.0015] * 12 + [0.0003],
UnitType.sample, epoch_size),
momentum=momentum_as_time_constant_schedule(minibatch_size /
-math.log(0.9)),
low_memory=True,
gradient_clipping_threshold_per_sample=15,
gradient_clipping_with_truncation=True)
# trainer
trainer = Trainer(None, criterion, learner)
# process minibatches and perform model training
log_number_of_parameters(model)
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
peek(model, epoch) # log some interesting info
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.
# TODO: change all examples to this pattern; then remove this comment
data = reader.next_minibatch(min(minibatch_size,
epoch_end - t)) # fetch minibatch
#trainer.train_minibatch(data[reader.streams.query], data[labels]) # update model with it
trainer.train_minibatch({
criterion.arguments[0]:
data[reader.streams.query],
criterion.arguments[1]:
data[labels]
}) # update model with it
t += data[labels].num_samples # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
loss, metric, actual_samples = progress_printer.epoch_summary(
with_metric=True)
return loss, metric # return values from last epoch
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,
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 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 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 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 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 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
def main(base_folder, training_mode='majority', model_name='VGG13', max_epochs = 100):
# create needed folders.
output_model_path = os.path.join(base_folder, R'models')
output_model_folder = os.path.join(output_model_path, model_name + '_' + training_mode)
if not os.path.exists(output_model_folder):
os.makedirs(output_model_folder)
# creating logging file
logging.basicConfig(filename = os.path.join(output_model_folder, "train.log"), filemode = 'w', level = logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
logging.info("Starting with training mode {} using {} model and max epochs {}.".format(training_mode, model_name, max_epochs))
# create the model
num_classes = len(emotion_table)
model = build_model(num_classes, model_name)
# set the input variables.
input_var = ct.input_variable((1, model.input_height, model.input_width), np.float32)
label_var = ct.input_variable((num_classes), np.float32)
# read FER+ dataset.
logging.info("Loading data...")
train_params = FERPlusParameters(num_classes, model.input_height, model.input_width, training_mode, False)
test_and_val_params = FERPlusParameters(num_classes, model.input_height, model.input_width, "majority", True)
train_data_reader = FERPlusReader.create(base_folder, train_folders, "label.csv", train_params)
val_data_reader = FERPlusReader.create(base_folder, valid_folders, "label.csv", test_and_val_params)
test_data_reader = FERPlusReader.create(base_folder, test_folders, "label.csv", test_and_val_params)
# print summary of the data.
display_summary(train_data_reader, val_data_reader, test_data_reader)
# get the probalistic output of the model.
z = model.model(input_var)
pred = ct.softmax(z)
epoch_size = train_data_reader.size()
minibatch_size = 32
# Training config
lr_per_minibatch = [model.learning_rate]*20 + [model.learning_rate / 2.0]*20 + [model.learning_rate / 10.0]
mm_time_constant = -minibatch_size/np.log(0.9)
lr_schedule = learning_rate_schedule(lr_per_minibatch, unit=UnitType.minibatch, epoch_size=epoch_size)
mm_schedule = momentum_as_time_constant_schedule(mm_time_constant)
# loss and error cost
train_loss = cost_func(training_mode, pred, label_var)
pe = classification_error(z, label_var)
# construct the trainer
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule)
trainer = Trainer(z, (train_loss, pe), learner)
# Get minibatches of images to train with and perform model training
max_val_accuracy = 0.0
final_test_accuracy = 0.0
best_test_accuracy = 0.0
logging.info("Start training...")
epoch = 0
best_epoch = 0
while epoch < max_epochs:
train_data_reader.reset()
val_data_reader.reset()
test_data_reader.reset()
# Training
start_time = time.time()
training_loss = 0
training_accuracy = 0
while train_data_reader.has_more():
images, labels, current_batch_size = train_data_reader.next_minibatch(minibatch_size)
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
trainer.train_minibatch({input_var : images, label_var : labels})
# keep track of statistics.
training_loss += trainer.previous_minibatch_loss_average * current_batch_size
training_accuracy += trainer.previous_minibatch_evaluation_average * current_batch_size
training_accuracy /= train_data_reader.size()
training_accuracy = 1.0 - training_accuracy
# Validation
val_accuracy = 0
while val_data_reader.has_more():
images, labels, current_batch_size = val_data_reader.next_minibatch(minibatch_size)
val_accuracy += trainer.test_minibatch({input_var : images, label_var : labels}) * current_batch_size
val_accuracy /= val_data_reader.size()
val_accuracy = 1.0 - val_accuracy
# if validation accuracy goes higher, we compute test accuracy
test_run = False
if val_accuracy > max_val_accuracy:
best_epoch = epoch
max_val_accuracy = val_accuracy
trainer.save_checkpoint(os.path.join(output_model_folder, "model_{}".format(best_epoch)))
test_run = True
test_accuracy = 0
while test_data_reader.has_more():
images, labels, current_batch_size = test_data_reader.next_minibatch(minibatch_size)
test_accuracy += trainer.test_minibatch({input_var : images, label_var : labels}) * current_batch_size
test_accuracy /= test_data_reader.size()
test_accuracy = 1.0 - test_accuracy
final_test_accuracy = test_accuracy
if final_test_accuracy > best_test_accuracy:
best_test_accuracy = final_test_accuracy
logging.info("Epoch {}: took {:.3f}s".format(epoch, time.time() - start_time))
logging.info(" training loss:\t{:e}".format(training_loss))
logging.info(" training accuracy:\t\t{:.2f} %".format(training_accuracy * 100))
logging.info(" validation accuracy:\t\t{:.2f} %".format(val_accuracy * 100))
if test_run:
logging.info(" test accuracy:\t\t{:.2f} %".format(test_accuracy * 100))
epoch += 1
logging.info("")
logging.info("Best validation accuracy:\t\t{:.2f} %, epoch {}".format(max_val_accuracy * 100, best_epoch))
logging.info("Test accuracy corresponding to best validation:\t\t{:.2f} %".format(final_test_accuracy * 100))
logging.info("Best test accuracy:\t\t{:.2f} %".format(best_test_accuracy * 100))
def train(train_reader, valid_reader, vocab, i2w, s2smodel, max_epochs, epoch_size):
# Note: We would like to set the signature of 's2smodel' (s2smodel.update_signature()), but that will cause
# an error since the training criterion uses a reduced sequence axis for the labels.
# This is because it removes the initial <s> symbol. Hence, we must leave the model
# with unspecified input shapes and axes.
# create the training wrapper for the s2smodel, as well as the criterion function
model_train = create_model_train(s2smodel)
criterion = create_criterion_function(model_train)
# also wire in a greedy decoder so that we can properly log progress on a validation example
# This is not used for the actual training process.
model_greedy = create_model_greedy(s2smodel)
# This does not need to be done in training generally though
# Instantiate the trainer object to drive the model training
minibatch_size = 72
lr = 0.001 if use_attention else 0.005 # TODO: can we use the same value for both?
learner = adam_sgd(model_train.parameters,
lr = learning_rate_schedule([lr]*2+[lr/2]*3+[lr/4], UnitType.sample, epoch_size),
momentum = momentum_as_time_constant_schedule(1100),
gradient_clipping_threshold_per_sample=2.3,
gradient_clipping_with_truncation=True)
trainer = Trainer(None, criterion, learner)
# Get minibatches of sequences to train with and perform model training
total_samples = 0
mbs = 0
eval_freq = 100
# print out some useful training information
log_number_of_parameters(model_train) ; print()
progress_printer = ProgressPrinter(freq=30, tag='Training')
#progress_printer = ProgressPrinter(freq=30, tag='Training', log_to_file=model_path_stem + ".log") # use this to log to file
sparse_to_dense = create_sparse_to_dense(input_vocab_dim)
for epoch in range(max_epochs):
print("Saving model to '%s'" % model_path(epoch))
s2smodel.save(model_path(epoch))
while total_samples < (epoch+1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size)
#trainer.train_minibatch(mb_train[train_reader.streams.features], mb_train[train_reader.streams.labels])
trainer.train_minibatch({criterion.arguments[0]: mb_train[train_reader.streams.features], criterion.arguments[1]: mb_train[train_reader.streams.labels]})
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % eval_freq == 0:
mb_valid = valid_reader.next_minibatch(1)
# run an eval on the decoder output model (i.e. don't use the groundtruth)
e = model_greedy(mb_valid[valid_reader.streams.features])
print(format_sequences(sparse_to_dense(mb_valid[valid_reader.streams.features]), i2w))
print("->")
print(format_sequences(e, i2w))
# debugging attention
if use_attention:
debug_attention(model_greedy, mb_valid[valid_reader.streams.features])
total_samples += mb_train[train_reader.streams.labels].num_samples
mbs += 1
# log a summary of the stats for the epoch
progress_printer.epoch_summary(with_metric=True)
# done: save the final model
print("Saving final model to '%s'" % model_path(max_epochs))
s2smodel.save(model_path(max_epochs))
print("%d epochs complete." % max_epochs)
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 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_minibatch = learning_rate_schedule([1]*80 + [0.1]*40 + [0.01], mb_size * num_mb_per_epoch, UnitType.minibatch)
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 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 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
