def resnet_cifar10(train_data,
test_data,
mean_data,
network_name,
epoch_size,
num_quantization_bits=32,
block_size=3200,
warm_up=0,
max_epochs=5,
restore=True,
log_to_file=None,
num_mbs_per_log=None,
gen_heartbeat=False,
scale_up=False,
profiling=False):
set_computation_network_trace_level(0)
# 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 * (Communicator.num_workers() if scale_up else 1)
progress_printer = ProgressPrinter(freq=num_mbs_per_log,
tag='Training',
log_to_file=log_to_file,
rank=Communicator.rank(),
gen_heartbeat=gen_heartbeat,
num_epochs=max_epochs)
network = create_resnet_network(network_name)
trainer = create_trainer(network, minibatch_size, epoch_size,
num_quantization_bits, block_size, warm_up,
progress_printer)
train_source = create_image_mb_source(train_data,
mean_data,
train=True,
total_number_of_samples=max_epochs *
epoch_size)
test_source = create_image_mb_source(
test_data,
mean_data,
train=False,
total_number_of_samples=cntk.io.FULL_DATA_SWEEP)
train_and_test(network, trainer, train_source, test_source, minibatch_size,
epoch_size, profiling)
def resnet_cifar10(train_data, test_data, mean_data, network_name, epoch_size, num_quantization_bits=32, block_size=3200, warm_up=0,
max_epochs=5, restore=True, log_to_file=None, num_mbs_per_log=None, gen_heartbeat=False, scale_up=False, profiling=False):
set_computation_network_trace_level(0)
# 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 * (Communicator.num_workers() if scale_up else 1)
progress_printer = ProgressPrinter(
freq=num_mbs_per_log,
tag='Training',
log_to_file=log_to_file,
rank=Communicator.rank(),
gen_heartbeat=gen_heartbeat,
num_epochs=max_epochs)
network = create_resnet_network(network_name)
trainer = create_trainer(network, minibatch_size, epoch_size, num_quantization_bits, block_size, warm_up, progress_printer)
train_source = create_image_mb_source(train_data, mean_data, train=True, total_number_of_samples=max_epochs * epoch_size)
test_source = create_image_mb_source(test_data, mean_data, train=False, total_number_of_samples=cntk.io.FULL_DATA_SWEEP)
train_and_test(network, trainer, train_source, test_source, minibatch_size, epoch_size, restore, profiling)
def train_and_evaluate(reader_train,
reader_test,
network_name,
epoch_size,
max_epochs,
profiler_dir=None,
model_dir=None,
log_dir=None,
tensorboard_logdir=None,
gen_heartbeat=False):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = C.input_variable((num_channels, image_height, image_width),
name='features')
label_var = C.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:
raise 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',
log_to_file=log_dir,
num_epochs=max_epochs,
gen_heartbeat=gen_heartbeat)
]
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 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 = C.input_variable((num_channels, sequence_length, image_height, image_width), np.float32)
label_var = C.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 C.default_options (activation=C.relu):
z = C.layers.Sequential([
C.layers.Convolution3D((3,3,3), 64, pad=True),
C.layers.MaxPooling((1,2,2), (1,2,2)),
C.layers.For(range(3), lambda i: [
C.layers.Convolution3D((3,3,3), [96, 128, 128][i], pad=True),
C.layers.Convolution3D((3,3,3), [96, 128, 128][i], pad=True),
C.layers.MaxPooling((2,2,2), (2,2,2))
]),
C.layers.For(range(2), lambda : [
C.layers.Dense(1024),
C.layers.Dropout(0.5)
]),
C.layers.Dense(num_output_classes, activation=None)
])(input_var)
# loss and classification error.
ce = C.cross_entropy_with_softmax(z, label_var)
pe = C.classification_error(z, label_var)
# training config
train_epoch_size = train_reader.size()
train_minibatch_size = 2
# Set learning parameters
lr_per_sample = [0.01]*10+[0.001]*10+[0.0001]
lr_schedule = C.learning_rate_schedule(lr_per_sample, epoch_size=train_epoch_size, unit=C.UnitType.sample)
momentum_time_constant = 4096
mm_schedule = C.momentum_as_time_constant_schedule([momentum_time_constant])
# Instantiate the trainer object to drive the model training
learner = C.momentum_sgd(z.parameters, lr_schedule, mm_schedule, True)
progress_printer = ProgressPrinter(tag='Training', num_epochs=max_epochs)
trainer = C.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(train_minibatch_size)
trainer.train_minibatch({input_var : videos, label_var : labels})
trainer.summarize_training_progress()
# Test data for trained model
epoch_size = test_reader.size()
test_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(test_minibatch_size)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch({input_var : videos, label_var : labels}) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(minibatch_index+1, (metric_numer*100.0)/metric_denom, metric_denom))
print("")
return metric_numer/metric_denom
def train_and_evaluate(reader_train, reader_test, network_name, epoch_size, max_epochs, profiler_dir=None,
model_dir=None, log_dir=None, tensorboard_logdir=None, gen_heartbeat=False):
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = C.input_variable((num_channels, image_height, image_width), name='features')
label_var = C.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:
raise 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', log_to_file=log_dir, num_epochs=max_epochs, gen_heartbeat=gen_heartbeat)]
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 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(
(num_channels, sequence_length, image_height, image_width), np.float32)
label_var = input(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
