def inception_v3_norm_model(input, labelDim, dropRate, bnTimeConst):
# Normalize inputs to -1 and 1.
featMean = 128
featScale = 1 / 128
input_subtracted = minus(input, featMean)
input_scaled = element_times(input_subtracted, featScale)
return inception_v3_model(input_scaled, labelDim, dropRate, bnTimeConst)
def inception_v3_norm_model(input, labelDim, dropRate, bnTimeConst):
# Normalize inputs to -1 and 1.
featMean = 128
featScale = 1/128
input_subtracted = minus(input, featMean)
input_scaled = element_times(input_subtracted, featScale)
return inception_v3_model(input_scaled, labelDim, dropRate, bnTimeConst)
def create_model(self):
mean_removed_features = minus(self.input,
constant(114),
name='mean_removed_input')
with default_options(activation=None, pad=True, bias=True):
self.model = Sequential([
Convolution2D((11, 11),
96,
init=normal(0.01),
pad=False,
name='conv1'),
Activation(activation=relu, name='relu1'),
self.__local_response_normalization(1.0,
2,
0.0001,
0.75,
name='norm1'),
MaxPooling((3, 3), (2, 2), name='pool1'),
Convolution2D((5, 5),
192,
init=normal(0.01),
init_bias=0.1,
name='conv2'),
Activation(activation=relu, name='relu2'),
self.__local_response_normalization(1.0,
2,
0.0001,
0.75,
name='norm2'),
MaxPooling((3, 3), (2, 2), name='pool2'),
Convolution2D((3, 3), 384, init=normal(0.01), name='conv3'),
Activation(activation=relu, name='relu3'),
Convolution2D((3, 3),
384,
init=normal(0.01),
init_bias=0.1,
name='conv4'),
Activation(activation=relu, name='relu4'),
Convolution2D((3, 3),
256,
init=normal(0.01),
init_bias=0.1,
name='conv5'),
Activation(activation=relu, name='relu5'),
MaxPooling((3, 3), (2, 2), name='pool5'),
Dense(4096, init=normal(0.005), init_bias=0.1, name='fc6'),
Activation(activation=relu, name='relu6'),
Dropout(0.5, name='drop6'),
Dense(4096, init=normal(0.005), init_bias=0.1, name='fc7'),
Activation(activation=relu, name='relu7'),
Dropout(0.5, name='drop7'),
Dense(self.number_labels, init=normal(0.01), name='fc8')
])(mean_removed_features)
def create_network(num_convolution_layers):
""" Create network
"""
# Input variables denoting the features and label data
input_var = cntk.input_variable(
(_NUM_CHANNELS, _IMAGE_HEIGHT, _IMAGE_WIDTH))
label_var = cntk.input_variable((_NUM_CLASSES))
# create model, and configure learning parameters
# Instantiate the feedforward classification model
input_removemean = minus(input_var, constant(128))
scaled_input = element_times(constant(0.00390625), input_removemean)
print('Creating NN model')
with layers.default_options(activation=relu, pad=True):
model = layers.Sequential([
layers.For(
range(num_convolution_layers), lambda: [
layers.Convolution2D((3, 3), 64),
layers.Convolution2D((3, 3), 64),
layers.MaxPooling((3, 3), (2, 2))
]),
layers.For(
range(2),
lambda i: [layers.Dense([256, 128][i]),
layers.Dropout(0.5)]),
layers.Dense(_NUM_CLASSES, activation=None)
])(scaled_input)
# loss and metric
ce = cross_entropy_with_softmax(model, label_var)
pe = classification_error(model, label_var)
return {
'name': 'convnet',
'feature': input_var,
'label': label_var,
'ce': ce,
'pe': pe,
'output': model
}
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 create_vgg19():
# Input variables denoting the features and label data
feature_var = C.input_variable((num_channels, image_height, image_width))
label_var = C.input_variable((num_classes))
# apply model to input
# remove mean value
input = minus(feature_var, constant([[[104]], [[117]], [[124]]]), name='mean_removed_input')
with default_options(activation=None, pad=True, bias=True):
z = Sequential([
# we separate Convolution and ReLU to name the output for feature extraction (usually before ReLU)
For(range(2), lambda i: [
Convolution2D((3,3), 64, name='conv1_{}'.format(i)),
Activation(activation=relu, name='relu1_{}'.format(i)),
]),
MaxPooling((2,2), (2,2), name='pool1'),
For(range(2), lambda i: [
Convolution2D((3,3), 128, name='conv2_{}'.format(i)),
Activation(activation=relu, name='relu2_{}'.format(i)),
]),
MaxPooling((2,2), (2,2), name='pool2'),
For(range(4), lambda i: [
Convolution2D((3,3), 256, name='conv3_{}'.format(i)),
Activation(activation=relu, name='relu3_{}'.format(i)),
]),
MaxPooling((2,2), (2,2), name='pool3'),
For(range(4), lambda i: [
Convolution2D((3,3), 512, name='conv4_{}'.format(i)),
Activation(activation=relu, name='relu4_{}'.format(i)),
]),
MaxPooling((2,2), (2,2), name='pool4'),
For(range(4), lambda i: [
Convolution2D((3,3), 512, name='conv5_{}'.format(i)),
Activation(activation=relu, name='relu5_{}'.format(i)),
]),
MaxPooling((2,2), (2,2), name='pool5'),
Dense(4096, name='fc6'),
Activation(activation=relu, name='relu6'),
Dropout(0.5, name='drop6'),
Dense(4096, name='fc7'),
Activation(activation=relu, name='relu7'),
Dropout(0.5, name='drop7'),
Dense(num_classes, name='fc8')
])(input)
# loss and metric
ce = C.cross_entropy_with_softmax(z, label_var)
pe = C.classification_error(z, label_var)
pe5 = C.classification_error(z, label_var, topN=5)
log_number_of_parameters(z) ; print()
return {
'feature': feature_var,
'label': label_var,
'ce' : ce,
'pe' : pe,
'pe5': pe5,
'output': z
}
def convnet_cifar10(debug_output=False):
set_computation_network_trace_level(0)
image_height = 32
image_width = 32
num_channels = 3
input_dim = image_height * image_width * num_channels
num_output_classes = 10
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width), np.float32)
label_var = input_variable(num_output_classes, np.float32)
# Instantiate the feedforward classification model
input_removemean = minus(input_var, constant(128))
scaled_input = element_times(constant(0.00390625), input_removemean)
with default_options (activation=relu, pad=True):
z = Sequential([
LayerStack(2, lambda : [
Convolution((3,3), 64),
Convolution((3,3), 64),
MaxPooling((3,3), (2,2))
]),
LayerStack(2, lambda i: [
Dense([256,128][i]),
Dropout(0.5)
]),
Dense(num_output_classes, activation=None)
])(scaled_input)
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
reader_train = create_reader(os.path.join(data_path, 'Train_cntk_text.txt'), True, input_dim, num_output_classes)
# training config
epoch_size = 50000 # for now we manually specify epoch size
minibatch_size = 64
# Set learning parameters
lr_per_sample = [0.0015625]*10+[0.00046875]*10+[0.00015625]
lr_schedule = learning_rate_schedule(lr_per_sample, epoch_size=epoch_size)
momentum_time_constant = [0]*20+[-minibatch_size/np.log(0.9)]
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant, epoch_size=epoch_size)
l2_reg_weight = 0.002
# Instantiate the trainer object to drive the model training
learner = momentum_sgd(z.parameters, lr_schedule, mm_schedule, l2_regularization_weight = l2_reg_weight)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var : reader_train.streams.features,
label_var : reader_train.streams.labels
}
log_number_of_parameters(z) ; print()
progress_printer = ProgressPrinter(tag='Training')
# Get minibatches of images to train with and perform model training
max_epochs = 30
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(min(minibatch_size, epoch_size - sample_count), input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += data[label_var].num_samples # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
persist.save_model(z, os.path.join(model_path, "ConvNet_CIFAR10_{}.dnn".format(epoch)))
# Load test data
reader_test = create_reader(os.path.join(data_path, 'Test_cntk_text.txt'), False, input_dim, num_output_classes)
input_map = {
input_var : reader_test.streams.features,
label_var : reader_test.streams.labels
}
# Test data for trained model
epoch_size = 10000
minibatch_size = 16
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch, input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(minibatch_index+1, (metric_numer*100.0)/metric_denom, metric_denom))
print("")
return metric_numer/metric_denom
def create_vgg19():
# Input variables denoting the features and label data
feature_var = C.input_variable((num_channels, image_height, image_width))
label_var = C.input_variable((num_classes))
# apply model to input
# remove mean value
input = minus(feature_var,
constant([[[104]], [[117]], [[124]]]),
name='mean_removed_input')
with default_options(activation=None, pad=True, bias=True):
z = Sequential([
# we separate Convolution and ReLU to name the output for feature extraction (usually before ReLU)
For(
range(2), lambda i: [
Convolution2D((3, 3), 64, name='conv1_{}'.format(i)),
Activation(activation=relu, name='relu1_{}'.format(i)),
]),
MaxPooling((2, 2), (2, 2), name='pool1'),
For(
range(2), lambda i: [
Convolution2D((3, 3), 128, name='conv2_{}'.format(i)),
Activation(activation=relu, name='relu2_{}'.format(i)),
]),
MaxPooling((2, 2), (2, 2), name='pool2'),
For(
range(4), lambda i: [
Convolution2D((3, 3), 256, name='conv3_{}'.format(i)),
Activation(activation=relu, name='relu3_{}'.format(i)),
]),
MaxPooling((2, 2), (2, 2), name='pool3'),
For(
range(4), lambda i: [
Convolution2D((3, 3), 512, name='conv4_{}'.format(i)),
Activation(activation=relu, name='relu4_{}'.format(i)),
]),
MaxPooling((2, 2), (2, 2), name='pool4'),
For(
range(4), lambda i: [
Convolution2D((3, 3), 512, name='conv5_{}'.format(i)),
Activation(activation=relu, name='relu5_{}'.format(i)),
]),
MaxPooling((2, 2), (2, 2), name='pool5'),
Dense(4096, name='fc6'),
Activation(activation=relu, name='relu6'),
Dropout(0.5, name='drop6'),
Dense(4096, name='fc7'),
Activation(activation=relu, name='relu7'),
Dropout(0.5, name='drop7'),
Dense(num_classes, name='fc8')
])(input)
# loss and metric
ce = C.cross_entropy_with_softmax(z, label_var)
pe = C.classification_error(z, label_var)
pe5 = C.classification_error(z, label_var, topN=5)
log_number_of_parameters(z)
print()
return {
'feature': feature_var,
'label': label_var,
'ce': ce,
'pe': pe,
'pe5': pe5,
'output': z
}
reader_test = create_reader(test_map_image,
data_mean_file,
False,
image_height=image_height,
image_width=image_width,
num_channels=num_channels,
num_classes=num_classes)
#==============================================================================
# ###
#==============================================================================
# Normalize the inputs
feature_scale = 1.0 / 256.0
input_var = input_variable((num_channels, image_height, image_width))
input_var_mean = minus(input_var, 128)
input_var_norm = element_times(feature_scale, input_var_mean)
label_var = input_variable((num_classes))
# apply model to input
z = model_func(input_var_norm, out_dims=num_classes)
"""
Training action
"""
# loss and metric
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
#pe5 = classification_error(z, label_var, topN=5)
def __init__(self,
state_dim,
action_dim,
gamma=0.99,
learning_rate=1e-4,
momentum=0.95):
self.state_dim = state_dim
self.action_dim = action_dim
self.gamma = gamma
with default_options(activation=relu, init=he_uniform()):
# Convolution filter counts were halved to save on memory, no gpu :(
self.model = Sequential([
Convolution2D((8, 8), 16, strides=4, name='conv1'),
Convolution2D((4, 4), 32, strides=2, name='conv2'),
Convolution2D((3, 3), 32, strides=1, name='conv3'),
Dense(256, init=he_uniform(scale=0.01), name='dense1'),
Dense(action_dim,
activation=None,
init=he_uniform(scale=0.01),
name='actions')
])
self.model.update_signature(Tensor[state_dim])
# Create the target model as a copy of the online model
self.target_model = None
self.update_target()
self.pre_states = input_variable(state_dim, name='pre_states')
self.actions = input_variable(action_dim, name='actions')
self.post_states = input_variable(state_dim, name='post_states')
self.rewards = input_variable((), name='rewards')
self.terminals = input_variable((), name='terminals')
self.is_weights = input_variable((), name='is_weights')
predicted_q = reduce_sum(self.model(self.pre_states) * self.actions,
axis=0)
# DQN - calculate target q values
# post_q = reduce_max(self.target_model(self.post_states), axis=0)
# DDQN - calculate target q values
online_selection = one_hot(
argmax(self.model(self.post_states), axis=0), self.action_dim)
post_q = reduce_sum(self.target_model(self.post_states) *
online_selection,
axis=0)
post_q = (1.0 - self.terminals) * post_q
target_q = stop_gradient(self.rewards + self.gamma * post_q)
# Huber loss
delta = 1.0
self.td_error = minus(predicted_q, target_q, name='td_error')
abs_error = abs(self.td_error)
errors = element_select(less(abs_error, delta),
square(self.td_error) * 0.5,
delta * (abs_error - 0.5 * delta))
loss = errors * self.is_weights
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_scheule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
self._learner = adam(self.model.parameters,
lr_schedule,
m_scheule,
variance_momentum=vm_schedule)
self.writer = TensorBoardProgressWriter(log_dir='metrics',
model=self.model)
self.trainer = Trainer(self.model, (loss, None), [self._learner],
self.writer)
