def run_mlp():
# # define the model layers
# layer1 = Dense(input_size=784, output_size=1000, activation='rectifier')
# layer2 = Dense(inputs_hook=(1000, layer1.get_outputs()), output_size=1000, activation='rectifier')
# classlayer3 = SoftmaxLayer(inputs_hook=(1000, layer2.get_outputs()), output_size=10, out_as_probs=False)
# # add the layers to the prototype
# mlp = Prototype(layers=[layer1, layer2, classlayer3])
# test the new way to automatically fill in inputs_hook for models
mlp = Prototype()
mlp.add(Dense(input_size=784, output_size=1000, activation='rectifier', noise='dropout'))
mlp.add(Dense(output_size=1500, activation='tanh', noise='dropout'))
mlp.add(SoftmaxLayer(output_size=10))
mnist = MNIST()
optimizer = AdaDelta(model=mlp, dataset=mnist, epochs=10)
optimizer.train()
test_data, test_labels = mnist.test_inputs, mnist.test_targets
test_data = test_data[:25]
test_labels = test_labels[:25]
# use the run function!
yhat = mlp.run(test_data)
print('-------')
print('Prediction: %s' % str(yhat))
print('Actual: %s' % str(test_labels.astype('int32')))
def build_model():
# add layers one-by-one to a Prototype container to build neural net
# inputs_hook created automatically by Prototype; thus, no need to specify
mlp = Prototype()
mlp.add(BasicLayer(input_size=28*28, output_size=512, activation='rectifier', noise='dropout'))
mlp.add(BasicLayer(output_size=512, activation='rectifier', noise='dropout'))
mlp.add(SoftmaxLayer(output_size=10))
return mlp
def main():
# First, let's create a simple feedforward MLP with one hidden layer as a Prototype.
mlp = Prototype()
mlp.add(
BasicLayer(input_size=28 * 28,
output_size=1000,
activation='rectifier',
noise='dropout'))
mlp.add(SoftmaxLayer(output_size=10))
# Now, we get to choose what values we want to monitor, and what datasets we would like to monitor on!
# Each Model (in our case, the Prototype), has a get_monitors method that will return a useful
# dictionary of {string_name: monitor_theano_expression} for various computations of the model we might
# care about. By default, this method returns an empty dictionary - it was the model creator's job to
# include potential monitor values.
mlp_monitors = mlp.get_monitors()
mlp_channel = MonitorsChannel(name="error")
for name, expression in mlp_monitors.items():
mlp_channel.add(
Monitor(name=name,
expression=expression,
train=True,
valid=True,
test=True))
# create some monitors for statistics about the hidden and output weights!
# let's look at the mean, variance, and standard deviation of the weights matrices.
weights_channel = MonitorsChannel(name="weights")
hiddens_1 = mlp[0].get_params()[0]
hiddens1_mean = T.mean(hiddens_1)
weights_channel.add(
Monitor(name="hiddens_mean", expression=hiddens1_mean, train=True))
hiddens_2 = mlp[1].get_params()[0]
hiddens2_mean = T.mean(hiddens_2)
weights_channel.add(
Monitor(name="out_mean", expression=hiddens2_mean, train=True))
# create our plot object to do live plotting!
plot = Plot(bokeh_doc_name="Monitor Tutorial",
monitor_channels=[mlp_channel, weights_channel],
open_browser=True)
# use SGD optimizer
optimizer = SGD(model=mlp,
dataset=MNIST(concat_train_valid=False),
n_epoch=500,
save_frequency=100,
batch_size=600,
learning_rate=.01,
lr_decay=False,
momentum=.9,
nesterov_momentum=True)
# train, with the plot!
optimizer.train(plot=plot)
def add_list_layers():
# You can also add lists of layers at a time (or as initialization) to a Prototype! This lets you specify
# more complex interactions between layers!
hidden1 = BasicLayer(input_size=28 * 28,
output_size=512,
activation='rectifier',
noise='dropout')
hidden2 = BasicLayer(inputs_hook=(512, hidden1.get_outputs()),
output_size=512,
activation='rectifier',
noise='dropout')
class_layer = SoftmaxLayer(inputs_hook=(512, hidden2.get_outputs()),
output_size=10)
mlp = Prototype([hidden1, hidden2, class_layer])
return mlp
def sequential_add_layers():
# This method is to demonstrate adding layers one-by-one to a Prototype container.
# As you can see, inputs_hook are created automatically by Prototype so we don't need to specify!
mlp = Prototype()
mlp.add(
BasicLayer(input_size=28 * 28,
output_size=1000,
activation='rectifier',
noise='dropout',
noise_level=0.5))
mlp.add(
BasicLayer(output_size=512,
activation='rectifier',
noise='dropout',
noise_level=0.5))
mlp.add(SoftmaxLayer(output_size=10))
return mlp
def create_mlp():
# define the model layers
relu_layer1 = BasicLayer(input_size=784, output_size=1000, activation='rectifier')
relu_layer2 = BasicLayer(inputs_hook=(1000, relu_layer1.get_outputs()), output_size=1000, activation='rectifier')
class_layer3 = SoftmaxLayer(inputs_hook=(1000, relu_layer2.get_outputs()), output_size=10, out_as_probs=False)
# add the layers as a Prototype
mlp = Prototype(layers=[relu_layer1, relu_layer2, class_layer3])
mnist = MNIST()
optimizer = AdaDelta(model=mlp, dataset=mnist, n_epoch=20)
optimizer.train()
test_data, test_labels = mnist.getSubset(TEST)
test_data = test_data[:25].eval()
test_labels = test_labels[:25].eval()
# use the run function!
preds = mlp.run(test_data)
log.info('-------')
log.info("predicted: %s",str(preds))
log.info("actual: %s",str(test_labels.astype('int32')))
def _build_computation_graph(self):
###################### BUILD NETWORK ##########################
# whether or not to mirror the input images before feeding them into the network
if self.flag_datalayer:
layer_1_input = mirror_images(input=self.x,
image_shape=(self.batch_size, 3, 256, 256), # bc01 format
cropsize=227,
rand=self.rand,
flag_rand=self.rand_crop)
else:
layer_1_input = self.x # 4D tensor (going to be in bc01 format)
# Start with 5 convolutional pooling layers
log.debug("convpool layer 1...")
convpool_layer1 = ConvPoolLayer(inputs_hook=((self.batch_size, 3, 227, 227), layer_1_input),
filter_shape=(96, 3, 11, 11),
convstride=4,
padsize=0,
group=1,
poolsize=3,
poolstride=2,
bias_init=0.0,
local_response_normalization=True)
# Add this layer's parameters!
self.params += convpool_layer1.get_params()
log.debug("convpool layer 2...")
convpool_layer2 = ConvPoolLayer(inputs_hook=((self.batch_size, 96, 27, 27, ), convpool_layer1.get_outputs()),
filter_shape=(256, 96, 5, 5),
convstride=1,
padsize=2,
group=2,
poolsize=3,
poolstride=2,
bias_init=0.1,
local_response_normalization=True)
# Add this layer's parameters!
self.params += convpool_layer2.get_params()
log.debug("convpool layer 3...")
convpool_layer3 = ConvPoolLayer(inputs_hook=((self.batch_size, 256, 13, 13), convpool_layer2.get_outputs()),
filter_shape=(384, 256, 3, 3),
convstride=1,
padsize=1,
group=1,
poolsize=1,
poolstride=0,
bias_init=0.0,
local_response_normalization=False)
# Add this layer's parameters!
self.params += convpool_layer3.get_params()
log.debug("convpool layer 4...")
convpool_layer4 = ConvPoolLayer(inputs_hook=((self.batch_size, 384, 13, 13), convpool_layer3.get_outputs()),
filter_shape=(384, 384, 3, 3),
convstride=1,
padsize=1,
group=2,
poolsize=1,
poolstride=0,
bias_init=0.1,
local_response_normalization=False)
# Add this layer's parameters!
self.params += convpool_layer4.get_params()
log.debug("convpool layer 5...")
convpool_layer5 = ConvPoolLayer(inputs_hook=((self.batch_size, 384, 13, 13), convpool_layer4.get_outputs()),
filter_shape=(256, 384, 3, 3),
convstride=1,
padsize=1,
group=2,
poolsize=3,
poolstride=2,
bias_init=0.0,
local_response_normalization=False)
# Add this layer's parameters!
self.params += convpool_layer5.get_params()
# Now onto the fully-connected layers!
fc_config = {
'activation': 'rectifier', # type of activation function to use for output
'weights_init': 'gaussian', # either 'gaussian' or 'uniform' - how to initialize weights
'weights_mean': 0.0, # mean for gaussian weights init
'weights_std': 0.005, # standard deviation for gaussian weights init
'bias_init': 0.0 # how to initialize the bias parameter
}
log.debug("fully connected layer 1 (model layer 6)...")
# we want to have dropout applied to the training version, but not the test version.
fc_layer6_input = T.flatten(convpool_layer5.get_outputs(), 2)
fc_layer6 = BasicLayer(inputs_hook=(9216, fc_layer6_input),
output_size=4096,
noise='dropout',
noise_level=0.5,
**fc_config)
# Add this layer's parameters!
self.params += fc_layer6.get_params()
# Add the dropout noise switch
self.noise_switches += fc_layer6.get_noise_switch()
log.debug("fully connected layer 2 (model layer 7)...")
fc_layer7 = BasicLayer(inputs_hook=(4096, fc_layer6.get_outputs()),
output_size=4096,
noise='dropout',
noise_level=0.5,
**fc_config)
# Add this layer's parameters!
self.params += fc_layer7.get_params()
# Add the dropout noise switch
self.noise_switches += fc_layer7.get_noise_switch()
# last layer is a softmax prediction output layer
softmax_config = {
'weights_init': 'gaussian',
'weights_mean': 0.0,
'weights_std': 0.005,
'bias_init': 0.0
}
log.debug("softmax classification layer (model layer 8)...")
softmax_layer8 = SoftmaxLayer(inputs_hook=(4096, fc_layer7.get_outputs()),
output_size=1000,
**softmax_config)
# Add this layer's parameters!
self.params += softmax_layer8.get_params()
# finally the softmax output from the whole thing!
self.output = softmax_layer8.get_outputs()
self.targets = softmax_layer8.get_targets()
#####################
# Cost and monitors #
#####################
self.train_cost = softmax_layer8.negative_log_likelihood()
cost = softmax_layer8.negative_log_likelihood()
errors = softmax_layer8.errors()
train_errors = softmax_layer8.errors()
self.monitors = OrderedDict([('cost', cost), ('errors', errors), ('dropout_errors', train_errors)])
#########################
# Compile the functions #
#########################
log.debug("Compiling functions!")
t = time.time()
log.debug("f_run...")
# use the actual argmax from the classification
self.f_run = function(inputs=[self.x], outputs=softmax_layer8.get_argmax_prediction())
log.debug("compilation took %s", make_time_units_string(time.time() - t))
from opendeep.data.standard_datasets.image.mnist import MNIST
from opendeep.optimization.adadelta import AdaDelta
if __name__ == '__main__':
# set up the logging environment to display outputs (optional)
# although this is recommended over print statements everywhere
import logging
from opendeep.log import config_root_logger
config_root_logger()
log = logging.getLogger(__name__)
log.info("Creating softmax!")
# grab the MNIST dataset
mnist = MNIST()
# create the softmax classifier
s = SoftmaxLayer(input_size=28 * 28, output_size=10, out_as_probs=False)
# make an optimizer to train it (AdaDelta is a good default)
optimizer = AdaDelta(model=s, dataset=mnist, epochs=20)
# perform training!
optimizer.train()
# test it on some images!
test_data, test_labels = mnist.test_inputs[:25], mnist.test_targets[:25]
# use the run function!
preds = s.run(test_data)
print('-------')
print(preds)
print(test_labels.astype('int32'))
print()
print()
del mnist
del s
from opendeep.optimization.adadelta import AdaDelta
if __name__ == '__main__':
# set up the logging environment to display outputs (optional)
# although this is recommended over print statements everywhere
import logging
import opendeep.log.logger as logger
logger.config_root_logger()
log = logging.getLogger(__name__)
log.info("Creating softmax!")
# grab the MNIST dataset
mnist = MNIST()
# create the softmax classifier
s = SoftmaxLayer(input_size=28 * 28, output_size=10, out_as_probs=False)
# make an optimizer to train it (AdaDelta is a good default)
optimizer = AdaDelta(model=s, dataset=mnist, n_epoch=20)
# perform training!
optimizer.train()
# test it on some images!
test_data = mnist.getDataByIndices(indices=range(25), subset=TEST)
# use the predict function!
preds = s.predict(test_data)
print '-------'
print preds
print mnist.getLabelsByIndices(indices=range(25), subset=TEST)
print
print
del mnist
del s
from opendeep.optimization.adadelta import AdaDelta
if __name__ == '__main__':
# set up the logging environment to display outputs (optional)
# although this is recommended over print statements everywhere
import logging
from opendeep.log import config_root_logger
config_root_logger()
log = logging.getLogger(__name__)
log.info("Creating softmax!")
# grab the MNIST dataset
mnist = MNIST()
# create the softmax classifier
s = SoftmaxLayer(input_size=28 * 28, output_size=10, out_as_probs=False)
# make an optimizer to train it (AdaDelta is a good default)
optimizer = AdaDelta(model=s, dataset=mnist, epochs=20)
# perform training!
optimizer.train()
# test it on some images!
test_data, test_labels = mnist.test_inputs[:25], mnist.test_targets[:25]
# use the run function!
preds = s.run(test_data)
print('-------')
print(preds)
print(test_labels.astype('int32'))
print()
print()
del mnist
del s
def _build_computation_graph(self):
###################### BUILD NETWORK ##########################
# whether or not to mirror the input images before feeding them into the network
if self.flag_datalayer:
layer_1_input = mirror_images(
input=self.x,
image_shape=(self.batch_size, 3, 256, 256), # bc01 format
cropsize=227,
rand=self.rand,
flag_rand=self.rand_crop)
else:
layer_1_input = self.x # 4D tensor (going to be in bc01 format)
# Start with 5 convolutional pooling layers
log.debug("convpool layer 1...")
convpool_layer1 = ConvPoolLayer(inputs_hook=((self.batch_size, 3, 227,
227), layer_1_input),
filter_shape=(96, 3, 11, 11),
convstride=4,
padsize=0,
group=1,
poolsize=3,
poolstride=2,
bias_init=0.0,
local_response_normalization=True)
# Add this layer's parameters!
self.params += convpool_layer1.get_params()
log.debug("convpool layer 2...")
convpool_layer2 = ConvPoolLayer(inputs_hook=((
self.batch_size,
96,
27,
27,
), convpool_layer1.get_outputs()),
filter_shape=(256, 96, 5, 5),
convstride=1,
padsize=2,
group=2,
poolsize=3,
poolstride=2,
bias_init=0.1,
local_response_normalization=True)
# Add this layer's parameters!
self.params += convpool_layer2.get_params()
log.debug("convpool layer 3...")
convpool_layer3 = ConvPoolLayer(
inputs_hook=((self.batch_size, 256, 13, 13),
convpool_layer2.get_outputs()),
filter_shape=(384, 256, 3, 3),
convstride=1,
padsize=1,
group=1,
poolsize=1,
poolstride=0,
bias_init=0.0,
local_response_normalization=False)
# Add this layer's parameters!
self.params += convpool_layer3.get_params()
log.debug("convpool layer 4...")
convpool_layer4 = ConvPoolLayer(
inputs_hook=((self.batch_size, 384, 13, 13),
convpool_layer3.get_outputs()),
filter_shape=(384, 384, 3, 3),
convstride=1,
padsize=1,
group=2,
poolsize=1,
poolstride=0,
bias_init=0.1,
local_response_normalization=False)
# Add this layer's parameters!
self.params += convpool_layer4.get_params()
log.debug("convpool layer 5...")
convpool_layer5 = ConvPoolLayer(
inputs_hook=((self.batch_size, 384, 13, 13),
convpool_layer4.get_outputs()),
filter_shape=(256, 384, 3, 3),
convstride=1,
padsize=1,
group=2,
poolsize=3,
poolstride=2,
bias_init=0.0,
local_response_normalization=False)
# Add this layer's parameters!
self.params += convpool_layer5.get_params()
# Now onto the fully-connected layers!
fc_config = {
'activation':
'rectifier', # type of activation function to use for output
'weights_init':
'gaussian', # either 'gaussian' or 'uniform' - how to initialize weights
'weights_mean': 0.0, # mean for gaussian weights init
'weights_std':
0.005, # standard deviation for gaussian weights init
'bias_init': 0.0 # how to initialize the bias parameter
}
log.debug("fully connected layer 1 (model layer 6)...")
# we want to have dropout applied to the training version, but not the test version.
fc_layer6_input = T.flatten(convpool_layer5.get_outputs(), 2)
fc_layer6 = BasicLayer(inputs_hook=(9216, fc_layer6_input),
output_size=4096,
noise='dropout',
noise_level=0.5,
**fc_config)
# Add this layer's parameters!
self.params += fc_layer6.get_params()
# Add the dropout noise switch
self.noise_switches += fc_layer6.get_noise_switch()
log.debug("fully connected layer 2 (model layer 7)...")
fc_layer7 = BasicLayer(inputs_hook=(4096, fc_layer6.get_outputs()),
output_size=4096,
noise='dropout',
noise_level=0.5,
**fc_config)
# Add this layer's parameters!
self.params += fc_layer7.get_params()
# Add the dropout noise switch
self.noise_switches += fc_layer7.get_noise_switch()
# last layer is a softmax prediction output layer
softmax_config = {
'weights_init': 'gaussian',
'weights_mean': 0.0,
'weights_std': 0.005,
'bias_init': 0.0
}
log.debug("softmax classification layer (model layer 8)...")
softmax_layer8 = SoftmaxLayer(inputs_hook=(4096,
fc_layer7.get_outputs()),
output_size=1000,
**softmax_config)
# Add this layer's parameters!
self.params += softmax_layer8.get_params()
# finally the softmax output from the whole thing!
self.output = softmax_layer8.get_outputs()
self.targets = softmax_layer8.get_targets()
#####################
# Cost and monitors #
#####################
self.train_cost = softmax_layer8.negative_log_likelihood()
cost = softmax_layer8.negative_log_likelihood()
errors = softmax_layer8.errors()
train_errors = softmax_layer8.errors()
self.monitors = OrderedDict([('cost', cost), ('errors', errors),
('dropout_errors', train_errors)])
#########################
# Compile the functions #
#########################
log.debug("Compiling functions!")
t = time.time()
log.debug("f_run...")
# use the actual argmax from the classification
self.f_run = function(inputs=[self.x],
outputs=softmax_layer8.get_argmax_prediction())
log.debug("compilation took %s",
make_time_units_string(time.time() - t))
# although this is recommended over print statements everywhere
import logging
import opendeep.log.logger as logger
logger.config_root_logger()
log = logging.getLogger(__name__)
log.info("Creating MLP!")
# grab the MNIST dataset
mnist = MNIST()
# create the basic layer
layer1 = BasicLayer(input_size=28 * 28,
output_size=1000,
activation='relu')
# create the softmax classifier
layer2 = SoftmaxLayer(inputs_hook=(1000, layer1.get_outputs()),
output_size=10,
out_as_probs=False)
# create the mlp from the two layers
mlp = Prototype(layers=[layer1, layer2])
# make an optimizer to train it (AdaDelta is a good default)
optimizer = AdaDelta(model=mlp, dataset=mnist, n_epoch=20)
# perform training!
optimizer.train()
# test it on some images!
test_data, test_labels = mnist.getSubset(subset=TEST)
test_data = test_data[:25].eval()
test_labels = test_labels[:25].eval()
# use the run function!
preds = mlp.run(test_data)
print('-------')
print(preds)