def feature_extraction():
decomp = DecompLayer([(2, 80, 30)] * len(bodyconf.groups))
column = MultiwayNeuralNet([
NeuralNet([
ConvPoolLayer((64, 2, 5, 5), (2, 2), (2, 80, 30), af.tanh,
False),
ConvPoolLayer((64, 64, 5, 5), (2, 2), None, af.tanh, True)
]) for __ in xrange(len(bodyconf.groups))
])
comp = CompLayer(strategy='Maxout')
return NeuralNet([decomp, column, comp])
def _train_model(nndata, load_from_cache=False, save_to_cache=False):
print "Training Model ..."
print "=================="
if load_from_cache:
with open(_cached_model, 'rb') as f:
model = cPickle.load(f)
else:
layers = [
ConvPoolLayer((20, 3, 5, 5), (2, 2), (3, 80, 40), actfuncs.tanh,
False),
ConvPoolLayer((50, 20, 5, 5), (2, 2), None, actfuncs.tanh, True),
FullConnLayer(5950, 500, actfuncs.tanh),
FullConnLayer(500, 3, actfuncs.softmax)
]
model = NeuralNet(layers)
sgd.train(model,
nndata,
costfuncs.mean_negative_loglikelihood,
costfuncs.mean_number_misclassified,
regularize=1e-3,
batch_size=500,
n_epoch=200,
learning_rate=1e-1,
momentum=0.9,
learning_rate_decr=0.95,
never_stop=True)
if save_to_cache:
with open(_cached_model, 'wb') as f:
cPickle.dump(model, f, protocol=cPickle.HIGHEST_PROTOCOL)
return model
def compute_result(model, dataset, data):
"""Compute output value of data samples by using the trained model
Args:
model: A NeuralNet model
dataset: A Dataset object returned by ``create_dataset``
data: list of pedestrian tuples returned by ``decomp_body``
Returns:
A tuple (train, valid, test) where each is three matrices
(image_tensor4D, output_matrix, target_matrix)
"""
result = cachem.load('result')
if result is None:
import theano
import theano.tensor as T
from reid.models.layers import CompLayer
from reid.models.neural_net import NeuralNet
model = NeuralNet([model, CompLayer()])
x = T.matrix()
f = theano.function(inputs=[x], outputs=model.get_output(x))
def compute_output(X):
outputs = [f(X[i:i+1, :]).ravel() for i in xrange(X.shape[0])]
return numpy.asarray(outputs)
images = numpy.asarray([p[0] for p in data])
train = (images[dataset.train_ind],
compute_output(dataset.train_x.get_value(borrow=True)),
dataset.train_y.get_value(borrow=True))
valid = (images[dataset.valid_ind],
compute_output(dataset.valid_x.get_value(borrow=True)),
dataset.valid_y.get_value(borrow=True))
test = (images[dataset.test_ind],
compute_output(dataset.test_x.get_value(borrow=True)),
dataset.test_y.get_value(borrow=True))
result = (train, valid, test)
return result
def attribute_classification():
fcl_1 = FullConnLayer(4352, 1024, af.tanh)
fcl_2 = FullConnLayer(1024, 104)
decomp = DecompLayer(
[(sz,) for sz in output_sizes],
[af.softmax] * len(attrconf.unival) + \
[af.sigmoid] * len(attrconf.multival)
)
return NeuralNet([fcl_1, fcl_2, decomp], through=True)
def _train_model(datasets, load_from_cache=False, save_to_cache=False):
if load_from_cache:
with open(_cached_model, 'rb') as f:
model = cPickle.load(f)
else:
# Build model
print "Building model ..."
# model = AutoEncoder(layer_sizes=[9600, 2400, 2400, 2400, 2400],
# active_funcs=[actfuncs.sigmoid, actfuncs.sigmoid, actfuncs.sigmoid, actfuncs.sigmoid],
# cost_func=costfuncs.mean_square_error,
# error_func=costfuncs.mean_square_error)
layers = [
ConvPoolLayer((128, 3, 5, 5), (2, 2), (3, 80, 40),
actfuncs.sigmoid, False),
ConvPoolLayer((64, 128, 5, 5), (1, 1), None, actfuncs.sigmoid,
False),
ConvPoolLayer((32, 64, 5, 5), (1, 1), None, actfuncs.sigmoid,
True),
FullConnLayer(9600, 9600, actfuncs.sigmoid),
FullConnLayer(9600, 9600, actfuncs.sigmoid)
]
model = NeuralNet(layers, costfuncs.mean_square_error,
costfuncs.mean_square_error)
import copy
pretrain_datasets = copy.copy(datasets)
pretrain_datasets.train_y = pretrain_datasets.train_x
pretrain_datasets.valid_y = pretrain_datasets.valid_x
pretrain_datasets.test_y = pretrain_datasets.test_x
sgd.train(model,
pretrain_datasets,
n_epoch=10,
learning_rate=1e-3,
learning_rate_decr=1.0)
sgd.train(model,
datasets,
n_epoch=100,
learning_rate=1e-3,
learning_rate_decr=1.0)
if save_to_cache:
with open(_cached_model, 'wb') as f:
cPickle.dump(model, f, protocol=cPickle.HIGHEST_PROTOCOL)
return model
def compute_attr(model, dataset, batch_size=100):
"""Compute the attributes matrix
Args:
model: The deep neural net model
dataset: The dataset returned by ``preprocess``
Returns:
The dataset ``(gA, gY, pA, pY)``
"""
data = cachem.load('attr')
if data is None:
import theano
import theano.tensor as T
from reid.models.layers import CompLayer
from reid.models.neural_net import NeuralNet
model = NeuralNet([model, CompLayer()])
def compute(X):
x = T.matrix('x')
i = T.lscalar('i')
func = theano.function(
inputs=[i],
outputs=model.get_output(x),
givens={x: X[i * batch_size:(i + 1) * batch_size]})
n_batches = X.get_value(borrow=True).shape[0] / batch_size + 1
return numpy.vstack([func(j) for j in xrange(n_batches)])
gX, gY, pX, pY = dataset
gA = compute(theano.shared(gX, borrow=True))
pA = compute(theano.shared(pX, borrow=True))
data = (gA, gY, pA, pY)
return data
def _train_model(datasets, load_from_cache=False, save_to_cache=False):
if load_from_cache:
with open(_cached_model, 'rb') as f:
model, threshold = cPickle.load(f)
else:
# Build model
print "Building model ..."
layers = [
FullConnLayer(input_size=38400,
output_size=1024,
active_func=actfuncs.sigmoid),
FullConnLayer(input_size=1024,
output_size=12800,
active_func=actfuncs.sigmoid)
]
model = NeuralNet(layers)
sgd.train(model,
datasets,
cost_func=costfuncs.mean_binary_cross_entropy,
error_func=costfuncs.mean_binary_cross_entropy,
n_epoch=100,
learning_rate=1e-3,
learning_rate_decr=1.0)
threshold = _choose_threshold(model, datasets, verbose=False)
if save_to_cache:
with open(_cached_model, 'wb') as f:
cPickle.dump((model, threshold),
f,
protocol=cPickle.HIGHEST_PROTOCOL)
return (model, threshold)
def target_adapter():
d1 = DecompLayer([(sum(target_sizes), ), (sum(target_sizes), ), (1, )])
d2 = DecompLayer([(sz, ) for sz in target_sizes])
return NeuralNet([d1, MultiwayNeuralNet([d2, d2, IdentityLayer()])])
def person_reidentification():
fp = FilterParingLayer((64, 17, 4), 4, (2, 2), True)
fcl_1 = FullConnLayer(1088, 256, af.tanh)
return NeuralNet([fp, fcl_1])
def train_model(batch_dir):
"""Train deep model
Args:
dataset: Dataset X = X1_X2, Y = A1_A2_(0/1)
Returns:
The trained deep model
"""
print "Training ..."
model = cachem.load('model')
if model is not None: return model
import reid.models.active_functions as af
import reid.models.cost_functions as cf
from reid.models.layers import ConvPoolLayer, FullConnLayer, IdentityLayer, FilterParingLayer
from reid.models.layers import CompLayer, DecompLayer, CloneLayer
from reid.models.neural_net import NeuralNet, MultiwayNeuralNet
from reid.models.evaluate import Evaluator
from reid.optimization import sgd
output_sizes = [len(grp) for grp in attrconf.unival + attrconf.multival]
target_sizes = [1] * len(attrconf.unival) + [
len(grp) for grp in attrconf.multival
]
# Feature extraction module
def feature_extraction():
decomp = DecompLayer([(2, 80, 30)] * len(bodyconf.groups))
column = MultiwayNeuralNet([
NeuralNet([
ConvPoolLayer((64, 2, 5, 5), (2, 2), (2, 80, 30), af.tanh,
False),
ConvPoolLayer((64, 64, 5, 5), (2, 2), None, af.tanh, True)
]) for __ in xrange(len(bodyconf.groups))
])
comp = CompLayer(strategy='Maxout')
return NeuralNet([decomp, column, comp])
fe = feature_extraction()
feat_module = NeuralNet([
DecompLayer([(2 * 80 * 30 * len(bodyconf.groups), )] * 2),
MultiwayNeuralNet([fe, fe]),
CompLayer()
])
# Attribute classification module
def attribute_classification():
fcl_1 = FullConnLayer(4352, 1024, af.tanh)
fcl_2 = FullConnLayer(1024, 104)
decomp = DecompLayer(
[(sz,) for sz in output_sizes],
[af.softmax] * len(attrconf.unival) + \
[af.sigmoid] * len(attrconf.multival)
)
return NeuralNet([fcl_1, fcl_2, decomp], through=True)
ac = NeuralNet([attribute_classification(), CompLayer()])
attr_module = NeuralNet([
DecompLayer([(4352, )] * 2),
MultiwayNeuralNet([ac, ac]),
CompLayer()
])
# Person re-identification module
def person_reidentification():
fp = FilterParingLayer((64, 17, 4), 4, (2, 2), True)
fcl_1 = FullConnLayer(1088, 256, af.tanh)
return NeuralNet([fp, fcl_1])
reid_module = person_reidentification()
# Combine them together
model = NeuralNet([
feat_module,
CloneLayer(2),
MultiwayNeuralNet([attr_module, reid_module]),
CompLayer(),
FullConnLayer(104 + 104 + 256, 256, af.tanh),
FullConnLayer(256, 2, af.softmax)
])
# Fine-tuning
def reid_cost(output, target):
return 6.0 * cf.mean_negative_loglikelihood(output, target)
def reid_error(output, target):
return 6.0 * cf.mean_negative_loglikelihood(output, target)
def target_adapter():
d1 = DecompLayer([(sum(target_sizes), ), (sum(target_sizes), ), (1, )])
d2 = DecompLayer([(sz, ) for sz in target_sizes])
return NeuralNet([d1, MultiwayNeuralNet([d2, d2, IdentityLayer()])])
cost_func = [
[cf.mean_negative_loglikelihood] * len(attrconf.unival) + \
[cf.mean_binary_cross_entropy] * len(attrconf.multival),
[cf.mean_negative_loglikelihood] * len(attrconf.unival) + \
[cf.mean_binary_cross_entropy] * len(attrconf.multival),
reid_cost
]
error_func = [
[cf.mean_number_misclassified] * len(attrconf.unival) + \
[cf.mean_zeroone_error_rate] * len(attrconf.multival),
[cf.mean_number_misclassified] * len(attrconf.unival) + \
[cf.mean_zeroone_error_rate] * len(attrconf.multival),
reid_error
]
evaluator = Evaluator(model,
cost_func,
error_func,
target_adapter(),
regularize=1e-3)
sgd.train_batch(evaluator,
batch_dir,
learning_rate=1e-4,
momentum=0.9,
batch_size=300,
n_epoch=100,
learning_rate_decr=1.0,
patience_incr=1.5)
return model
W1 = theano.shared(numpy.asarray([[1,2], [3,4]], dtype=numpy.float32), borrow=True)
W2 = theano.shared(numpy.asarray([[4,3], [2,1]], dtype=numpy.float32), borrow=True)
# Build up model
decomp = DecompLayer([(2,), (2,)])
columns = MultiwayNeuralNet([
FullConnLayer(2, 2, W=W1),
FullConnLayer(2, 2, W=W2)
])
comp = CompLayer()
multitask = DecompLayer([(1,), (3,)])
model = NeuralNet([decomp, columns, comp, multitask])
# Build up the target value adapter
adapter = DecompLayer([(1,), (3,)])
# Build up evaluator
evaluator = Evaluator(model, [mse, mse], [mse, mse], adapter)
# Compute the expression by using the model
x = T.matrix('x')
y = T.matrix('y')
output = model.get_output(x)
cost, updates = evaluator.get_cost_updates(x, y, 1.0)
error = evaluator.get_error(x, y)
def train_model(dataset):
"""Train deep model
This function will build up a deep neural network and train it using given
dataset.
Args:
dataset: A Dataset object returned by ``create_dataset``
Returns:
The trained deep model.
"""
model = cachem.load('model')
if model is None:
import reid.models.active_functions as actfuncs
import reid.models.cost_functions as costfuncs
from reid.models.layers import ConvPoolLayer, FullConnLayer
from reid.models.layers import CompLayer, DecompLayer
from reid.models.neural_net import NeuralNet, MultiwayNeuralNet
from reid.models.evaluate import Evaluator
from reid.optimization import sgd
output_sizes = [len(grp) for grp in attrconf.unival + attrconf.multival]
target_sizes = [1] * len(attrconf.unival) + [len(grp) for grp in attrconf.multival]
# Build up model
input_decomp = DecompLayer([(3,80,30)] * len(bodyconf.groups))
columns = MultiwayNeuralNet([NeuralNet([
ConvPoolLayer((64,3,3,3), (2,2), (3,80,30), actfuncs.tanh, False),
ConvPoolLayer((64,64,3,3), (2,2), None, actfuncs.tanh, True)
]) for __ in xrange(len(bodyconf.groups))])
feature_comp = CompLayer(strategy='Maxout')
classify_1 = FullConnLayer(6912, 99)
classify_2 = FullConnLayer(99, 99)
attr_decomp = DecompLayer(
[(sz,) for sz in output_sizes],
[actfuncs.softmax] * len(attrconf.unival) + \
[actfuncs.sigmoid] * len(attrconf.multival)
)
model = NeuralNet([input_decomp, columns, feature_comp, classify_1, classify_2, attr_decomp])
# Build up adapter
adapter = DecompLayer([(sz,) for sz in target_sizes])
# Build up evaluator
cost_functions = [costfuncs.mean_negative_loglikelihood] * len(attrconf.unival) + \
[costfuncs.mean_binary_cross_entropy] * len(attrconf.multival)
error_functions = [costfuncs.mean_number_misclassified] * len(attrconf.unival) + \
[costfuncs.mean_zeroone_error_rate] * len(attrconf.multival)
evaluator = Evaluator(model, cost_functions, error_functions, adapter,
regularize=1e-3)
# Train the feature extraction model
sgd.train(evaluator, dataset,
learning_rate=1e-3, momentum=0.9,
batch_size=300, n_epoch=200,
learning_rate_decr=1.0, patience_incr=1.5)
return model
train_set, valid_set, test_set = cPickle.load(f)
train_set = (train_set[0], train_set[1].reshape(train_set[1].shape[0], 1))
valid_set = (valid_set[0], valid_set[1].reshape(valid_set[1].shape[0], 1))
test_set = (test_set[0], test_set[1].reshape(test_set[1].shape[0], 1))
datasets = Datasets(train_set=train_set,
valid_set=valid_set,
test_set=test_set)
# Build up the model and evaluator
layers = [
ConvPoolLayer((20, 1, 5, 5), (2, 2), (1, 28, 28), actfuncs.tanh, False),
ConvPoolLayer((50, 20, 5, 5), (2, 2), None, actfuncs.tanh, True),
FullConnLayer(800, 500, actfuncs.tanh),
FullConnLayer(500, 10, actfuncs.softmax)
]
model = NeuralNet(layers)
evaluator = Evaluator(model, costfuncs.mean_negative_loglikelihood,
costfuncs.mean_number_misclassified)
# Train the model
sgd.train(evaluator,
datasets,
learning_rate=0.1,
batch_size=500,
n_epoch=200,
learning_rate_decr=1.0)
borrow=True)
W2 = theano.shared(numpy.asarray([[4, 3], [2, 1]], dtype=numpy.float32),
borrow=True)
# Build up model
decomp = DecompLayer([(2, ), (2, )])
columns = MultiwayNeuralNet(
[FullConnLayer(2, 2, W=W1),
FullConnLayer(2, 2, W=W2)])
comp = CompLayer()
multitask = DecompLayer([(1, ), (3, )])
model = NeuralNet([decomp, columns, comp, multitask])
# Build up the target value adapter
adapter = DecompLayer([(1, ), (3, )])
# Build up evaluator
evaluator = Evaluator(model, [mse, mse], [mse, mse], adapter)
# Compute the expression by using the model
x = T.matrix('x')
y = T.matrix('y')
output = model.get_output(x)
cost, updates = evaluator.get_cost_updates(x, y, 1.0)
error = evaluator.get_error(x, y)
