def __init__(self, **kwargs):
"""
Initializes geometric mean layer parameters.
Parameters
----------
: dict like object
Contains hyperparameters
"""
# Set parameter dictionary
self.__params = dd()
self.__params.encoder_weights = None
if 'import_path' in kwargs:
for param in self.__params.values:
fn = os.path.join(kwargs.import_path, "{}.npy".format(str(param)))
param.set_value(np.load(fn))
print 'Loading parameter {} from file with shape {}'.format(str(param), param.get_value().shape)
# Hyperparameters
self.hp = dd(kwargs)
# Parameter not to be optimized
init = binomial(1, self.hp.selection_prob, size=(self.hp.nb_out, self.hp.nb_out)).astype(th.config.floatX)
init /= np.sqrt((init**2).sum(0))
self.encoder_selection = th.shared(init, name='encoder_selection')
# Format layer_id
self.layer_id = '_' + str(self.hp.layer_id) if 'layer_id' in self.hp else ''
self.output_debug = False
def __init__(self, hp, layer_id=''):
"""
Initializes maxout parameters.
Parameters
----------
hp dict like object
Contains hyperparameters
.nb_inp = number of layer inputs
.nb_max = number of inputs per maxout units
.nb_out = number of maxout units
.inp_corruption_level
.inp_corruption_type
.hid_corruption_level
.hid_corruption_type
"""
# Format layer_id
self.layer_id = '_' + str(layer_id)
# Hyperparameters
self.hp = hp
# Set parameter dictionary
self.__params = [('encoder_weights', None)]
self.__params = dd(self.__params)
def __init__(self, hp, layer_id=''):
"""
Initializes geometric mean layer parameters.
Parameters
----------
hp: dict like object
Contains hyperparameters
.nb_inp = number of layer inputs
.nb_geo = number of elements per geometric mean
.nb_out = number of output units
.inp_corruption_level
.inp_corruption_type
.hid_corruption_level
.hid_corruption_type
.debug_path
.patch_sz = input patch size (if coming from image), for debug purpose
"""
self.output_debug = False
# Format layer_id
self.layer_id = '_' + str(layer_id)
# Hyperparameters
self.hp = hp
# Set parameter dictionary
self.__params = dd()
self.__params.encoder_weights = None
self.__params.selection_biases = None
self.__params.selection_factor = None
"""
for param in self.__params.values:
fn = os.path.join(path, "{}.npy".format(str(param)))
param.set_value(np.load(fn))
print 'Loading parameter {} from file with shape {}'.format(
str(param),
param.get_value().shape)
if __name__ == '__main__':
# For testing the maxout model
x = T.matrix('tensor2', dtype=th.config.floatX)
hp = dd()
hp.nb_inp = 5
hp.nb_max = 3
hp.nb_out = 7
hp.inp_corruption_level = 0
hp.inp_corruption_type = None
hp.hid_corruption_level = 0
hp.hid_corruption_type = None
params = maxout_params(hp)
layer = maxout(x, params)
h = layer.enc(x)
d = layer.dec(h)
c = layer.cost(d, x)
enc_fn = th.function(inputs=[x], outputs=h)
dec_fn = th.function(inputs=[x], outputs=d)
make_dir(join(debug_path, 'histograms'))
self.__pca_weights = pca_weights
self.__patch_sz = patch_sz
self.__debug_path = debug_path
self.__prefix = '' if prefix == None else prefix
self.output_debug = True
if __name__ == '__main__':
from learn.utils.activation import relu, sigmoid, linear
# For testing the maxout model
x = T.matrix('tensor2', dtype=th.config.floatX)
hp = dd()
hp.nb_inp = 5
hp.nb_geo = 3
hp.nb_out = 7
hp.inp_corruption_level = 0
hp.inp_corruption_type = None
hp.hid_corruption_level = 0
hp.hid_corruption_type = None
hp = dd()
hp.type = 'geometric'
hp.nb_inp = 25
hp.nb_geo = 3
hp.nb_out = 7
hp.variant = 'product'
hp.inp_act = relu if hp.variant == 'product' else linear