def __init__(self,
input,
W,
b,
filter_shape,
poolsize=(2, 2, 2),
activation=relu):
self.input = input
# self.W = theano.shared(value = W, borrow = True)
# self.b = theano.shared(value = b, borrow = True)
self.W = sharedata(W)
self.b = sharedata(b)
# self.W = theano.shared(value = numpy.asarray(W), borrow = True)
# self.b = theano.shared(value = numpy.asarray(b), borrow = True)
conv = T.nnet.conv3d2d.conv3d(signals=self.input,
filters=self.W,
signals_shape=None,
filters_shape=filter_shape,
border_mode='valid')
pooled_out = max_pool_3d(input=conv, ds=poolsize, ignore_border=True)
# non-linear function
self.output = activation(pooled_out + self.b.dimshuffle(
'x', 'x', 0, 'x', 'x')) # arg activitaion function
# store parameters of this layer
self.params = [self.W, self.b]
def __init__(self, rng, input, input_shape, filter, filter_shape, base, activation, poolsize, dtype = theano.config.floatX):
self.input = input
self.W = filter
self.b = base
# do the convolution with flip
conv_out = theano.tensor.nnet.conv3d2d.conv3d(
signals = self.input,
filters = self.W, #
signals_shape = input_shape,
filters_shape = filter_shape,
border_mode = 'valid') # the conv stride is 1
conv = conv_out + self.b.dimshuffle('x','x',0,'x','x')
if poolsize is None:
pooled_out = conv
else:
pooled_out = max_pool_3d( input = conv, ds = poolsize, ignore_border = True)
# non-linear
self.output = (
pooled_out if activation is None
else activation(pooled_out)
)
# store parameters of this layer
self.params = [self.W, self.b]
def __init__(self,
input,
W,
b,
filter_shape,
poolsize=(2, 2, 2),
activation=relu):
self.input = input #Lo mismo el shape.eval() aqui tampoco funciona y es del tipo theano.tensor.var.TensorVariable
#hacen esto para convertir lo que llega en W que es un CudaArray no una shared variable en
#una shared variable como era params_L0[0] Osea hacen una cosa para desacerla no muy sense...
self.W = theano.shared(value=W, borrow=True)
self.b = theano.shared(value=b, borrow=True)
# conv tampoco funciona el shape y es...<class 'theano.tensor.var.TensorVariable'>
conv = T.nnet.conv3d2d.conv3d(
signals=self.
input, #( batch_size, time, in_channels, height, width )
filters=self.
W, #( num_of_filters, flt_time, in_channels, flt_height, flt_width)
signals_shape=None,
filters_shape=filter_shape,
border_mode='valid')
pooled_out = max_pool_3d(input=conv, ds=poolsize, ignore_border=True)
# non-linear function
#Aqui esto está bien lo que no se es si esta bien hacer lo de sumar la b antes del pooling
# por supuerto en output tampoco va shape.eval() <class 'theano.tensor.var.TensorVariable'>
self.output = activation(pooled_out + self.b.dimshuffle(
'x', 'x', 0, 'x', 'x')) # arg activitaion function
# store parameters of this layer
self.params = [self.W, self.b]
def __init__(self,
input,
filter,
base,
activation,
poolsize,
dtype=theano.config.floatX):
"""
Allocate a Conv3dLayer with shared variable internal parameters.
:type input: theano.tensor
:param input: 5D matrix -- (batch_size, time, in_channels, height, width)
:type filter:
:param filter: 5D matrix -- (num_of_filters, flt_time, in_channels, flt_height, flt_width)
:type filters_shape: tuple or list of length 5
:param filter_shape:(number_of_filters, flt_time,in_channels,flt_height,flt_width)
:type base: tuple or list of length number_of_filters
:param base:(number_of_filters)
:param activation: non-linear activation function, typically relu or tanh
:poolsize: tuple or list of length 3
:param poolsize: the pooling stride, typically (2,2,2)
"""
self.input = input
self.W = filter
self.b = base
# do the 3d convolution --- have flip
conv_out = theano.tensor.nnet.conv3d2d.conv3d(
signals=self.input,
filters=self.W,
signals_shape=None,
filters_shape=None,
border_mode='valid') # the convolution stride is 1
conv = conv_out + self.b.dimshuffle('x', 'x', 0, 'x', 'x')
if poolsize is None:
pooled_out = conv
else:
pooled_out = max_pool_3d(input=conv,
ds=poolsize,
ignore_border=True)
# non-linear function
self.output = (pooled_out
if activation is None else activation(pooled_out))
# store parameters of this layer
self.params = [self.W, self.b]
class myConvPool3dLayer(object):
def __init(self,input,W,b,filter_shape,poolsize=(1,1,1),activation=relu):
self.input = input
self.W = theano.shared(numpy.asarray(W,dtype = theano.config.floatX),borrow=True)
self.b = theano.shared(numpy.asarray(W,dtype = theano.config.floatX),borrow=True)
conv = T.nnet.conv3d2d.conv3d(
signals = self.input,
filters = self.W,
signals_shape = None,
filters_shape = filter_shape,
border_mode = 'valid')
pooled_out = max_pool_3d(
input = conv,
ds = poolsize,
ignore_border = True)
#After convolution and pooling if needed (if pooling = (1,1,1)) is like doing nothing
self.output = activation( pooled_out + self.b.dimshuffle('x','x',0,'x','x'))
self.params = [self.W, self.b]
def __init__(self,
input,
filter,
base,
activation,
poolsize,
dtype=theano.config.floatX):
"""
Allocate a Conv3dLayer with shared variable internal parameters.
:type input: theano.tensor
:param input: 5D matrix -- (batch_size, time, in_channels, height, width)
:type filter:
:param filter: 5D matrix -- (num_of_filters, flt_time, in_channels, flt_height, flt_width)
:type filters_shape: tuple or list of length 5
:param filter_shape:(number_of_filters, flt_time,in_channels,flt_height,flt_width)
:type base: tuple or list of length number_of_filters
:param base:(number_of_filters)
:param activation: non-linear activation function, typically relu or tanh
:poolsize: tuple or list of length 3
:param poolsize: the pooling stride, typically (2,2,2)
"""
self.input = input #shared variable el self.input #solo la primera vez.
self.W = filter #Theano <class 'theano.tensor.var.TensorVariable'>
self.b = base #Shared
# do the 3d convolution --- have flip
conv_out = theano.tensor.nnet.conv3d2d.conv3d(
signals=self.input,
filters=self.
W, #Tensor Variable ya no es shared anymore DISTINTO QE EN CNN3D
signals_shape=None,
filters_shape=None,
border_mode='valid') # the convolution stride is 1
#conv_out = array([ 1, 76, 64, 176, 176]) <class 'theano.tensor.var.TensorVariable'>
#Aqui convout se puede evaluar lul!
#2 array([ 1, 36, 64, 86, 86])
#3 array([ 1, 36, 64, 84, 84])
#4 array([ 1, 35, 150,83, 83])
#5 array([ 1, 35, 2, 83, 83])
conv = conv_out + self.b.dimshuffle('x', 'x', 0, 'x', 'x')
if poolsize is None:
pooled_out = conv
else:
pooled_out = max_pool_3d(input=conv,
ds=poolsize,
ignore_border=True)
#array([ 1, 38, 64, 88, 88])
#<class 'theano.tensor.var.TensorVariable'>
#2array([ 1, 36, 64, 86, 86])
# non-linear function
self.output = (pooled_out
if activation is None else activation(pooled_out))
# store parameters of this layer
self.params = [self.W, self.b]
