def lazy_init(self):
if self.lazy:
(prev_features,n_features,filter_heigth,
filter_width,border_mode,reg,init)=self.lazy
self.weight = SharedVar(self,'weight',init((n_features,prev_features,
filter_heigth,filter_width)))
self.bias = SharedVar(self,'bias',init((n_features,)))
self.param = (self.weight,self.bias)
self.add_reg(self.weight,reg)
self.mode = border_mode
self.lazy = None
class Linear(Module):
def __init__(self,input_size,output_size,reg=None,init=Init.Uniform):
super(Linear,self).__init__()
if isinstance(init,(dict)):
self.bias = dict['bias']
self.weight=dict['weight']
self.lazy = None
else:
self.lazy = (input_size,output_size,reg,init)
self.input = T.matrix(dtype=dtype)
def params(self):
self.lazy_init()
return self.param
def lazy_init(self):
if self.lazy:
lazy = self.lazy
self.weight = SharedVar(self,
'weigth',lazy[3]((lazy[0],lazy[1])))
self.bias = SharedVar(self,'bias',lazy[3](lazy[1]))
self.param = (self.weight,self.bias)
self.add_reg(self.weight,lazy[2])
self.lazy=None
def get_output(self,input):
self.lazy_init()
return T.dot(input,self.weight)+self.bias
def __str__(self):
return "Linear({}->{})".format(*(self.weight.get_value(borrow=True).shape))
def __init__(self,prev_features,n_features,filter_heigth,
filter_width,border_mode=Valid,reg=None,init=Init.Uniform):
super(Conv2D, self).__init__()
if isinstance(init,(dict)):
self.bias = dict['bias']
self.weight =dict['weigth']
else:
self.lazy = (prev_features,n_features,filter_heigth,
filter_width,border_mode,reg,init)
self.weight = SharedVar(self,'weight',init((n_features,prev_features,
filter_heigth,filter_width)))
self.bias = SharedVar(self,'bias',init((n_features,)))
self.input = T.tensor4()
self.param = (self.weight,self.bias)
self.add_reg(self.weight,reg)
self.mode = border_mode
def lazy_init(self):
if self.lazy:
lazy = self.lazy
self.weight = SharedVar(self,
'weigth',lazy[3]((lazy[0],lazy[1])))
self.bias = SharedVar(self,'bias',lazy[3](lazy[1]))
self.param = (self.weight,self.bias)
self.add_reg(self.weight,lazy[2])
self.lazy=None
class Conv2D(Module):
Full = "full"
Valid= 'valid'
def __init__(self,prev_features,n_features,filter_heigth,
filter_width,border_mode=Valid,reg=None,init=Init.Uniform):
super(Conv2D, self).__init__()
if isinstance(init,(dict)):
self.bias = dict['bias']
self.weight =dict['weigth']
else:
self.lazy = (prev_features,n_features,filter_heigth,
filter_width,border_mode,reg,init)
self.weight = SharedVar(self,'weight',init((n_features,prev_features,
filter_heigth,filter_width)))
self.bias = SharedVar(self,'bias',init((n_features,)))
self.input = T.tensor4()
self.param = (self.weight,self.bias)
self.add_reg(self.weight,reg)
self.mode = border_mode
def lazy_init(self):
if self.lazy:
(prev_features,n_features,filter_heigth,
filter_width,border_mode,reg,init)=self.lazy
self.weight = SharedVar(self,'weight',init((n_features,prev_features,
filter_heigth,filter_width)))
self.bias = SharedVar(self,'bias',init((n_features,)))
self.param = (self.weight,self.bias)
self.add_reg(self.weight,reg)
self.mode = border_mode
self.lazy = None
def params(self):
self.lazy_init()
return self.param
def get_output(self,input):
self.lazy_init()
return conv2d(input,self.weight,border_mode=self.mode) + \
self.bias.dimshuffle('x',0,'x','x')
def __str__(self):
shp = self.weight.get_value(borrow=True).shape
return "Conv2D(inMap:{},outMap:{},filter:({},{}))".format(shp[1],shp[0],shp[2],shp[3])
