def run(self, test):
p = self.get_distance_matrix(test)
p *= -1.0 # no argmin supported yet
idx = cp.dev_tensor_uint(test.shape[0])
cp.reduce_to_row(idx, p, cp.reduce_functor.ARGMAX)
hidx = idx.np.reshape(idx.shape[0])
return self.data_l.reshape(self.data.shape[0])[hidx]
def run(self,test):
p = self.get_distance_matrix(test)
p *= -1. # no argmin supported yet
idx = cp.dev_tensor_uint(test.shape[0])
cp.reduce_to_row(idx, p, cp.reduce_functor.ARGMAX)
hidx = idx.np.reshape(idx.shape[0])
return self.data_l.reshape(self.data.shape[0])[hidx]
def delta_outputSoftMax(self, calculated, correct):
derivative = calculated.copy()
cp.apply_scalar_functor(derivative, cp.scalar_functor.EXP)
sums = cp.dev_tensor_float(calculated.shape[1])
cp.fill(sums,0)
cp.reduce_to_row(sums, derivative, cp.reduce_functor.ADD)
cp.apply_scalar_functor(sums,cp.scalar_functor.ADD,0.1/derivative.shape[0])
rv = cp.transposed_view(derivative)
cp.matrix_divide_col(rv,sums)
cp.apply_binary_functor(derivative, correct, cp.binary_functor.AXPBY, -1.,1.)
sums.dealloc()
return derivative
def delta_outputSoftMax(self, calculated, correct):
derivative = calculated.copy()
cp.apply_scalar_functor(derivative, cp.scalar_functor.EXP)
sums = cp.dev_tensor_float(calculated.shape[1])
cp.fill(sums, 0)
cp.reduce_to_row(sums, derivative, cp.reduce_functor.ADD)
cp.apply_scalar_functor(sums, cp.scalar_functor.ADD,
0.1 / derivative.shape[0])
rv = cp.transposed_view(derivative)
cp.matrix_divide_col(rv, sums)
cp.apply_binary_functor(derivative, correct, cp.binary_functor.AXPBY,
-1., 1.)
sums.dealloc()
return derivative
def p_k(self,beta,tmp,tmp2,collect):
cp.prod(tmp,self.v,self.baserate_bias,'t','n')
cp.apply_scalar_functor(tmp,cp.scalar_functor.MULT,(1-beta))
collect(tmp)
cp.prod(tmp2,self.w,self.v,'t','n')
cp.matrix_plus_col(tmp2,self.bias_hi)
cp.apply_scalar_functor(tmp2,cp.scalar_functor.MULT,beta)
# RECT computes log(1+exp(x))
cp.apply_scalar_functor(tmp2,cp.scalar_functor.RECT,1)
cp.reduce_to_row(tmp.T,tmp2,cp.reduce_functor.ADD) # tmp.T is an evil hack. it makes tmp into row major, which doesn't change anything since it's a vector any way. But vectors are always assumed to be row major.
collect(tmp)
cp.prod(tmp,self.v,self.bias_lo.T,'t','n')
cp.apply_scalar_functor(tmp,cp.scalar_functor.MULT,beta)
collect(tmp)
def partialsumV(self, actv, acth, row):
"""
sums out hidden variables for given v
exp( log(exp(bh + actv*W)+1).sum(axis=0) + (v*bv).sum(axis=0) )
"""
# acth = bv + actv*W
cp.prod(acth, self.weight, actv, "t", "n")
cp.matrix_plus_col(acth, self.bh)
# acth = log(exp(acth)+1)
cp.apply_scalar_functor(acth, cp.scalar_functor.RECT, 1.0)
# row = actv.sum(axis=0)
cp.reduce_to_row(row, acth, cp.reduce_functor.ADD)
# row += h*bh
cp.matrix_times_col(actv, self.bv)
cp.reduce_to_row(row, actv, cp.reduce_functor.ADD, 1.0, 1.0)
# exp(row)
m = row.np.astype("float64")
return math.fsum(m.flatten()) / actv.shape[1]
def partialsumV(self, actv, acth, row):
"""
sums out hidden variables for given v
exp( log(exp(bh + actv*W)+1).sum(axis=0) + (v*bv).sum(axis=0) )
"""
# acth = bv + actv*W
cp.prod(acth, self.weight, actv, 't', 'n')
cp.matrix_plus_col(acth, self.bh)
# acth = log(exp(acth)+1)
cp.apply_scalar_functor(acth, cp.scalar_functor.RECT, 1.0)
# row = actv.sum(axis=0)
cp.reduce_to_row(row, acth, cp.reduce_functor.ADD)
# row += h*bh
cp.matrix_times_col(actv, self.bv)
cp.reduce_to_row(row, actv, cp.reduce_functor.ADD, 1.0, 1.0)
# exp(row)
m = row.np.astype("float64")
return math.fsum(m.flatten()) / actv.shape[1]
def partialsum(self, acth, actv, row):
"""
sums out visible variables for given hidden variables
exp( log(exp(bv + acth*W)+1).sum(axis=0) + (h*bh).sum(axis=0) )
"""
# actv = bv + acth*W
cp.prod(actv, self.weight, acth, "n", "n")
cp.matrix_plus_col(actv, self.bv)
# actv = log(exp(actv)+1)
cp.apply_scalar_functor(actv, cp.scalar_functor.RECT, 1.0)
# row = actv.sum(axis=0)
cp.reduce_to_row(row, actv, cp.reduce_functor.ADD)
# row += h*bh
cp.matrix_times_col(acth, self.bh)
cp.reduce_to_row(row, acth, cp.reduce_functor.ADD, 1.0, 1.0)
# cp.prod(row,self.bv,actv,'t','n',1.0,1.0)
# exp(row)
m = row.np.astype("float64")
return math.fsum(np.exp(m).flatten())
def partialsum(self, acth, actv, row):
"""
sums out visible variables for given hidden variables
exp( log(exp(bv + acth*W)+1).sum(axis=0) + (h*bh).sum(axis=0) )
"""
# actv = bv + acth*W
cp.prod(actv, self.weight, acth, 'n', 'n')
cp.matrix_plus_col(actv, self.bv)
# actv = log(exp(actv)+1)
cp.apply_scalar_functor(actv, cp.scalar_functor.RECT, 1.0)
# row = actv.sum(axis=0)
cp.reduce_to_row(row, actv, cp.reduce_functor.ADD)
# row += h*bh
cp.matrix_times_col(acth, self.bh)
cp.reduce_to_row(row, acth, cp.reduce_functor.ADD, 1.0, 1.0)
#cp.prod(row,self.bv,actv,'t','n',1.0,1.0)
# exp(row)
m = row.np.astype("float64")
return math.fsum(np.exp(m).flatten())