def trl(x, weight, bias=None, **kwargs):
"""Tensor Regression Layer
Parameters
----------
x : torch.tensor
batch of inputs
weight : FactorizedTensor
factorized weights of the TRL
bias : torch.Tensor, optional
1D tensor, by default None
Returns
-------
result
input x contracted with regression weights
"""
if isinstance(weight, TuckerTensor):
return tucker_trl(x, weight, bias=bias, **kwargs)
else:
if bias is None:
return tenalg.inner(x, weight.to_tensor(), n_modes=tl.ndim(x) - 1)
else:
return tenalg.inner(x, weight.to_tensor(),
n_modes=tl.ndim(x) - 1) + bias
def tucker_trl(x, weight, project_input=False, bias=None):
n_input = tl.ndim(x) - 1
if project_input:
x = tenalg.multi_mode_dot(x,
weight.factors[:n_input],
modes=range(1, n_input + 1),
transpose=True)
regression_weights = tenalg.multi_mode_dot(weight.core,
weight.factors[n_input:],
modes=range(
n_input, weight.order))
else:
regression_weights = weight.to_tensor()
if bias is None:
return tenalg.inner(x, regression_weights, n_modes=tl.ndim(x) - 1)
else:
return tenalg.inner(x, regression_weights,
n_modes=tl.ndim(x) - 1) + bias
ALLDELTA = [delta1, delta2, delta3]
def nd_id(n, d):
out = np.zeros((n, ) * d)
out[[np.arange(n)] * d] = 1
return out
ER = nd_id(3, 3)
#print("fdghkjdfhgkdj")
#identity tesnor that i should find
#print(ER)
ans = ten.inner(ALLU[0], ER, 1)
print(ans)
print("jksfkjsdf")
for i in range(0, P):
for j in range(1, M):
ans = ten.inner(ans, ALLU[j], j)
b = w**P - ans
sum2 = sum2 + LA.norm(b, 'fro')**2
sum3 = 0
Gamma = [2, 1, 3]
for n in range(0, M):
sum3 = sum3 + Gamma[n] * LA.norm(ALLU[n], 1)
#print(LA.norm(b, 'fro'))
for m in range(1, M):
for k in range(1, Nprim):
def forward(self, x):
regression_weights = tl.tucker_to_tensor((self.core, self.factors))
return inner(x, regression_weights, n_modes=tl.ndim(x) - 1) + self.bias