`tril` function for tensors

serhii · January 12, 2018, 7:09pm

Hello everyone.
Does anyone know how to make tril work on tensor without using a loop as suggested here? I would appreciate any help.

SimonW · January 12, 2018, 7:48pm

Do you mean batched tril?

tom · January 12, 2018, 7:51pm

I used numpy’s triu_indices to set things to 0 above the diagonal (see below). An alternative can be you could also use a mask.

github.com

t-vi/candlegp/blob/59e03ce263a1ccb5972c4eca80617e3698db049a/candlegp/conditionals.py#L21


# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.




import torch
from torch.autograd import Variable
import numpy


def batch_tril(A):
B = A.clone()
ii,jj = numpy.triu_indices(B.size(-2), k=1, m=B.size(-1))
B[...,ii,jj] = 0
return B


def batch_diag(A):
ii,jj = numpy.diag_indices(min(A.size(-2),A.size(-1)))
return A[...,ii,jj]

Best regards

Thomas

serhii · January 12, 2018, 7:52pm

Yes, that is what I meant.

serhii · January 12, 2018, 7:55pm

Nice, this is a good workaround. Thank you a lot, Thomas!