a28601751e
* chore: Absolute imports * feat: Add new mesh util * chore: replace bumpversion original fork no longer maintained, move config * ci: remove old configuration files * ci: update github action * ci: add python 3.10 test * chore: update pre-commit hooks * ci: update supported python versions supported are 3.7, 3.8 and 3.9. 3.6 had EOL in december 2021. 3.10 has no pytorch distribution yet. * ci: add windows test * ci: update action less windows tests, pre commit * ci: fix typo * chore: run precommit for all files * ci: two step tests * ci: compatibility waits for style * fix: init file had missing imports * ci: add deployment script * ci: skip complete publish step * ci: cleanup readme
94 lines
2.2 KiB
Python
94 lines
2.2 KiB
Python
"""ProtoTorch competitions"""
|
|
|
|
import torch
|
|
|
|
|
|
def wtac(distances: torch.Tensor, labels: torch.LongTensor):
|
|
"""Winner-Takes-All-Competition.
|
|
|
|
Returns the labels corresponding to the winners.
|
|
|
|
"""
|
|
winning_indices = torch.min(distances, dim=1).indices
|
|
winning_labels = labels[winning_indices].squeeze()
|
|
return winning_labels
|
|
|
|
|
|
def knnc(distances: torch.Tensor, labels: torch.LongTensor, k: int = 1):
|
|
"""K-Nearest-Neighbors-Competition.
|
|
|
|
Returns the labels corresponding to the winners.
|
|
|
|
"""
|
|
winning_indices = torch.topk(-distances, k=k, dim=1).indices
|
|
winning_labels = torch.mode(labels[winning_indices], dim=1).values
|
|
return winning_labels
|
|
|
|
|
|
def cbcc(detections: torch.Tensor, reasonings: torch.Tensor):
|
|
"""Classification-By-Components Competition.
|
|
|
|
Returns probability distributions over the classes.
|
|
|
|
`detections` must be of shape [batch_size, num_components].
|
|
`reasonings` must be of shape [num_components, num_classes, 2].
|
|
|
|
"""
|
|
A, B = reasonings.permute(2, 1, 0).clamp(0, 1)
|
|
pk = A
|
|
nk = (1 - A) * B
|
|
numerator = (detections @ (pk - nk).T) + nk.sum(1)
|
|
probs = numerator / (pk + nk).sum(1)
|
|
return probs
|
|
|
|
|
|
class WTAC(torch.nn.Module):
|
|
"""Winner-Takes-All-Competition Layer.
|
|
|
|
Thin wrapper over the `wtac` function.
|
|
|
|
"""
|
|
|
|
def forward(self, distances, labels): # pylint: disable=no-self-use
|
|
return wtac(distances, labels)
|
|
|
|
|
|
class LTAC(torch.nn.Module):
|
|
"""Loser-Takes-All-Competition Layer.
|
|
|
|
Thin wrapper over the `wtac` function.
|
|
|
|
"""
|
|
|
|
def forward(self, probs, labels): # pylint: disable=no-self-use
|
|
return wtac(-1.0 * probs, labels)
|
|
|
|
|
|
class KNNC(torch.nn.Module):
|
|
"""K-Nearest-Neighbors-Competition.
|
|
|
|
Thin wrapper over the `knnc` function.
|
|
|
|
"""
|
|
|
|
def __init__(self, k=1, **kwargs):
|
|
super().__init__(**kwargs)
|
|
self.k = k
|
|
|
|
def forward(self, distances, labels):
|
|
return knnc(distances, labels, k=self.k)
|
|
|
|
def extra_repr(self):
|
|
return f"k: {self.k}"
|
|
|
|
|
|
class CBCC(torch.nn.Module):
|
|
"""Classification-By-Components Competition.
|
|
|
|
Thin wrapper over the `cbcc` function.
|
|
|
|
"""
|
|
|
|
def forward(self, detections, reasonings): # pylint: disable=no-self-use
|
|
return cbcc(detections, reasonings)
|