lawlict/eend_loss

End-to-end diarization loss

eend_loss's Issues

raise HungarianError("Unable to find results. Algorithm has failed.")

Traceback (most recent call last):
hungarian.py line 141, in calculate raise HungarianError("Unable to find results. Algorithm has failed.")
Will finalize trainer extensions and updater before reraising the exception.
eend.chainer_backend.hungarian.HungarianError: Unable to find results. Algorithm has failed.

I use OPTM loss in EEND_EDA, and this question comes up. It is strange that the program runs for some time and then terminates. How can I solve the error?

[J total [..................................................] 0.09%
this epoch [#.................................................] 2.21%
100 iter, 0 epoch / 25 epochs
inf iters/sec. Estimated time to finish: 0:00:00.
�[4A�[J total [..................................................] 0.18%
this epoch [##................................................] 4.42%
200 iter, 0 epoch / 25 epochs
0.39093 iters/sec. Estimated time to finish: 3 days, 8:12:04.111484.
�[4A�[J total [..................................................] 0.27%
this epoch [###...............................................] 6.63%
300 iter, 0 epoch / 25 epochs
0.3951 iters/sec. Estimated time to finish: 3 days, 7:17:04.064572.
�[4A�[J total [..................................................] 0.35%
this epoch [####..............................................] 8.84%
400 iter, 0 epoch / 25 epochs
0.39668 iters/sec. Estimated time to finish: 3 days, 6:53:54.777250.
�[4A�[J total [..................................................] 0.44%
this epoch [#####.............................................] 11.06%
500 iter, 0 epoch / 25 epochs
0.39872 iters/sec. Estimated time to finish: 3 days, 6:25:33.341658.
�[4A�[J total [..................................................] 0.53%
this epoch [######............................................] 13.27%
600 iter, 0 epoch / 25 epochs
0.39889 iters/sec. Estimated time to finish: 3 days, 6:19:22.559402.
�[4A�[J total [..................................................] 0.62%
this epoch [#######...........................................] 15.48%
700 iter, 0 epoch / 25 epochs
0.39926 iters/sec. Estimated time to finish: 3 days, 6:10:47.791780.
�[4A�[JException in main training loop: Unable to find results. Algorithm has failed.
Traceback (most recent call last):
hungarian.py line 141, in calculate raise HungarianError("Unable to find results. Algorithm has failed.")
Will finalize trainer extensions and updater before reraising the exception.
eend.chainer_backend.hungarian.HungarianError: Unable to find results. Algorithm has failed.

raise HungarianError("Unable to find results. Algorithm has failed.")

When I input :profit_matrix = np.array([[0.47341415,0.50526756,0.5631776 ,0.55596864,0.5414672,0.5414672],
[0.6781792,0.23801588,0.20288624,0.55276686,0.109804,0.109804],
[0.42393875,0.370307,0.37129858,0.5507781,0.2596858,0.2596858],
[0.52666223,0.57174313,0.5873437,0.6438369,0.5166061,0.5166061],
[0.6931475,0.6931475,0.6931475,0.6931475,0.6931475,0.6931475],
[0.6931475,0.6931475,0.6931475,0.6931475,0.6931475,0.6931475]])
a error will appear as follows.
File "models.py", line 41, in minimize_loss
H.calculate(np_loss_mat)
File "hungarian.py", line 151, in calculate
raise HungarianError("Unable to find results. Algorithm has failed.")
I do not know the reasons. Look forward to your reply.

Is OPTM really faster?

Thanks a lot for the implementation, it's great to have to have all three possibilities in the same code-base to be able to benchmark those.

We have something similar in Asteroid, (PITLossWrapper)[https://github.com/mpariente/asteroid/blob/master/asteroid/losses/pit_wrapper.py#L8], which can convert any loss function into a PIT loss function.
I've had a look at the Hungarian algorithm in the past and thought about integrating it but I thought the time spent on CPU-GPU transfers would outweigh the gain of the Hungarian algorithm.

I cannot access your paper so I don't see if you made the experiments, but I guess yes. Did you actually gain some time or lose some time when using OPTMLoss against FastPITLoss?

Thanks,
Manu