Giters

yandex-research / rtdl

Research on Tabular Deep Learning: Papers & Packages

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

# bug, located in rtdl.data.piecewise_linear_encoding line #618

grayicon opened this issue · comments

commented

"is_last_bin = bin_indices + 1 == torch.as_tensor(list(map(len, bin_edges)))is_last_bin = bin_indices + 1 == torch.as_tensor(list(map(len, bin_edges)))" should be "is_last_bin = bin_indices + 1 == torch.as_tensor(list(map(len, bin_edges)))is_last_bin = bin_indices + 1 == torch.as_tensor(list(map(len, bin_edges))).unsqueeze(-1))", to use tensor broadcast func

commented

Thank you for the report. Currently, the main branch contains significant amount of unfinished work not released to PyPI, so bugs are expected, and use it at your own risk.

P.S. The formatted version of the above bug report:

instead of this:

is_last_bin = bin_indices + 1 == torch.as_tensor(list(map(len, bin_edges)))

it should be this:

is_last_bin = bin_indices + 1 == torch.as_tensor(list(map(len, bin_edges))).unsqueeze(-1))
commented

@grayicon is it actually a bug?

  • bin_indices + 1 has the shape (n_objects, n_features)
  • as_tensor(list(map(len, bin_edges))) has the shape (n_features,)
  • According to broadcasting rules, the result is_last_bin has the shape (n_objects, n_features)
commented

Feel free to reopen the issue if needed.

commented

@grayicon is it actually a bug?

  • bin_indices + 1 has the shape (n_objects, n_features)
  • as_tensor(list(map(len, bin_edges))) has the shape (n_features,)
  • According to broadcasting rules, the result is_last_bin has the shape (n_objects, n_features)

The above desc is right in theory, but in fact i found the real bug lie in the line #590 of rtdl.data.py "bin_edges = torch.as_tensor(bin_ratios)", which cause the state "as_tensor(list(map(len, bin_edges)))" has the shape (n_objects,)

commented

So it seems to be the issue reported here, right?

#47

commented

So it should be now resolved by 7929497

Please, let me know if the issue still there

commented

So it should be now resolved by 7929497

Please, let me know if the issue still there

The bin_egdes is a list of torch tensor, So it doesn't support the opera of 'torch.as_tensor'. Consequently, I just comment this line "bin_edges = torch.as_tensor(bin_ratios)", and it will works when call of "map(len, bin_edges) below".

commented

@grayicon oh, I think now I understand. I removed this line in this commit. Does it look good now?

commented

@grayicon oh, I think now I understand. I removed this line in this commit. Does it look good now?

Yes，It can work now, Thanks for your effort.