diff options
author | Marcin Chrzanowski <m@m-chrzan.xyz> | 2021-05-27 19:55:59 +0200 |
---|---|---|
committer | Marcin Chrzanowski <m@m-chrzan.xyz> | 2021-05-27 19:55:59 +0200 |
commit | 3e4924d0dcec2eba4f3019b55178cd1c3b70a474 (patch) | |
tree | 07e37709e52ce7f69e7c676b04e150038d3b94b6 | |
parent | bb6a6004bfb5398b5cdf8f8a973b47e5659aec79 (diff) |
Fix position indexing in encoding
-rw-r--r-- | util/util.py | 7 |
1 files changed, 4 insertions, 3 deletions
diff --git a/util/util.py b/util/util.py index 419c23a..65f1838 100644 --- a/util/util.py +++ b/util/util.py @@ -6,11 +6,12 @@ positional_encoding = None def get_positional_encoding(n_positions, n_dimensions, device='cpu'): global positional_encoding if positional_encoding is None: - numerators = torch.tensor(range(n_positions)).repeat(n_dimensions, 1).T + # Number positions from 1 instead of 0, to avoid repeated values in + # first row of encoding + numerators = 1 + torch.tensor(range(n_positions)).repeat(n_dimensions, 1).T denominators = 10000 ** (torch.tensor(range(n_dimensions)) // 2 * 2 / n_dimensions) - print('denoms:', denominators) positional_encoding = numerators / denominators positional_encoding[:, ::2] = torch.sin(positional_encoding[:, ::2]) positional_encoding[:, 1::2] = torch.cos(positional_encoding[:, 1::2]) # output shape: (seqlen, hiddendim) - return torch.tensor(positional_encoding, dtype=torch.float, device=device) + return positional_encoding |