diff options
author | Marcin Chrzanowski <m@m-chrzan.xyz> | 2021-05-25 17:37:48 +0200 |
---|---|---|
committer | Marcin Chrzanowski <m@m-chrzan.xyz> | 2021-05-25 17:37:48 +0200 |
commit | fdf4bff14c1d694dbfae4e82c05226a7c9135ca5 (patch) | |
tree | 23abb00e9ebe0a2515e5fcbb8815835cd707ceaa | |
parent | b60ec2754313b013a29aa068bea7a55ebe00453c (diff) |
Implement positional encoding
-rw-r--r-- | util/util.py | 17 |
1 files changed, 12 insertions, 5 deletions
diff --git a/util/util.py b/util/util.py index 102ad93..419c23a 100644 --- a/util/util.py +++ b/util/util.py @@ -1,9 +1,16 @@ import torch import numpy as np +positional_encoding = None + def get_positional_encoding(n_positions, n_dimensions, device='cpu'): - # TODO: implement positional encoding - positional_encoding = np.zeros((n_positions, n_dimensions)) # placeholder - pass - # output shape: (seqlen, hiddendim) - return torch.tensor(positional_encoding, dtype=torch.float, device=device) + global positional_encoding + if positional_encoding is None: + numerators = 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) |