Randomly crop a sequence from each sequence in xs. Args: xs: (seq_len, batch, *) input sequences valid_lens: (batch,) valid sequence length of xs min_ratio: min ratio in valid sequence length max_ratio: max r
(
xs: torch.Tensor,
valid_lens: T.Union[torch.Tensor, T.List[int]],
min_ratio: float = 0.1,
max_ratio: float = 1.0,
pad_val: float = 0.,
min_seq_len: int = 1,
)
| 43 | |
| 44 | |
| 45 | def random_crop( |
| 46 | xs: torch.Tensor, |
| 47 | valid_lens: T.Union[torch.Tensor, T.List[int]], |
| 48 | min_ratio: float = 0.1, |
| 49 | max_ratio: float = 1.0, |
| 50 | pad_val: float = 0., |
| 51 | min_seq_len: int = 1, |
| 52 | ) -> T.Tuple[torch.Tensor, T.Union[torch.Tensor, T.List[int]]]: |
| 53 | """ |
| 54 | Randomly crop a sequence from each sequence in xs. |
| 55 | |
| 56 | Args: |
| 57 | xs: |
| 58 | (seq_len, batch, *) input sequences |
| 59 | valid_lens: |
| 60 | (batch,) valid sequence length of xs |
| 61 | min_ratio: |
| 62 | min ratio in valid sequence length |
| 63 | max_ratio: |
| 64 | max ratio in valid sequence length |
| 65 | pad_val: |
| 66 | value used to pad when forming a tensor |
| 67 | min_seq_len: |
| 68 | minimum sequence length in the cropped sequence |
| 69 | |
| 70 | Returns: |
| 71 | xs_cropped: |
| 72 | (seq_len_2, batch, *). cropped sequences |
| 73 | new_valid_lens: |
| 74 | (batch,) valid sequence length of xs_cropped. |
| 75 | """ |
| 76 | if valid_lens is None: |
| 77 | valid_lens = [xs.size(0)] * xs.size(1) |
| 78 | |
| 79 | if isinstance(valid_lens, (list, tuple)): |
| 80 | valid_lens = torch.tensor(valid_lens) |
| 81 | is_list = True |
| 82 | valid_lens_device = None |
| 83 | else: |
| 84 | valid_lens_device = valid_lens.device |
| 85 | is_list = False |
| 86 | |
| 87 | # determine new len |
| 88 | rs = torch.rand(xs.size(1), device=xs.device) * (max_ratio - min_ratio) + min_ratio # (batch,) |
| 89 | new_valid_lens = torch.maximum( |
| 90 | torch.minimum((valid_lens * rs).long(), valid_lens), |
| 91 | min_seq_len * torch.ones(1, dtype=torch.long, device=rs.device), |
| 92 | ) |
| 93 | new_valid_lens = torch.minimum(new_valid_lens, valid_lens) |
| 94 | |
| 95 | # determine starting point |
| 96 | last_idxs = (valid_lens - new_valid_lens) # (included) |
| 97 | s_idxs = (torch.rand(xs.size(1), device=xs.device) * (last_idxs + 1)).long() |
| 98 | xs_cropped = [] |
| 99 | cropped_valid_lens = [] |
| 100 | for i in range(xs.size(1)): |
| 101 | x_cropped = xs[s_idxs[i]:s_idxs[i] + new_valid_lens[i], i] |
| 102 | assert x_cropped.size(0) > 0 |