This is a more standard version of the position embedding, very similar to the one used by the Attention is all you need paper, generalized to work on images.
| 121 | |
| 122 | |
| 123 | class PositionEmbeddingSine_1D(nn.Module): |
| 124 | """This is a more standard version of the position embedding, very similar |
| 125 | to the one used by the Attention is all you need paper, generalized to work |
| 126 | on images.""" |
| 127 | |
| 128 | def __init__(self, |
| 129 | num_pos_feats=64, |
| 130 | temperature=10000, |
| 131 | normalize=True, |
| 132 | scale=None): |
| 133 | super().__init__() |
| 134 | self.num_pos_feats = num_pos_feats |
| 135 | self.temperature = temperature |
| 136 | self.normalize = normalize |
| 137 | if scale is not None and normalize is False: |
| 138 | raise ValueError('normalize should be True if scale is passed') |
| 139 | if scale is None: |
| 140 | scale = 2 * math.pi |
| 141 | self.scale = scale |
| 142 | |
| 143 | def forward(self, B, L): |
| 144 | |
| 145 | position = torch.arange(0, L, dtype=torch.float32).unsqueeze(0) |
| 146 | position = position.repeat(B, 1) |
| 147 | |
| 148 | if self.normalize: |
| 149 | eps = 1e-6 |
| 150 | position = position / (position[:, -1:] + eps) * self.scale |
| 151 | |
| 152 | dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32) |
| 153 | dim_t = self.temperature**(2 * (torch.div(dim_t, 1)) / |
| 154 | self.num_pos_feats) |
| 155 | |
| 156 | pe = torch.zeros(B, L, self.num_pos_feats * 2) |
| 157 | pe[:, :, 0::2] = torch.sin(position[:, :, None] / dim_t) |
| 158 | pe[:, :, 1::2] = torch.cos(position[:, :, None] / dim_t) |
| 159 | |
| 160 | pe = pe.permute(1, 0, 2) |
| 161 | |
| 162 | return pe |
| 163 | |
| 164 | |
| 165 | class DeciWatch(nn.Module): |
no outgoing calls
no test coverage detected