Sorobot/Document-modify-Toolkit

1import math2from typing import Optional3 4import paddle5import paddle.nn as nn6import paddle.nn.initializer as init7 8 9class NestedTensor(object):10    def __init__(self, tensors, mask: Optional[paddle.Tensor]):11        self.tensors = tensors12        self.mask = mask13 14    def decompose(self):15        return self.tensors, self.mask16 17    def __repr__(self):18        return str(self.tensors)19 20 21class PositionEmbeddingSine(nn.Layer):22    """23    This is a more standard version of the position embedding, very similar to the one24    used by the Attention is all you need paper, generalized to work on images.25    """26 27    def __init__(28        self, num_pos_feats=64, temperature=10000, normalize=False, scale=None29    ):30        super().__init__()31        self.num_pos_feats = num_pos_feats32        self.temperature = temperature33        self.normalize = normalize34 35        if scale is not None and normalize is False:36            raise ValueError("normalize should be True if scale is passed")37 38        if scale is None:39            scale = 2 * math.pi40 41        self.scale = scale42 43    def forward(self, mask):44        assert mask is not None45 46        y_embed = mask.cumsum(axis=1, dtype="float32")47        x_embed = mask.cumsum(axis=2, dtype="float32")48 49        if self.normalize:50            eps = 1e-0651            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale52            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale53 54        dim_t = paddle.arange(end=self.num_pos_feats, dtype="float32")55        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)56 57        pos_x = x_embed[:, :, :, None] / dim_t58        pos_y = y_embed[:, :, :, None] / dim_t59 60        pos_x = paddle.stack(61            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), axis=462        ).flatten(3)63        pos_y = paddle.stack(64            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), axis=465        ).flatten(3)66 67        pos = paddle.concat((pos_y, pos_x), axis=3).transpose(perm=[0, 3, 1, 2])68 69        return pos70 71 72class PositionEmbeddingLearned(nn.Layer):73    """74    Absolute pos embedding, learned.75    """76 77    def __init__(self, num_pos_feats=256):78        super().__init__()79        self.row_embed = nn.Embedding(50, num_pos_feats)80        self.col_embed = nn.Embedding(50, num_pos_feats)81        self.reset_parameters()82 83    def reset_parameters(self):84        init_Constant = init.Uniform()85        init_Constant(self.row_embed.weight)86        init_Constant(self.col_embed.weight)87 88    def forward(self, tensor_list: NestedTensor):89        x = tensor_list.tensors90 91        h, w = x.shape[-2:]92 93        i = paddle.arange(end=w)94        j = paddle.arange(end=h)95 96        x_emb = self.col_embed(i)97        y_emb = self.row_embed(j)98 99        pos = (100            paddle.concat(101                [102                    x_emb.unsqueeze(0).tile([h, 1, 1]),103                    y_emb.unsqueeze(1).tile([1, w, 1]),104                ],105                axis=-1,106            )107            .transpose([2, 0, 1])108            .unsqueeze(0)109            .tile([x.shape[0], 1, 1, 1])110        )111 112        return pos113 114 115def build_position_encoding(hidden_dim=512, position_embedding="sine"):116    N_steps = hidden_dim // 2117 118    if position_embedding in ("v2", "sine"):119        position_embedding = PositionEmbeddingSine(N_steps, normalize=True)120    elif position_embedding in ("v3", "learned"):121        position_embedding = PositionEmbeddingLearned(N_steps)122    else:123        raise ValueError(f"not supported {position_embedding}")124    return position_embedding125