Method forward

architecture/embeddings.py:2051–2093

r""" Args: x (`torch.Tensor`): Tensor of shape `(B, S, D)` of input tokens. mask (`torch.Tensor`): Boolean ensor of shape `(B, S)` indicating which tokens are not padding. Returns: `torch.Tensor`: `(

(self, x: torch.Tensor, mask: torch.BoolTensor)

        return pooled

    def forward(self, x: torch.Tensor, mask: torch.BoolTensor) -> torch.Tensor:
        r"""
        Args:
            x (`torch.Tensor`):
                Tensor of shape `(B, S, D)` of input tokens.
            mask (`torch.Tensor`):
                Boolean ensor of shape `(B, S)` indicating which tokens are not padding.

        Returns:
            `torch.Tensor`:
                `(B, D)` tensor of pooled tokens.
        """
        D = x.size(2)

        # Construct attention mask, shape: (B, 1, num_queries=1, num_keys=1+L).
        attn_mask = mask[:, None, None, :].bool()  # (B, 1, 1, L).
        attn_mask = F.pad(attn_mask, (1, 0), value=True)  # (B, 1, 1, 1+L).

        # Average non-padding token features. These will be used as the query.
        x_pool = self.pool_tokens(x, mask, keepdim=True)  # (B, 1, D)

        # Concat pooled features to input sequence.
        x = torch.cat([x_pool, x], dim=1)  # (B, L+1, D)

        # Compute queries, keys, values. Only the mean token is used to create a query.
        kv = self.to_kv(x)  # (B, L+1, 2 * D)
        q = self.to_q(x[:, 0])  # (B, D)

        # Extract heads.
        head_dim = D // self.num_attention_heads
        kv = kv.unflatten(2, (2, self.num_attention_heads, head_dim))  # (B, 1+L, 2, H, head_dim)
        kv = kv.transpose(1, 3)  # (B, H, 2, 1+L, head_dim)
        k, v = kv.unbind(2)  # (B, H, 1+L, head_dim)
        q = q.unflatten(1, (self.num_attention_heads, head_dim))  # (B, H, head_dim)
        q = q.unsqueeze(2)  # (B, H, 1, head_dim)

        # Compute attention.
        x = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, dropout_p=0.0)  # (B, H, 1, head_dim)

        # Concatenate heads and run output.
        x = x.squeeze(2).flatten(1, 2)  # (B, D = H * head_dim)
        x = self.to_out(x)
        return x


def get_fourier_embeds_from_boundingbox(embed_dim, box):