Reference for ultralytics/models/sam/modules/decoders.py

def forward(
    self,
    image_embeddings: torch.Tensor,
    image_pe: torch.Tensor,
    sparse_prompt_embeddings: torch.Tensor,
    dense_prompt_embeddings: torch.Tensor,
    multimask_output: bool,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Predicts masks given image and prompt embeddings.

    Args:
        image_embeddings (torch.Tensor): Embeddings from the image encoder.
        image_pe (torch.Tensor): Positional encoding with the shape of image_embeddings.
        sparse_prompt_embeddings (torch.Tensor): Embeddings of the points and boxes.
        dense_prompt_embeddings (torch.Tensor): Embeddings of the mask inputs.
        multimask_output (bool): Whether to return multiple masks or a single mask.

    Returns:
        (Tuple[torch.Tensor, torch.Tensor]): A tuple containing:
            - masks (torch.Tensor): Batched predicted masks.
            - iou_pred (torch.Tensor): Batched predictions of mask quality.

    Examples:
        >>> decoder = MaskDecoder(transformer_dim=256, transformer=transformer_module)
        >>> image_emb = torch.rand(1, 256, 64, 64)
        >>> image_pe = torch.rand(1, 256, 64, 64)
        >>> sparse_emb = torch.rand(1, 2, 256)
        >>> dense_emb = torch.rand(1, 256, 64, 64)
        >>> masks, iou_pred = decoder(image_emb, image_pe, sparse_emb, dense_emb, multimask_output=True)
        >>> print(f"Masks shape: {masks.shape}, IoU predictions shape: {iou_pred.shape}")
    """
    masks, iou_pred = self.predict_masks(
        image_embeddings=image_embeddings,
        image_pe=image_pe,
        sparse_prompt_embeddings=sparse_prompt_embeddings,
        dense_prompt_embeddings=dense_prompt_embeddings,
    )

    # Select the correct mask or masks for output
    mask_slice = slice(1, None) if multimask_output else slice(0, 1)
    masks = masks[:, mask_slice, :, :]
    iou_pred = iou_pred[:, mask_slice]

    # Prepare output
    return masks, iou_pred

def predict_masks(
    self,
    image_embeddings: torch.Tensor,
    image_pe: torch.Tensor,
    sparse_prompt_embeddings: torch.Tensor,
    dense_prompt_embeddings: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """Predicts masks and quality scores using image and prompt embeddings via transformer architecture."""
    # Concatenate output tokens
    output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
    output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.shape[0], -1, -1)
    tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)

    # Expand per-image data in batch direction to be per-mask
    src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
    src = src + dense_prompt_embeddings
    pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
    b, c, h, w = src.shape

    # Run the transformer
    hs, src = self.transformer(src, pos_src, tokens)
    iou_token_out = hs[:, 0, :]
    mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]

    # Upscale mask embeddings and predict masks using the mask tokens
    src = src.transpose(1, 2).view(b, c, h, w)
    upscaled_embedding = self.output_upscaling(src)
    hyper_in_list: List[torch.Tensor] = [
        self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]) for i in range(self.num_mask_tokens)
    ]
    hyper_in = torch.stack(hyper_in_list, dim=1)
    b, c, h, w = upscaled_embedding.shape
    masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)

    # Generate mask quality predictions
    iou_pred = self.iou_prediction_head(iou_token_out)

    return masks, iou_pred

def forward(
    self,
    image_embeddings: torch.Tensor,
    image_pe: torch.Tensor,
    sparse_prompt_embeddings: torch.Tensor,
    dense_prompt_embeddings: torch.Tensor,
    multimask_output: bool,
    repeat_image: bool,
    high_res_features: Optional[List[torch.Tensor]] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Predicts masks given image and prompt embeddings.

    Args:
        image_embeddings (torch.Tensor): Embeddings from the image encoder with shape (B, C, H, W).
        image_pe (torch.Tensor): Positional encoding with the shape of image_embeddings (B, C, H, W).
        sparse_prompt_embeddings (torch.Tensor): Embeddings of the points and boxes with shape (B, N, C).
        dense_prompt_embeddings (torch.Tensor): Embeddings of the mask inputs with shape (B, C, H, W).
        multimask_output (bool): Whether to return multiple masks or a single mask.
        repeat_image (bool): Flag to repeat the image embeddings.
        high_res_features (List[torch.Tensor] | None): Optional high-resolution features.

    Returns:
        (Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]): A tuple containing:
            - masks (torch.Tensor): Batched predicted masks with shape (B, N, H, W).
            - iou_pred (torch.Tensor): Batched predictions of mask quality with shape (B, N).
            - sam_tokens_out (torch.Tensor): Batched SAM token for mask output with shape (B, N, C).
            - object_score_logits (torch.Tensor): Batched object score logits with shape (B, 1).

    Examples:
        >>> image_embeddings = torch.rand(1, 256, 64, 64)
        >>> image_pe = torch.rand(1, 256, 64, 64)
        >>> sparse_prompt_embeddings = torch.rand(1, 2, 256)
        >>> dense_prompt_embeddings = torch.rand(1, 256, 64, 64)
        >>> decoder = SAM2MaskDecoder(256, transformer)
        >>> masks, iou_pred, sam_tokens_out, obj_score_logits = decoder.forward(
        ...     image_embeddings, image_pe, sparse_prompt_embeddings, dense_prompt_embeddings, True, False
        ... )
    """
    masks, iou_pred, mask_tokens_out, object_score_logits = self.predict_masks(
        image_embeddings=image_embeddings,
        image_pe=image_pe,
        sparse_prompt_embeddings=sparse_prompt_embeddings,
        dense_prompt_embeddings=dense_prompt_embeddings,
        repeat_image=repeat_image,
        high_res_features=high_res_features,
    )

    # Select the correct mask or masks for output
    if multimask_output:
        masks = masks[:, 1:, :, :]
        iou_pred = iou_pred[:, 1:]
    elif self.dynamic_multimask_via_stability and not self.training:
        masks, iou_pred = self._dynamic_multimask_via_stability(masks, iou_pred)
    else:
        masks = masks[:, 0:1, :, :]
        iou_pred = iou_pred[:, 0:1]

    if multimask_output and self.use_multimask_token_for_obj_ptr:
        sam_tokens_out = mask_tokens_out[:, 1:]  # [b, 3, c] shape
    else:
        # Take the mask output token. Here we *always* use the token for single mask output.
        # At test time, even if we track after 1-click (and using multimask_output=True),
        # we still take the single mask token here. The rationale is that we always track
        # after multiple clicks during training, so the past tokens seen during training
        # are always the single mask token (and we'll let it be the object-memory token).
        sam_tokens_out = mask_tokens_out[:, 0:1]  # [b, 1, c] shape

    # Prepare output
    return masks, iou_pred, sam_tokens_out, object_score_logits

def predict_masks(
    self,
    image_embeddings: torch.Tensor,
    image_pe: torch.Tensor,
    sparse_prompt_embeddings: torch.Tensor,
    dense_prompt_embeddings: torch.Tensor,
    repeat_image: bool,
    high_res_features: Optional[List[torch.Tensor]] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """Predicts instance segmentation masks from image and prompt embeddings using a transformer."""
    # Concatenate output tokens
    s = 0
    if self.pred_obj_scores:
        output_tokens = torch.cat(
            [
                self.obj_score_token.weight,
                self.iou_token.weight,
                self.mask_tokens.weight,
            ],
            dim=0,
        )
        s = 1
    else:
        output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
    output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)
    tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)

    # Expand per-image data in batch direction to be per-mask
    if repeat_image:
        src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
    else:
        assert image_embeddings.shape[0] == tokens.shape[0]
        src = image_embeddings
    src = src + dense_prompt_embeddings
    assert image_pe.size(0) == 1, "image_pe should have size 1 in batch dim (from `get_dense_pe()`)"
    pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
    b, c, h, w = src.shape

    # Run the transformer
    hs, src = self.transformer(src, pos_src, tokens)
    iou_token_out = hs[:, s, :]
    mask_tokens_out = hs[:, s + 1 : (s + 1 + self.num_mask_tokens), :]

    # Upscale mask embeddings and predict masks using the mask tokens
    src = src.transpose(1, 2).view(b, c, h, w)
    if not self.use_high_res_features:
        upscaled_embedding = self.output_upscaling(src)
    else:
        dc1, ln1, act1, dc2, act2 = self.output_upscaling
        feat_s0, feat_s1 = high_res_features
        upscaled_embedding = act1(ln1(dc1(src) + feat_s1))
        upscaled_embedding = act2(dc2(upscaled_embedding) + feat_s0)

    hyper_in_list: List[torch.Tensor] = [
        self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]) for i in range(self.num_mask_tokens)
    ]
    hyper_in = torch.stack(hyper_in_list, dim=1)
    b, c, h, w = upscaled_embedding.shape
    masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)

    # Generate mask quality predictions
    iou_pred = self.iou_prediction_head(iou_token_out)
    if self.pred_obj_scores:
        assert s == 1
        object_score_logits = self.pred_obj_score_head(hs[:, 0, :])
    else:
        # Obj scores logits - default to 10.0, i.e. assuming the object is present, sigmoid(10)=1
        object_score_logits = 10.0 * iou_pred.new_ones(iou_pred.shape[0], 1)

    return masks, iou_pred, mask_tokens_out, object_score_logits