added sam_hq

app/sam_hq

@@ -1 +0,0 @@
-Subproject commit 759948f24f5524e5946bc274b4086ff1cc13b676

app/sam_hq/__init__.py

@@ -0,0 +1,2 @@
+from .build_sam import build_sam
+from .predictor import SamPredictor

app/sam_hq/build_sam.py

@@ -0,0 +1,117 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from functools import partial
+
+from .image_encoder import ImageEncoderViT
+from .mask_decoder_hq import MaskDecoderHQ
+from .prompt_encoder import PromptEncoder
+from .transformer import TwoWayTransformer
+from .sam import Sam
+
+
+def build_sam_vit_h(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=1280,
+        encoder_depth=32,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[7, 15, 23, 31],
+        checkpoint=checkpoint,
+    )
+
+
+build_sam = build_sam_vit_h
+
+
+def build_sam_vit_l(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=1024,
+        encoder_depth=24,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[5, 11, 17, 23],
+        checkpoint=checkpoint,
+    )
+
+
+def build_sam_vit_b(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=768,
+        encoder_depth=12,
+        encoder_num_heads=12,
+        encoder_global_attn_indexes=[2, 5, 8, 11],
+        checkpoint=checkpoint,
+    )
+
+
+sam_model_registry = {
+    "default": build_sam_vit_h,
+    "vit_h": build_sam_vit_h,
+    "vit_l": build_sam_vit_l,
+    "vit_b": build_sam_vit_b,
+}
+
+
+def _build_sam(
+    encoder_embed_dim,
+    encoder_depth,
+    encoder_num_heads,
+    encoder_global_attn_indexes,
+    checkpoint=None,
+):
+    prompt_embed_dim = 256
+    image_size = 1024
+    vit_patch_size = 16
+    image_embedding_size = image_size // vit_patch_size
+    sam = Sam(
+        image_encoder=ImageEncoderViT(
+            depth=encoder_depth,
+            embed_dim=encoder_embed_dim,
+            img_size=image_size,
+            mlp_ratio=4,
+            norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
+            num_heads=encoder_num_heads,
+            patch_size=vit_patch_size,
+            qkv_bias=True,
+            use_rel_pos=True,
+            global_attn_indexes=encoder_global_attn_indexes,
+            window_size=14,
+            out_chans=prompt_embed_dim,
+        ),
+        prompt_encoder=PromptEncoder(
+            embed_dim=prompt_embed_dim,
+            image_embedding_size=(image_embedding_size, image_embedding_size),
+            input_image_size=(image_size, image_size),
+            mask_in_chans=16,
+        ),
+        mask_decoder=MaskDecoderHQ(
+            num_multimask_outputs=3,
+            transformer=TwoWayTransformer(
+                depth=2,
+                embedding_dim=prompt_embed_dim,
+                mlp_dim=2048,
+                num_heads=8,
+            ),
+            transformer_dim=prompt_embed_dim,
+            iou_head_depth=3,
+            iou_head_hidden_dim=256,
+            vit_dim=encoder_embed_dim,
+        ),
+        pixel_mean=[123.675, 116.28, 103.53],
+        pixel_std=[58.395, 57.12, 57.375],
+    )
+    # sam.eval()
+    if checkpoint is not None:
+        with open(checkpoint, "rb") as f:
+            state_dict = torch.load(f, map_location=torch.device("cpu"))
+        info = sam.load_state_dict(state_dict, strict=False)
+        print(info)
+    for n, p in sam.named_parameters():
+        if 'hf_token' not in n and 'hf_mlp' not in n and 'compress_vit_feat' not in n and 'embedding_encoder' not in n and 'embedding_maskfeature' not in n:
+            p.requires_grad = False
+
+    return sam

app/sam_hq/common.py

@@ -0,0 +1,43 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+
+from typing import Type
+
+
+class MLPBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        mlp_dim: int,
+        act: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        super().__init__()
+        self.lin1 = nn.Linear(embedding_dim, mlp_dim)
+        self.lin2 = nn.Linear(mlp_dim, embedding_dim)
+        self.act = act()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.lin2(self.act(self.lin1(x)))
+
+
+# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
+# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x

app/sam_hq/image_encoder.py

@@ -0,0 +1,398 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from typing import Optional, Tuple, Type
+
+from .common import LayerNorm2d, MLPBlock
+
+
+# This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa
+class ImageEncoderViT(nn.Module):
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        out_chans: int = 256,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        global_attn_indexes: Tuple[int, ...] = (),
+    ) -> None:
+        """
+        Args:
+            img_size (int): Input image size.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_attn_indexes (list): Indexes for blocks using global attention.
+        """
+        super().__init__()
+        self.img_size = img_size
+
+        self.patch_embed = PatchEmbed(
+            kernel_size=(patch_size, patch_size),
+            stride=(patch_size, patch_size),
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+        )
+
+        self.pos_embed: Optional[nn.Parameter] = None
+        if use_abs_pos:
+            # Initialize absolute positional embedding with pretrain image size.
+            self.pos_embed = nn.Parameter(
+                torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim)
+            )
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            block = Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=use_rel_pos,
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=window_size if i not in global_attn_indexes else 0,
+                input_size=(img_size // patch_size, img_size // patch_size),
+            )
+            self.blocks.append(block)
+
+        self.neck = nn.Sequential(
+            nn.Conv2d(
+                embed_dim,
+                out_chans,
+                kernel_size=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+            nn.Conv2d(
+                out_chans,
+                out_chans,
+                kernel_size=3,
+                padding=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patch_embed(x)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+
+        interm_embeddings=[]
+        for blk in self.blocks:
+            x = blk(x)
+            if blk.window_size == 0:
+                interm_embeddings.append(x)
+
+        x = self.neck(x.permute(0, 3, 1, 2))
+
+        return x, interm_embeddings
+
+
+class Block(nn.Module):
+    """Transformer blocks with support of window attention and residual propagation blocks"""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then
+                use global attention.
+            input_size (tuple(int, int) or None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size, window_size),
+        )
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
+
+        self.window_size = window_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x
+        x = self.norm1(x)
+        # Window partition
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        x = self.attn(x)
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+
+        return x
+
+
+class Attention(nn.Module):
+    """Multi-head Attention block with relative position embeddings."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool):  If True, add a learnable bias to query, key, value.
+            rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            input_size (tuple(int, int) or None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        self.use_rel_pos = use_rel_pos
+        if self.use_rel_pos:
+            assert (
+                input_size is not None
+            ), "Input size must be provided if using relative positional encoding."
+            # initialize relative positional embeddings
+            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
+            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, H, W, _ = x.shape
+        # qkv with shape (3, B, nHead, H * W, C)
+        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        # q, k, v with shape (B * nHead, H * W, C)
+        q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
+
+        attn = (q * self.scale) @ k.transpose(-2, -1)
+
+        if self.use_rel_pos:
+            attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
+
+        attn = attn.softmax(dim=-1)
+        x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
+        x = self.proj(x)
+
+        return x
+
+
+def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(
+    windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
+) -> torch.Tensor:
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        windows (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
+    """
+    Get relative positional embeddings according to the relative positions of
+        query and key sizes.
+    Args:
+        q_size (int): size of query q.
+        k_size (int): size of key k.
+        rel_pos (Tensor): relative position embeddings (L, C).
+
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel pos.
+        rel_pos_resized = F.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+        )
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def add_decomposed_rel_pos(
+    attn: torch.Tensor,
+    q: torch.Tensor,
+    rel_pos_h: torch.Tensor,
+    rel_pos_w: torch.Tensor,
+    q_size: Tuple[int, int],
+    k_size: Tuple[int, int],
+) -> torch.Tensor:
+    """
+    Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
+    https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py   # noqa B950
+    Args:
+        attn (Tensor): attention map.
+        q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
+        rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
+        rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
+        q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
+        k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
+
+    Returns:
+        attn (Tensor): attention map with added relative positional embeddings.
+    """
+    q_h, q_w = q_size
+    k_h, k_w = k_size
+    Rh = get_rel_pos(q_h, k_h, rel_pos_h)
+    Rw = get_rel_pos(q_w, k_w, rel_pos_w)
+
+    B, _, dim = q.shape
+    r_q = q.reshape(B, q_h, q_w, dim)
+    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
+    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
+
+    attn = (
+        attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
+    ).view(B, q_h * q_w, k_h * k_w)
+
+    return attn
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        padding: Tuple[int, int] = (0, 0),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ) -> None:
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x

app/sam_hq/mask_decoder_hq.py

@@ -0,0 +1,264 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# Modified by HQ-SAM team
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from typing import List, Tuple, Type
+from .common import LayerNorm2d
+
+
+class MaskDecoderHQ(nn.Module):
+    def __init__(
+        self,
+        *,
+        transformer_dim: int,
+        transformer: nn.Module,
+        num_multimask_outputs: int = 3,
+        activation: Type[nn.Module] = nn.GELU,
+        iou_head_depth: int = 3,
+        iou_head_hidden_dim: int = 256,
+        vit_dim: int = 1024,
+    ) -> None:
+        """
+        Predicts masks given an image and prompt embeddings, using a
+        transformer architecture.
+
+        Arguments:
+          transformer_dim (int): the channel dimension of the transformer
+          transformer (nn.Module): the transformer used to predict masks
+          num_multimask_outputs (int): the number of masks to predict
+            when disambiguating masks
+          activation (nn.Module): the type of activation to use when
+            upscaling masks
+          iou_head_depth (int): the depth of the MLP used to predict
+            mask quality
+          iou_head_hidden_dim (int): the hidden dimension of the MLP
+            used to predict mask quality
+        """
+        super().__init__()
+        self.transformer_dim = transformer_dim
+        self.transformer = transformer
+
+        self.num_multimask_outputs = num_multimask_outputs
+
+        self.iou_token = nn.Embedding(1, transformer_dim)
+        self.num_mask_tokens = num_multimask_outputs + 1
+        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
+
+        self.output_upscaling = nn.Sequential(
+            nn.ConvTranspose2d(
+                transformer_dim, transformer_dim // 4, kernel_size=2, stride=2
+            ),
+            LayerNorm2d(transformer_dim // 4),
+            activation(),
+            nn.ConvTranspose2d(
+                transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2
+            ),
+            activation(),
+        )
+        self.output_hypernetworks_mlps = nn.ModuleList(
+            [
+                MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
+                for i in range(self.num_mask_tokens)
+            ]
+        )
+
+        self.iou_prediction_head = MLP(
+            transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth
+        )
+
+        # HQ-SAM parameters
+        self.hf_token = nn.Embedding(1, transformer_dim)  # HQ-Ouptput-Token
+        self.hf_mlp = MLP(
+            transformer_dim, transformer_dim, transformer_dim // 8, 3
+        )  # corresponding new MLP layer for HQ-Ouptput-Token
+        self.num_mask_tokens = self.num_mask_tokens + 1
+
+        # three conv fusion layers for obtaining HQ-Feature
+        self.compress_vit_feat = nn.Sequential(
+            nn.ConvTranspose2d(vit_dim, transformer_dim, kernel_size=2, stride=2),
+            LayerNorm2d(transformer_dim),
+            nn.GELU(),
+            nn.ConvTranspose2d(
+                transformer_dim, transformer_dim // 8, kernel_size=2, stride=2
+            ),
+        )
+
+        self.embedding_encoder = nn.Sequential(
+            nn.ConvTranspose2d(
+                transformer_dim, transformer_dim // 4, kernel_size=2, stride=2
+            ),
+            LayerNorm2d(transformer_dim // 4),
+            nn.GELU(),
+            nn.ConvTranspose2d(
+                transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2
+            ),
+        )
+        self.embedding_maskfeature = nn.Sequential(
+            nn.Conv2d(transformer_dim // 8, transformer_dim // 4, 3, 1, 1),
+            LayerNorm2d(transformer_dim // 4),
+            nn.GELU(),
+            nn.Conv2d(transformer_dim // 4, transformer_dim // 8, 3, 1, 1),
+        )
+
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        multimask_output: bool,
+        hq_token_only: bool,
+        interm_embeddings: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predict masks given image and prompt embeddings.
+
+        Arguments:
+          image_embeddings (torch.Tensor): the embeddings from the ViT image encoder
+          image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
+          sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
+          dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
+          multimask_output (bool): Whether to return multiple masks or a single
+            mask.
+
+        Returns:
+          torch.Tensor: batched predicted masks
+          torch.Tensor: batched predictions of mask quality
+        """
+        vit_features = interm_embeddings[0].permute(
+            0, 3, 1, 2
+        )  # early-layer ViT feature, after 1st global attention block in ViT
+        hq_features = self.embedding_encoder(image_embeddings) + self.compress_vit_feat(
+            vit_features
+        )
+
+        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,
+            hq_features=hq_features,
+        )
+
+        # Select the correct mask or masks for output
+        if multimask_output:
+            # mask with highest score
+            mask_slice = slice(1, self.num_mask_tokens - 1)
+            iou_pred = iou_pred[:, mask_slice]
+            iou_pred, max_iou_idx = torch.max(iou_pred, dim=1)
+            iou_pred = iou_pred.unsqueeze(1)
+            masks_multi = masks[:, mask_slice, :, :]
+            masks_sam = masks_multi[
+                torch.arange(masks_multi.size(0)), max_iou_idx
+            ].unsqueeze(1)
+        else:
+            # singale mask output, default
+            mask_slice = slice(0, 1)
+            iou_pred = iou_pred[:, mask_slice]
+            masks_sam = masks[:, mask_slice]
+
+        masks_hq = masks[:, slice(self.num_mask_tokens - 1, self.num_mask_tokens)]
+        if hq_token_only:
+            masks = masks_hq
+        else:
+            masks = masks_sam + masks_hq
+        # 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,
+        hq_features: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Predicts masks. See 'forward' for more details."""
+        # Concatenate output tokens
+        output_tokens = torch.cat(
+            [self.iou_token.weight, self.mask_tokens.weight, self.hf_token.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
+        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_sam = self.output_upscaling(src)
+        upscaled_embedding_hq = self.embedding_maskfeature(
+            upscaled_embedding_sam
+        ) + hq_features.repeat(b, 1, 1, 1)
+
+        hyper_in_list: List[torch.Tensor] = []
+        for i in range(self.num_mask_tokens):
+            if i < self.num_mask_tokens - 1:
+                hyper_in_list.append(
+                    self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :])
+                )
+            else:
+                hyper_in_list.append(self.hf_mlp(mask_tokens_out[:, i, :]))
+
+        hyper_in = torch.stack(hyper_in_list, dim=1)
+        b, c, h, w = upscaled_embedding_sam.shape
+
+        masks_sam = (
+            hyper_in[:, : self.num_mask_tokens - 1]
+            @ upscaled_embedding_sam.view(b, c, h * w)
+        ).view(b, -1, h, w)
+        masks_sam_hq = (
+            hyper_in[:, self.num_mask_tokens - 1 :]
+            @ upscaled_embedding_hq.view(b, c, h * w)
+        ).view(b, -1, h, w)
+        masks = torch.cat([masks_sam, masks_sam_hq], dim=1)
+        # Generate mask quality predictions
+        iou_pred = self.iou_prediction_head(iou_token_out)
+
+        return masks, iou_pred
+
+
+# Lightly adapted from
+# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
+class MLP(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        sigmoid_output: bool = False,
+    ) -> None:
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(
+            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
+        )
+        self.sigmoid_output = sigmoid_output
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        if self.sigmoid_output:
+            x = F.sigmoid(x)
+        return x

app/sam_hq/predictor.py

@@ -0,0 +1,291 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+
+
+from typing import Optional, Tuple
+from .sam import Sam
+from .transforms import ResizeLongestSide
+
+
+class SamPredictor:
+    def __init__(
+        self,
+        sam_model: Sam,
+    ) -> None:
+        """
+        Uses SAM to calculate the image embedding for an image, and then
+        allow repeated, efficient mask prediction given prompts.
+
+        Arguments:
+          sam_model (Sam): The model to use for mask prediction.
+        """
+        super().__init__()
+        self.model = sam_model
+        self.transform = ResizeLongestSide(sam_model.image_encoder.img_size)
+        self.reset_image()
+
+    def set_image(
+        self,
+        image: np.ndarray,
+        image_format: str = "RGB",
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method.
+
+        Arguments:
+          image (np.ndarray): The image for calculating masks. Expects an
+            image in HWC uint8 format, with pixel values in [0, 255].
+          image_format (str): The color format of the image, in ['RGB', 'BGR'].
+        """
+        assert image_format in [
+            "RGB",
+            "BGR",
+        ], f"image_format must be in ['RGB', 'BGR'], is {image_format}."
+        # import pdb;pdb.set_trace()
+        if image_format != self.model.image_format:
+            image = image[..., ::-1]
+
+        # Transform the image to the form expected by the model
+        # import pdb;pdb.set_trace()
+        input_image = self.transform.apply_image(image)
+        input_image_torch = torch.as_tensor(input_image, device=self.device)
+        input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[
+            None, :, :, :
+        ]
+
+        self.set_torch_image(input_image_torch, image.shape[:2])
+
+    @torch.no_grad()
+    def set_torch_image(
+        self,
+        transformed_image: torch.Tensor,
+        original_image_size: Tuple[int, ...],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method. Expects the input
+        image to be already transformed to the format expected by the model.
+
+        Arguments:
+          transformed_image (torch.Tensor): The input image, with shape
+            1x3xHxW, which has been transformed with ResizeLongestSide.
+          original_image_size (tuple(int, int)): The size of the image
+            before transformation, in (H, W) format.
+        """
+        assert (
+            len(transformed_image.shape) == 4
+            and transformed_image.shape[1] == 3
+            and max(*transformed_image.shape[2:]) == self.model.image_encoder.img_size
+        ), f"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}."
+        self.reset_image()
+
+        self.original_size = original_image_size
+        self.input_size = tuple(transformed_image.shape[-2:])
+        input_image = self.model.preprocess(transformed_image)
+        self.features, self.interm_features = self.model.image_encoder(input_image)
+        self.is_image_set = True
+
+    def predict(
+        self,
+        point_coords: Optional[np.ndarray] = None,
+        point_labels: Optional[np.ndarray] = None,
+        box: Optional[np.ndarray] = None,
+        mask_input: Optional[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        hq_token_only: bool = False,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+
+        Arguments:
+          point_coords (np.ndarray or None): A Nx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (np.ndarray or None): A length N array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A length 4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form 1xHxW, where
+            for SAM, H=W=256.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (np.ndarray): The output masks in CxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (np.ndarray): An array of length C containing the model's
+            predictions for the quality of each mask.
+          (np.ndarray): An array of shape CxHxW, where C is the number
+            of masks and H=W=256. These low resolution logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self.is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        # Transform input prompts
+        coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None
+        if point_coords is not None:
+            assert (
+                point_labels is not None
+            ), "point_labels must be supplied if point_coords is supplied."
+            point_coords = self.transform.apply_coords(point_coords, self.original_size)
+            coords_torch = torch.as_tensor(
+                point_coords, dtype=torch.float, device=self.device
+            )
+            labels_torch = torch.as_tensor(
+                point_labels, dtype=torch.int, device=self.device
+            )
+            coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
+        if box is not None:
+            box = self.transform.apply_boxes(box, self.original_size)
+            box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)
+            box_torch = box_torch[None, :]
+        if mask_input is not None:
+            mask_input_torch = torch.as_tensor(
+                mask_input, dtype=torch.float, device=self.device
+            )
+            mask_input_torch = mask_input_torch[None, :, :, :]
+
+        masks, iou_predictions, low_res_masks = self.predict_torch(
+            coords_torch,
+            labels_torch,
+            box_torch,
+            mask_input_torch,
+            multimask_output,
+            return_logits=return_logits,
+            hq_token_only=hq_token_only,
+        )
+
+        masks_np = masks[0].detach().cpu().numpy()
+        iou_predictions_np = iou_predictions[0].detach().cpu().numpy()
+        low_res_masks_np = low_res_masks[0].detach().cpu().numpy()
+        return masks_np, iou_predictions_np, low_res_masks_np
+
+    @torch.no_grad()
+    def predict_torch(
+        self,
+        point_coords: Optional[torch.Tensor],
+        point_labels: Optional[torch.Tensor],
+        boxes: Optional[torch.Tensor] = None,
+        mask_input: Optional[torch.Tensor] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        hq_token_only: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+        Input prompts are batched torch tensors and are expected to already be
+        transformed to the input frame using ResizeLongestSide.
+
+        Arguments:
+          point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (torch.Tensor or None): A BxN array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          boxes (np.ndarray or None): A Bx4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form Bx1xHxW, where
+            for SAM, H=W=256. Masks returned by a previous iteration of the
+            predict method do not need further transformation.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (torch.Tensor): The output masks in BxCxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (torch.Tensor): An array of shape BxC containing the model's
+            predictions for the quality of each mask.
+          (torch.Tensor): An array of shape BxCxHxW, where C is the number
+            of masks and H=W=256. These low res logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self.is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        if point_coords is not None:
+            points = (point_coords, point_labels)
+        else:
+            points = None
+
+        # Embed prompts
+        sparse_embeddings, dense_embeddings = self.model.prompt_encoder(
+            points=points,
+            boxes=boxes,
+            masks=mask_input,
+        )
+
+        # Predict masks
+        low_res_masks, iou_predictions = self.model.mask_decoder(
+            image_embeddings=self.features,
+            image_pe=self.model.prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            hq_token_only=hq_token_only,
+            interm_embeddings=self.interm_features,
+        )
+
+        # Upscale the masks to the original image resolution
+        masks = self.model.postprocess_masks(
+            low_res_masks, self.input_size, self.original_size
+        )
+
+        if not return_logits:
+            masks = masks > self.model.mask_threshold
+
+        return masks, iou_predictions, low_res_masks
+
+    def get_image_embedding(self) -> torch.Tensor:
+        """
+        Returns the image embeddings for the currently set image, with
+        shape 1xCxHxW, where C is the embedding dimension and (H,W) are
+        the embedding spatial dimension of SAM (typically C=256, H=W=64).
+        """
+        if not self.is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) to generate an embedding."
+            )
+        assert (
+            self.features is not None
+        ), "Features must exist if an image has been set."
+        return self.features
+
+    @property
+    def device(self) -> torch.device:
+        return self.model.device
+
+    def reset_image(self) -> None:
+        """Resets the currently set image."""
+        self.is_image_set = False
+        self.features = None
+        self.orig_h = None
+        self.orig_w = None
+        self.input_h = None
+        self.input_w = None

app/sam_hq/prompt_encoder.py

@@ -0,0 +1,213 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+from torch import nn
+
+from typing import Any, Optional, Tuple, Type
+from .common import LayerNorm2d
+
+
+class PromptEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        image_embedding_size: Tuple[int, int],
+        input_image_size: Tuple[int, int],
+        mask_in_chans: int,
+        activation: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        """
+        Encodes prompts for input to SAM's mask decoder.
+
+        Arguments:
+          embed_dim (int): The prompts' embedding dimension
+          image_embedding_size (tuple(int, int)): The spatial size of the
+            image embedding, as (H, W).
+          input_image_size (int): The padded size of the image as input
+            to the image encoder, as (H, W).
+          mask_in_chans (int): The number of hidden channels used for
+            encoding input masks.
+          activation (nn.Module): The activation to use when encoding
+            input masks.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.input_image_size = input_image_size
+        self.image_embedding_size = image_embedding_size
+        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
+
+        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
+        point_embeddings = [nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)]
+        self.point_embeddings = nn.ModuleList(point_embeddings)
+        self.not_a_point_embed = nn.Embedding(1, embed_dim)
+
+        self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])
+        self.mask_downscaling = nn.Sequential(
+            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans // 4),
+            activation(),
+            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans),
+            activation(),
+            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
+        )
+        self.no_mask_embed = nn.Embedding(1, embed_dim)
+
+    def get_dense_pe(self) -> torch.Tensor:
+        """
+        Returns the positional encoding used to encode point prompts,
+        applied to a dense set of points the shape of the image encoding.
+
+        Returns:
+          torch.Tensor: Positional encoding with shape
+            1x(embed_dim)x(embedding_h)x(embedding_w)
+        """
+        return self.pe_layer(self.image_embedding_size).unsqueeze(0)
+
+    def _embed_points(
+        self,
+        points: torch.Tensor,
+        labels: torch.Tensor,
+        pad: bool,
+    ) -> torch.Tensor:
+        """Embeds point prompts."""
+        points = points + 0.5  # Shift to center of pixel
+        if pad:
+            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
+            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
+            points = torch.cat([points, padding_point], dim=1)
+            labels = torch.cat([labels, padding_label], dim=1)
+        point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)
+        point_embedding[labels == -1] = 0.0
+        point_embedding[labels == -1] += self.not_a_point_embed.weight
+        point_embedding[labels == 0] += self.point_embeddings[0].weight
+        point_embedding[labels == 1] += self.point_embeddings[1].weight
+        return point_embedding
+
+    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
+        """Embeds box prompts."""
+        boxes = boxes + 0.5  # Shift to center of pixel
+        coords = boxes.reshape(-1, 2, 2)
+        corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)
+        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
+        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
+        return corner_embedding
+
+    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
+        """Embeds mask inputs."""
+        mask_embedding = self.mask_downscaling(masks)
+        return mask_embedding
+
+    def _get_batch_size(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> int:
+        """
+        Gets the batch size of the output given the batch size of the input prompts.
+        """
+        if points is not None:
+            return points[0].shape[0]
+        elif boxes is not None:
+            return boxes.shape[0]
+        elif masks is not None:
+            return masks.shape[0]
+        else:
+            return 1
+
+    def _get_device(self) -> torch.device:
+        return self.point_embeddings[0].weight.device
+
+    def forward(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Embeds different types of prompts, returning both sparse and dense
+        embeddings.
+
+        Arguments:
+          points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
+            and labels to embed.
+          boxes (torch.Tensor or none): boxes to embed
+          masks (torch.Tensor or none): masks to embed
+
+        Returns:
+          torch.Tensor: sparse embeddings for the points and boxes, with shape
+            BxNx(embed_dim), where N is determined by the number of input points
+            and boxes.
+          torch.Tensor: dense embeddings for the masks, in the shape
+            Bx(embed_dim)x(embed_H)x(embed_W)
+        """
+        bs = self._get_batch_size(points, boxes, masks)
+        sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())
+        if points is not None:
+            coords, labels = points
+            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
+            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
+        if boxes is not None:
+            box_embeddings = self._embed_boxes(boxes)
+            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
+
+        if masks is not None:
+            dense_embeddings = self._embed_masks(masks)
+        else:
+            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
+                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
+            )
+
+        return sparse_embeddings, dense_embeddings
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """
+    Positional encoding using random spatial frequencies.
+    """
+
+    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
+        super().__init__()
+        if scale is None or scale <= 0.0:
+            scale = 1.0
+        self.register_buffer(
+            "positional_encoding_gaussian_matrix",
+            scale * torch.randn((2, num_pos_feats)),
+        )
+
+    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
+        """Positionally encode points that are normalized to [0,1]."""
+        # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
+        coords = 2 * coords - 1
+        coords = coords @ self.positional_encoding_gaussian_matrix
+        coords = 2 * np.pi * coords
+        # outputs d_1 x ... x d_n x C shape
+        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
+
+    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
+        """Generate positional encoding for a grid of the specified size."""
+        h, w = size
+        device: Any = self.positional_encoding_gaussian_matrix.device
+        grid = torch.ones((h, w), device=device, dtype=torch.float32)
+        y_embed = grid.cumsum(dim=0) - 0.5
+        x_embed = grid.cumsum(dim=1) - 0.5
+        y_embed = y_embed / h
+        x_embed = x_embed / w
+
+        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
+        return pe.permute(2, 0, 1)  # C x H x W
+
+    def forward_with_coords(
+        self, coords_input: torch.Tensor, image_size: Tuple[int, int]
+    ) -> torch.Tensor:
+        """Positionally encode points that are not normalized to [0,1]."""
+        coords = coords_input.clone()
+        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
+        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
+        return self._pe_encoding(coords.to(torch.float))  # B x N x C

app/sam_hq/sam.py

@@ -0,0 +1,177 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from typing import Any, Dict, List, Tuple
+
+from .image_encoder import ImageEncoderViT
+from .mask_decoder_hq import MaskDecoderHQ
+from .prompt_encoder import PromptEncoder
+
+
+class Sam(nn.Module):
+    mask_threshold: float = 0.0
+    image_format: str = "RGB"
+
+    def __init__(
+        self,
+        image_encoder: ImageEncoderViT,
+        prompt_encoder: PromptEncoder,
+        mask_decoder: MaskDecoderHQ,
+        pixel_mean: List[float] = [123.675, 116.28, 103.53],
+        pixel_std: List[float] = [58.395, 57.12, 57.375],
+    ) -> None:
+        """
+        SAM predicts object masks from an image and input prompts.
+
+        Arguments:
+          image_encoder (ImageEncoderViT): The backbone used to encode the
+            image into image embeddings that allow for efficient mask prediction.
+          prompt_encoder (PromptEncoder): Encodes various types of input prompts.
+          mask_decoder (MaskDecoder): Predicts masks from the image embeddings
+            and encoded prompts.
+          pixel_mean (list(float)): Mean values for normalizing pixels in the input image.
+          pixel_std (list(float)): Std values for normalizing pixels in the input image.
+        """
+        super().__init__()
+        self.image_encoder = image_encoder
+        self.prompt_encoder = prompt_encoder
+        self.mask_decoder = mask_decoder
+        self.register_buffer("pixel_mean", torch.Tensor(pixel_mean).view(-1, 1, 1), False)
+        self.register_buffer("pixel_std", torch.Tensor(pixel_std).view(-1, 1, 1), False)
+
+    @property
+    def device(self) -> Any:
+        return self.pixel_mean.device
+
+    def forward(
+        self,
+        batched_input: List[Dict[str, Any]],
+        multimask_output: bool,
+        hq_token_only: bool =False,
+    ) -> List[Dict[str, torch.Tensor]]:
+        """
+        Predicts masks end-to-end from provided images and prompts.
+        If prompts are not known in advance, using SamPredictor is
+        recommended over calling the model directly.
+
+        Arguments:
+          batched_input (list(dict)): A list over input images, each a
+            dictionary with the following keys. A prompt key can be
+            excluded if it is not present.
+              'image': The image as a torch tensor in 3xHxW format,
+                already transformed for input to the model.
+              'original_size': (tuple(int, int)) The original size of
+                the image before transformation, as (H, W).
+              'point_coords': (torch.Tensor) Batched point prompts for
+                this image, with shape BxNx2. Already transformed to the
+                input frame of the model.
+              'point_labels': (torch.Tensor) Batched labels for point prompts,
+                with shape BxN.
+              'boxes': (torch.Tensor) Batched box inputs, with shape Bx4.
+                Already transformed to the input frame of the model.
+              'mask_inputs': (torch.Tensor) Batched mask inputs to the model,
+                in the form Bx1xHxW.
+          multimask_output (bool): Whether the model should predict multiple
+            disambiguating masks, or return a single mask.
+
+        Returns:
+          (list(dict)): A list over input images, where each element is
+            as dictionary with the following keys.
+              'masks': (torch.Tensor) Batched binary mask predictions,
+                with shape BxCxHxW, where B is the number of input prompts,
+                C is determined by multimask_output, and (H, W) is the
+                original size of the image.
+              'iou_predictions': (torch.Tensor) The model's predictions
+                of mask quality, in shape BxC.
+              'low_res_logits': (torch.Tensor) Low resolution logits with
+                shape BxCxHxW, where H=W=256. Can be passed as mask input
+                to subsequent iterations of prediction.
+        """
+        input_images = torch.stack([self.preprocess(x["image"]) for x in batched_input], dim=0)
+        image_embeddings, interm_embeddings = self.image_encoder(input_images)
+        interm_embeddings = interm_embeddings[0] # early layer
+
+        outputs = []
+        for image_record, curr_embedding, curr_interm in zip(batched_input, image_embeddings, interm_embeddings):
+            if "point_coords" in image_record:
+                points = (image_record["point_coords"], image_record["point_labels"])
+            else:
+                points = None
+            sparse_embeddings, dense_embeddings = self.prompt_encoder(
+                points=points,
+                boxes=image_record.get("boxes", None),
+                masks=image_record.get("mask_inputs", None),
+            )
+            low_res_masks, iou_predictions = self.mask_decoder(
+                image_embeddings=curr_embedding.unsqueeze(0),
+                image_pe=self.prompt_encoder.get_dense_pe(),
+                sparse_prompt_embeddings=sparse_embeddings,
+                dense_prompt_embeddings=dense_embeddings,
+                multimask_output=multimask_output,
+                hq_token_only=hq_token_only,
+                interm_embeddings=curr_interm.unsqueeze(0).unsqueeze(0),
+            )
+            masks = self.postprocess_masks(
+                low_res_masks,
+                input_size=image_record["image"].shape[-2:],
+                original_size=image_record["original_size"],
+            )
+            masks = masks > self.mask_threshold
+            outputs.append(
+                {
+                    "masks": masks,
+                    "iou_predictions": iou_predictions,
+                    "low_res_logits": low_res_masks,
+                }
+            )
+        return outputs
+
+    def postprocess_masks(
+        self,
+        masks: torch.Tensor,
+        input_size: Tuple[int, ...],
+        original_size: Tuple[int, ...],
+    ) -> torch.Tensor:
+        """
+        Remove padding and upscale masks to the original image size.
+
+        Arguments:
+          masks (torch.Tensor): Batched masks from the mask_decoder,
+            in BxCxHxW format.
+          input_size (tuple(int, int)): The size of the image input to the
+            model, in (H, W) format. Used to remove padding.
+          original_size (tuple(int, int)): The original size of the image
+            before resizing for input to the model, in (H, W) format.
+
+        Returns:
+          (torch.Tensor): Batched masks in BxCxHxW format, where (H, W)
+            is given by original_size.
+        """
+        masks = F.interpolate(
+            masks,
+            (self.image_encoder.img_size, self.image_encoder.img_size),
+            mode="bilinear",
+            align_corners=False,
+        )
+        masks = masks[..., : input_size[0], : input_size[1]]
+        masks = F.interpolate(masks, original_size, mode="bilinear", align_corners=False)
+        return masks
+
+    def preprocess(self, x: torch.Tensor) -> torch.Tensor:
+        """Normalize pixel values and pad to a square input."""
+        # Normalize colors
+        x = (x - self.pixel_mean) / self.pixel_std
+
+        # Pad
+        h, w = x.shape[-2:]
+        padh = self.image_encoder.img_size - h
+        padw = self.image_encoder.img_size - w
+        x = F.pad(x, (0, padw, 0, padh))
+        return x

app/sam_hq/transformer.py

@@ -0,0 +1,239 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import math
+
+from torch import Tensor, nn
+from typing import Tuple, Type
+from .common import MLPBlock
+
+
+class TwoWayTransformer(nn.Module):
+    def __init__(
+        self,
+        depth: int,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+    ) -> None:
+        """
+        A transformer decoder that attends to an input image using
+        queries whose positional embedding is supplied.
+
+        Args:
+          depth (int): number of layers in the transformer
+          embedding_dim (int): the channel dimension for the input embeddings
+          num_heads (int): the number of heads for multihead attention. Must
+            divide embedding_dim
+          mlp_dim (int): the channel dimension internal to the MLP block
+          activation (nn.Module): the activation to use in the MLP block
+        """
+        super().__init__()
+        self.depth = depth
+        self.embedding_dim = embedding_dim
+        self.num_heads = num_heads
+        self.mlp_dim = mlp_dim
+        self.layers = nn.ModuleList()
+
+        for i in range(depth):
+            self.layers.append(
+                TwoWayAttentionBlock(
+                    embedding_dim=embedding_dim,
+                    num_heads=num_heads,
+                    mlp_dim=mlp_dim,
+                    activation=activation,
+                    attention_downsample_rate=attention_downsample_rate,
+                    skip_first_layer_pe=(i == 0),
+                )
+            )
+
+        self.final_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm_final_attn = nn.LayerNorm(embedding_dim)
+
+    def forward(
+        self,
+        image_embedding: Tensor,
+        image_pe: Tensor,
+        point_embedding: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+          image_embedding (torch.Tensor): image to attend to. Should be shape
+            B x embedding_dim x h x w for any h and w.
+          image_pe (torch.Tensor): the positional encoding to add to the image. Must
+            have the same shape as image_embedding.
+          point_embedding (torch.Tensor): the embedding to add to the query points.
+            Must have shape B x N_points x embedding_dim for any N_points.
+
+        Returns:
+          torch.Tensor: the processed point_embedding
+          torch.Tensor: the processed image_embedding
+        """
+        # BxCxHxW -> BxHWxC == B x N_image_tokens x C
+        bs, c, h, w = image_embedding.shape
+        image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
+        image_pe = image_pe.flatten(2).permute(0, 2, 1)
+
+        # Prepare queries
+        queries = point_embedding
+        keys = image_embedding
+
+        # Apply transformer blocks and final layernorm
+        for layer in self.layers:
+            queries, keys = layer(
+                queries=queries,
+                keys=keys,
+                query_pe=point_embedding,
+                key_pe=image_pe,
+            )
+
+        # Apply the final attention layer from the points to the image
+        q = queries + point_embedding
+        k = keys + image_pe
+        attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm_final_attn(queries)
+
+        return queries, keys
+
+
+class TwoWayAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int = 2048,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+        skip_first_layer_pe: bool = False,
+    ) -> None:
+        """
+        A transformer block with four layers: (1) self-attention of sparse
+        inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
+        block on sparse inputs, and (4) cross attention of dense inputs to sparse
+        inputs.
+
+        Arguments:
+          embedding_dim (int): the channel dimension of the embeddings
+          num_heads (int): the number of heads in the attention layers
+          mlp_dim (int): the hidden dimension of the mlp block
+          activation (nn.Module): the activation of the mlp block
+          skip_first_layer_pe (bool): skip the PE on the first layer
+        """
+        super().__init__()
+        self.self_attn = Attention(embedding_dim, num_heads)
+        self.norm1 = nn.LayerNorm(embedding_dim)
+
+        self.cross_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm2 = nn.LayerNorm(embedding_dim)
+
+        self.mlp = MLPBlock(embedding_dim, mlp_dim, activation)
+        self.norm3 = nn.LayerNorm(embedding_dim)
+
+        self.norm4 = nn.LayerNorm(embedding_dim)
+        self.cross_attn_image_to_token = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+
+        self.skip_first_layer_pe = skip_first_layer_pe
+
+    def forward(
+        self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        # Self attention block
+        if self.skip_first_layer_pe:
+            queries = self.self_attn(q=queries, k=queries, v=queries)
+        else:
+            q = queries + query_pe
+            attn_out = self.self_attn(q=q, k=q, v=queries)
+            queries = queries + attn_out
+        queries = self.norm1(queries)
+
+        # Cross attention block, tokens attending to image embedding
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm2(queries)
+
+        # MLP block
+        mlp_out = self.mlp(queries)
+        queries = queries + mlp_out
+        queries = self.norm3(queries)
+
+        # Cross attention block, image embedding attending to tokens
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
+        keys = keys + attn_out
+        keys = self.norm4(keys)
+
+        return queries, keys
+
+
+class Attention(nn.Module):
+    """
+    An attention layer that allows for downscaling the size of the embedding
+    after projection to queries, keys, and values.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        downsample_rate: int = 1,
+    ) -> None:
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.internal_dim = embedding_dim // downsample_rate
+        self.num_heads = num_heads
+        assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."
+
+        self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.k_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.v_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
+
+    def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
+        b, n, c = x.shape
+        x = x.reshape(b, n, num_heads, c // num_heads)
+        return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_head
+
+    def _recombine_heads(self, x: Tensor) -> Tensor:
+        b, n_heads, n_tokens, c_per_head = x.shape
+        x = x.transpose(1, 2)
+        return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x C
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        # Attention
+        _, _, _, c_per_head = q.shape
+        attn = q @ k.permute(0, 1, 3, 2)  # B x N_heads x N_tokens x N_tokens
+        attn = attn / math.sqrt(c_per_head)
+        attn = torch.softmax(attn, dim=-1)
+
+        # Get output
+        out = attn @ v
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out

app/sam_hq/transforms.py

@@ -0,0 +1,102 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+from torch.nn import functional as F
+from torchvision.transforms.functional import resize, to_pil_image  # type: ignore
+
+from copy import deepcopy
+from typing import Tuple
+
+
+class ResizeLongestSide:
+    """
+    Resizes images to the longest side 'target_length', as well as provides
+    methods for resizing coordinates and boxes. Provides methods for
+    transforming both numpy array and batched torch tensors.
+    """
+
+    def __init__(self, target_length: int) -> None:
+        self.target_length = target_length
+
+    def apply_image(self, image: np.ndarray) -> np.ndarray:
+        """
+        Expects a numpy array with shape HxWxC in uint8 format.
+        """
+        target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
+        return np.array(resize(to_pil_image(image), target_size))
+
+    def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
+        """
+        Expects a numpy array of length 2 in the final dimension. Requires the
+        original image size in (H, W) format.
+        """
+        old_h, old_w = original_size
+        new_h, new_w = self.get_preprocess_shape(
+            original_size[0], original_size[1], self.target_length
+        )
+        coords = deepcopy(coords).astype(float)
+        coords[..., 0] = coords[..., 0] * (new_w / old_w)
+        coords[..., 1] = coords[..., 1] * (new_h / old_h)
+        return coords
+
+    def apply_boxes(self, boxes: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
+        """
+        Expects a numpy array shape Bx4. Requires the original image size
+        in (H, W) format.
+        """
+        boxes = self.apply_coords(boxes.reshape(-1, 2, 2), original_size)
+        return boxes.reshape(-1, 4)
+
+    def apply_image_torch(self, image: torch.Tensor) -> torch.Tensor:
+        """
+        Expects batched images with shape BxCxHxW and float format. This
+        transformation may not exactly match apply_image. apply_image is
+        the transformation expected by the model.
+        """
+        # Expects an image in BCHW format. May not exactly match apply_image.
+        target_size = self.get_preprocess_shape(image.shape[2], image.shape[3], self.target_length)
+        return F.interpolate(
+            image, target_size, mode="bilinear", align_corners=False, antialias=True
+        )
+
+    def apply_coords_torch(
+        self, coords: torch.Tensor, original_size: Tuple[int, ...]
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with length 2 in the last dimension. Requires the
+        original image size in (H, W) format.
+        """
+        old_h, old_w = original_size
+        new_h, new_w = self.get_preprocess_shape(
+            original_size[0], original_size[1], self.target_length
+        )
+        coords = deepcopy(coords).to(torch.float)
+        coords[..., 0] = coords[..., 0] * (new_w / old_w)
+        coords[..., 1] = coords[..., 1] * (new_h / old_h)
+        return coords
+
+    def apply_boxes_torch(
+        self, boxes: torch.Tensor, original_size: Tuple[int, ...]
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with shape Bx4. Requires the original image
+        size in (H, W) format.
+        """
+        boxes = self.apply_coords_torch(boxes.reshape(-1, 2, 2), original_size)
+        return boxes.reshape(-1, 4)
+
+    @staticmethod
+    def get_preprocess_shape(oldh: int, oldw: int, long_side_length: int) -> Tuple[int, int]:
+        """
+        Compute the output size given input size and target long side length.
+        """
+        scale = long_side_length * 1.0 / max(oldh, oldw)
+        newh, neww = oldh * scale, oldw * scale
+        neww = int(neww + 0.5)
+        newh = int(newh + 0.5)
+        return (newh, neww)