This commit is contained in:
Qing 2024-08-12 10:10:24 +08:00
parent 9afdbd1c0a
commit 2f833029aa
23 changed files with 3801 additions and 3 deletions

View File

@ -9,6 +9,8 @@ from iopaint.helper import download_model
from iopaint.plugins.base_plugin import BasePlugin from iopaint.plugins.base_plugin import BasePlugin
from iopaint.plugins.segment_anything import SamPredictor, sam_model_registry from iopaint.plugins.segment_anything import SamPredictor, sam_model_registry
from iopaint.plugins.segment_anything.predictor_hq import SamHQPredictor from iopaint.plugins.segment_anything.predictor_hq import SamHQPredictor
from iopaint.plugins.segment_anything2.build_sam import build_sam2
from iopaint.plugins.segment_anything2.sam2_image_predictor import SAM2ImagePredictor
from iopaint.schema import RunPluginRequest from iopaint.schema import RunPluginRequest
# 从小到大 # 从小到大
@ -41,6 +43,22 @@ SEGMENT_ANYTHING_MODELS = {
"url": "https://huggingface.co/lkeab/hq-sam/resolve/main/sam_hq_vit_h.pth", "url": "https://huggingface.co/lkeab/hq-sam/resolve/main/sam_hq_vit_h.pth",
"md5": "3560f6b6a5a6edacd814a1325c39640a", "md5": "3560f6b6a5a6edacd814a1325c39640a",
}, },
"sam2_tiny": {
"url": "https://dl.fbaipublicfiles.com/segment_anything_2/072824/sam2_hiera_tiny.pt",
"md5": "99eacccce4ada0b35153d4fd7af05297",
},
"sam2_small": {
"url": "https://dl.fbaipublicfiles.com/segment_anything_2/072824/sam2_hiera_small.pt",
"md5": "7f320dbeb497330a2472da5a16c7324d",
},
"sam2_base": {
"url": "https://dl.fbaipublicfiles.com/segment_anything_2/072824/sam2_hiera_base_plus.pt",
"md5": "09dc5a3d7719f64aaea1d37341ef26f2",
},
"sam2_large": {
"url": "https://dl.fbaipublicfiles.com/segment_anything_2/072824/sam2_hiera_large.pt",
"md5": "08083462423be3260cd6a5eef94dc01c",
},
} }
@ -64,6 +82,11 @@ class InteractiveSeg(BasePlugin):
self.predictor = SamHQPredictor( self.predictor = SamHQPredictor(
sam_model_registry[model_name](checkpoint=model_path).to(self.device) sam_model_registry[model_name](checkpoint=model_path).to(self.device)
) )
elif model_name.startswith("sam2"):
sam2_model = build_sam2(
model_name, ckpt_path=model_path, device=self.device
)
self.predictor = SAM2ImagePredictor(sam2_model)
else: else:
self.predictor = SamPredictor( self.predictor = SamPredictor(
sam_model_registry[model_name](checkpoint=model_path).to(self.device) sam_model_registry[model_name](checkpoint=model_path).to(self.device)
@ -98,7 +121,7 @@ class InteractiveSeg(BasePlugin):
self.prev_img_md5 = img_md5 self.prev_img_md5 = img_md5
self.predictor.set_image(rgb_np_img) self.predictor.set_image(rgb_np_img)
masks, scores, _ = self.predictor.predict( masks, _, _ = self.predictor.predict(
point_coords=np.array(input_point), point_coords=np.array(input_point),
point_labels=np.array(input_label), point_labels=np.array(input_label),
multimask_output=False, multimask_output=False,

View File

@ -0,0 +1,5 @@
# 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.

View File

@ -0,0 +1,262 @@
# 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 logging
import torch
from pathlib import Path
from .modeling.backbones.hieradet import Hiera
from .modeling.backbones.image_encoder import ImageEncoder, FpnNeck
from .modeling.memory_attention import MemoryAttention, MemoryAttentionLayer
from .modeling.memory_encoder import MemoryEncoder, MaskDownSampler, Fuser, CXBlock
from .modeling.position_encoding import PositionEmbeddingSine
from .modeling.sam.transformer import RoPEAttention
from .modeling.sam2_base import SAM2Base
CURRENT_DIR = Path(__file__).parent
CONFIG_DIR = CURRENT_DIR / "sam2_configs"
common_kwargs = dict(
num_maskmem=7,
image_size=1024,
sigmoid_scale_for_mem_enc=20.0,
sigmoid_bias_for_mem_enc=-10.0,
use_mask_input_as_output_without_sam=True,
directly_add_no_mem_embed=True,
use_high_res_features_in_sam=True,
multimask_output_in_sam=True,
iou_prediction_use_sigmoid=True,
use_obj_ptrs_in_encoder=True,
add_tpos_enc_to_obj_ptrs=False,
only_obj_ptrs_in_the_past_for_eval=True,
pred_obj_scores=True,
pred_obj_scores_mlp=True,
fixed_no_obj_ptr=True,
multimask_output_for_tracking=True,
use_multimask_token_for_obj_ptr=True,
multimask_min_pt_num=0,
multimask_max_pt_num=1,
use_mlp_for_obj_ptr_proj=True,
compile_image_encoder=False,
)
def build_memory_attention():
return MemoryAttention(
d_model=256,
pos_enc_at_input=True,
layer=MemoryAttentionLayer(
activation="relu",
dim_feedforward=2048,
dropout=0.1,
pos_enc_at_attn=False,
self_attention=RoPEAttention(
rope_theta=10000.0,
feat_sizes=[32, 32],
embedding_dim=256,
num_heads=1,
downsample_rate=1,
dropout=0.1,
),
d_model=256,
pos_enc_at_cross_attn_keys=True,
pos_enc_at_cross_attn_queries=False,
cross_attention=RoPEAttention(
rope_theta=10000.0,
feat_sizes=[32, 32],
embedding_dim=256,
num_heads=1,
downsample_rate=1,
dropout=0.1,
kv_in_dim=64,
),
),
num_layers=4,
)
def build_memory_encoder():
return MemoryEncoder(
out_dim=64,
position_encoding=PositionEmbeddingSine(
num_pos_feats=64, normalize=True, scale=None, temperature=10000
),
mask_downsampler=MaskDownSampler(
kernel_size=3,
stride=2,
padding=1,
),
fuser=Fuser(
layer=CXBlock(
dim=256,
kernel_size=7,
padding=3,
layer_scale_init_value=1e-6,
use_dwconv=True,
),
num_layers=2,
),
)
def build_sam2_tiny():
return SAM2Base(
**common_kwargs,
image_encoder=ImageEncoder(
scalp=1,
trunk=Hiera(
embed_dim=96,
num_heads=1,
stages=(1, 2, 7, 2),
global_att_blocks=(5, 7, 9),
window_pos_embed_bkg_spatial_size=(7, 7),
window_spec=(8, 4, 14, 7),
),
neck=FpnNeck(
position_encoding=PositionEmbeddingSine(
num_pos_feats=256,
normalize=True,
scale=None,
temperature=10000,
),
d_model=256,
backbone_channel_list=[768, 384, 192, 96],
fpn_top_down_levels=[2, 3],
fpn_interp_model="nearest",
),
),
memory_attention=build_memory_attention(),
memory_encoder=build_memory_encoder(),
)
def build_sam2_small():
return SAM2Base(
**common_kwargs,
image_encoder=ImageEncoder(
scalp=1,
trunk=Hiera(
embed_dim=96,
num_heads=1,
stages=(1, 2, 11, 2),
global_att_blocks=(7, 10, 13),
window_pos_embed_bkg_spatial_size=(7, 7),
window_spec=(8, 4, 14, 7),
),
neck=FpnNeck(
position_encoding=PositionEmbeddingSine(
num_pos_feats=256,
normalize=True,
scale=None,
temperature=10000,
),
d_model=256,
backbone_channel_list=[768, 384, 192, 96],
fpn_top_down_levels=[2, 3],
fpn_interp_model="nearest",
),
),
memory_attention=build_memory_attention(),
memory_encoder=build_memory_encoder(),
)
def build_sam2_base():
return SAM2Base(
**common_kwargs,
image_encoder=ImageEncoder(
scalp=1,
trunk=Hiera(
embed_dim=112,
num_heads=2,
stages=(2, 3, 16, 3),
global_att_blocks=(12, 16, 20),
window_pos_embed_bkg_spatial_size=(14, 14),
window_spec=(8, 4, 14, 7),
),
neck=FpnNeck(
position_encoding=PositionEmbeddingSine(
num_pos_feats=256,
normalize=True,
scale=None,
temperature=10000,
),
d_model=256,
backbone_channel_list=[896, 448, 224, 112],
fpn_top_down_levels=[2, 3],
fpn_interp_model="nearest",
),
),
memory_attention=build_memory_attention(),
memory_encoder=build_memory_encoder(),
)
def build_sam2_large():
return SAM2Base(
**common_kwargs,
image_encoder=ImageEncoder(
scalp=1,
trunk=Hiera(
embed_dim=144,
num_heads=2,
stages=(2, 6, 36, 4),
global_att_blocks=(23, 33, 43),
window_pos_embed_bkg_spatial_size=(7, 7),
window_spec=(8, 4, 16, 8),
),
neck=FpnNeck(
position_encoding=PositionEmbeddingSine(
num_pos_feats=256,
normalize=True,
scale=None,
temperature=10000,
),
d_model=256,
backbone_channel_list=[1152, 576, 288, 144],
fpn_top_down_levels=[2, 3],
fpn_interp_model="nearest",
),
),
memory_attention=build_memory_attention(),
memory_encoder=build_memory_encoder(),
)
sam2_model_registry = {
"sam2_tiny": build_sam2_tiny,
"sam2_small": build_sam2_small,
"sam2_base": build_sam2_base,
"sam2_large": build_sam2_large,
}
def build_sam2(
name,
ckpt_path=None,
device="cuda",
mode="eval",
):
model = sam2_model_registry[name]()
_load_checkpoint(model, ckpt_path)
model = model.to(device)
if mode == "eval":
model.eval()
return model
def _load_checkpoint(model, ckpt_path):
if ckpt_path is not None:
sd = torch.load(ckpt_path, map_location="cpu")["model"]
missing_keys, unexpected_keys = model.load_state_dict(sd)
if missing_keys:
logging.error(missing_keys)
raise RuntimeError()
if unexpected_keys:
logging.error(unexpected_keys)
raise RuntimeError()
logging.info("Loaded checkpoint sucessfully")

View File

@ -0,0 +1,5 @@
# 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.

View File

@ -0,0 +1,5 @@
# 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.

View File

@ -0,0 +1,295 @@
# 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.
from functools import partial
from typing import List, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..backbones.utils import (
PatchEmbed,
window_partition,
window_unpartition,
)
from ..sam2_utils import DropPath, MLP
def do_pool(x: torch.Tensor, pool: nn.Module, norm: nn.Module = None) -> torch.Tensor:
if pool is None:
return x
# (B, H, W, C) -> (B, C, H, W)
x = x.permute(0, 3, 1, 2)
x = pool(x)
# (B, C, H', W') -> (B, H', W', C)
x = x.permute(0, 2, 3, 1)
if norm:
x = norm(x)
return x
class MultiScaleAttention(nn.Module):
def __init__(
self,
dim: int,
dim_out: int,
num_heads: int,
q_pool: nn.Module = None,
):
super().__init__()
self.dim = dim
self.dim_out = dim_out
self.num_heads = num_heads
head_dim = dim_out // num_heads
self.scale = head_dim**-0.5
self.q_pool = q_pool
self.qkv = nn.Linear(dim, dim_out * 3)
self.proj = nn.Linear(dim_out, dim_out)
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, H, W, _ = x.shape
# qkv with shape (B, H * W, 3, nHead, C)
qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1)
# q, k, v with shape (B, H * W, nheads, C)
q, k, v = torch.unbind(qkv, 2)
# Q pooling (for downsample at stage changes)
if self.q_pool:
q = do_pool(q.reshape(B, H, W, -1), self.q_pool)
H, W = q.shape[1:3] # downsampled shape
q = q.reshape(B, H * W, self.num_heads, -1)
# Torch's SDPA expects [B, nheads, H*W, C] so we transpose
x = F.scaled_dot_product_attention(
q.transpose(1, 2),
k.transpose(1, 2),
v.transpose(1, 2),
)
# Transpose back
x = x.transpose(1, 2)
x = x.reshape(B, H, W, -1)
x = self.proj(x)
return x
class MultiScaleBlock(nn.Module):
def __init__(
self,
dim: int,
dim_out: int,
num_heads: int,
mlp_ratio: float = 4.0,
drop_path: float = 0.0,
norm_layer: Union[nn.Module, str] = "LayerNorm",
q_stride: Tuple[int, int] = None,
act_layer: nn.Module = nn.GELU,
window_size: int = 0,
):
super().__init__()
if isinstance(norm_layer, str):
norm_layer = partial(getattr(nn, norm_layer), eps=1e-6)
self.dim = dim
self.dim_out = dim_out
self.norm1 = norm_layer(dim)
self.window_size = window_size
self.pool, self.q_stride = None, q_stride
if self.q_stride:
self.pool = nn.MaxPool2d(
kernel_size=q_stride, stride=q_stride, ceil_mode=False
)
self.attn = MultiScaleAttention(
dim,
dim_out,
num_heads=num_heads,
q_pool=self.pool,
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.norm2 = norm_layer(dim_out)
self.mlp = MLP(
dim_out,
int(dim_out * mlp_ratio),
dim_out,
num_layers=2,
activation=act_layer,
)
if dim != dim_out:
self.proj = nn.Linear(dim, dim_out)
def forward(self, x: torch.Tensor) -> torch.Tensor:
shortcut = x # B, H, W, C
x = self.norm1(x)
# Skip connection
if self.dim != self.dim_out:
shortcut = do_pool(self.proj(x), self.pool)
# Window partition
window_size = self.window_size
if window_size > 0:
H, W = x.shape[1], x.shape[2]
x, pad_hw = window_partition(x, window_size)
# Window Attention + Q Pooling (if stage change)
x = self.attn(x)
if self.q_stride:
# Shapes have changed due to Q pooling
window_size = self.window_size // self.q_stride[0]
H, W = shortcut.shape[1:3]
pad_h = (window_size - H % window_size) % window_size
pad_w = (window_size - W % window_size) % window_size
pad_hw = (H + pad_h, W + pad_w)
# Reverse window partition
if self.window_size > 0:
x = window_unpartition(x, window_size, pad_hw, (H, W))
x = shortcut + self.drop_path(x)
# MLP
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
class Hiera(nn.Module):
"""
Reference: https://arxiv.org/abs/2306.00989
"""
def __init__(
self,
embed_dim: int = 96, # initial embed dim
num_heads: int = 1, # initial number of heads
drop_path_rate: float = 0.0, # stochastic depth
q_pool: int = 3, # number of q_pool stages
q_stride: Tuple[int, int] = (2, 2), # downsample stride bet. stages
stages: Tuple[int, ...] = (2, 3, 16, 3), # blocks per stage
dim_mul: float = 2.0, # dim_mul factor at stage shift
head_mul: float = 2.0, # head_mul factor at stage shift
window_pos_embed_bkg_spatial_size: Tuple[int, int] = (14, 14),
# window size per stage, when not using global att.
window_spec: Tuple[int, ...] = (
8,
4,
14,
7,
),
# global attn in these blocks
global_att_blocks: Tuple[int, ...] = (
12,
16,
20,
),
return_interm_layers=True, # return feats from every stage
):
super().__init__()
assert len(stages) == len(window_spec)
self.window_spec = window_spec
depth = sum(stages)
self.q_stride = q_stride
self.stage_ends = [sum(stages[:i]) - 1 for i in range(1, len(stages) + 1)]
assert 0 <= q_pool <= len(self.stage_ends[:-1])
self.q_pool_blocks = [x + 1 for x in self.stage_ends[:-1]][:q_pool]
self.return_interm_layers = return_interm_layers
self.patch_embed = PatchEmbed(
embed_dim=embed_dim,
)
# Which blocks have global att?
self.global_att_blocks = global_att_blocks
# Windowed positional embedding (https://arxiv.org/abs/2311.05613)
self.window_pos_embed_bkg_spatial_size = window_pos_embed_bkg_spatial_size
self.pos_embed = nn.Parameter(
torch.zeros(1, embed_dim, *self.window_pos_embed_bkg_spatial_size)
)
self.pos_embed_window = nn.Parameter(
torch.zeros(1, embed_dim, self.window_spec[0], self.window_spec[0])
)
dpr = [
x.item() for x in torch.linspace(0, drop_path_rate, depth)
] # stochastic depth decay rule
cur_stage = 1
self.blocks = nn.ModuleList()
for i in range(depth):
dim_out = embed_dim
# lags by a block, so first block of
# next stage uses an initial window size
# of previous stage and final window size of current stage
window_size = self.window_spec[cur_stage - 1]
if self.global_att_blocks is not None:
window_size = 0 if i in self.global_att_blocks else window_size
if i - 1 in self.stage_ends:
dim_out = int(embed_dim * dim_mul)
num_heads = int(num_heads * head_mul)
cur_stage += 1
block = MultiScaleBlock(
dim=embed_dim,
dim_out=dim_out,
num_heads=num_heads,
drop_path=dpr[i],
q_stride=self.q_stride if i in self.q_pool_blocks else None,
window_size=window_size,
)
embed_dim = dim_out
self.blocks.append(block)
self.channel_list = (
[self.blocks[i].dim_out for i in self.stage_ends[::-1]]
if return_interm_layers
else [self.blocks[-1].dim_out]
)
def _get_pos_embed(self, hw: Tuple[int, int]) -> torch.Tensor:
h, w = hw
window_embed = self.pos_embed_window
pos_embed = F.interpolate(self.pos_embed, size=(h, w), mode="bicubic")
pos_embed = pos_embed + window_embed.tile(
[x // y for x, y in zip(pos_embed.shape, window_embed.shape)]
)
pos_embed = pos_embed.permute(0, 2, 3, 1)
return pos_embed
def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
x = self.patch_embed(x)
# x: (B, H, W, C)
# Add pos embed
x = x + self._get_pos_embed(x.shape[1:3])
outputs = []
for i, blk in enumerate(self.blocks):
x = blk(x)
if (i == self.stage_ends[-1]) or (
i in self.stage_ends and self.return_interm_layers
):
feats = x.permute(0, 3, 1, 2)
outputs.append(feats)
return outputs

View File

@ -0,0 +1,133 @@
# 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.
from typing import List, Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
class ImageEncoder(nn.Module):
def __init__(
self,
trunk: nn.Module,
neck: nn.Module,
scalp: int = 0,
):
super().__init__()
self.trunk = trunk
self.neck = neck
self.scalp = scalp
assert (
self.trunk.channel_list == self.neck.backbone_channel_list
), f"Channel dims of trunk and neck do not match. Trunk: {self.trunk.channel_list}, neck: {self.neck.backbone_channel_list}"
def forward(self, sample: torch.Tensor):
# Forward through backbone
features, pos = self.neck(self.trunk(sample))
if self.scalp > 0:
# Discard the lowest resolution features
features, pos = features[: -self.scalp], pos[: -self.scalp]
src = features[-1]
output = {
"vision_features": src,
"vision_pos_enc": pos,
"backbone_fpn": features,
}
return output
class FpnNeck(nn.Module):
"""
A modified variant of Feature Pyramid Network (FPN) neck
(we remove output conv and also do bicubic interpolation similar to ViT
pos embed interpolation)
"""
def __init__(
self,
position_encoding: nn.Module,
d_model: int,
backbone_channel_list: List[int],
kernel_size: int = 1,
stride: int = 1,
padding: int = 0,
fpn_interp_model: str = "bilinear",
fuse_type: str = "sum",
fpn_top_down_levels: Optional[List[int]] = None,
):
"""Initialize the neck
:param trunk: the backbone
:param position_encoding: the positional encoding to use
:param d_model: the dimension of the model
:param neck_norm: the normalization to use
"""
super().__init__()
self.position_encoding = position_encoding
self.convs = nn.ModuleList()
self.backbone_channel_list = backbone_channel_list
for dim in backbone_channel_list:
current = nn.Sequential()
current.add_module(
"conv",
nn.Conv2d(
in_channels=dim,
out_channels=d_model,
kernel_size=kernel_size,
stride=stride,
padding=padding,
),
)
self.convs.append(current)
self.fpn_interp_model = fpn_interp_model
assert fuse_type in ["sum", "avg"]
self.fuse_type = fuse_type
# levels to have top-down features in its outputs
# e.g. if fpn_top_down_levels is [2, 3], then only outputs of level 2 and 3
# have top-down propagation, while outputs of level 0 and level 1 have only
# lateral features from the same backbone level.
if fpn_top_down_levels is None:
# default is to have top-down features on all levels
fpn_top_down_levels = range(len(self.convs))
self.fpn_top_down_levels = list(fpn_top_down_levels)
def forward(self, xs: List[torch.Tensor]):
out = [None] * len(self.convs)
pos = [None] * len(self.convs)
assert len(xs) == len(self.convs)
# fpn forward pass
# see https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/fpn.py
prev_features = None
# forward in top-down order (from low to high resolution)
n = len(self.convs) - 1
for i in range(n, -1, -1):
x = xs[i]
lateral_features = self.convs[n - i](x)
if i in self.fpn_top_down_levels and prev_features is not None:
top_down_features = F.interpolate(
prev_features.to(dtype=torch.float32),
scale_factor=2.0,
mode=self.fpn_interp_model,
align_corners=(
None if self.fpn_interp_model == "nearest" else False
),
antialias=False,
)
prev_features = lateral_features + top_down_features
if self.fuse_type == "avg":
prev_features /= 2
else:
prev_features = lateral_features
x_out = prev_features
out[i] = x_out
pos[i] = self.position_encoding(x_out).to(x_out.dtype)
return out, pos

View File

@ -0,0 +1,95 @@
# 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.
"""Some utilities for backbones, in particular for windowing"""
from typing import Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
def window_partition(x, window_size):
"""
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, window_size, pad_hw, hw):
"""
Window unpartition into original sequences and removing padding.
Args:
x (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
class PatchEmbed(nn.Module):
"""
Image to Patch Embedding.
"""
def __init__(
self,
kernel_size: Tuple[int, ...] = (7, 7),
stride: Tuple[int, ...] = (4, 4),
padding: Tuple[int, ...] = (3, 3),
in_chans: int = 3,
embed_dim: int = 768,
):
"""
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): 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

View File

@ -0,0 +1,169 @@
# 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.
from typing import Optional
import torch
from torch import nn, Tensor
from .sam.transformer import RoPEAttention
from .sam2_utils import get_activation_fn, get_clones
class MemoryAttentionLayer(nn.Module):
def __init__(
self,
activation: str,
cross_attention: nn.Module,
d_model: int,
dim_feedforward: int,
dropout: float,
pos_enc_at_attn: bool,
pos_enc_at_cross_attn_keys: bool,
pos_enc_at_cross_attn_queries: bool,
self_attention: nn.Module,
):
super().__init__()
self.d_model = d_model
self.dim_feedforward = dim_feedforward
self.dropout_value = dropout
self.self_attn = self_attention
self.cross_attn_image = cross_attention
# Implementation of Feedforward model
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
self.activation_str = activation
self.activation = get_activation_fn(activation)
# Where to add pos enc
self.pos_enc_at_attn = pos_enc_at_attn
self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys
def _forward_sa(self, tgt, query_pos):
# Self-Attention
tgt2 = self.norm1(tgt)
q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
tgt2 = self.self_attn(q, k, v=tgt2)
tgt = tgt + self.dropout1(tgt2)
return tgt
def _forward_ca(self, tgt, memory, query_pos, pos, num_k_exclude_rope=0):
kwds = {}
if num_k_exclude_rope > 0:
assert isinstance(self.cross_attn_image, RoPEAttention)
kwds = {"num_k_exclude_rope": num_k_exclude_rope}
# Cross-Attention
tgt2 = self.norm2(tgt)
tgt2 = self.cross_attn_image(
q=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
k=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
v=memory,
**kwds,
)
tgt = tgt + self.dropout2(tgt2)
return tgt
def forward(
self,
tgt,
memory,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None,
num_k_exclude_rope: int = 0,
) -> torch.Tensor:
# Self-Attn, Cross-Attn
tgt = self._forward_sa(tgt, query_pos)
tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
# MLP
tgt2 = self.norm3(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
tgt = tgt + self.dropout3(tgt2)
return tgt
class MemoryAttention(nn.Module):
def __init__(
self,
d_model: int,
pos_enc_at_input: bool,
layer: nn.Module,
num_layers: int,
batch_first: bool = True, # Do layers expect batch first input?
):
super().__init__()
self.d_model = d_model
self.layers = get_clones(layer, num_layers)
self.num_layers = num_layers
self.norm = nn.LayerNorm(d_model)
self.pos_enc_at_input = pos_enc_at_input
self.batch_first = batch_first
def forward(
self,
curr: torch.Tensor, # self-attention inputs
memory: torch.Tensor, # cross-attention inputs
curr_pos: Optional[Tensor] = None, # pos_enc for self-attention inputs
memory_pos: Optional[Tensor] = None, # pos_enc for cross-attention inputs
num_obj_ptr_tokens: int = 0, # number of object pointer *tokens*
):
if isinstance(curr, list):
assert isinstance(curr_pos, list)
assert len(curr) == len(curr_pos) == 1
curr, curr_pos = (
curr[0],
curr_pos[0],
)
assert (
curr.shape[1] == memory.shape[1]
), "Batch size must be the same for curr and memory"
output = curr
if self.pos_enc_at_input and curr_pos is not None:
output = output + 0.1 * curr_pos
if self.batch_first:
# Convert to batch first
output = output.transpose(0, 1)
curr_pos = curr_pos.transpose(0, 1)
memory = memory.transpose(0, 1)
memory_pos = memory_pos.transpose(0, 1)
for layer in self.layers:
kwds = {}
if isinstance(layer.cross_attn_image, RoPEAttention):
kwds = {"num_k_exclude_rope": num_obj_ptr_tokens}
output = layer(
tgt=output,
memory=memory,
pos=memory_pos,
query_pos=curr_pos,
**kwds,
)
normed_output = self.norm(output)
if self.batch_first:
# Convert back to seq first
normed_output = normed_output.transpose(0, 1)
curr_pos = curr_pos.transpose(0, 1)
return normed_output

View File

@ -0,0 +1,181 @@
# 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 math
from typing import Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from .sam2_utils import DropPath, get_clones, LayerNorm2d
class MaskDownSampler(nn.Module):
"""
Progressively downsample a mask by total_stride, each time by stride.
Note that LayerNorm is applied per *token*, like in ViT.
With each downsample (by a factor stride**2), channel capacity increases by the same factor.
In the end, we linearly project to embed_dim channels.
"""
def __init__(
self,
embed_dim=256,
kernel_size=4,
stride=4,
padding=0,
total_stride=16,
activation=nn.GELU,
):
super().__init__()
num_layers = int(math.log2(total_stride) // math.log2(stride))
assert stride**num_layers == total_stride
self.encoder = nn.Sequential()
mask_in_chans, mask_out_chans = 1, 1
for _ in range(num_layers):
mask_out_chans = mask_in_chans * (stride**2)
self.encoder.append(
nn.Conv2d(
mask_in_chans,
mask_out_chans,
kernel_size=kernel_size,
stride=stride,
padding=padding,
)
)
self.encoder.append(LayerNorm2d(mask_out_chans))
self.encoder.append(activation())
mask_in_chans = mask_out_chans
self.encoder.append(nn.Conv2d(mask_out_chans, embed_dim, kernel_size=1))
def forward(self, x):
return self.encoder(x)
# Lightly adapted from ConvNext (https://github.com/facebookresearch/ConvNeXt)
class CXBlock(nn.Module):
r"""ConvNeXt Block. There are two equivalent implementations:
(1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
(2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
We use (2) as we find it slightly faster in PyTorch
Args:
dim (int): Number of input channels.
drop_path (float): Stochastic depth rate. Default: 0.0
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
"""
def __init__(
self,
dim,
kernel_size=7,
padding=3,
drop_path=0.0,
layer_scale_init_value=1e-6,
use_dwconv=True,
):
super().__init__()
self.dwconv = nn.Conv2d(
dim,
dim,
kernel_size=kernel_size,
padding=padding,
groups=dim if use_dwconv else 1,
) # depthwise conv
self.norm = LayerNorm2d(dim, eps=1e-6)
self.pwconv1 = nn.Linear(
dim, 4 * dim
) # pointwise/1x1 convs, implemented with linear layers
self.act = nn.GELU()
self.pwconv2 = nn.Linear(4 * dim, dim)
self.gamma = (
nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)
if layer_scale_init_value > 0
else None
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
def forward(self, x):
input = x
x = self.dwconv(x)
x = self.norm(x)
x = x.permute(0, 2, 3, 1) # (N, C, H, W) -> (N, H, W, C)
x = self.pwconv1(x)
x = self.act(x)
x = self.pwconv2(x)
if self.gamma is not None:
x = self.gamma * x
x = x.permute(0, 3, 1, 2) # (N, H, W, C) -> (N, C, H, W)
x = input + self.drop_path(x)
return x
class Fuser(nn.Module):
def __init__(self, layer, num_layers, dim=None, input_projection=False):
super().__init__()
self.proj = nn.Identity()
self.layers = get_clones(layer, num_layers)
if input_projection:
assert dim is not None
self.proj = nn.Conv2d(dim, dim, kernel_size=1)
def forward(self, x):
# normally x: (N, C, H, W)
x = self.proj(x)
for layer in self.layers:
x = layer(x)
return x
class MemoryEncoder(nn.Module):
def __init__(
self,
out_dim,
mask_downsampler,
fuser,
position_encoding,
in_dim=256, # in_dim of pix_feats
):
super().__init__()
self.mask_downsampler = mask_downsampler
self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
self.fuser = fuser
self.position_encoding = position_encoding
self.out_proj = nn.Identity()
if out_dim != in_dim:
self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)
def forward(
self,
pix_feat: torch.Tensor,
masks: torch.Tensor,
skip_mask_sigmoid: bool = False,
) -> Tuple[torch.Tensor, torch.Tensor]:
## Process masks
# sigmoid, so that less domain shift from gt masks which are bool
if not skip_mask_sigmoid:
masks = F.sigmoid(masks)
masks = self.mask_downsampler(masks)
## Fuse pix_feats and downsampled masks
# in case the visual features are on CPU, cast them to CUDA
pix_feat = pix_feat.to(masks.device)
x = self.pix_feat_proj(pix_feat)
x = x + masks
x = self.fuser(x)
x = self.out_proj(x)
pos = self.position_encoding(x).to(x.dtype)
return {"vision_features": x, "vision_pos_enc": [pos]}

View File

@ -0,0 +1,216 @@
# 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 math
from typing import Any, Optional, Tuple
import numpy as np
import torch
from torch import nn
class PositionEmbeddingSine(nn.Module):
"""
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.
"""
def __init__(
self,
num_pos_feats,
temperature: int = 10000,
normalize: bool = True,
scale: Optional[float] = None,
):
super().__init__()
assert num_pos_feats % 2 == 0, "Expecting even model width"
self.num_pos_feats = num_pos_feats // 2
self.temperature = temperature
self.normalize = normalize
if scale is not None and normalize is False:
raise ValueError("normalize should be True if scale is passed")
if scale is None:
scale = 2 * math.pi
self.scale = scale
self.cache = {}
def _encode_xy(self, x, y):
# The positions are expected to be normalized
assert len(x) == len(y) and x.ndim == y.ndim == 1
x_embed = x * self.scale
y_embed = y * self.scale
dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
pos_x = x_embed[:, None] / dim_t
pos_y = y_embed[:, None] / dim_t
pos_x = torch.stack(
(pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2
).flatten(1)
pos_y = torch.stack(
(pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
).flatten(1)
return pos_x, pos_y
@torch.no_grad()
def encode_boxes(self, x, y, w, h):
pos_x, pos_y = self._encode_xy(x, y)
pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
return pos
encode = encode_boxes # Backwards compatibility
@torch.no_grad()
def encode_points(self, x, y, labels):
(bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
assert bx == by and nx == ny and bx == bl and nx == nl
pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
return pos
@torch.no_grad()
def forward(self, x: torch.Tensor):
cache_key = (x.shape[-2], x.shape[-1])
if cache_key in self.cache:
return self.cache[cache_key][None].repeat(x.shape[0], 1, 1, 1)
y_embed = (
torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
.view(1, -1, 1)
.repeat(x.shape[0], 1, x.shape[-1])
)
x_embed = (
torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
.view(1, 1, -1)
.repeat(x.shape[0], x.shape[-2], 1)
)
if self.normalize:
eps = 1e-6
y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
pos_x = x_embed[:, :, :, None] / dim_t
pos_y = y_embed[:, :, :, None] / dim_t
pos_x = torch.stack(
(pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
).flatten(3)
pos_y = torch.stack(
(pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
).flatten(3)
pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
self.cache[cache_key] = pos[0]
return pos
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
# Rotary Positional Encoding, adapted from:
# 1. https://github.com/meta-llama/codellama/blob/main/llama/model.py
# 2. https://github.com/naver-ai/rope-vit
# 3. https://github.com/lucidrains/rotary-embedding-torch
def init_t_xy(end_x: int, end_y: int):
t = torch.arange(end_x * end_y, dtype=torch.float32)
t_x = (t % end_x).float()
t_y = torch.div(t, end_x, rounding_mode="floor").float()
return t_x, t_y
def compute_axial_cis(dim: int, end_x: int, end_y: int, theta: float = 10000.0):
freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
t_x, t_y = init_t_xy(end_x, end_y)
freqs_x = torch.outer(t_x, freqs_x)
freqs_y = torch.outer(t_y, freqs_y)
freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)
def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
ndim = x.ndim
assert 0 <= 1 < ndim
assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
return freqs_cis.view(*shape)
def apply_rotary_enc(
xq: torch.Tensor,
xk: torch.Tensor,
freqs_cis: torch.Tensor,
repeat_freqs_k: bool = False,
):
xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
xk_ = (
torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
if xk.shape[-2] != 0
else None
)
freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
if xk_ is None:
# no keys to rotate, due to dropout
return xq_out.type_as(xq).to(xq.device), xk
# repeat freqs along seq_len dim to match k seq_len
if repeat_freqs_k:
r = xk_.shape[-2] // xq_.shape[-2]
freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)

View File

@ -0,0 +1,5 @@
# 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.

View File

@ -0,0 +1,295 @@
# 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.
from typing import List, Optional, Tuple, Type
import torch
from torch import nn
from ..sam2_utils import LayerNorm2d, MLP
class MaskDecoder(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,
use_high_res_features: bool = False,
iou_prediction_use_sigmoid=False,
dynamic_multimask_via_stability=False,
dynamic_multimask_stability_delta=0.05,
dynamic_multimask_stability_thresh=0.98,
pred_obj_scores: bool = False,
pred_obj_scores_mlp: bool = False,
use_multimask_token_for_obj_ptr: bool = False,
) -> 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.pred_obj_scores = pred_obj_scores
if self.pred_obj_scores:
self.obj_score_token = nn.Embedding(1, transformer_dim)
self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
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.use_high_res_features = use_high_res_features
if use_high_res_features:
self.conv_s0 = nn.Conv2d(
transformer_dim, transformer_dim // 8, kernel_size=1, stride=1
)
self.conv_s1 = nn.Conv2d(
transformer_dim, transformer_dim // 4, kernel_size=1, stride=1
)
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,
sigmoid_output=iou_prediction_use_sigmoid,
)
if self.pred_obj_scores:
self.pred_obj_score_head = nn.Linear(transformer_dim, 1)
if pred_obj_scores_mlp:
self.pred_obj_score_head = MLP(transformer_dim, transformer_dim, 1, 3)
# When outputting a single mask, optionally we can dynamically fall back to the best
# multimask output token if the single mask output token gives low stability scores.
self.dynamic_multimask_via_stability = dynamic_multimask_via_stability
self.dynamic_multimask_stability_delta = dynamic_multimask_stability_delta
self.dynamic_multimask_stability_thresh = dynamic_multimask_stability_thresh
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]:
"""
Predict masks given image and prompt embeddings.
Arguments:
image_embeddings (torch.Tensor): the embeddings from the 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
torch.Tensor: batched SAM token for mask output
"""
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 masks. See 'forward' for more details."""
# 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] = []
for i in range(self.num_mask_tokens):
hyper_in_list.append(
self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :])
)
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
def _get_stability_scores(self, mask_logits):
"""
Compute stability scores of the mask logits based on the IoU between upper and
lower thresholds, similar to https://github.com/fairinternal/onevision/pull/568.
"""
mask_logits = mask_logits.flatten(-2)
stability_delta = self.dynamic_multimask_stability_delta
area_i = torch.sum(mask_logits > stability_delta, dim=-1).float()
area_u = torch.sum(mask_logits > -stability_delta, dim=-1).float()
stability_scores = torch.where(area_u > 0, area_i / area_u, 1.0)
return stability_scores
def _dynamic_multimask_via_stability(self, all_mask_logits, all_iou_scores):
"""
When outputting a single mask, if the stability score from the current single-mask
output (based on output token 0) falls below a threshold, we instead select from
multi-mask outputs (based on output token 1~3) the mask with the highest predicted
IoU score. This is intended to ensure a valid mask for both clicking and tracking.
"""
# The best mask from multimask output tokens (1~3)
multimask_logits = all_mask_logits[:, 1:, :, :]
multimask_iou_scores = all_iou_scores[:, 1:]
best_scores_inds = torch.argmax(multimask_iou_scores, dim=-1)
batch_inds = torch.arange(
multimask_iou_scores.size(0), device=all_iou_scores.device
)
best_multimask_logits = multimask_logits[batch_inds, best_scores_inds]
best_multimask_logits = best_multimask_logits.unsqueeze(1)
best_multimask_iou_scores = multimask_iou_scores[batch_inds, best_scores_inds]
best_multimask_iou_scores = best_multimask_iou_scores.unsqueeze(1)
# The mask from singlemask output token 0 and its stability score
singlemask_logits = all_mask_logits[:, 0:1, :, :]
singlemask_iou_scores = all_iou_scores[:, 0:1]
stability_scores = self._get_stability_scores(singlemask_logits)
is_stable = stability_scores >= self.dynamic_multimask_stability_thresh
# Dynamically fall back to best multimask output upon low stability scores.
mask_logits_out = torch.where(
is_stable[..., None, None].expand_as(singlemask_logits),
singlemask_logits,
best_multimask_logits,
)
iou_scores_out = torch.where(
is_stable.expand_as(singlemask_iou_scores),
singlemask_iou_scores,
best_multimask_iou_scores,
)
return mask_logits_out, iou_scores_out

View File

@ -0,0 +1,182 @@
# 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.
from typing import Optional, Tuple, Type
import torch
from torch import nn
from ..position_encoding import PositionEmbeddingRandom
from ..sam2_utils 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
point_embedding[labels == 2] += self.point_embeddings[2].weight
point_embedding[labels == 3] += self.point_embeddings[3].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

View File

@ -0,0 +1,327 @@
# 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 math
import warnings
from functools import partial
from typing import Tuple, Type
import torch
import torch.nn.functional as F
from torch import nn, Tensor
from ..position_encoding import apply_rotary_enc, compute_axial_cis
from ..sam2_utils import MLP
from ...utils.misc import get_sdpa_settings
warnings.simplefilter(action="ignore", category=FutureWarning)
OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = get_sdpa_settings()
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 = MLP(
embedding_dim, mlp_dim, embedding_dim, num_layers=2, activation=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,
dropout: float = 0.0,
kv_in_dim: int = None,
) -> None:
super().__init__()
self.embedding_dim = embedding_dim
self.kv_in_dim = kv_in_dim if kv_in_dim is not None else 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(self.kv_in_dim, self.internal_dim)
self.v_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
self.dropout_p = dropout
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)
dropout_p = self.dropout_p if self.training else 0.0
# Attention
with torch.backends.cuda.sdp_kernel(
enable_flash=USE_FLASH_ATTN,
# if Flash attention kernel is off, then math kernel needs to be enabled
enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
enable_mem_efficient=OLD_GPU,
):
out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
out = self._recombine_heads(out)
out = self.out_proj(out)
return out
class RoPEAttention(Attention):
"""Attention with rotary position encoding."""
def __init__(
self,
*args,
rope_theta=10000.0,
# whether to repeat q rope to match k length
# this is needed for cross-attention to memories
rope_k_repeat=False,
feat_sizes=(32, 32), # [w, h] for stride 16 feats at 512 resolution
**kwargs,
):
super().__init__(*args, **kwargs)
self.compute_cis = partial(
compute_axial_cis, dim=self.internal_dim // self.num_heads, theta=rope_theta
)
freqs_cis = self.compute_cis(end_x=feat_sizes[0], end_y=feat_sizes[1])
self.freqs_cis = freqs_cis
self.rope_k_repeat = rope_k_repeat
def forward(
self, q: Tensor, k: Tensor, v: Tensor, num_k_exclude_rope: int = 0
) -> 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)
# Apply rotary position encoding
w = h = math.sqrt(q.shape[-2])
self.freqs_cis = self.freqs_cis.to(q.device)
if self.freqs_cis.shape[0] != q.shape[-2]:
self.freqs_cis = self.compute_cis(end_x=w, end_y=h).to(q.device)
if q.shape[-2] != k.shape[-2]:
assert self.rope_k_repeat
num_k_rope = k.size(-2) - num_k_exclude_rope
q, k[:, :, :num_k_rope] = apply_rotary_enc(
q,
k[:, :, :num_k_rope],
freqs_cis=self.freqs_cis,
repeat_freqs_k=self.rope_k_repeat,
)
dropout_p = self.dropout_p if self.training else 0.0
# Attention
with torch.backends.cuda.sdp_kernel(
enable_flash=USE_FLASH_ATTN,
# if Flash attention kernel is off, then math kernel needs to be enabled
enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
enable_mem_efficient=OLD_GPU,
):
out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
out = self._recombine_heads(out)
out = self.out_proj(out)
return out

View File

@ -0,0 +1,832 @@
# 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.distributed
import torch.nn.functional as F
from torch.nn.init import trunc_normal_
from .sam.mask_decoder import MaskDecoder
from .sam.prompt_encoder import PromptEncoder
from .sam.transformer import TwoWayTransformer
from .sam2_utils import get_1d_sine_pe, MLP, select_closest_cond_frames
# a large negative value as a placeholder score for missing objects
NO_OBJ_SCORE = -1024.0
class SAM2Base(torch.nn.Module):
def __init__(
self,
image_encoder,
memory_attention,
memory_encoder,
num_maskmem=7, # default 1 input frame + 6 previous frames
image_size=512,
backbone_stride=16, # stride of the image backbone output
sigmoid_scale_for_mem_enc=1.0, # scale factor for mask sigmoid prob
sigmoid_bias_for_mem_enc=0.0, # bias factor for mask sigmoid prob
# During evaluation, whether to binarize the sigmoid mask logits on interacted frames with clicks
binarize_mask_from_pts_for_mem_enc=False,
use_mask_input_as_output_without_sam=False,
# on frames with mask input, whether to directly output the input mask without using a SAM prompt encoder + mask decoder
# The maximum number of conditioning frames to participate in the memory attention (-1 means no limit; if there are more conditioning frames than this limit,
# we only cross-attend to the temporally closest `max_cond_frames_in_attn` conditioning frames in the encoder when tracking each frame). This gives the model
# a temporal locality when handling a large number of annotated frames (since closer frames should be more important) and also avoids GPU OOM.
max_cond_frames_in_attn=-1,
# on the first frame, whether to directly add the no-memory embedding to the image feature
# (instead of using the transformer encoder)
directly_add_no_mem_embed=False,
# whether to use high-resolution feature maps in the SAM mask decoder
use_high_res_features_in_sam=False,
# whether to output multiple (3) masks for the first click on initial conditioning frames
multimask_output_in_sam=False,
# the minimum and maximum number of clicks to use multimask_output_in_sam (only relevant when `multimask_output_in_sam=True`;
# default is 1 for both, meaning that only the first click gives multimask output; also note that a box counts as two points)
multimask_min_pt_num=1,
multimask_max_pt_num=1,
# whether to also use multimask output for tracking (not just for the first click on initial conditioning frames; only relevant when `multimask_output_in_sam=True`)
multimask_output_for_tracking=False,
# Whether to use multimask tokens for obj ptr; Only relevant when both
# use_obj_ptrs_in_encoder=True and multimask_output_for_tracking=True
use_multimask_token_for_obj_ptr: bool = False,
# whether to use sigmoid to restrict ious prediction to [0-1]
iou_prediction_use_sigmoid=False,
# The memory bank's temporal stride during evaluation (i.e. the `r` parameter in XMem and Cutie; XMem and Cutie use r=5).
# For r>1, the (self.num_maskmem - 1) non-conditioning memory frames consist of
# (self.num_maskmem - 2) nearest frames from every r-th frames, plus the last frame.
memory_temporal_stride_for_eval=1,
# if `add_all_frames_to_correct_as_cond` is True, we also append to the conditioning frame list any frame that receives a later correction click
# if `add_all_frames_to_correct_as_cond` is False, we conditioning frame list to only use those initial conditioning frames
add_all_frames_to_correct_as_cond=False,
# whether to apply non-overlapping constraints on the object masks in the memory encoder during evaluation (to avoid/alleviate superposing masks)
non_overlap_masks_for_mem_enc=False,
# whether to cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
use_obj_ptrs_in_encoder=False,
# the maximum number of object pointers from other frames in encoder cross attention (only relevant when `use_obj_ptrs_in_encoder=True`)
max_obj_ptrs_in_encoder=16,
# whether to add temporal positional encoding to the object pointers in the encoder (only relevant when `use_obj_ptrs_in_encoder=True`)
add_tpos_enc_to_obj_ptrs=True,
# whether to add an extra linear projection layer for the temporal positional encoding in the object pointers to avoid potential interference
# with spatial positional encoding (only relevant when both `use_obj_ptrs_in_encoder=True` and `add_tpos_enc_to_obj_ptrs=True`)
proj_tpos_enc_in_obj_ptrs=False,
# whether to only attend to object pointers in the past (before the current frame) in the encoder during evaluation
# (only relevant when `use_obj_ptrs_in_encoder=True`; this might avoid pointer information too far in the future to distract the initial tracking)
only_obj_ptrs_in_the_past_for_eval=False,
# Whether to predict if there is an object in the frame
pred_obj_scores: bool = False,
# Whether to use an MLP to predict object scores
pred_obj_scores_mlp: bool = False,
# Only relevant if pred_obj_scores=True and use_obj_ptrs_in_encoder=True;
# Whether to have a fixed no obj pointer when there is no object present
# or to use it as an additive embedding with obj_ptr produced by decoder
fixed_no_obj_ptr: bool = False,
# Soft no object, i.e. mix in no_obj_ptr softly,
# hope to make recovery easier if there is a mistake and mitigate accumulation of errors
soft_no_obj_ptr: bool = False,
use_mlp_for_obj_ptr_proj: bool = False,
# extra arguments used to construct the SAM mask decoder; if not None, it should be a dict of kwargs to be passed into `MaskDecoder` class.
sam_mask_decoder_extra_args=None,
compile_image_encoder: bool = False,
):
super().__init__()
# Part 1: the image backbone
self.image_encoder = image_encoder
# Use level 0, 1, 2 for high-res setting, or just level 2 for the default setting
self.use_high_res_features_in_sam = use_high_res_features_in_sam
self.num_feature_levels = 3 if use_high_res_features_in_sam else 1
self.use_obj_ptrs_in_encoder = use_obj_ptrs_in_encoder
self.max_obj_ptrs_in_encoder = max_obj_ptrs_in_encoder
if use_obj_ptrs_in_encoder:
# A conv layer to downsample the mask prompt to stride 4 (the same stride as
# low-res SAM mask logits) and to change its scales from 0~1 to SAM logit scale,
# so that it can be fed into the SAM mask decoder to generate a pointer.
self.mask_downsample = torch.nn.Conv2d(1, 1, kernel_size=4, stride=4)
self.add_tpos_enc_to_obj_ptrs = add_tpos_enc_to_obj_ptrs
if proj_tpos_enc_in_obj_ptrs:
assert add_tpos_enc_to_obj_ptrs # these options need to be used together
self.proj_tpos_enc_in_obj_ptrs = proj_tpos_enc_in_obj_ptrs
self.only_obj_ptrs_in_the_past_for_eval = only_obj_ptrs_in_the_past_for_eval
# Part 2: memory attention to condition current frame's visual features
# with memories (and obj ptrs) from past frames
self.memory_attention = memory_attention
self.hidden_dim = memory_attention.d_model
# Part 3: memory encoder for the previous frame's outputs
self.memory_encoder = memory_encoder
self.mem_dim = self.hidden_dim
if hasattr(self.memory_encoder, "out_proj") and hasattr(
self.memory_encoder.out_proj, "weight"
):
# if there is compression of memories along channel dim
self.mem_dim = self.memory_encoder.out_proj.weight.shape[0]
self.num_maskmem = num_maskmem # Number of memories accessible
# Temporal encoding of the memories
self.maskmem_tpos_enc = torch.nn.Parameter(
torch.zeros(num_maskmem, 1, 1, self.mem_dim)
)
trunc_normal_(self.maskmem_tpos_enc, std=0.02)
# a single token to indicate no memory embedding from previous frames
self.no_mem_embed = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
self.no_mem_pos_enc = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
trunc_normal_(self.no_mem_embed, std=0.02)
trunc_normal_(self.no_mem_pos_enc, std=0.02)
self.directly_add_no_mem_embed = directly_add_no_mem_embed
# Apply sigmoid to the output raw mask logits (to turn them from
# range (-inf, +inf) to range (0, 1)) before feeding them into the memory encoder
self.sigmoid_scale_for_mem_enc = sigmoid_scale_for_mem_enc
self.sigmoid_bias_for_mem_enc = sigmoid_bias_for_mem_enc
self.binarize_mask_from_pts_for_mem_enc = binarize_mask_from_pts_for_mem_enc
self.non_overlap_masks_for_mem_enc = non_overlap_masks_for_mem_enc
self.memory_temporal_stride_for_eval = memory_temporal_stride_for_eval
# On frames with mask input, whether to directly output the input mask without
# using a SAM prompt encoder + mask decoder
self.use_mask_input_as_output_without_sam = use_mask_input_as_output_without_sam
self.multimask_output_in_sam = multimask_output_in_sam
self.multimask_min_pt_num = multimask_min_pt_num
self.multimask_max_pt_num = multimask_max_pt_num
self.multimask_output_for_tracking = multimask_output_for_tracking
self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
self.iou_prediction_use_sigmoid = iou_prediction_use_sigmoid
# Part 4: SAM-style prompt encoder (for both mask and point inputs)
# and SAM-style mask decoder for the final mask output
self.image_size = image_size
self.backbone_stride = backbone_stride
self.sam_mask_decoder_extra_args = sam_mask_decoder_extra_args
self.pred_obj_scores = pred_obj_scores
self.pred_obj_scores_mlp = pred_obj_scores_mlp
self.fixed_no_obj_ptr = fixed_no_obj_ptr
self.soft_no_obj_ptr = soft_no_obj_ptr
if self.fixed_no_obj_ptr:
assert self.pred_obj_scores
assert self.use_obj_ptrs_in_encoder
if self.pred_obj_scores and self.use_obj_ptrs_in_encoder:
self.no_obj_ptr = torch.nn.Parameter(torch.zeros(1, self.hidden_dim))
trunc_normal_(self.no_obj_ptr, std=0.02)
self.use_mlp_for_obj_ptr_proj = use_mlp_for_obj_ptr_proj
self._build_sam_heads()
self.add_all_frames_to_correct_as_cond = add_all_frames_to_correct_as_cond
self.max_cond_frames_in_attn = max_cond_frames_in_attn
# Model compilation
if compile_image_encoder:
# Compile the forward function (not the full module) to allow loading checkpoints.
print(
"Image encoder compilation is enabled. First forward pass will be slow."
)
self.image_encoder.forward = torch.compile(
self.image_encoder.forward,
mode="max-autotune",
fullgraph=True,
dynamic=False,
)
@property
def device(self):
return next(self.parameters()).device
def forward(self, *args, **kwargs):
raise NotImplementedError(
"Please use the corresponding methods in SAM2VideoPredictor for inference."
"See notebooks/video_predictor_example.ipynb for an example."
)
def _build_sam_heads(self):
"""Build SAM-style prompt encoder and mask decoder."""
self.sam_prompt_embed_dim = self.hidden_dim
self.sam_image_embedding_size = self.image_size // self.backbone_stride
# build PromptEncoder and MaskDecoder from SAM
# (their hyperparameters like `mask_in_chans=16` are from SAM code)
self.sam_prompt_encoder = PromptEncoder(
embed_dim=self.sam_prompt_embed_dim,
image_embedding_size=(
self.sam_image_embedding_size,
self.sam_image_embedding_size,
),
input_image_size=(self.image_size, self.image_size),
mask_in_chans=16,
)
self.sam_mask_decoder = MaskDecoder(
num_multimask_outputs=3,
transformer=TwoWayTransformer(
depth=2,
embedding_dim=self.sam_prompt_embed_dim,
mlp_dim=2048,
num_heads=8,
),
transformer_dim=self.sam_prompt_embed_dim,
iou_head_depth=3,
iou_head_hidden_dim=256,
use_high_res_features=self.use_high_res_features_in_sam,
iou_prediction_use_sigmoid=self.iou_prediction_use_sigmoid,
pred_obj_scores=self.pred_obj_scores,
pred_obj_scores_mlp=self.pred_obj_scores_mlp,
use_multimask_token_for_obj_ptr=self.use_multimask_token_for_obj_ptr,
**(self.sam_mask_decoder_extra_args or {}),
)
if self.use_obj_ptrs_in_encoder:
# a linear projection on SAM output tokens to turn them into object pointers
self.obj_ptr_proj = torch.nn.Linear(self.hidden_dim, self.hidden_dim)
if self.use_mlp_for_obj_ptr_proj:
self.obj_ptr_proj = MLP(
self.hidden_dim, self.hidden_dim, self.hidden_dim, 3
)
else:
self.obj_ptr_proj = torch.nn.Identity()
if self.proj_tpos_enc_in_obj_ptrs:
# a linear projection on temporal positional encoding in object pointers to
# avoid potential interference with spatial positional encoding
self.obj_ptr_tpos_proj = torch.nn.Linear(self.hidden_dim, self.mem_dim)
else:
self.obj_ptr_tpos_proj = torch.nn.Identity()
def _forward_sam_heads(
self,
backbone_features,
point_inputs=None,
mask_inputs=None,
high_res_features=None,
multimask_output=False,
):
"""
Forward SAM prompt encoders and mask heads.
Inputs:
- backbone_features: image features of [B, C, H, W] shape
- point_inputs: a dictionary with "point_coords" and "point_labels", where
1) "point_coords" has [B, P, 2] shape and float32 dtype and contains the
absolute pixel-unit coordinate in (x, y) format of the P input points
2) "point_labels" has shape [B, P] and int32 dtype, where 1 means
positive clicks, 0 means negative clicks, and -1 means padding
- mask_inputs: a mask of [B, 1, H*16, W*16] shape, float or bool, with the
same spatial size as the image.
- high_res_features: either 1) None or 2) or a list of length 2 containing
two feature maps of [B, C, 4*H, 4*W] and [B, C, 2*H, 2*W] shapes respectively,
which will be used as high-resolution feature maps for SAM decoder.
- multimask_output: if it's True, we output 3 candidate masks and their 3
corresponding IoU estimates, and if it's False, we output only 1 mask and
its corresponding IoU estimate.
Outputs:
- low_res_multimasks: [B, M, H*4, W*4] shape (where M = 3 if
`multimask_output=True` and M = 1 if `multimask_output=False`), the SAM
output mask logits (before sigmoid) for the low-resolution masks, with 4x
the resolution (1/4 stride) of the input backbone_features.
- high_res_multimasks: [B, M, H*16, W*16] shape (where M = 3
if `multimask_output=True` and M = 1 if `multimask_output=False`),
upsampled from the low-resolution masks, with shape size as the image
(stride is 1 pixel).
- ious, [B, M] shape, where (where M = 3 if `multimask_output=True` and M = 1
if `multimask_output=False`), the estimated IoU of each output mask.
- low_res_masks: [B, 1, H*4, W*4] shape, the best mask in `low_res_multimasks`.
If `multimask_output=True`, it's the mask with the highest IoU estimate.
If `multimask_output=False`, it's the same as `low_res_multimasks`.
- high_res_masks: [B, 1, H*16, W*16] shape, the best mask in `high_res_multimasks`.
If `multimask_output=True`, it's the mask with the highest IoU estimate.
If `multimask_output=False`, it's the same as `high_res_multimasks`.
- obj_ptr: [B, C] shape, the object pointer vector for the output mask, extracted
based on the output token from the SAM mask decoder.
"""
B = backbone_features.size(0)
device = backbone_features.device
assert backbone_features.size(1) == self.sam_prompt_embed_dim
assert backbone_features.size(2) == self.sam_image_embedding_size
assert backbone_features.size(3) == self.sam_image_embedding_size
# a) Handle point prompts
if point_inputs is not None:
sam_point_coords = point_inputs["point_coords"]
sam_point_labels = point_inputs["point_labels"]
assert sam_point_coords.size(0) == B and sam_point_labels.size(0) == B
else:
# If no points are provide, pad with an empty point (with label -1)
sam_point_coords = torch.zeros(B, 1, 2, device=device)
sam_point_labels = -torch.ones(B, 1, dtype=torch.int32, device=device)
# b) Handle mask prompts
if mask_inputs is not None:
# If mask_inputs is provided, downsize it into low-res mask input if needed
# and feed it as a dense mask prompt into the SAM mask encoder
assert len(mask_inputs.shape) == 4 and mask_inputs.shape[:2] == (B, 1)
if mask_inputs.shape[-2:] != self.sam_prompt_encoder.mask_input_size:
sam_mask_prompt = F.interpolate(
mask_inputs.float(),
size=self.sam_prompt_encoder.mask_input_size,
align_corners=False,
mode="bilinear",
antialias=True, # use antialias for downsampling
)
else:
sam_mask_prompt = mask_inputs
else:
# Otherwise, simply feed None (and SAM's prompt encoder will add
# a learned `no_mask_embed` to indicate no mask input in this case).
sam_mask_prompt = None
sparse_embeddings, dense_embeddings = self.sam_prompt_encoder(
points=(sam_point_coords, sam_point_labels),
boxes=None,
masks=sam_mask_prompt,
)
(
low_res_multimasks,
ious,
sam_output_tokens,
object_score_logits,
) = self.sam_mask_decoder(
image_embeddings=backbone_features,
image_pe=self.sam_prompt_encoder.get_dense_pe(),
sparse_prompt_embeddings=sparse_embeddings,
dense_prompt_embeddings=dense_embeddings,
multimask_output=multimask_output,
repeat_image=False, # the image is already batched
high_res_features=high_res_features,
)
if self.pred_obj_scores:
is_obj_appearing = object_score_logits > 0
# Mask used for spatial memories is always a *hard* choice between obj and no obj,
# consistent with the actual mask prediction
low_res_multimasks = torch.where(
is_obj_appearing[:, None, None],
low_res_multimasks,
NO_OBJ_SCORE,
)
# convert masks from possibly bfloat16 (or float16) to float32
# (older PyTorch versions before 2.1 don't support `interpolate` on bf16)
low_res_multimasks = low_res_multimasks.float()
high_res_multimasks = F.interpolate(
low_res_multimasks,
size=(self.image_size, self.image_size),
mode="bilinear",
align_corners=False,
)
sam_output_token = sam_output_tokens[:, 0]
if multimask_output:
# take the best mask prediction (with the highest IoU estimation)
best_iou_inds = torch.argmax(ious, dim=-1)
batch_inds = torch.arange(B, device=device)
low_res_masks = low_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
high_res_masks = high_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
if sam_output_tokens.size(1) > 1:
sam_output_token = sam_output_tokens[batch_inds, best_iou_inds]
else:
low_res_masks, high_res_masks = low_res_multimasks, high_res_multimasks
# Extract object pointer from the SAM output token (with occlusion handling)
obj_ptr = self.obj_ptr_proj(sam_output_token)
if self.pred_obj_scores:
# Allow *soft* no obj ptr, unlike for masks
if self.soft_no_obj_ptr:
# Only hard possible with gt
assert not self.teacher_force_obj_scores_for_mem
lambda_is_obj_appearing = object_score_logits.sigmoid()
else:
lambda_is_obj_appearing = is_obj_appearing.float()
if self.fixed_no_obj_ptr:
obj_ptr = lambda_is_obj_appearing * obj_ptr
obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
return (
low_res_multimasks,
high_res_multimasks,
ious,
low_res_masks,
high_res_masks,
obj_ptr,
object_score_logits,
)
def _use_mask_as_output(self, backbone_features, high_res_features, mask_inputs):
"""
Directly turn binary `mask_inputs` into a output mask logits without using SAM.
(same input and output shapes as in _forward_sam_heads above).
"""
# Use -10/+10 as logits for neg/pos pixels (very close to 0/1 in prob after sigmoid).
out_scale, out_bias = 20.0, -10.0 # sigmoid(-10.0)=4.5398e-05
mask_inputs_float = mask_inputs.float()
high_res_masks = mask_inputs_float * out_scale + out_bias
low_res_masks = F.interpolate(
high_res_masks,
size=(high_res_masks.size(-2) // 4, high_res_masks.size(-1) // 4),
align_corners=False,
mode="bilinear",
antialias=True, # use antialias for downsampling
)
# a dummy IoU prediction of all 1's under mask input
ious = mask_inputs.new_ones(mask_inputs.size(0), 1).float()
if not self.use_obj_ptrs_in_encoder:
# all zeros as a dummy object pointer (of shape [B, C])
obj_ptr = torch.zeros(
mask_inputs.size(0), self.hidden_dim, device=mask_inputs.device
)
else:
# produce an object pointer using the SAM decoder from the mask input
_, _, _, _, _, obj_ptr, _ = self._forward_sam_heads(
backbone_features=backbone_features,
mask_inputs=self.mask_downsample(mask_inputs_float),
high_res_features=high_res_features,
)
# In this method, we are treating mask_input as output, e.g. using it directly to create spatial mem;
# Below, we follow the same design axiom to use mask_input to decide if obj appears or not instead of relying
# on the object_scores from the SAM decoder.
is_obj_appearing = torch.any(mask_inputs.flatten(1).float() > 0.0, dim=1)
is_obj_appearing = is_obj_appearing[..., None]
lambda_is_obj_appearing = is_obj_appearing.float()
object_score_logits = out_scale * lambda_is_obj_appearing + out_bias
if self.pred_obj_scores:
if self.fixed_no_obj_ptr:
obj_ptr = lambda_is_obj_appearing * obj_ptr
obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
return (
low_res_masks,
high_res_masks,
ious,
low_res_masks,
high_res_masks,
obj_ptr,
object_score_logits,
)
def forward_image(self, img_batch: torch.Tensor):
"""Get the image feature on the input batch."""
backbone_out = self.image_encoder(img_batch)
if self.use_high_res_features_in_sam:
# precompute projected level 0 and level 1 features in SAM decoder
# to avoid running it again on every SAM click
backbone_out["backbone_fpn"][0] = self.sam_mask_decoder.conv_s0(
backbone_out["backbone_fpn"][0]
)
backbone_out["backbone_fpn"][1] = self.sam_mask_decoder.conv_s1(
backbone_out["backbone_fpn"][1]
)
return backbone_out
def _prepare_backbone_features(self, backbone_out):
"""Prepare and flatten visual features."""
backbone_out = backbone_out.copy()
assert len(backbone_out["backbone_fpn"]) == len(backbone_out["vision_pos_enc"])
assert len(backbone_out["backbone_fpn"]) >= self.num_feature_levels
feature_maps = backbone_out["backbone_fpn"][-self.num_feature_levels :]
vision_pos_embeds = backbone_out["vision_pos_enc"][-self.num_feature_levels :]
feat_sizes = [(x.shape[-2], x.shape[-1]) for x in vision_pos_embeds]
# flatten NxCxHxW to HWxNxC
vision_feats = [x.flatten(2).permute(2, 0, 1) for x in feature_maps]
vision_pos_embeds = [x.flatten(2).permute(2, 0, 1) for x in vision_pos_embeds]
return backbone_out, vision_feats, vision_pos_embeds, feat_sizes
def _prepare_memory_conditioned_features(
self,
frame_idx,
is_init_cond_frame,
current_vision_feats,
current_vision_pos_embeds,
feat_sizes,
output_dict,
num_frames,
track_in_reverse=False, # tracking in reverse time order (for demo usage)
):
"""Fuse the current frame's visual feature map with previous memory."""
B = current_vision_feats[-1].size(1) # batch size on this frame
C = self.hidden_dim
H, W = feat_sizes[-1] # top-level (lowest-resolution) feature size
device = current_vision_feats[-1].device
# The case of `self.num_maskmem == 0` below is primarily used for reproducing SAM on images.
# In this case, we skip the fusion with any memory.
if self.num_maskmem == 0: # Disable memory and skip fusion
pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
return pix_feat
num_obj_ptr_tokens = 0
# Step 1: condition the visual features of the current frame on previous memories
if not is_init_cond_frame:
# Retrieve the memories encoded with the maskmem backbone
to_cat_memory, to_cat_memory_pos_embed = [], []
# Add conditioning frames's output first (all cond frames have t_pos=0 for
# when getting temporal positional embedding below)
assert len(output_dict["cond_frame_outputs"]) > 0
# Select a maximum number of temporally closest cond frames for cross attention
cond_outputs = output_dict["cond_frame_outputs"]
selected_cond_outputs, unselected_cond_outputs = select_closest_cond_frames(
frame_idx, cond_outputs, self.max_cond_frames_in_attn
)
t_pos_and_prevs = [(0, out) for out in selected_cond_outputs.values()]
# Add last (self.num_maskmem - 1) frames before current frame for non-conditioning memory
# the earliest one has t_pos=1 and the latest one has t_pos=self.num_maskmem-1
# We also allow taking the memory frame non-consecutively (with r>1), in which case
# we take (self.num_maskmem - 2) frames among every r-th frames plus the last frame.
r = self.memory_temporal_stride_for_eval
for t_pos in range(1, self.num_maskmem):
t_rel = self.num_maskmem - t_pos # how many frames before current frame
if t_rel == 1:
# for t_rel == 1, we take the last frame (regardless of r)
if not track_in_reverse:
# the frame immediately before this frame (i.e. frame_idx - 1)
prev_frame_idx = frame_idx - t_rel
else:
# the frame immediately after this frame (i.e. frame_idx + 1)
prev_frame_idx = frame_idx + t_rel
else:
# for t_rel >= 2, we take the memory frame from every r-th frames
if not track_in_reverse:
# first find the nearest frame among every r-th frames before this frame
# for r=1, this would be (frame_idx - 2)
prev_frame_idx = ((frame_idx - 2) // r) * r
# then seek further among every r-th frames
prev_frame_idx = prev_frame_idx - (t_rel - 2) * r
else:
# first find the nearest frame among every r-th frames after this frame
# for r=1, this would be (frame_idx + 2)
prev_frame_idx = -(-(frame_idx + 2) // r) * r
# then seek further among every r-th frames
prev_frame_idx = prev_frame_idx + (t_rel - 2) * r
out = output_dict["non_cond_frame_outputs"].get(prev_frame_idx, None)
if out is None:
# If an unselected conditioning frame is among the last (self.num_maskmem - 1)
# frames, we still attend to it as if it's a non-conditioning frame.
out = unselected_cond_outputs.get(prev_frame_idx, None)
t_pos_and_prevs.append((t_pos, out))
for t_pos, prev in t_pos_and_prevs:
if prev is None:
continue # skip padding frames
# "maskmem_features" might have been offloaded to CPU in demo use cases,
# so we load it back to GPU (it's a no-op if it's already on GPU).
feats = prev["maskmem_features"].cuda(non_blocking=True)
to_cat_memory.append(feats.flatten(2).permute(2, 0, 1))
# Spatial positional encoding (it might have been offloaded to CPU in eval)
maskmem_enc = prev["maskmem_pos_enc"][-1].cuda()
maskmem_enc = maskmem_enc.flatten(2).permute(2, 0, 1)
# Temporal positional encoding
maskmem_enc = (
maskmem_enc + self.maskmem_tpos_enc[self.num_maskmem - t_pos - 1]
)
to_cat_memory_pos_embed.append(maskmem_enc)
# Construct the list of past object pointers
if self.use_obj_ptrs_in_encoder:
max_obj_ptrs_in_encoder = min(num_frames, self.max_obj_ptrs_in_encoder)
# First add those object pointers from selected conditioning frames
# (optionally, only include object pointers in the past during evaluation)
if not self.training and self.only_obj_ptrs_in_the_past_for_eval:
ptr_cond_outputs = {
t: out
for t, out in selected_cond_outputs.items()
if (t >= frame_idx if track_in_reverse else t <= frame_idx)
}
else:
ptr_cond_outputs = selected_cond_outputs
pos_and_ptrs = [
# Temporal pos encoding contains how far away each pointer is from current frame
(abs(frame_idx - t), out["obj_ptr"])
for t, out in ptr_cond_outputs.items()
]
# Add up to (max_obj_ptrs_in_encoder - 1) non-conditioning frames before current frame
for t_diff in range(1, max_obj_ptrs_in_encoder):
t = frame_idx + t_diff if track_in_reverse else frame_idx - t_diff
if t < 0 or (num_frames is not None and t >= num_frames):
break
out = output_dict["non_cond_frame_outputs"].get(
t, unselected_cond_outputs.get(t, None)
)
if out is not None:
pos_and_ptrs.append((t_diff, out["obj_ptr"]))
# If we have at least one object pointer, add them to the across attention
if len(pos_and_ptrs) > 0:
pos_list, ptrs_list = zip(*pos_and_ptrs)
# stack object pointers along dim=0 into [ptr_seq_len, B, C] shape
obj_ptrs = torch.stack(ptrs_list, dim=0)
# a temporal positional embedding based on how far each object pointer is from
# the current frame (sine embedding normalized by the max pointer num).
if self.add_tpos_enc_to_obj_ptrs:
t_diff_max = max_obj_ptrs_in_encoder - 1
tpos_dim = C if self.proj_tpos_enc_in_obj_ptrs else self.mem_dim
obj_pos = torch.tensor(pos_list, device=device)
obj_pos = get_1d_sine_pe(obj_pos / t_diff_max, dim=tpos_dim)
obj_pos = self.obj_ptr_tpos_proj(obj_pos)
obj_pos = obj_pos.unsqueeze(1).expand(-1, B, self.mem_dim)
else:
obj_pos = obj_ptrs.new_zeros(len(pos_list), B, self.mem_dim)
if self.mem_dim < C:
# split a pointer into (C // self.mem_dim) tokens for self.mem_dim < C
obj_ptrs = obj_ptrs.reshape(
-1, B, C // self.mem_dim, self.mem_dim
)
obj_ptrs = obj_ptrs.permute(0, 2, 1, 3).flatten(0, 1)
obj_pos = obj_pos.repeat_interleave(C // self.mem_dim, dim=0)
to_cat_memory.append(obj_ptrs)
to_cat_memory_pos_embed.append(obj_pos)
num_obj_ptr_tokens = obj_ptrs.shape[0]
else:
num_obj_ptr_tokens = 0
else:
# for initial conditioning frames, encode them without using any previous memory
if self.directly_add_no_mem_embed:
# directly add no-mem embedding (instead of using the transformer encoder)
pix_feat_with_mem = current_vision_feats[-1] + self.no_mem_embed
pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
return pix_feat_with_mem
# Use a dummy token on the first frame (to avoid emtpy memory input to tranformer encoder)
to_cat_memory = [self.no_mem_embed.expand(1, B, self.mem_dim)]
to_cat_memory_pos_embed = [self.no_mem_pos_enc.expand(1, B, self.mem_dim)]
# Step 2: Concatenate the memories and forward through the transformer encoder
memory = torch.cat(to_cat_memory, dim=0)
memory_pos_embed = torch.cat(to_cat_memory_pos_embed, dim=0)
pix_feat_with_mem = self.memory_attention(
curr=current_vision_feats,
curr_pos=current_vision_pos_embeds,
memory=memory,
memory_pos=memory_pos_embed,
num_obj_ptr_tokens=num_obj_ptr_tokens,
)
# reshape the output (HW)BC => BCHW
pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
return pix_feat_with_mem
def _encode_new_memory(
self,
current_vision_feats,
feat_sizes,
pred_masks_high_res,
is_mask_from_pts,
):
"""Encode the current image and its prediction into a memory feature."""
B = current_vision_feats[-1].size(1) # batch size on this frame
C = self.hidden_dim
H, W = feat_sizes[-1] # top-level (lowest-resolution) feature size
# top-level feature, (HW)BC => BCHW
pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
if self.non_overlap_masks_for_mem_enc and not self.training:
# optionally, apply non-overlapping constraints to the masks (it's applied
# in the batch dimension and should only be used during eval, where all
# the objects come from the same video under batch size 1).
pred_masks_high_res = self._apply_non_overlapping_constraints(
pred_masks_high_res
)
# scale the raw mask logits with a temperature before applying sigmoid
binarize = self.binarize_mask_from_pts_for_mem_enc and is_mask_from_pts
if binarize and not self.training:
mask_for_mem = (pred_masks_high_res > 0).float()
else:
# apply sigmoid on the raw mask logits to turn them into range (0, 1)
mask_for_mem = torch.sigmoid(pred_masks_high_res)
# apply scale and bias terms to the sigmoid probabilities
if self.sigmoid_scale_for_mem_enc != 1.0:
mask_for_mem = mask_for_mem * self.sigmoid_scale_for_mem_enc
if self.sigmoid_bias_for_mem_enc != 0.0:
mask_for_mem = mask_for_mem + self.sigmoid_bias_for_mem_enc
maskmem_out = self.memory_encoder(
pix_feat,
mask_for_mem,
skip_mask_sigmoid=True, # sigmoid already applied
)
maskmem_features = maskmem_out["vision_features"]
maskmem_pos_enc = maskmem_out["vision_pos_enc"]
return maskmem_features, maskmem_pos_enc
def track_step(
self,
frame_idx,
is_init_cond_frame,
current_vision_feats,
current_vision_pos_embeds,
feat_sizes,
point_inputs,
mask_inputs,
output_dict,
num_frames,
track_in_reverse=False, # tracking in reverse time order (for demo usage)
# Whether to run the memory encoder on the predicted masks. Sometimes we might want
# to skip the memory encoder with `run_mem_encoder=False`. For example,
# in demo we might call `track_step` multiple times for each user click,
# and only encode the memory when the user finalizes their clicks. And in ablation
# settings like SAM training on static images, we don't need the memory encoder.
run_mem_encoder=True,
# The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
prev_sam_mask_logits=None,
):
current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
# High-resolution feature maps for the SAM head, reshape (HW)BC => BCHW
if len(current_vision_feats) > 1:
high_res_features = [
x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
for x, s in zip(current_vision_feats[:-1], feat_sizes[:-1])
]
else:
high_res_features = None
if mask_inputs is not None and self.use_mask_input_as_output_without_sam:
# When use_mask_input_as_output_without_sam=True, we directly output the mask input
# (see it as a GT mask) without using a SAM prompt encoder + mask decoder.
pix_feat = current_vision_feats[-1].permute(1, 2, 0)
pix_feat = pix_feat.view(-1, self.hidden_dim, *feat_sizes[-1])
sam_outputs = self._use_mask_as_output(
pix_feat, high_res_features, mask_inputs
)
else:
# fused the visual feature with previous memory features in the memory bank
pix_feat_with_mem = self._prepare_memory_conditioned_features(
frame_idx=frame_idx,
is_init_cond_frame=is_init_cond_frame,
current_vision_feats=current_vision_feats[-1:],
current_vision_pos_embeds=current_vision_pos_embeds[-1:],
feat_sizes=feat_sizes[-1:],
output_dict=output_dict,
num_frames=num_frames,
track_in_reverse=track_in_reverse,
)
# apply SAM-style segmentation head
# here we might feed previously predicted low-res SAM mask logits into the SAM mask decoder,
# e.g. in demo where such logits come from earlier interaction instead of correction sampling
# (in this case, any `mask_inputs` shouldn't reach here as they are sent to _use_mask_as_output instead)
if prev_sam_mask_logits is not None:
assert point_inputs is not None and mask_inputs is None
mask_inputs = prev_sam_mask_logits
multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
sam_outputs = self._forward_sam_heads(
backbone_features=pix_feat_with_mem,
point_inputs=point_inputs,
mask_inputs=mask_inputs,
high_res_features=high_res_features,
multimask_output=multimask_output,
)
(
_,
_,
_,
low_res_masks,
high_res_masks,
obj_ptr,
_,
) = sam_outputs
current_out["pred_masks"] = low_res_masks
current_out["pred_masks_high_res"] = high_res_masks
current_out["obj_ptr"] = obj_ptr
# Finally run the memory encoder on the predicted mask to encode
# it into a new memory feature (that can be used in future frames)
if run_mem_encoder and self.num_maskmem > 0:
high_res_masks_for_mem_enc = high_res_masks
maskmem_features, maskmem_pos_enc = self._encode_new_memory(
current_vision_feats=current_vision_feats,
feat_sizes=feat_sizes,
pred_masks_high_res=high_res_masks_for_mem_enc,
is_mask_from_pts=(point_inputs is not None),
)
current_out["maskmem_features"] = maskmem_features
current_out["maskmem_pos_enc"] = maskmem_pos_enc
else:
current_out["maskmem_features"] = None
current_out["maskmem_pos_enc"] = None
return current_out
def _use_multimask(self, is_init_cond_frame, point_inputs):
"""Whether to use multimask output in the SAM head."""
num_pts = 0 if point_inputs is None else point_inputs["point_labels"].size(1)
multimask_output = (
self.multimask_output_in_sam
and (is_init_cond_frame or self.multimask_output_for_tracking)
and (self.multimask_min_pt_num <= num_pts <= self.multimask_max_pt_num)
)
return multimask_output
def _apply_non_overlapping_constraints(self, pred_masks):
"""
Apply non-overlapping constraints to the object scores in pred_masks. Here we
keep only the highest scoring object at each spatial location in pred_masks.
"""
batch_size = pred_masks.size(0)
if batch_size == 1:
return pred_masks
device = pred_masks.device
# "max_obj_inds": object index of the object with the highest score at each location
max_obj_inds = torch.argmax(pred_masks, dim=0, keepdim=True)
# "batch_obj_inds": object index of each object slice (along dim 0) in `pred_masks`
batch_obj_inds = torch.arange(batch_size, device=device)[:, None, None, None]
keep = max_obj_inds == batch_obj_inds
# suppress overlapping regions' scores below -10.0 so that the foreground regions
# don't overlap (here sigmoid(-10.0)=4.5398e-05)
pred_masks = torch.where(keep, pred_masks, torch.clamp(pred_masks, max=-10.0))
return pred_masks

View File

@ -0,0 +1,149 @@
# 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 copy
import torch
import torch.nn as nn
import torch.nn.functional as F
def select_closest_cond_frames(frame_idx, cond_frame_outputs, max_cond_frame_num):
"""
Select up to `max_cond_frame_num` conditioning frames from `cond_frame_outputs`
that are temporally closest to the current frame at `frame_idx`. Here, we take
- a) the closest conditioning frame before `frame_idx` (if any);
- b) the closest conditioning frame after `frame_idx` (if any);
- c) any other temporally closest conditioning frames until reaching a total
of `max_cond_frame_num` conditioning frames.
Outputs:
- selected_outputs: selected items (keys & values) from `cond_frame_outputs`.
- unselected_outputs: items (keys & values) not selected in `cond_frame_outputs`.
"""
if max_cond_frame_num == -1 or len(cond_frame_outputs) <= max_cond_frame_num:
selected_outputs = cond_frame_outputs
unselected_outputs = {}
else:
assert max_cond_frame_num >= 2, "we should allow using 2+ conditioning frames"
selected_outputs = {}
# the closest conditioning frame before `frame_idx` (if any)
idx_before = max((t for t in cond_frame_outputs if t < frame_idx), default=None)
if idx_before is not None:
selected_outputs[idx_before] = cond_frame_outputs[idx_before]
# the closest conditioning frame after `frame_idx` (if any)
idx_after = min((t for t in cond_frame_outputs if t >= frame_idx), default=None)
if idx_after is not None:
selected_outputs[idx_after] = cond_frame_outputs[idx_after]
# add other temporally closest conditioning frames until reaching a total
# of `max_cond_frame_num` conditioning frames.
num_remain = max_cond_frame_num - len(selected_outputs)
inds_remain = sorted(
(t for t in cond_frame_outputs if t not in selected_outputs),
key=lambda x: abs(x - frame_idx),
)[:num_remain]
selected_outputs.update((t, cond_frame_outputs[t]) for t in inds_remain)
unselected_outputs = {
t: v for t, v in cond_frame_outputs.items() if t not in selected_outputs
}
return selected_outputs, unselected_outputs
def get_1d_sine_pe(pos_inds, dim, temperature=10000):
"""
Get 1D sine positional embedding as in the original Transformer paper.
"""
pe_dim = dim // 2
dim_t = torch.arange(pe_dim, dtype=torch.float32, device=pos_inds.device)
dim_t = temperature ** (2 * (dim_t // 2) / pe_dim)
pos_embed = pos_inds.unsqueeze(-1) / dim_t
pos_embed = torch.cat([pos_embed.sin(), pos_embed.cos()], dim=-1)
return pos_embed
def get_activation_fn(activation):
"""Return an activation function given a string"""
if activation == "relu":
return F.relu
if activation == "gelu":
return F.gelu
if activation == "glu":
return F.glu
raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
def get_clones(module, N):
return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
class DropPath(nn.Module):
# adapted from https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/drop.py
def __init__(self, drop_prob=0.0, scale_by_keep=True):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
self.scale_by_keep = scale_by_keep
def forward(self, x):
if self.drop_prob == 0.0 or not self.training:
return x
keep_prob = 1 - self.drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
if keep_prob > 0.0 and self.scale_by_keep:
random_tensor.div_(keep_prob)
return x * random_tensor
# 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,
activation: nn.Module = nn.ReLU,
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
self.act = activation()
def forward(self, x):
for i, layer in enumerate(self.layers):
x = self.act(layer(x)) if i < self.num_layers - 1 else layer(x)
if self.sigmoid_output:
x = F.sigmoid(x)
return 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

View File

@ -0,0 +1,445 @@
# 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 logging
from typing import List, Optional, Tuple, Union
import numpy as np
import torch
from PIL.Image import Image
from .modeling.sam2_base import SAM2Base
from .utils.transforms import SAM2Transforms
class SAM2ImagePredictor:
def __init__(
self,
sam_model: SAM2Base,
mask_threshold=0.0,
max_hole_area=0.0,
max_sprinkle_area=0.0,
) -> None:
"""
Uses SAM-2 to calculate the image embedding for an image, and then
allow repeated, efficient mask prediction given prompts.
Arguments:
sam_model (Sam-2): The model to use for mask prediction.
mask_threshold (float): The threshold to use when converting mask logits
to binary masks. Masks are thresholded at 0 by default.
fill_hole_area (int): If fill_hole_area > 0, we fill small holes in up to
the maximum area of fill_hole_area in low_res_masks.
"""
super().__init__()
self.model = sam_model
self._transforms = SAM2Transforms(
resolution=self.model.image_size,
mask_threshold=mask_threshold,
max_hole_area=max_hole_area,
max_sprinkle_area=max_sprinkle_area,
)
# Predictor state
self._is_image_set = False
self._features = None
self._orig_hw = None
# Whether the predictor is set for single image or a batch of images
self._is_batch = False
# Predictor config
self.mask_threshold = mask_threshold
# Spatial dim for backbone feature maps
self._bb_feat_sizes = [
(256, 256),
(128, 128),
(64, 64),
]
@torch.no_grad()
def set_image(
self,
image: Union[np.ndarray, Image],
) -> None:
"""
Calculates the image embeddings for the provided image, allowing
masks to be predicted with the 'predict' method.
Arguments:
image (np.ndarray or PIL Image): The input image to embed in RGB format. The image should be in HWC format if np.ndarray, or WHC format if PIL Image
with pixel values in [0, 255].
image_format (str): The color format of the image, in ['RGB', 'BGR'].
"""
self.reset_predictor()
# Transform the image to the form expected by the model
if isinstance(image, np.ndarray):
logging.info("For numpy array image, we assume (HxWxC) format")
self._orig_hw = [image.shape[:2]]
elif isinstance(image, Image):
w, h = image.size
self._orig_hw = [(h, w)]
else:
raise NotImplementedError("Image format not supported")
input_image = self._transforms(image)
input_image = input_image[None, ...].to(self.device)
assert (
len(input_image.shape) == 4 and input_image.shape[1] == 3
), f"input_image must be of size 1x3xHxW, got {input_image.shape}"
logging.info("Computing image embeddings for the provided image...")
backbone_out = self.model.forward_image(input_image)
_, vision_feats, _, _ = self.model._prepare_backbone_features(backbone_out)
# Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
if self.model.directly_add_no_mem_embed:
vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
feats = [
feat.permute(1, 2, 0).view(1, -1, *feat_size)
for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
][::-1]
self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
self._is_image_set = True
logging.info("Image embeddings computed.")
@torch.no_grad()
def set_image_batch(
self,
image_list: List[Union[np.ndarray]],
) -> None:
"""
Calculates the image embeddings for the provided image batch, allowing
masks to be predicted with the 'predict_batch' method.
Arguments:
image_list (List[np.ndarray]): The input images to embed in RGB format. The image should be in HWC format if np.ndarray
with pixel values in [0, 255].
"""
self.reset_predictor()
assert isinstance(image_list, list)
self._orig_hw = []
for image in image_list:
assert isinstance(
image, np.ndarray
), "Images are expected to be an np.ndarray in RGB format, and of shape HWC"
self._orig_hw.append(image.shape[:2])
# Transform the image to the form expected by the model
img_batch = self._transforms.forward_batch(image_list)
img_batch = img_batch.to(self.device)
batch_size = img_batch.shape[0]
assert (
len(img_batch.shape) == 4 and img_batch.shape[1] == 3
), f"img_batch must be of size Bx3xHxW, got {img_batch.shape}"
logging.info("Computing image embeddings for the provided images...")
backbone_out = self.model.forward_image(img_batch)
_, vision_feats, _, _ = self.model._prepare_backbone_features(backbone_out)
# Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
if self.model.directly_add_no_mem_embed:
vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
feats = [
feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
][::-1]
self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
self._is_image_set = True
self._is_batch = True
logging.info("Image embeddings computed.")
def predict_batch(
self,
point_coords_batch: List[np.ndarray] = None,
point_labels_batch: List[np.ndarray] = None,
box_batch: List[np.ndarray] = None,
mask_input_batch: List[np.ndarray] = None,
multimask_output: bool = True,
return_logits: bool = False,
normalize_coords=True,
) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
"""This function is very similar to predict(...), however it is used for batched mode, when the model is expected to generate predictions on multiple images.
It returns a tupele of lists of masks, ious, and low_res_masks_logits.
"""
assert self._is_batch, "This function should only be used when in batched mode"
if not self._is_image_set:
raise RuntimeError(
"An image must be set with .set_image_batch(...) before mask prediction."
)
num_images = len(self._features["image_embed"])
all_masks = []
all_ious = []
all_low_res_masks = []
for img_idx in range(num_images):
# Transform input prompts
point_coords = (
point_coords_batch[img_idx] if point_coords_batch is not None else None
)
point_labels = (
point_labels_batch[img_idx] if point_labels_batch is not None else None
)
box = box_batch[img_idx] if box_batch is not None else None
mask_input = (
mask_input_batch[img_idx] if mask_input_batch is not None else None
)
mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
point_coords,
point_labels,
box,
mask_input,
normalize_coords,
img_idx=img_idx,
)
masks, iou_predictions, low_res_masks = self._predict(
unnorm_coords,
labels,
unnorm_box,
mask_input,
multimask_output,
return_logits=return_logits,
img_idx=img_idx,
)
masks_np = masks.squeeze(0).float().detach().cpu().numpy()
iou_predictions_np = (
iou_predictions.squeeze(0).float().detach().cpu().numpy()
)
low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
all_masks.append(masks_np)
all_ious.append(iou_predictions_np)
all_low_res_masks.append(low_res_masks_np)
return all_masks, all_ious, all_low_res_masks
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,
normalize_coords=True,
) -> 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.
normalize_coords (bool): If true, the point coordinates will be normalized to the range [0,1] and point_coords is expected to be wrt. image dimensions.
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
mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
point_coords, point_labels, box, mask_input, normalize_coords
)
masks, iou_predictions, low_res_masks = self._predict(
unnorm_coords,
labels,
unnorm_box,
mask_input,
multimask_output,
return_logits=return_logits,
)
masks_np = masks.squeeze(0).float().detach().cpu().numpy()
iou_predictions_np = iou_predictions.squeeze(0).float().detach().cpu().numpy()
low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
return masks_np, iou_predictions_np, low_res_masks_np
def _prep_prompts(
self, point_coords, point_labels, box, mask_logits, normalize_coords, img_idx=-1
):
unnorm_coords, labels, unnorm_box, mask_input = 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 = torch.as_tensor(
point_coords, dtype=torch.float, device=self.device
)
unnorm_coords = self._transforms.transform_coords(
point_coords, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
)
labels = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
if len(unnorm_coords.shape) == 2:
unnorm_coords, labels = unnorm_coords[None, ...], labels[None, ...]
if box is not None:
box = torch.as_tensor(box, dtype=torch.float, device=self.device)
unnorm_box = self._transforms.transform_boxes(
box, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
) # Bx2x2
if mask_logits is not None:
mask_input = torch.as_tensor(
mask_logits, dtype=torch.float, device=self.device
)
if len(mask_input.shape) == 3:
mask_input = mask_input[None, :, :, :]
return mask_input, unnorm_coords, labels, unnorm_box
@torch.no_grad()
def _predict(
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,
img_idx: int = -1,
) -> 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 SAM2Transforms.
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:
concat_points = (point_coords, point_labels)
else:
concat_points = None
# Embed prompts
if boxes is not None:
box_coords = boxes.reshape(-1, 2, 2)
box_labels = torch.tensor([[2, 3]], dtype=torch.int, device=boxes.device)
box_labels = box_labels.repeat(boxes.size(0), 1)
# we merge "boxes" and "points" into a single "concat_points" input (where
# boxes are added at the beginning) to sam_prompt_encoder
if concat_points is not None:
concat_coords = torch.cat([box_coords, concat_points[0]], dim=1)
concat_labels = torch.cat([box_labels, concat_points[1]], dim=1)
concat_points = (concat_coords, concat_labels)
else:
concat_points = (box_coords, box_labels)
sparse_embeddings, dense_embeddings = self.model.sam_prompt_encoder(
points=concat_points,
boxes=None,
masks=mask_input,
)
# Predict masks
batched_mode = (
concat_points is not None and concat_points[0].shape[0] > 1
) # multi object prediction
high_res_features = [
feat_level[img_idx].unsqueeze(0)
for feat_level in self._features["high_res_feats"]
]
low_res_masks, iou_predictions, _, _ = self.model.sam_mask_decoder(
image_embeddings=self._features["image_embed"][img_idx].unsqueeze(0),
image_pe=self.model.sam_prompt_encoder.get_dense_pe(),
sparse_prompt_embeddings=sparse_embeddings,
dense_prompt_embeddings=dense_embeddings,
multimask_output=multimask_output,
repeat_image=batched_mode,
high_res_features=high_res_features,
)
# Upscale the masks to the original image resolution
masks = self._transforms.postprocess_masks(
low_res_masks, self._orig_hw[img_idx]
)
low_res_masks = torch.clamp(low_res_masks, -32.0, 32.0)
if not return_logits:
masks = masks > self.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["image_embed"]
@property
def device(self) -> torch.device:
return self.model.device
def reset_predictor(self) -> None:
"""
Resets the image embeddings and other state variables.
"""
self._is_image_set = False
self._features = None
self._orig_hw = None
self._is_batch = False

View File

@ -0,0 +1,5 @@
# 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.

View File

@ -0,0 +1,90 @@
# 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 warnings
import numpy as np
import torch
from PIL import Image
def get_sdpa_settings():
if torch.cuda.is_available():
old_gpu = torch.cuda.get_device_properties(0).major < 7
# only use Flash Attention on Ampere (8.0) or newer GPUs
use_flash_attn = torch.cuda.get_device_properties(0).major >= 8
if not use_flash_attn:
warnings.warn(
"Flash Attention is disabled as it requires a GPU with Ampere (8.0) CUDA capability.",
category=UserWarning,
stacklevel=2,
)
# keep math kernel for PyTorch versions before 2.2 (Flash Attention v2 is only
# available on PyTorch 2.2+, while Flash Attention v1 cannot handle all cases)
pytorch_version = tuple(int(v) for v in torch.__version__.split(".")[:2])
if pytorch_version < (2, 2):
warnings.warn(
f"You are using PyTorch {torch.__version__} without Flash Attention v2 support. "
"Consider upgrading to PyTorch 2.2+ for Flash Attention v2 (which could be faster).",
category=UserWarning,
stacklevel=2,
)
math_kernel_on = pytorch_version < (2, 2) or not use_flash_attn
else:
old_gpu = True
use_flash_attn = False
math_kernel_on = True
return old_gpu, use_flash_attn, math_kernel_on
def mask_to_box(masks: torch.Tensor):
"""
compute bounding box given an input mask
Inputs:
- masks: [B, 1, H, W] boxes, dtype=torch.Tensor
Returns:
- box_coords: [B, 1, 4], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.Tensor
"""
B, _, h, w = masks.shape
device = masks.device
xs = torch.arange(w, device=device, dtype=torch.int32)
ys = torch.arange(h, device=device, dtype=torch.int32)
grid_xs, grid_ys = torch.meshgrid(xs, ys, indexing="xy")
grid_xs = grid_xs[None, None, ...].expand(B, 1, h, w)
grid_ys = grid_ys[None, None, ...].expand(B, 1, h, w)
min_xs, _ = torch.min(torch.where(masks, grid_xs, w).flatten(-2), dim=-1)
max_xs, _ = torch.max(torch.where(masks, grid_xs, -1).flatten(-2), dim=-1)
min_ys, _ = torch.min(torch.where(masks, grid_ys, h).flatten(-2), dim=-1)
max_ys, _ = torch.max(torch.where(masks, grid_ys, -1).flatten(-2), dim=-1)
bbox_coords = torch.stack((min_xs, min_ys, max_xs, max_ys), dim=-1)
return bbox_coords
def _load_img_as_tensor(img_path, image_size):
img_pil = Image.open(img_path)
img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
if img_np.dtype == np.uint8: # np.uint8 is expected for JPEG images
img_np = img_np / 255.0
else:
raise RuntimeError(f"Unknown image dtype: {img_np.dtype} on {img_path}")
img = torch.from_numpy(img_np).permute(2, 0, 1)
video_width, video_height = img_pil.size # the original video size
return img, video_height, video_width
def concat_points(old_point_inputs, new_points, new_labels):
"""Add new points and labels to previous point inputs (add at the end)."""
if old_point_inputs is None:
points, labels = new_points, new_labels
else:
points = torch.cat([old_point_inputs["point_coords"], new_points], dim=1)
labels = torch.cat([old_point_inputs["point_labels"], new_labels], dim=1)
return {"point_coords": points, "point_labels": labels}

View File

@ -0,0 +1,77 @@
# 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 torchvision.transforms import Normalize, Resize, ToTensor
class SAM2Transforms(nn.Module):
def __init__(
self, resolution, mask_threshold, max_hole_area=0.0, max_sprinkle_area=0.0
):
"""
Transforms for SAM2.
"""
super().__init__()
self.resolution = resolution
self.mask_threshold = mask_threshold
self.max_hole_area = max_hole_area
self.max_sprinkle_area = max_sprinkle_area
self.mean = [0.485, 0.456, 0.406]
self.std = [0.229, 0.224, 0.225]
self.to_tensor = ToTensor()
self.transforms = torch.jit.script(
nn.Sequential(
Resize((self.resolution, self.resolution)),
Normalize(self.mean, self.std),
)
)
def __call__(self, x):
x = self.to_tensor(x)
return self.transforms(x)
def forward_batch(self, img_list):
img_batch = [self.transforms(self.to_tensor(img)) for img in img_list]
img_batch = torch.stack(img_batch, dim=0)
return img_batch
def transform_coords(
self, coords: torch.Tensor, normalize=False, orig_hw=None
) -> torch.Tensor:
"""
Expects a torch tensor with length 2 in the last dimension. The coordinates can be in absolute image or normalized coordinates,
If the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
Returns
Un-normalized coordinates in the range of [0, 1] which is expected by the SAM2 model.
"""
if normalize:
assert orig_hw is not None
h, w = orig_hw
coords = coords.clone()
coords[..., 0] = coords[..., 0] / w
coords[..., 1] = coords[..., 1] / h
coords = coords * self.resolution # unnormalize coords
return coords
def transform_boxes(
self, boxes: torch.Tensor, normalize=False, orig_hw=None
) -> torch.Tensor:
"""
Expects a tensor of shape Bx4. The coordinates can be in absolute image or normalized coordinates,
if the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
"""
boxes = self.transform_coords(boxes.reshape(-1, 2, 2), normalize, orig_hw)
return boxes
def postprocess_masks(self, masks: torch.Tensor, orig_hw) -> torch.Tensor:
"""
Perform PostProcessing on output masks.
"""
return masks

View File

@ -167,6 +167,10 @@ class InteractiveSegModel(Choices):
sam_hq_vit_l = "sam_hq_vit_l" sam_hq_vit_l = "sam_hq_vit_l"
sam_hq_vit_h = "sam_hq_vit_h" sam_hq_vit_h = "sam_hq_vit_h"
mobile_sam = "mobile_sam" mobile_sam = "mobile_sam"
sam2_tiny = "sam2_tiny"
sam2_small = "sam2_small"
sam2_base = "sam2_base"
sam2_large = "sam2_large"
class PluginInfo(BaseModel): class PluginInfo(BaseModel):

View File

@ -1,6 +1,4 @@
import hashlib
import os import os
import time
from PIL import Image from PIL import Image
from iopaint.helper import encode_pil_to_base64, gen_frontend_mask from iopaint.helper import encode_pil_to_base64, gen_frontend_mask