File size: 13,132 Bytes
7cc9dda | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 | # Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
from typing import Dict, List
import numpy as np
import PIL
import torch
from .box_ops import box_cxcywh_to_xyxy
from .data_misc import FindStage, interpolate
from torchvision.transforms import v2
class Sam3Processor:
""" """
def __init__(self, model, resolution=1008, device="cuda", confidence_threshold=0.5):
self.model = model
self.resolution = resolution
self.device = device
self.transform = v2.Compose(
[
v2.ToDtype(torch.uint8, scale=True),
v2.Resize(size=(resolution, resolution)),
v2.ToDtype(torch.float32, scale=True),
v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
]
)
self.confidence_threshold = confidence_threshold
self.find_stage = FindStage(
img_ids=torch.tensor([0], device=device, dtype=torch.long),
text_ids=torch.tensor([0], device=device, dtype=torch.long),
input_boxes=None,
input_boxes_mask=None,
input_boxes_label=None,
input_points=None,
input_points_mask=None,
)
@torch.inference_mode()
def set_image(self, image, state=None):
"""Sets the image on which we want to do predictions."""
if state is None:
state = {}
if isinstance(image, PIL.Image.Image):
width, height = image.size
elif isinstance(image, (torch.Tensor, np.ndarray)):
height, width = image.shape[-2:]
else:
raise ValueError("Image must be a PIL image or a tensor")
image = v2.functional.to_image(image).to(self.device)
image = self.transform(image).unsqueeze(0)
state["original_height"] = height
state["original_width"] = width
state["backbone_out"] = self.model.backbone.forward_image(image)
inst_interactivity_en = self.model.inst_interactive_predictor is not None
if inst_interactivity_en and "sam2_backbone_out" in state["backbone_out"]:
sam2_backbone_out = state["backbone_out"]["sam2_backbone_out"]
sam2_backbone_out["backbone_fpn"][0] = (
self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s0(
sam2_backbone_out["backbone_fpn"][0]
)
)
sam2_backbone_out["backbone_fpn"][1] = (
self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s1(
sam2_backbone_out["backbone_fpn"][1]
)
)
return state
@torch.inference_mode()
def set_image_batch(self, images: List[np.ndarray], state=None):
"""Sets the image batch on which we want to do predictions."""
if state is None:
state = {}
if not isinstance(images, list):
raise ValueError("Images must be a list of PIL images or tensors")
assert len(images) > 0, "Images list must not be empty"
assert isinstance(
images[0], PIL.Image.Image
), "Images must be a list of PIL images"
state["original_heights"] = [image.height for image in images]
state["original_widths"] = [image.width for image in images]
images = [
self.transform(v2.functional.to_image(image).to(self.device))
for image in images
]
images = torch.stack(images, dim=0)
state["backbone_out"] = self.model.backbone.forward_image(images)
inst_interactivity_en = self.model.inst_interactive_predictor is not None
if inst_interactivity_en and "sam2_backbone_out" in state["backbone_out"]:
sam2_backbone_out = state["backbone_out"]["sam2_backbone_out"]
sam2_backbone_out["backbone_fpn"][0] = (
self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s0(
sam2_backbone_out["backbone_fpn"][0]
)
)
sam2_backbone_out["backbone_fpn"][1] = (
self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s1(
sam2_backbone_out["backbone_fpn"][1]
)
)
return state
@torch.inference_mode()
def set_text_prompt(self, prompt: str, state: Dict):
"""Sets the text prompt and run the inference"""
if "backbone_out" not in state:
raise ValueError("You must call set_image before set_text_prompt")
text_outputs = self.model.backbone.forward_text([prompt], device=self.device)
# will erase the previous text prompt if any
state["backbone_out"].update(text_outputs)
if "geometric_prompt" not in state:
state["geometric_prompt"] = self.model._get_dummy_prompt()
return self._forward_grounding(state)
@torch.inference_mode()
def add_geometric_prompt(self, box: List, label: bool, state: Dict):
"""Adds a box prompt and run the inference.
The image needs to be set, but not necessarily the text prompt.
The box is assumed to be in [center_x, center_y, width, height] format and normalized in [0, 1] range.
The label is True for a positive box, False for a negative box.
"""
if "backbone_out" not in state:
raise ValueError("You must call set_image before add_geometric_prompt")
if "language_features" not in state["backbone_out"]:
# Looks like we don't have a text prompt yet. This is allowed, but we need to set the text prompt to "visual" for the model to rely only on the geometric prompt
dummy_text_outputs = self.model.backbone.forward_text(
["visual"], device=self.device
)
state["backbone_out"].update(dummy_text_outputs)
if "geometric_prompt" not in state:
state["geometric_prompt"] = self.model._get_dummy_prompt()
# adding a batch and sequence dimension
boxes = torch.tensor(box, device=self.device, dtype=torch.float32).view(1, 1, 4)
labels = torch.tensor([label], device=self.device, dtype=torch.bool).view(1, 1)
state["geometric_prompt"].append_boxes(boxes, labels)
return self._forward_grounding(state)
@torch.inference_mode()
def add_boxes_prompts(self, boxes: List[List], labels: List[bool], state: Dict):
"""Adds multiple box prompts and run the inference.
The image needs to be set, but not necessarily the text prompt.
Each box is assumed to be in [center_x, center_y, width, height] format and normalized in [0, 1] range.
Each label is True for a positive box, False for a negative box.
"""
if "backbone_out" not in state:
raise ValueError("You must call set_image before add_boxes_prompt")
if "language_features" not in state["backbone_out"]:
dummy_text_outputs = self.model.backbone.forward_text(
["visual"], device=self.device
)
state["backbone_out"].update(dummy_text_outputs)
if "geometric_prompt" not in state:
state["geometric_prompt"] = self.model._get_dummy_prompt()
# Convert to [seq_len, batch_size, 4] format
boxes_tensor = torch.tensor(boxes, device=self.device, dtype=torch.float32).view(len(boxes), 1, 4)
labels_tensor = torch.tensor(labels, device=self.device, dtype=torch.bool).view(len(labels), 1)
state["geometric_prompt"].append_boxes(boxes_tensor, labels_tensor)
return self._forward_grounding(state)
@torch.inference_mode()
def add_point_prompt(self, points: List[List], labels: List[int], state: Dict):
"""Adds point prompts and run the inference.
The image needs to be set, but not necessarily the text prompt.
Points should be in [x, y] format, normalized in [0, 1] range.
Labels should be 1 for foreground points, 0 for background points.
"""
if "backbone_out" not in state:
raise ValueError("You must call set_image before add_point_prompt")
if "language_features" not in state["backbone_out"]:
dummy_text_outputs = self.model.backbone.forward_text(
["visual"], device=self.device
)
state["backbone_out"].update(dummy_text_outputs)
if "geometric_prompt" not in state:
state["geometric_prompt"] = self.model._get_dummy_prompt()
# Convert to [seq_len, batch_size, 2] format
points_tensor = torch.tensor(points, device=self.device, dtype=torch.float32).view(len(points), 1, 2)
labels_tensor = torch.tensor(labels, device=self.device, dtype=torch.long).view(len(labels), 1)
state["geometric_prompt"].append_points(points_tensor, labels_tensor)
return self._forward_grounding(state)
@torch.inference_mode()
def add_mask_prompt(self, mask: torch.Tensor, state: Dict):
"""Adds a mask prompt and run the inference.
The mask should be a binary tensor with shape matching the model's expected input.
This is typically used for iterative refinement.
"""
if "backbone_out" not in state:
raise ValueError("You must call set_image before add_mask_prompt")
if "language_features" not in state["backbone_out"]:
dummy_text_outputs = self.model.backbone.forward_text(
["visual"], device=self.device
)
state["backbone_out"].update(dummy_text_outputs)
if "geometric_prompt" not in state:
state["geometric_prompt"] = self.model._get_dummy_prompt()
# Ensure mask is on correct device and has batch dimension
if mask.device != self.device:
mask = mask.to(self.device)
# Add sequence and batch dimensions if needed: [seq_len, batch_size, H, W]
if len(mask.shape) == 2: # [H, W]
mask = mask.unsqueeze(0).unsqueeze(0)
elif len(mask.shape) == 3: # [1, H, W] or [batch, H, W]
mask = mask.unsqueeze(0)
state["geometric_prompt"].append_masks(mask)
return self._forward_grounding(state)
def reset_all_prompts(self, state: Dict):
"""Removes all the prompts and results"""
if "backbone_out" in state:
backbone_keys_to_del = [
"language_features",
"language_mask",
"language_embeds",
]
for key in backbone_keys_to_del:
if key in state["backbone_out"]:
del state["backbone_out"][key]
keys_to_del = ["geometric_prompt", "boxes", "masks", "masks_logits", "scores"]
for key in keys_to_del:
if key in state:
del state[key]
@torch.inference_mode()
def set_confidence_threshold(self, threshold: float, state=None):
"""Sets the confidence threshold for the masks"""
self.confidence_threshold = threshold
if state is not None and "boxes" in state:
# we need to filter the boxes again
# In principle we could do this more efficiently since we would only need
# to rerun the heads. But this is simpler and not too inefficient
return self._forward_grounding(state)
return state
@torch.inference_mode()
def _forward_grounding(self, state: Dict):
outputs = self.model.forward_grounding(
backbone_out=state["backbone_out"],
find_input=self.find_stage,
geometric_prompt=state["geometric_prompt"],
find_target=None,
)
out_bbox = outputs["pred_boxes"]
out_logits = outputs["pred_logits"]
out_masks = outputs["pred_masks"]
out_probs = out_logits.sigmoid()
presence_score = outputs["presence_logit_dec"].sigmoid().unsqueeze(1)
out_probs = (out_probs * presence_score).squeeze(-1)
keep = out_probs > self.confidence_threshold
out_probs = out_probs[keep]
out_masks = out_masks[keep]
out_bbox = out_bbox[keep]
# convert to [x0, y0, x1, y1] format
boxes = box_cxcywh_to_xyxy(out_bbox)
img_h = state["original_height"]
img_w = state["original_width"]
scale_fct = torch.tensor([img_w, img_h, img_w, img_h]).to(self.device)
boxes = boxes * scale_fct[None, :]
out_masks = interpolate(
out_masks.unsqueeze(1),
(img_h, img_w),
mode="bilinear",
align_corners=False,
).sigmoid()
state["masks_logits"] = out_masks
state["masks"] = out_masks > 0.5
state["boxes"] = boxes
state["scores"] = out_probs
return state
|