[ { "path": ".gitignore", "content": "*.pth\n" }, { "path": "README.md", "content": "# LoRA for SAM (meta's segment-anything)\n\n## Usage\n```\nfrom segment_anything import build_sam, SamAutomaticMaskGenerator \nfrom segment_anything import sam_model_registry\nfrom sam_lora import LoRA_Sam\nimport torch\nsam = sam_model_registry[\"vit_b\"](checkpoint=\"sam_vit_b_01ec64.pth\")\nlora_sam = LoRA_Sam(sam,r = 4)\nresult = lora_sam.sam.image_encoder(torch.rand(size=(1,3,1024,1024)))\nprint(result.shape)\n```\n\n## Train\nComing soon and welcome pull request.\n\n## Thanks\nThe code for LoRA ViT comes form\nhttps://github.com/JamesQFreeman/LoRA-ViT" }, { "path": "demo.ipynb", "content": "{\n \"cells\": [\n {\n \"cell_type\": \"code\",\n \"execution_count\": 1,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stderr\",\n \"output_type\": \"stream\",\n \"text\": [\n \"/Users/wangsheng/miniforge3/envs/torch/lib/python3.9/site-packages/tqdm/auto.py:22: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\\n\",\n \" from .autonotebook import tqdm as notebook_tqdm\\n\"\n ]\n }\n ],\n \"source\": [\n \"from segment_anything import build_sam, SamAutomaticMaskGenerator \\n\",\n \"from segment_anything import sam_model_registry\\n\",\n \"from sam_lora import LoRA_Sam\\n\",\n \"import torch\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 3,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"torch.Size([1, 256, 64, 64])\\n\"\n ]\n }\n ],\n \"source\": [\n \"sam = sam_model_registry[\\\"vit_b\\\"](checkpoint=\\\"sam_vit_b_01ec64.pth\\\")\\n\",\n \"lora_sam = LoRA_Sam(sam,r = 4)\\n\",\n \"result = lora_sam.sam.image_encoder(torch.rand(size=(1,3,1024,1024)))\\n\",\n \"print(result.shape)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n }\n ],\n \"metadata\": {\n \"kernelspec\": {\n \"display_name\": \"Python 3\",\n \"language\": \"python\",\n \"name\": \"python3\"\n },\n \"language_info\": {\n \"codemirror_mode\": {\n \"name\": \"ipython\",\n \"version\": 3\n },\n \"file_extension\": \".py\",\n \"mimetype\": \"text/x-python\",\n \"name\": \"python\",\n \"nbconvert_exporter\": \"python\",\n \"pygments_lexer\": \"ipython3\",\n \"version\": \"3.9.2\"\n },\n \"orig_nbformat\": 4\n },\n \"nbformat\": 4,\n \"nbformat_minor\": 2\n}\n" }, { "path": "sam_lora.py", "content": "# Sheng Wang at Apr 6 2023\n# What a time to be alive (first half of 2023)\n\nfrom segment_anything import build_sam, SamPredictor\nfrom segment_anything import sam_model_registry\n\nimport math\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch import Tensor\nfrom torch.nn.parameter import Parameter\nfrom segment_anything.modeling import Sam\nfrom safetensors import safe_open\nfrom safetensors.torch import save_file\n\n\nclass _LoRA_qkv(nn.Module):\n \"\"\"In Sam it is implemented as\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n B, N, C = x.shape\n qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)\n q, k, v = qkv.unbind(0)\n \"\"\"\n\n def __init__(\n self,\n qkv: nn.Module,\n linear_a_q: nn.Module,\n linear_b_q: nn.Module,\n linear_a_v: nn.Module,\n linear_b_v: nn.Module,\n ):\n super().__init__()\n self.qkv = qkv\n self.linear_a_q = linear_a_q\n self.linear_b_q = linear_b_q\n self.linear_a_v = linear_a_v\n self.linear_b_v = linear_b_v\n self.dim = qkv.in_features\n self.w_identity = torch.eye(qkv.in_features)\n\n def forward(self, x):\n qkv = self.qkv(x) # B,N,N,3*org_C\n new_q = self.linear_b_q(self.linear_a_q(x))\n new_v = self.linear_b_v(self.linear_a_v(x))\n qkv[:, :, :, : self.dim] += new_q\n qkv[:, :, :, -self.dim :] += new_v\n return qkv\n\nclass LoRA_Sam(nn.Module):\n \"\"\"Applies low-rank adaptation to a Sam model's image encoder.\n\n Args:\n sam_model: a vision transformer model, see base_vit.py\n r: rank of LoRA\n num_classes: how many classes the model output, default to the vit model\n lora_layer: which layer we apply LoRA.\n\n Examples::\n >>> model = ViT('B_16_imagenet1k')\n >>> lora_model = LoRA_ViT(model, r=4)\n >>> preds = lora_model(img)\n >>> print(preds.shape)\n torch.Size([1, 1000])\n \"\"\"\n\n def __init__(self, sam_model: Sam, r: int, lora_layer=None):\n super(LoRA_Sam, self).__init__()\n\n assert r > 0\n # base_vit_dim = sam_model.image_encoder.patch_embed.proj.out_channels\n # dim = base_vit_dim\n if lora_layer:\n self.lora_layer = lora_layer\n else:\n self.lora_layer = list(range(len(sam_model.image_encoder.blocks)))\n # create for storage, then we can init them or load weights\n self.w_As = [] # These are linear layers\n self.w_Bs = []\n\n # lets freeze first\n for param in sam_model.image_encoder.parameters():\n param.requires_grad = False\n\n # Here, we do the surgery\n for t_layer_i, blk in enumerate(sam_model.image_encoder.blocks):\n # If we only want few lora layer instead of all\n if t_layer_i not in self.lora_layer:\n continue\n w_qkv_linear = blk.attn.qkv\n self.dim = w_qkv_linear.in_features\n w_a_linear_q = nn.Linear(self.dim, r, bias=False)\n w_b_linear_q = nn.Linear(r, self.dim, bias=False)\n w_a_linear_v = nn.Linear(self.dim, r, bias=False)\n w_b_linear_v = nn.Linear(r, self.dim, bias=False)\n self.w_As.append(w_a_linear_q)\n self.w_Bs.append(w_b_linear_q)\n self.w_As.append(w_a_linear_v)\n self.w_Bs.append(w_b_linear_v)\n blk.attn.qkv = _LoRA_qkv(\n w_qkv_linear,\n w_a_linear_q,\n w_b_linear_q,\n w_a_linear_v,\n w_b_linear_v,\n )\n self.reset_parameters()\n self.sam = sam_model\n\n def load_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n with safe_open(filename, framework=\"pt\") as f:\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.head.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def save_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \n save both lora and fc parameters.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n num_layer = len(self.w_As) # actually, it is half\n a_tensors = {f\"w_a_{i:03d}\": self.w_As[i].weight for i in range(num_layer)}\n b_tensors = {f\"w_b_{i:03d}\": self.w_Bs[i].weight for i in range(num_layer)}\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.head.weight}\n \n merged_dict = {**a_tensors, **b_tensors, **fc_tensors}\n save_file(merged_dict, filename)\n\n def load_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\\\n \n load both lora and fc parameters.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n with safe_open(filename, framework=\"pt\") as f:\n for i, w_A_linear in enumerate(self.w_As):\n saved_key = f\"w_a_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_A_linear.weight = Parameter(saved_tensor)\n\n for i, w_B_linear in enumerate(self.w_Bs):\n saved_key = f\"w_b_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_B_linear.weight = Parameter(saved_tensor)\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.head.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def reset_parameters(self) -> None:\n for w_A in self.w_As:\n nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n for w_B in self.w_Bs:\n nn.init.zeros_(w_B.weight)\n\n # def forward(self, x: Tensor) -> Tensor:\n # return self.lora_vit(x)\n\n\nif __name__ == \"__main__\":\n sam = sam_model_registry[\"vit_b\"](checkpoint=\"sam_vit_b_01ec64.pth\")\n lora_sam = LoRA_Sam(sam,4)\n lora_sam.sam.image_encoder(torch.rand(size=(1,3,1024,1024)))" }, { "path": "segment_anything/__init__.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nfrom .build_sam import (\n build_sam,\n build_sam_vit_h,\n build_sam_vit_l,\n build_sam_vit_b,\n sam_model_registry,\n)\nfrom .predictor import SamPredictor\nfrom .automatic_mask_generator import SamAutomaticMaskGenerator\n" }, { "path": "segment_anything/automatic_mask_generator.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport numpy as np\nimport torch\nfrom torchvision.ops.boxes import batched_nms, box_area # type: ignore\n\nfrom typing import Any, Dict, List, Optional, Tuple\n\nfrom .modeling import Sam\nfrom .predictor import SamPredictor\nfrom .utils.amg import (\n MaskData,\n area_from_rle,\n batch_iterator,\n batched_mask_to_box,\n box_xyxy_to_xywh,\n build_all_layer_point_grids,\n calculate_stability_score,\n coco_encode_rle,\n generate_crop_boxes,\n is_box_near_crop_edge,\n mask_to_rle_pytorch,\n remove_small_regions,\n rle_to_mask,\n uncrop_boxes_xyxy,\n uncrop_masks,\n uncrop_points,\n)\n\n\nclass SamAutomaticMaskGenerator:\n def __init__(\n self,\n model: Sam,\n points_per_side: Optional[int] = 32,\n points_per_batch: int = 64,\n pred_iou_thresh: float = 0.88,\n stability_score_thresh: float = 0.95,\n stability_score_offset: float = 1.0,\n box_nms_thresh: float = 0.7,\n crop_n_layers: int = 0,\n crop_nms_thresh: float = 0.7,\n crop_overlap_ratio: float = 512 / 1500,\n crop_n_points_downscale_factor: int = 1,\n point_grids: Optional[List[np.ndarray]] = None,\n min_mask_region_area: int = 0,\n output_mode: str = \"binary_mask\",\n ) -> None:\n \"\"\"\n Using a SAM model, generates masks for the entire image.\n Generates a grid of point prompts over the image, then filters\n low quality and duplicate masks. The default settings are chosen\n for SAM with a ViT-H backbone.\n\n Arguments:\n model (Sam): The SAM model to use for mask prediction.\n points_per_side (int or None): The number of points to be sampled\n along one side of the image. The total number of points is\n points_per_side**2. If None, 'point_grids' must provide explicit\n point sampling.\n points_per_batch (int): Sets the number of points run simultaneously\n by the model. Higher numbers may be faster but use more GPU memory.\n pred_iou_thresh (float): A filtering threshold in [0,1], using the\n model's predicted mask quality.\n stability_score_thresh (float): A filtering threshold in [0,1], using\n the stability of the mask under changes to the cutoff used to binarize\n the model's mask predictions.\n stability_score_offset (float): The amount to shift the cutoff when\n calculated the stability score.\n box_nms_thresh (float): The box IoU cutoff used by non-maximal\n suppression to filter duplicate masks.\n crops_n_layers (int): If >0, mask prediction will be run again on\n crops of the image. Sets the number of layers to run, where each\n layer has 2**i_layer number of image crops.\n crops_nms_thresh (float): The box IoU cutoff used by non-maximal\n suppression to filter duplicate masks between different crops.\n crop_overlap_ratio (float): Sets the degree to which crops overlap.\n In the first crop layer, crops will overlap by this fraction of\n the image length. Later layers with more crops scale down this overlap.\n crop_n_points_downscale_factor (int): The number of points-per-side\n sampled in layer n is scaled down by crop_n_points_downscale_factor**n.\n point_grids (list(np.ndarray) or None): A list over explicit grids\n of points used for sampling, normalized to [0,1]. The nth grid in the\n list is used in the nth crop layer. Exclusive with points_per_side.\n min_mask_region_area (int): If >0, postprocessing will be applied\n to remove disconnected regions and holes in masks with area smaller\n than min_mask_region_area. Requires opencv.\n output_mode (str): The form masks are returned in. Can be 'binary_mask',\n 'uncompressed_rle', or 'coco_rle'. 'coco_rle' requires pycocotools.\n For large resolutions, 'binary_mask' may consume large amounts of\n memory.\n \"\"\"\n\n assert (points_per_side is None) != (\n point_grids is None\n ), \"Exactly one of points_per_side or point_grid must be provided.\"\n if points_per_side is not None:\n self.point_grids = build_all_layer_point_grids(\n points_per_side,\n crop_n_layers,\n crop_n_points_downscale_factor,\n )\n elif point_grids is not None:\n self.point_grids = point_grids\n else:\n raise ValueError(\"Can't have both points_per_side and point_grid be None.\")\n\n assert output_mode in [\n \"binary_mask\",\n \"uncompressed_rle\",\n \"coco_rle\",\n ], f\"Unknown output_mode {output_mode}.\"\n if output_mode == \"coco_rle\":\n from pycocotools import mask as mask_utils # type: ignore # noqa: F401\n\n if min_mask_region_area > 0:\n import cv2 # type: ignore # noqa: F401\n\n self.predictor = SamPredictor(model)\n self.points_per_batch = points_per_batch\n self.pred_iou_thresh = pred_iou_thresh\n self.stability_score_thresh = stability_score_thresh\n self.stability_score_offset = stability_score_offset\n self.box_nms_thresh = box_nms_thresh\n self.crop_n_layers = crop_n_layers\n self.crop_nms_thresh = crop_nms_thresh\n self.crop_overlap_ratio = crop_overlap_ratio\n self.crop_n_points_downscale_factor = crop_n_points_downscale_factor\n self.min_mask_region_area = min_mask_region_area\n self.output_mode = output_mode\n\n @torch.no_grad()\n def generate(self, image: np.ndarray) -> List[Dict[str, Any]]:\n \"\"\"\n Generates masks for the given image.\n\n Arguments:\n image (np.ndarray): The image to generate masks for, in HWC uint8 format.\n\n Returns:\n list(dict(str, any)): A list over records for masks. Each record is\n a dict containing the following keys:\n segmentation (dict(str, any) or np.ndarray): The mask. If\n output_mode='binary_mask', is an array of shape HW. Otherwise,\n is a dictionary containing the RLE.\n bbox (list(float)): The box around the mask, in XYWH format.\n area (int): The area in pixels of the mask.\n predicted_iou (float): The model's own prediction of the mask's\n quality. This is filtered by the pred_iou_thresh parameter.\n point_coords (list(list(float))): The point coordinates input\n to the model to generate this mask.\n stability_score (float): A measure of the mask's quality. This\n is filtered on using the stability_score_thresh parameter.\n crop_box (list(float)): The crop of the image used to generate\n the mask, given in XYWH format.\n \"\"\"\n\n # Generate masks\n mask_data = self._generate_masks(image)\n\n # Filter small disconnected regions and holes in masks\n if self.min_mask_region_area > 0:\n mask_data = self.postprocess_small_regions(\n mask_data,\n self.min_mask_region_area,\n max(self.box_nms_thresh, self.crop_nms_thresh),\n )\n\n # Encode masks\n if self.output_mode == \"coco_rle\":\n mask_data[\"segmentations\"] = [coco_encode_rle(rle) for rle in mask_data[\"rles\"]]\n elif self.output_mode == \"binary_mask\":\n mask_data[\"segmentations\"] = [rle_to_mask(rle) for rle in mask_data[\"rles\"]]\n else:\n mask_data[\"segmentations\"] = mask_data[\"rles\"]\n\n # Write mask records\n curr_anns = []\n for idx in range(len(mask_data[\"segmentations\"])):\n ann = {\n \"segmentation\": mask_data[\"segmentations\"][idx],\n \"area\": area_from_rle(mask_data[\"rles\"][idx]),\n \"bbox\": box_xyxy_to_xywh(mask_data[\"boxes\"][idx]).tolist(),\n \"predicted_iou\": mask_data[\"iou_preds\"][idx].item(),\n \"point_coords\": [mask_data[\"points\"][idx].tolist()],\n \"stability_score\": mask_data[\"stability_score\"][idx].item(),\n \"crop_box\": box_xyxy_to_xywh(mask_data[\"crop_boxes\"][idx]).tolist(),\n }\n curr_anns.append(ann)\n\n return curr_anns\n\n def _generate_masks(self, image: np.ndarray) -> MaskData:\n orig_size = image.shape[:2]\n crop_boxes, layer_idxs = generate_crop_boxes(\n orig_size, self.crop_n_layers, self.crop_overlap_ratio\n )\n\n # Iterate over image crops\n data = MaskData()\n for crop_box, layer_idx in zip(crop_boxes, layer_idxs):\n crop_data = self._process_crop(image, crop_box, layer_idx, orig_size)\n data.cat(crop_data)\n\n # Remove duplicate masks between crops\n if len(crop_boxes) > 1:\n # Prefer masks from smaller crops\n scores = 1 / box_area(data[\"crop_boxes\"])\n scores = scores.to(data[\"boxes\"].device)\n keep_by_nms = batched_nms(\n data[\"boxes\"].float(),\n scores,\n torch.zeros(len(data[\"boxes\"])), # categories\n iou_threshold=self.crop_nms_thresh,\n )\n data.filter(keep_by_nms)\n\n data.to_numpy()\n return data\n\n def _process_crop(\n self,\n image: np.ndarray,\n crop_box: List[int],\n crop_layer_idx: int,\n orig_size: Tuple[int, ...],\n ) -> MaskData:\n # Crop the image and calculate embeddings\n x0, y0, x1, y1 = crop_box\n cropped_im = image[y0:y1, x0:x1, :]\n cropped_im_size = cropped_im.shape[:2]\n self.predictor.set_image(cropped_im)\n\n # Get points for this crop\n points_scale = np.array(cropped_im_size)[None, ::-1]\n points_for_image = self.point_grids[crop_layer_idx] * points_scale\n\n # Generate masks for this crop in batches\n data = MaskData()\n for (points,) in batch_iterator(self.points_per_batch, points_for_image):\n batch_data = self._process_batch(points, cropped_im_size, crop_box, orig_size)\n data.cat(batch_data)\n del batch_data\n self.predictor.reset_image()\n\n # Remove duplicates within this crop.\n keep_by_nms = batched_nms(\n data[\"boxes\"].float(),\n data[\"iou_preds\"],\n torch.zeros(len(data[\"boxes\"])), # categories\n iou_threshold=self.box_nms_thresh,\n )\n data.filter(keep_by_nms)\n\n # Return to the original image frame\n data[\"boxes\"] = uncrop_boxes_xyxy(data[\"boxes\"], crop_box)\n data[\"points\"] = uncrop_points(data[\"points\"], crop_box)\n data[\"crop_boxes\"] = torch.tensor([crop_box for _ in range(len(data[\"rles\"]))])\n\n return data\n\n def _process_batch(\n self,\n points: np.ndarray,\n im_size: Tuple[int, ...],\n crop_box: List[int],\n orig_size: Tuple[int, ...],\n ) -> MaskData:\n orig_h, orig_w = orig_size\n\n # Run model on this batch\n transformed_points = self.predictor.transform.apply_coords(points, im_size)\n in_points = torch.as_tensor(transformed_points, device=self.predictor.device)\n in_labels = torch.ones(in_points.shape[0], dtype=torch.int, device=in_points.device)\n masks, iou_preds, _ = self.predictor.predict_torch(\n in_points[:, None, :],\n in_labels[:, None],\n multimask_output=True,\n return_logits=True,\n )\n\n # Serialize predictions and store in MaskData\n data = MaskData(\n masks=masks.flatten(0, 1),\n iou_preds=iou_preds.flatten(0, 1),\n points=torch.as_tensor(points.repeat(masks.shape[1], axis=0)),\n )\n del masks\n\n # Filter by predicted IoU\n if self.pred_iou_thresh > 0.0:\n keep_mask = data[\"iou_preds\"] > self.pred_iou_thresh\n data.filter(keep_mask)\n\n # Calculate stability score\n data[\"stability_score\"] = calculate_stability_score(\n data[\"masks\"], self.predictor.model.mask_threshold, self.stability_score_offset\n )\n if self.stability_score_thresh > 0.0:\n keep_mask = data[\"stability_score\"] >= self.stability_score_thresh\n data.filter(keep_mask)\n\n # Threshold masks and calculate boxes\n data[\"masks\"] = data[\"masks\"] > self.predictor.model.mask_threshold\n data[\"boxes\"] = batched_mask_to_box(data[\"masks\"])\n\n # Filter boxes that touch crop boundaries\n keep_mask = ~is_box_near_crop_edge(data[\"boxes\"], crop_box, [0, 0, orig_w, orig_h])\n if not torch.all(keep_mask):\n data.filter(keep_mask)\n\n # Compress to RLE\n data[\"masks\"] = uncrop_masks(data[\"masks\"], crop_box, orig_h, orig_w)\n data[\"rles\"] = mask_to_rle_pytorch(data[\"masks\"])\n del data[\"masks\"]\n\n return data\n\n @staticmethod\n def postprocess_small_regions(\n mask_data: MaskData, min_area: int, nms_thresh: float\n ) -> MaskData:\n \"\"\"\n Removes small disconnected regions and holes in masks, then reruns\n box NMS to remove any new duplicates.\n\n Edits mask_data in place.\n\n Requires open-cv as a dependency.\n \"\"\"\n if len(mask_data[\"rles\"]) == 0:\n return mask_data\n\n # Filter small disconnected regions and holes\n new_masks = []\n scores = []\n for rle in mask_data[\"rles\"]:\n mask = rle_to_mask(rle)\n\n mask, changed = remove_small_regions(mask, min_area, mode=\"holes\")\n unchanged = not changed\n mask, changed = remove_small_regions(mask, min_area, mode=\"islands\")\n unchanged = unchanged and not changed\n\n new_masks.append(torch.as_tensor(mask).unsqueeze(0))\n # Give score=0 to changed masks and score=1 to unchanged masks\n # so NMS will prefer ones that didn't need postprocessing\n scores.append(float(unchanged))\n\n # Recalculate boxes and remove any new duplicates\n masks = torch.cat(new_masks, dim=0)\n boxes = batched_mask_to_box(masks)\n keep_by_nms = batched_nms(\n boxes.float(),\n torch.as_tensor(scores),\n torch.zeros(len(boxes)), # categories\n iou_threshold=nms_thresh,\n )\n\n # Only recalculate RLEs for masks that have changed\n for i_mask in keep_by_nms:\n if scores[i_mask] == 0.0:\n mask_torch = masks[i_mask].unsqueeze(0)\n mask_data[\"rles\"][i_mask] = mask_to_rle_pytorch(mask_torch)[0]\n mask_data[\"boxes\"][i_mask] = boxes[i_mask] # update res directly\n mask_data.filter(keep_by_nms)\n\n return mask_data\n" }, { "path": "segment_anything/build_sam.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport torch\n\nfrom functools import partial\n\nfrom .modeling import ImageEncoderViT, MaskDecoder, PromptEncoder, Sam, TwoWayTransformer\n\n\ndef build_sam_vit_h(checkpoint=None):\n return _build_sam(\n encoder_embed_dim=1280,\n encoder_depth=32,\n encoder_num_heads=16,\n encoder_global_attn_indexes=[7, 15, 23, 31],\n checkpoint=checkpoint,\n )\n\n\nbuild_sam = build_sam_vit_h\n\n\ndef build_sam_vit_l(checkpoint=None):\n return _build_sam(\n encoder_embed_dim=1024,\n encoder_depth=24,\n encoder_num_heads=16,\n encoder_global_attn_indexes=[5, 11, 17, 23],\n checkpoint=checkpoint,\n )\n\n\ndef build_sam_vit_b(checkpoint=None):\n return _build_sam(\n encoder_embed_dim=768,\n encoder_depth=12,\n encoder_num_heads=12,\n encoder_global_attn_indexes=[2, 5, 8, 11],\n checkpoint=checkpoint,\n )\n\n\nsam_model_registry = {\n \"default\": build_sam,\n \"vit_h\": build_sam,\n \"vit_l\": build_sam_vit_l,\n \"vit_b\": build_sam_vit_b,\n}\n\n\ndef _build_sam(\n encoder_embed_dim,\n encoder_depth,\n encoder_num_heads,\n encoder_global_attn_indexes,\n checkpoint=None,\n):\n prompt_embed_dim = 256\n image_size = 1024\n vit_patch_size = 16\n image_embedding_size = image_size // vit_patch_size\n sam = Sam(\n image_encoder=ImageEncoderViT(\n depth=encoder_depth,\n embed_dim=encoder_embed_dim,\n img_size=image_size,\n mlp_ratio=4,\n norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),\n num_heads=encoder_num_heads,\n patch_size=vit_patch_size,\n qkv_bias=True,\n use_rel_pos=True,\n global_attn_indexes=encoder_global_attn_indexes,\n window_size=14,\n out_chans=prompt_embed_dim,\n ),\n prompt_encoder=PromptEncoder(\n embed_dim=prompt_embed_dim,\n image_embedding_size=(image_embedding_size, image_embedding_size),\n input_image_size=(image_size, image_size),\n mask_in_chans=16,\n ),\n mask_decoder=MaskDecoder(\n num_multimask_outputs=3,\n transformer=TwoWayTransformer(\n depth=2,\n embedding_dim=prompt_embed_dim,\n mlp_dim=2048,\n num_heads=8,\n ),\n transformer_dim=prompt_embed_dim,\n iou_head_depth=3,\n iou_head_hidden_dim=256,\n ),\n pixel_mean=[123.675, 116.28, 103.53],\n pixel_std=[58.395, 57.12, 57.375],\n )\n sam.eval()\n if checkpoint is not None:\n with open(checkpoint, \"rb\") as f:\n state_dict = torch.load(f)\n sam.load_state_dict(state_dict)\n return sam\n" }, { "path": "segment_anything/modeling/__init__.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nfrom .sam import Sam\nfrom .image_encoder import ImageEncoderViT\nfrom .mask_decoder import MaskDecoder\nfrom .prompt_encoder import PromptEncoder\nfrom .transformer import TwoWayTransformer\n" }, { "path": "segment_anything/modeling/common.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport torch\nimport torch.nn as nn\n\nfrom typing import Type\n\n\nclass MLPBlock(nn.Module):\n def __init__(\n self,\n embedding_dim: int,\n mlp_dim: int,\n act: Type[nn.Module] = nn.GELU,\n ) -> None:\n super().__init__()\n self.lin1 = nn.Linear(embedding_dim, mlp_dim)\n self.lin2 = nn.Linear(mlp_dim, embedding_dim)\n self.act = act()\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n return self.lin2(self.act(self.lin1(x)))\n\n\n# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa\n# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119 # noqa\nclass LayerNorm2d(nn.Module):\n def __init__(self, num_channels: int, eps: float = 1e-6) -> None:\n super().__init__()\n self.weight = nn.Parameter(torch.ones(num_channels))\n self.bias = nn.Parameter(torch.zeros(num_channels))\n self.eps = eps\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n u = x.mean(1, keepdim=True)\n s = (x - u).pow(2).mean(1, keepdim=True)\n x = (x - u) / torch.sqrt(s + self.eps)\n x = self.weight[:, None, None] * x + self.bias[:, None, None]\n return x\n" }, { "path": "segment_anything/modeling/image_encoder.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom typing import Optional, Tuple, Type\n\nfrom .common import LayerNorm2d, MLPBlock\n\n\n# This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa\nclass ImageEncoderViT(nn.Module):\n def __init__(\n self,\n img_size: int = 1024,\n patch_size: int = 16,\n in_chans: int = 3,\n embed_dim: int = 768,\n depth: int = 12,\n num_heads: int = 12,\n mlp_ratio: float = 4.0,\n out_chans: int = 256,\n qkv_bias: bool = True,\n norm_layer: Type[nn.Module] = nn.LayerNorm,\n act_layer: Type[nn.Module] = nn.GELU,\n use_abs_pos: bool = True,\n use_rel_pos: bool = False,\n rel_pos_zero_init: bool = True,\n window_size: int = 0,\n global_attn_indexes: Tuple[int, ...] = (),\n ) -> None:\n \"\"\"\n Args:\n img_size (int): Input image size.\n patch_size (int): Patch size.\n in_chans (int): Number of input image channels.\n embed_dim (int): Patch embedding dimension.\n depth (int): Depth of ViT.\n num_heads (int): Number of attention heads in each ViT block.\n mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.\n qkv_bias (bool): If True, add a learnable bias to query, key, value.\n norm_layer (nn.Module): Normalization layer.\n act_layer (nn.Module): Activation layer.\n use_abs_pos (bool): If True, use absolute positional embeddings.\n use_rel_pos (bool): If True, add relative positional embeddings to the attention map.\n rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.\n window_size (int): Window size for window attention blocks.\n global_attn_indexes (list): Indexes for blocks using global attention.\n \"\"\"\n super().__init__()\n self.img_size = img_size\n\n self.patch_embed = PatchEmbed(\n kernel_size=(patch_size, patch_size),\n stride=(patch_size, patch_size),\n in_chans=in_chans,\n embed_dim=embed_dim,\n )\n\n self.pos_embed: Optional[nn.Parameter] = None\n if use_abs_pos:\n # Initialize absolute positional embedding with pretrain image size.\n self.pos_embed = nn.Parameter(\n torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim)\n )\n\n self.blocks = nn.ModuleList()\n for i in range(depth):\n block = Block(\n dim=embed_dim,\n num_heads=num_heads,\n mlp_ratio=mlp_ratio,\n qkv_bias=qkv_bias,\n norm_layer=norm_layer,\n act_layer=act_layer,\n use_rel_pos=use_rel_pos,\n rel_pos_zero_init=rel_pos_zero_init,\n window_size=window_size if i not in global_attn_indexes else 0,\n input_size=(img_size // patch_size, img_size // patch_size),\n )\n self.blocks.append(block)\n\n self.neck = nn.Sequential(\n nn.Conv2d(\n embed_dim,\n out_chans,\n kernel_size=1,\n bias=False,\n ),\n LayerNorm2d(out_chans),\n nn.Conv2d(\n out_chans,\n out_chans,\n kernel_size=3,\n padding=1,\n bias=False,\n ),\n LayerNorm2d(out_chans),\n )\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n x = self.patch_embed(x)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n\n for blk in self.blocks:\n x = blk(x)\n\n x = self.neck(x.permute(0, 3, 1, 2))\n\n return x\n\n\nclass Block(nn.Module):\n \"\"\"Transformer blocks with support of window attention and residual propagation blocks\"\"\"\n\n def __init__(\n self,\n dim: int,\n num_heads: int,\n mlp_ratio: float = 4.0,\n qkv_bias: bool = True,\n norm_layer: Type[nn.Module] = nn.LayerNorm,\n act_layer: Type[nn.Module] = nn.GELU,\n use_rel_pos: bool = False,\n rel_pos_zero_init: bool = True,\n window_size: int = 0,\n input_size: Optional[Tuple[int, int]] = None,\n ) -> None:\n \"\"\"\n Args:\n dim (int): Number of input channels.\n num_heads (int): Number of attention heads in each ViT block.\n mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.\n qkv_bias (bool): If True, add a learnable bias to query, key, value.\n norm_layer (nn.Module): Normalization layer.\n act_layer (nn.Module): Activation layer.\n use_rel_pos (bool): If True, add relative positional embeddings to the attention map.\n rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.\n window_size (int): Window size for window attention blocks. If it equals 0, then\n use global attention.\n input_size (int or None): Input resolution for calculating the relative positional\n parameter size.\n \"\"\"\n super().__init__()\n self.norm1 = norm_layer(dim)\n self.attn = Attention(\n dim,\n num_heads=num_heads,\n qkv_bias=qkv_bias,\n use_rel_pos=use_rel_pos,\n rel_pos_zero_init=rel_pos_zero_init,\n input_size=input_size if window_size == 0 else (window_size, window_size),\n )\n\n self.norm2 = norm_layer(dim)\n self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)\n\n self.window_size = window_size\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n shortcut = x\n x = self.norm1(x)\n # Window partition\n if self.window_size > 0:\n H, W = x.shape[1], x.shape[2]\n x, pad_hw = window_partition(x, self.window_size)\n\n x = self.attn(x)\n # Reverse window partition\n if self.window_size > 0:\n x = window_unpartition(x, self.window_size, pad_hw, (H, W))\n\n x = shortcut + x\n x = x + self.mlp(self.norm2(x))\n\n return x\n\n\nclass Attention(nn.Module):\n \"\"\"Multi-head Attention block with relative position embeddings.\"\"\"\n\n def __init__(\n self,\n dim: int,\n num_heads: int = 8,\n qkv_bias: bool = True,\n use_rel_pos: bool = False,\n rel_pos_zero_init: bool = True,\n input_size: Optional[Tuple[int, int]] = None,\n ) -> None:\n \"\"\"\n Args:\n dim (int): Number of input channels.\n num_heads (int): Number of attention heads.\n qkv_bias (bool: If True, add a learnable bias to query, key, value.\n rel_pos (bool): If True, add relative positional embeddings to the attention map.\n rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.\n input_size (int or None): Input resolution for calculating the relative positional\n parameter size.\n \"\"\"\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n self.scale = head_dim**-0.5\n\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.proj = nn.Linear(dim, dim)\n\n self.use_rel_pos = use_rel_pos\n if self.use_rel_pos:\n assert (\n input_size is not None\n ), \"Input size must be provided if using relative positional encoding.\"\n # initialize relative positional embeddings\n self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))\n self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n B, H, W, _ = x.shape\n # qkv with shape (3, B, nHead, H * W, C)\n qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)\n # q, k, v with shape (B * nHead, H * W, C)\n q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)\n\n attn = (q * self.scale) @ k.transpose(-2, -1)\n\n if self.use_rel_pos:\n attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))\n\n attn = attn.softmax(dim=-1)\n x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)\n x = self.proj(x)\n\n return x\n\n\ndef window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:\n \"\"\"\n Partition into non-overlapping windows with padding if needed.\n Args:\n x (tensor): input tokens with [B, H, W, C].\n window_size (int): window size.\n\n Returns:\n windows: windows after partition with [B * num_windows, window_size, window_size, C].\n (Hp, Wp): padded height and width before partition\n \"\"\"\n B, H, W, C = x.shape\n\n pad_h = (window_size - H % window_size) % window_size\n pad_w = (window_size - W % window_size) % window_size\n if pad_h > 0 or pad_w > 0:\n x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))\n Hp, Wp = H + pad_h, W + pad_w\n\n x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)\n windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)\n return windows, (Hp, Wp)\n\n\ndef window_unpartition(\n windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]\n) -> torch.Tensor:\n \"\"\"\n Window unpartition into original sequences and removing padding.\n Args:\n x (tensor): input tokens with [B * num_windows, window_size, window_size, C].\n window_size (int): window size.\n pad_hw (Tuple): padded height and width (Hp, Wp).\n hw (Tuple): original height and width (H, W) before padding.\n\n Returns:\n x: unpartitioned sequences with [B, H, W, C].\n \"\"\"\n Hp, Wp = pad_hw\n H, W = hw\n B = windows.shape[0] // (Hp * Wp // window_size // window_size)\n x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)\n x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)\n\n if Hp > H or Wp > W:\n x = x[:, :H, :W, :].contiguous()\n return x\n\n\ndef get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:\n \"\"\"\n Get relative positional embeddings according to the relative positions of\n query and key sizes.\n Args:\n q_size (int): size of query q.\n k_size (int): size of key k.\n rel_pos (Tensor): relative position embeddings (L, C).\n\n Returns:\n Extracted positional embeddings according to relative positions.\n \"\"\"\n max_rel_dist = int(2 * max(q_size, k_size) - 1)\n # Interpolate rel pos if needed.\n if rel_pos.shape[0] != max_rel_dist:\n # Interpolate rel pos.\n rel_pos_resized = F.interpolate(\n rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),\n size=max_rel_dist,\n mode=\"linear\",\n )\n rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)\n else:\n rel_pos_resized = rel_pos\n\n # Scale the coords with short length if shapes for q and k are different.\n q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)\n k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)\n relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)\n\n return rel_pos_resized[relative_coords.long()]\n\n\ndef add_decomposed_rel_pos(\n attn: torch.Tensor,\n q: torch.Tensor,\n rel_pos_h: torch.Tensor,\n rel_pos_w: torch.Tensor,\n q_size: Tuple[int, int],\n k_size: Tuple[int, int],\n) -> torch.Tensor:\n \"\"\"\n Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.\n https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py # noqa B950\n Args:\n attn (Tensor): attention map.\n q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).\n rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.\n rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.\n q_size (Tuple): spatial sequence size of query q with (q_h, q_w).\n k_size (Tuple): spatial sequence size of key k with (k_h, k_w).\n\n Returns:\n attn (Tensor): attention map with added relative positional embeddings.\n \"\"\"\n q_h, q_w = q_size\n k_h, k_w = k_size\n Rh = get_rel_pos(q_h, k_h, rel_pos_h)\n Rw = get_rel_pos(q_w, k_w, rel_pos_w)\n\n B, _, dim = q.shape\n r_q = q.reshape(B, q_h, q_w, dim)\n rel_h = torch.einsum(\"bhwc,hkc->bhwk\", r_q, Rh)\n rel_w = torch.einsum(\"bhwc,wkc->bhwk\", r_q, Rw)\n\n attn = (\n attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]\n ).view(B, q_h * q_w, k_h * k_w)\n\n return attn\n\n\nclass PatchEmbed(nn.Module):\n \"\"\"\n Image to Patch Embedding.\n \"\"\"\n\n def __init__(\n self,\n kernel_size: Tuple[int, int] = (16, 16),\n stride: Tuple[int, int] = (16, 16),\n padding: Tuple[int, int] = (0, 0),\n in_chans: int = 3,\n embed_dim: int = 768,\n ) -> None:\n \"\"\"\n Args:\n kernel_size (Tuple): kernel size of the projection layer.\n stride (Tuple): stride of the projection layer.\n padding (Tuple): padding size of the projection layer.\n in_chans (int): Number of input image channels.\n embed_dim (int): embed_dim (int): Patch embedding dimension.\n \"\"\"\n super().__init__()\n\n self.proj = nn.Conv2d(\n in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding\n )\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n x = self.proj(x)\n # B C H W -> B H W C\n x = x.permute(0, 2, 3, 1)\n return x\n" }, { "path": "segment_anything/modeling/mask_decoder.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport torch\nfrom torch import nn\nfrom torch.nn import functional as F\n\nfrom typing import List, Tuple, Type\n\nfrom .common import LayerNorm2d\n\n\nclass MaskDecoder(nn.Module):\n def __init__(\n self,\n *,\n transformer_dim: int,\n transformer: nn.Module,\n num_multimask_outputs: int = 3,\n activation: Type[nn.Module] = nn.GELU,\n iou_head_depth: int = 3,\n iou_head_hidden_dim: int = 256,\n ) -> None:\n \"\"\"\n Predicts masks given an image and prompt embeddings, using a\n tranformer architecture.\n\n Arguments:\n transformer_dim (int): the channel dimension of the transformer\n transformer (nn.Module): the transformer used to predict masks\n num_multimask_outputs (int): the number of masks to predict\n when disambiguating masks\n activation (nn.Module): the type of activation to use when\n upscaling masks\n iou_head_depth (int): the depth of the MLP used to predict\n mask quality\n iou_head_hidden_dim (int): the hidden dimension of the MLP\n used to predict mask quality\n \"\"\"\n super().__init__()\n self.transformer_dim = transformer_dim\n self.transformer = transformer\n\n self.num_multimask_outputs = num_multimask_outputs\n\n self.iou_token = nn.Embedding(1, transformer_dim)\n self.num_mask_tokens = num_multimask_outputs + 1\n self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)\n\n self.output_upscaling = nn.Sequential(\n nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),\n LayerNorm2d(transformer_dim // 4),\n activation(),\n nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),\n activation(),\n )\n self.output_hypernetworks_mlps = nn.ModuleList(\n [\n MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)\n for i in range(self.num_mask_tokens)\n ]\n )\n\n self.iou_prediction_head = MLP(\n transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth\n )\n\n def forward(\n self,\n image_embeddings: torch.Tensor,\n image_pe: torch.Tensor,\n sparse_prompt_embeddings: torch.Tensor,\n dense_prompt_embeddings: torch.Tensor,\n multimask_output: bool,\n ) -> Tuple[torch.Tensor, torch.Tensor]:\n \"\"\"\n Predict masks given image and prompt embeddings.\n\n Arguments:\n image_embeddings (torch.Tensor): the embeddings from the image encoder\n image_pe (torch.Tensor): positional encoding with the shape of image_embeddings\n sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes\n dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs\n multimask_output (bool): Whether to return multiple masks or a single\n mask.\n\n Returns:\n torch.Tensor: batched predicted masks\n torch.Tensor: batched predictions of mask quality\n \"\"\"\n masks, iou_pred = self.predict_masks(\n image_embeddings=image_embeddings,\n image_pe=image_pe,\n sparse_prompt_embeddings=sparse_prompt_embeddings,\n dense_prompt_embeddings=dense_prompt_embeddings,\n )\n\n # Select the correct mask or masks for outptu\n if multimask_output:\n mask_slice = slice(1, None)\n else:\n mask_slice = slice(0, 1)\n masks = masks[:, mask_slice, :, :]\n iou_pred = iou_pred[:, mask_slice]\n\n # Prepare output\n return masks, iou_pred\n\n def predict_masks(\n self,\n image_embeddings: torch.Tensor,\n image_pe: torch.Tensor,\n sparse_prompt_embeddings: torch.Tensor,\n dense_prompt_embeddings: torch.Tensor,\n ) -> Tuple[torch.Tensor, torch.Tensor]:\n \"\"\"Predicts masks. See 'forward' for more details.\"\"\"\n # Concatenate output tokens\n output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)\n output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)\n tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)\n\n # Expand per-image data in batch direction to be per-mask\n src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)\n src = src + dense_prompt_embeddings\n pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)\n b, c, h, w = src.shape\n\n # Run the transformer\n hs, src = self.transformer(src, pos_src, tokens)\n iou_token_out = hs[:, 0, :]\n mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]\n\n # Upscale mask embeddings and predict masks using the mask tokens\n src = src.transpose(1, 2).view(b, c, h, w)\n upscaled_embedding = self.output_upscaling(src)\n hyper_in_list: List[torch.Tensor] = []\n for i in range(self.num_mask_tokens):\n hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))\n hyper_in = torch.stack(hyper_in_list, dim=1)\n b, c, h, w = upscaled_embedding.shape\n masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)\n\n # Generate mask quality predictions\n iou_pred = self.iou_prediction_head(iou_token_out)\n\n return masks, iou_pred\n\n\n# Lightly adapted from\n# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa\nclass MLP(nn.Module):\n def __init__(\n self,\n input_dim: int,\n hidden_dim: int,\n output_dim: int,\n num_layers: int,\n sigmoid_output: bool = False,\n ) -> None:\n super().__init__()\n self.num_layers = num_layers\n h = [hidden_dim] * (num_layers - 1)\n self.layers = nn.ModuleList(\n nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])\n )\n self.sigmoid_output = sigmoid_output\n\n def forward(self, x):\n for i, layer in enumerate(self.layers):\n x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)\n if self.sigmoid_output:\n x = F.sigmoid(x)\n return x\n" }, { "path": "segment_anything/modeling/prompt_encoder.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport numpy as np\nimport torch\nfrom torch import nn\n\nfrom typing import Any, Optional, Tuple, Type\n\nfrom .common import LayerNorm2d\n\n\nclass PromptEncoder(nn.Module):\n def __init__(\n self,\n embed_dim: int,\n image_embedding_size: Tuple[int, int],\n input_image_size: Tuple[int, int],\n mask_in_chans: int,\n activation: Type[nn.Module] = nn.GELU,\n ) -> None:\n \"\"\"\n Encodes prompts for input to SAM's mask decoder.\n\n Arguments:\n embed_dim (int): The prompts' embedding dimension\n image_embedding_size (tuple(int, int)): The spatial size of the\n image embedding, as (H, W).\n input_image_size (int): The padded size of the image as input\n to the image encoder, as (H, W).\n mask_in_chans (int): The number of hidden channels used for\n encoding input masks.\n activation (nn.Module): The activation to use when encoding\n input masks.\n \"\"\"\n super().__init__()\n self.embed_dim = embed_dim\n self.input_image_size = input_image_size\n self.image_embedding_size = image_embedding_size\n self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)\n\n self.num_point_embeddings: int = 4 # pos/neg point + 2 box corners\n point_embeddings = [nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)]\n self.point_embeddings = nn.ModuleList(point_embeddings)\n self.not_a_point_embed = nn.Embedding(1, embed_dim)\n\n self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])\n self.mask_downscaling = nn.Sequential(\n nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),\n LayerNorm2d(mask_in_chans // 4),\n activation(),\n nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),\n LayerNorm2d(mask_in_chans),\n activation(),\n nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),\n )\n self.no_mask_embed = nn.Embedding(1, embed_dim)\n\n def get_dense_pe(self) -> torch.Tensor:\n \"\"\"\n Returns the positional encoding used to encode point prompts,\n applied to a dense set of points the shape of the image encoding.\n\n Returns:\n torch.Tensor: Positional encoding with shape\n 1x(embed_dim)x(embedding_h)x(embedding_w)\n \"\"\"\n return self.pe_layer(self.image_embedding_size).unsqueeze(0)\n\n def _embed_points(\n self,\n points: torch.Tensor,\n labels: torch.Tensor,\n pad: bool,\n ) -> torch.Tensor:\n \"\"\"Embeds point prompts.\"\"\"\n points = points + 0.5 # Shift to center of pixel\n if pad:\n padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)\n padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)\n points = torch.cat([points, padding_point], dim=1)\n labels = torch.cat([labels, padding_label], dim=1)\n point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)\n point_embedding[labels == -1] = 0.0\n point_embedding[labels == -1] += self.not_a_point_embed.weight\n point_embedding[labels == 0] += self.point_embeddings[0].weight\n point_embedding[labels == 1] += self.point_embeddings[1].weight\n return point_embedding\n\n def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:\n \"\"\"Embeds box prompts.\"\"\"\n boxes = boxes + 0.5 # Shift to center of pixel\n coords = boxes.reshape(-1, 2, 2)\n corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)\n corner_embedding[:, 0, :] += self.point_embeddings[2].weight\n corner_embedding[:, 1, :] += self.point_embeddings[3].weight\n return corner_embedding\n\n def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:\n \"\"\"Embeds mask inputs.\"\"\"\n mask_embedding = self.mask_downscaling(masks)\n return mask_embedding\n\n def _get_batch_size(\n self,\n points: Optional[Tuple[torch.Tensor, torch.Tensor]],\n boxes: Optional[torch.Tensor],\n masks: Optional[torch.Tensor],\n ) -> int:\n \"\"\"\n Gets the batch size of the output given the batch size of the input prompts.\n \"\"\"\n if points is not None:\n return points[0].shape[0]\n elif boxes is not None:\n return boxes.shape[0]\n elif masks is not None:\n return masks.shape[0]\n else:\n return 1\n\n def _get_device(self) -> torch.device:\n return self.point_embeddings[0].weight.device\n\n def forward(\n self,\n points: Optional[Tuple[torch.Tensor, torch.Tensor]],\n boxes: Optional[torch.Tensor],\n masks: Optional[torch.Tensor],\n ) -> Tuple[torch.Tensor, torch.Tensor]:\n \"\"\"\n Embeds different types of prompts, returning both sparse and dense\n embeddings.\n\n Arguments:\n points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates\n and labels to embed.\n boxes (torch.Tensor or none): boxes to embed\n masks (torch.Tensor or none): masks to embed\n\n Returns:\n torch.Tensor: sparse embeddings for the points and boxes, with shape\n BxNx(embed_dim), where N is determined by the number of input points\n and boxes.\n torch.Tensor: dense embeddings for the masks, in the shape\n Bx(embed_dim)x(embed_H)x(embed_W)\n \"\"\"\n bs = self._get_batch_size(points, boxes, masks)\n sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())\n if points is not None:\n coords, labels = points\n point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))\n sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)\n if boxes is not None:\n box_embeddings = self._embed_boxes(boxes)\n sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)\n\n if masks is not None:\n dense_embeddings = self._embed_masks(masks)\n else:\n dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(\n bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]\n )\n\n return sparse_embeddings, dense_embeddings\n\n\nclass PositionEmbeddingRandom(nn.Module):\n \"\"\"\n Positional encoding using random spatial frequencies.\n \"\"\"\n\n def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:\n super().__init__()\n if scale is None or scale <= 0.0:\n scale = 1.0\n self.register_buffer(\n \"positional_encoding_gaussian_matrix\",\n scale * torch.randn((2, num_pos_feats)),\n )\n\n def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:\n \"\"\"Positionally encode points that are normalized to [0,1].\"\"\"\n # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape\n coords = 2 * coords - 1\n coords = coords @ self.positional_encoding_gaussian_matrix\n coords = 2 * np.pi * coords\n # outputs d_1 x ... x d_n x C shape\n return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)\n\n def forward(self, size: Tuple[int, int]) -> torch.Tensor:\n \"\"\"Generate positional encoding for a grid of the specified size.\"\"\"\n h, w = size\n device: Any = self.positional_encoding_gaussian_matrix.device\n grid = torch.ones((h, w), device=device, dtype=torch.float32)\n y_embed = grid.cumsum(dim=0) - 0.5\n x_embed = grid.cumsum(dim=1) - 0.5\n y_embed = y_embed / h\n x_embed = x_embed / w\n\n pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))\n return pe.permute(2, 0, 1) # C x H x W\n\n def forward_with_coords(\n self, coords_input: torch.Tensor, image_size: Tuple[int, int]\n ) -> torch.Tensor:\n \"\"\"Positionally encode points that are not normalized to [0,1].\"\"\"\n coords = coords_input.clone()\n coords[:, :, 0] = coords[:, :, 0] / image_size[1]\n coords[:, :, 1] = coords[:, :, 1] / image_size[0]\n return self._pe_encoding(coords.to(torch.float)) # B x N x C\n" }, { "path": "segment_anything/modeling/sam.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport torch\nfrom torch import nn\nfrom torch.nn import functional as F\n\nfrom typing import Any, Dict, List, Tuple\n\nfrom .image_encoder import ImageEncoderViT\nfrom .mask_decoder import MaskDecoder\nfrom .prompt_encoder import PromptEncoder\n\n\nclass Sam(nn.Module):\n mask_threshold: float = 0.0\n image_format: str = \"RGB\"\n\n def __init__(\n self,\n image_encoder: ImageEncoderViT,\n prompt_encoder: PromptEncoder,\n mask_decoder: MaskDecoder,\n pixel_mean: List[float] = [123.675, 116.28, 103.53],\n pixel_std: List[float] = [58.395, 57.12, 57.375],\n ) -> None:\n \"\"\"\n SAM predicts object masks from an image and input prompts.\n\n Arguments:\n image_encoder (ImageEncoderViT): The backbone used to encode the\n image into image embeddings that allow for efficient mask prediction.\n prompt_encoder (PromptEncoder): Encodes various types of input prompts.\n mask_decoder (MaskDecoder): Predicts masks from the image embeddings\n and encoded prompts.\n pixel_mean (list(float)): Mean values for normalizing pixels in the input image.\n pixel_std (list(float)): Std values for normalizing pixels in the input image.\n \"\"\"\n super().__init__()\n self.image_encoder = image_encoder\n self.prompt_encoder = prompt_encoder\n self.mask_decoder = mask_decoder\n self.register_buffer(\"pixel_mean\", torch.Tensor(pixel_mean).view(-1, 1, 1), False)\n self.register_buffer(\"pixel_std\", torch.Tensor(pixel_std).view(-1, 1, 1), False)\n\n @property\n def device(self) -> Any:\n return self.pixel_mean.device\n\n @torch.no_grad()\n def forward(\n self,\n batched_input: List[Dict[str, Any]],\n multimask_output: bool,\n ) -> List[Dict[str, torch.Tensor]]:\n \"\"\"\n Predicts masks end-to-end from provided images and prompts.\n If prompts are not known in advance, using SamPredictor is\n recommended over calling the model directly.\n\n Arguments:\n batched_input (list(dict)): A list over input images, each a\n dictionary with the following keys. A prompt key can be\n excluded if it is not present.\n 'image': The image as a torch tensor in 3xHxW format,\n already transformed for input to the model.\n 'original_size': (tuple(int, int)) The original size of\n the image before transformation, as (H, W).\n 'point_coords': (torch.Tensor) Batched point prompts for\n this image, with shape BxNx2. Already transformed to the\n input frame of the model.\n 'point_labels': (torch.Tensor) Batched labels for point prompts,\n with shape BxN.\n 'boxes': (torch.Tensor) Batched box inputs, with shape Bx4.\n Already transformed to the input frame of the model.\n 'mask_inputs': (torch.Tensor) Batched mask inputs to the model,\n in the form Bx1xHxW.\n multimask_output (bool): Whether the model should predict multiple\n disambiguating masks, or return a single mask.\n\n Returns:\n (list(dict)): A list over input images, where each element is\n as dictionary with the following keys.\n 'masks': (torch.Tensor) Batched binary mask predictions,\n with shape BxCxHxW, where B is the number of input promts,\n C is determiend by multimask_output, and (H, W) is the\n original size of the image.\n 'iou_predictions': (torch.Tensor) The model's predictions\n of mask quality, in shape BxC.\n 'low_res_logits': (torch.Tensor) Low resolution logits with\n shape BxCxHxW, where H=W=256. Can be passed as mask input\n to subsequent iterations of prediction.\n \"\"\"\n input_images = torch.stack([self.preprocess(x[\"image\"]) for x in batched_input], dim=0)\n image_embeddings = self.image_encoder(input_images)\n\n outputs = []\n for image_record, curr_embedding in zip(batched_input, image_embeddings):\n if \"point_coords\" in image_record:\n points = (image_record[\"point_coords\"], image_record[\"point_labels\"])\n else:\n points = None\n sparse_embeddings, dense_embeddings = self.prompt_encoder(\n points=points,\n boxes=image_record.get(\"boxes\", None),\n masks=image_record.get(\"mask_inputs\", None),\n )\n low_res_masks, iou_predictions = self.mask_decoder(\n image_embeddings=curr_embedding.unsqueeze(0),\n image_pe=self.prompt_encoder.get_dense_pe(),\n sparse_prompt_embeddings=sparse_embeddings,\n dense_prompt_embeddings=dense_embeddings,\n multimask_output=multimask_output,\n )\n masks = self.postprocess_masks(\n low_res_masks,\n input_size=image_record[\"image\"].shape[-2:],\n original_size=image_record[\"original_size\"],\n )\n masks = masks > self.mask_threshold\n outputs.append(\n {\n \"masks\": masks,\n \"iou_predictions\": iou_predictions,\n \"low_res_logits\": low_res_masks,\n }\n )\n return outputs\n\n def postprocess_masks(\n self,\n masks: torch.Tensor,\n input_size: Tuple[int, ...],\n original_size: Tuple[int, ...],\n ) -> torch.Tensor:\n \"\"\"\n Remove padding and upscale masks to the original image size.\n\n Arguments:\n masks (torch.Tensor): Batched masks from the mask_decoder,\n in BxCxHxW format.\n input_size (tuple(int, int)): The size of the image input to the\n model, in (H, W) format. Used to remove padding.\n original_size (tuple(int, int)): The original size of the image\n before resizing for input to the model, in (H, W) format.\n\n Returns:\n (torch.Tensor): Batched masks in BxCxHxW format, where (H, W)\n is given by original_size.\n \"\"\"\n masks = F.interpolate(\n masks,\n (self.image_encoder.img_size, self.image_encoder.img_size),\n mode=\"bilinear\",\n align_corners=False,\n )\n masks = masks[..., : input_size[0], : input_size[1]]\n masks = F.interpolate(masks, original_size, mode=\"bilinear\", align_corners=False)\n return masks\n\n def preprocess(self, x: torch.Tensor) -> torch.Tensor:\n \"\"\"Normalize pixel values and pad to a square input.\"\"\"\n # Normalize colors\n x = (x - self.pixel_mean) / self.pixel_std\n\n # Pad\n h, w = x.shape[-2:]\n padh = self.image_encoder.img_size - h\n padw = self.image_encoder.img_size - w\n x = F.pad(x, (0, padw, 0, padh))\n return x\n" }, { "path": "segment_anything/modeling/transformer.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport torch\nfrom torch import Tensor, nn\n\nimport math\nfrom typing import Tuple, Type\n\nfrom .common import MLPBlock\n\n\nclass TwoWayTransformer(nn.Module):\n def __init__(\n self,\n depth: int,\n embedding_dim: int,\n num_heads: int,\n mlp_dim: int,\n activation: Type[nn.Module] = nn.ReLU,\n attention_downsample_rate: int = 2,\n ) -> None:\n \"\"\"\n A transformer decoder that attends to an input image using\n queries whose positional embedding is supplied.\n\n Args:\n depth (int): number of layers in the transformer\n embedding_dim (int): the channel dimension for the input embeddings\n num_heads (int): the number of heads for multihead attention. Must\n divide embedding_dim\n mlp_dim (int): the channel dimension internal to the MLP block\n activation (nn.Module): the activation to use in the MLP block\n \"\"\"\n super().__init__()\n self.depth = depth\n self.embedding_dim = embedding_dim\n self.num_heads = num_heads\n self.mlp_dim = mlp_dim\n self.layers = nn.ModuleList()\n\n for i in range(depth):\n self.layers.append(\n TwoWayAttentionBlock(\n embedding_dim=embedding_dim,\n num_heads=num_heads,\n mlp_dim=mlp_dim,\n activation=activation,\n attention_downsample_rate=attention_downsample_rate,\n skip_first_layer_pe=(i == 0),\n )\n )\n\n self.final_attn_token_to_image = Attention(\n embedding_dim, num_heads, downsample_rate=attention_downsample_rate\n )\n self.norm_final_attn = nn.LayerNorm(embedding_dim)\n\n def forward(\n self,\n image_embedding: Tensor,\n image_pe: Tensor,\n point_embedding: Tensor,\n ) -> Tuple[Tensor, Tensor]:\n \"\"\"\n Args:\n image_embedding (torch.Tensor): image to attend to. Should be shape\n B x embedding_dim x h x w for any h and w.\n image_pe (torch.Tensor): the positional encoding to add to the image. Must\n have the same shape as image_embedding.\n point_embedding (torch.Tensor): the embedding to add to the query points.\n Must have shape B x N_points x embedding_dim for any N_points.\n\n Returns:\n torch.Tensor: the processed point_embedding\n torch.Tensor: the processed image_embedding\n \"\"\"\n # BxCxHxW -> BxHWxC == B x N_image_tokens x C\n bs, c, h, w = image_embedding.shape\n image_embedding = image_embedding.flatten(2).permute(0, 2, 1)\n image_pe = image_pe.flatten(2).permute(0, 2, 1)\n\n # Prepare queries\n queries = point_embedding\n keys = image_embedding\n\n # Apply transformer blocks and final layernorm\n for layer in self.layers:\n queries, keys = layer(\n queries=queries,\n keys=keys,\n query_pe=point_embedding,\n key_pe=image_pe,\n )\n\n # Apply the final attenion layer from the points to the image\n q = queries + point_embedding\n k = keys + image_pe\n attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)\n queries = queries + attn_out\n queries = self.norm_final_attn(queries)\n\n return queries, keys\n\n\nclass TwoWayAttentionBlock(nn.Module):\n def __init__(\n self,\n embedding_dim: int,\n num_heads: int,\n mlp_dim: int = 2048,\n activation: Type[nn.Module] = nn.ReLU,\n attention_downsample_rate: int = 2,\n skip_first_layer_pe: bool = False,\n ) -> None:\n \"\"\"\n A transformer block with four layers: (1) self-attention of sparse\n inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp\n block on sparse inputs, and (4) cross attention of dense inputs to sparse\n inputs.\n\n Arguments:\n embedding_dim (int): the channel dimension of the embeddings\n num_heads (int): the number of heads in the attention layers\n mlp_dim (int): the hidden dimension of the mlp block\n activation (nn.Module): the activation of the mlp block\n skip_first_layer_pe (bool): skip the PE on the first layer\n \"\"\"\n super().__init__()\n self.self_attn = Attention(embedding_dim, num_heads)\n self.norm1 = nn.LayerNorm(embedding_dim)\n\n self.cross_attn_token_to_image = Attention(\n embedding_dim, num_heads, downsample_rate=attention_downsample_rate\n )\n self.norm2 = nn.LayerNorm(embedding_dim)\n\n self.mlp = MLPBlock(embedding_dim, mlp_dim, activation)\n self.norm3 = nn.LayerNorm(embedding_dim)\n\n self.norm4 = nn.LayerNorm(embedding_dim)\n self.cross_attn_image_to_token = Attention(\n embedding_dim, num_heads, downsample_rate=attention_downsample_rate\n )\n\n self.skip_first_layer_pe = skip_first_layer_pe\n\n def forward(\n self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor\n ) -> Tuple[Tensor, Tensor]:\n # Self attention block\n if self.skip_first_layer_pe:\n queries = self.self_attn(q=queries, k=queries, v=queries)\n else:\n q = queries + query_pe\n attn_out = self.self_attn(q=q, k=q, v=queries)\n queries = queries + attn_out\n queries = self.norm1(queries)\n\n # Cross attention block, tokens attending to image embedding\n q = queries + query_pe\n k = keys + key_pe\n attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)\n queries = queries + attn_out\n queries = self.norm2(queries)\n\n # MLP block\n mlp_out = self.mlp(queries)\n queries = queries + mlp_out\n queries = self.norm3(queries)\n\n # Cross attention block, image embedding attending to tokens\n q = queries + query_pe\n k = keys + key_pe\n attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)\n keys = keys + attn_out\n keys = self.norm4(keys)\n\n return queries, keys\n\n\nclass Attention(nn.Module):\n \"\"\"\n An attention layer that allows for downscaling the size of the embedding\n after projection to queries, keys, and values.\n \"\"\"\n\n def __init__(\n self,\n embedding_dim: int,\n num_heads: int,\n downsample_rate: int = 1,\n ) -> None:\n super().__init__()\n self.embedding_dim = embedding_dim\n self.internal_dim = embedding_dim // downsample_rate\n self.num_heads = num_heads\n assert self.internal_dim % num_heads == 0, \"num_heads must divide embedding_dim.\"\n\n self.q_proj = nn.Linear(embedding_dim, self.internal_dim)\n self.k_proj = nn.Linear(embedding_dim, self.internal_dim)\n self.v_proj = nn.Linear(embedding_dim, self.internal_dim)\n self.out_proj = nn.Linear(self.internal_dim, embedding_dim)\n\n def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:\n b, n, c = x.shape\n x = x.reshape(b, n, num_heads, c // num_heads)\n return x.transpose(1, 2) # B x N_heads x N_tokens x C_per_head\n\n def _recombine_heads(self, x: Tensor) -> Tensor:\n b, n_heads, n_tokens, c_per_head = x.shape\n x = x.transpose(1, 2)\n return x.reshape(b, n_tokens, n_heads * c_per_head) # B x N_tokens x C\n\n def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:\n # Input projections\n q = self.q_proj(q)\n k = self.k_proj(k)\n v = self.v_proj(v)\n\n # Separate into heads\n q = self._separate_heads(q, self.num_heads)\n k = self._separate_heads(k, self.num_heads)\n v = self._separate_heads(v, self.num_heads)\n\n # Attention\n _, _, _, c_per_head = q.shape\n attn = q @ k.permute(0, 1, 3, 2) # B x N_heads x N_tokens x N_tokens\n attn = attn / math.sqrt(c_per_head)\n attn = torch.softmax(attn, dim=-1)\n\n # Get output\n out = attn @ v\n out = self._recombine_heads(out)\n out = self.out_proj(out)\n\n return out\n" }, { "path": "segment_anything/predictor.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport numpy as np\nimport torch\n\nfrom segment_anything.modeling import Sam\n\nfrom typing import Optional, Tuple\n\nfrom .utils.transforms import ResizeLongestSide\n\n\nclass SamPredictor:\n def __init__(\n self,\n sam_model: Sam,\n ) -> None:\n \"\"\"\n Uses SAM to calculate the image embedding for an image, and then\n allow repeated, efficient mask prediction given prompts.\n\n Arguments:\n sam_model (Sam): The model to use for mask prediction.\n \"\"\"\n super().__init__()\n self.model = sam_model\n self.transform = ResizeLongestSide(sam_model.image_encoder.img_size)\n self.reset_image()\n\n def set_image(\n self,\n image: np.ndarray,\n image_format: str = \"RGB\",\n ) -> None:\n \"\"\"\n Calculates the image embeddings for the provided image, allowing\n masks to be predicted with the 'predict' method.\n\n Arguments:\n image (np.ndarray): The image for calculating masks. Expects an\n image in HWC uint8 format, with pixel values in [0, 255].\n image_format (str): The color format of the image, in ['RGB', 'BGR'].\n \"\"\"\n assert image_format in [\n \"RGB\",\n \"BGR\",\n ], f\"image_format must be in ['RGB', 'BGR'], is {image_format}.\"\n if image_format != self.model.image_format:\n image = image[..., ::-1]\n\n # Transform the image to the form expected by the model\n input_image = self.transform.apply_image(image)\n input_image_torch = torch.as_tensor(input_image, device=self.device)\n input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]\n\n self.set_torch_image(input_image_torch, image.shape[:2])\n\n @torch.no_grad()\n def set_torch_image(\n self,\n transformed_image: torch.Tensor,\n original_image_size: Tuple[int, ...],\n ) -> None:\n \"\"\"\n Calculates the image embeddings for the provided image, allowing\n masks to be predicted with the 'predict' method. Expects the input\n image to be already transformed to the format expected by the model.\n\n Arguments:\n transformed_image (torch.Tensor): The input image, with shape\n 1x3xHxW, which has been transformed with ResizeLongestSide.\n original_image_size (tuple(int, int)): The size of the image\n before transformation, in (H, W) format.\n \"\"\"\n assert (\n len(transformed_image.shape) == 4\n and transformed_image.shape[1] == 3\n and max(*transformed_image.shape[2:]) == self.model.image_encoder.img_size\n ), f\"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}.\"\n self.reset_image()\n\n self.original_size = original_image_size\n self.input_size = tuple(transformed_image.shape[-2:])\n input_image = self.model.preprocess(transformed_image)\n self.features = self.model.image_encoder(input_image)\n self.is_image_set = True\n\n def predict(\n self,\n point_coords: Optional[np.ndarray] = None,\n point_labels: Optional[np.ndarray] = None,\n box: Optional[np.ndarray] = None,\n mask_input: Optional[np.ndarray] = None,\n multimask_output: bool = True,\n return_logits: bool = False,\n ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:\n \"\"\"\n Predict masks for the given input prompts, using the currently set image.\n\n Arguments:\n point_coords (np.ndarray or None): A Nx2 array of point prompts to the\n model. Each point is in (X,Y) in pixels.\n point_labels (np.ndarray or None): A length N array of labels for the\n point prompts. 1 indicates a foreground point and 0 indicates a\n background point.\n box (np.ndarray or None): A length 4 array given a box prompt to the\n model, in XYXY format.\n mask_input (np.ndarray): A low resolution mask input to the model, typically\n coming from a previous prediction iteration. Has form 1xHxW, where\n for SAM, H=W=256.\n multimask_output (bool): If true, the model will return three masks.\n For ambiguous input prompts (such as a single click), this will often\n produce better masks than a single prediction. If only a single\n mask is needed, the model's predicted quality score can be used\n to select the best mask. For non-ambiguous prompts, such as multiple\n input prompts, multimask_output=False can give better results.\n return_logits (bool): If true, returns un-thresholded masks logits\n instead of a binary mask.\n\n Returns:\n (np.ndarray): The output masks in CxHxW format, where C is the\n number of masks, and (H, W) is the original image size.\n (np.ndarray): An array of length C containing the model's\n predictions for the quality of each mask.\n (np.ndarray): An array of shape CxHxW, where C is the number\n of masks and H=W=256. These low resolution logits can be passed to\n a subsequent iteration as mask input.\n \"\"\"\n if not self.is_image_set:\n raise RuntimeError(\"An image must be set with .set_image(...) before mask prediction.\")\n\n # Transform input prompts\n coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None\n if point_coords is not None:\n assert (\n point_labels is not None\n ), \"point_labels must be supplied if point_coords is supplied.\"\n point_coords = self.transform.apply_coords(point_coords, self.original_size)\n coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=self.device)\n labels_torch = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)\n coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]\n if box is not None:\n box = self.transform.apply_boxes(box, self.original_size)\n box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)\n box_torch = box_torch[None, :]\n if mask_input is not None:\n mask_input_torch = torch.as_tensor(mask_input, dtype=torch.float, device=self.device)\n mask_input_torch = mask_input_torch[None, :, :, :]\n\n masks, iou_predictions, low_res_masks = self.predict_torch(\n coords_torch,\n labels_torch,\n box_torch,\n mask_input_torch,\n multimask_output,\n return_logits=return_logits,\n )\n\n masks = masks[0].detach().cpu().numpy()\n iou_predictions = iou_predictions[0].detach().cpu().numpy()\n low_res_masks = low_res_masks[0].detach().cpu().numpy()\n return masks, iou_predictions, low_res_masks\n\n @torch.no_grad()\n def predict_torch(\n self,\n point_coords: Optional[torch.Tensor],\n point_labels: Optional[torch.Tensor],\n boxes: Optional[torch.Tensor] = None,\n mask_input: Optional[torch.Tensor] = None,\n multimask_output: bool = True,\n return_logits: bool = False,\n ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:\n \"\"\"\n Predict masks for the given input prompts, using the currently set image.\n Input prompts are batched torch tensors and are expected to already be\n transformed to the input frame using ResizeLongestSide.\n\n Arguments:\n point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the\n model. Each point is in (X,Y) in pixels.\n point_labels (torch.Tensor or None): A BxN array of labels for the\n point prompts. 1 indicates a foreground point and 0 indicates a\n background point.\n box (np.ndarray or None): A Bx4 array given a box prompt to the\n model, in XYXY format.\n mask_input (np.ndarray): A low resolution mask input to the model, typically\n coming from a previous prediction iteration. Has form Bx1xHxW, where\n for SAM, H=W=256. Masks returned by a previous iteration of the\n predict method do not need further transformation.\n multimask_output (bool): If true, the model will return three masks.\n For ambiguous input prompts (such as a single click), this will often\n produce better masks than a single prediction. If only a single\n mask is needed, the model's predicted quality score can be used\n to select the best mask. For non-ambiguous prompts, such as multiple\n input prompts, multimask_output=False can give better results.\n return_logits (bool): If true, returns un-thresholded masks logits\n instead of a binary mask.\n\n Returns:\n (torch.Tensor): The output masks in BxCxHxW format, where C is the\n number of masks, and (H, W) is the original image size.\n (torch.Tensor): An array of shape BxC containing the model's\n predictions for the quality of each mask.\n (torch.Tensor): An array of shape BxCxHxW, where C is the number\n of masks and H=W=256. These low res logits can be passed to\n a subsequent iteration as mask input.\n \"\"\"\n if not self.is_image_set:\n raise RuntimeError(\"An image must be set with .set_image(...) before mask prediction.\")\n\n if point_coords is not None:\n points = (point_coords, point_labels)\n else:\n points = None\n\n # Embed prompts\n sparse_embeddings, dense_embeddings = self.model.prompt_encoder(\n points=points,\n boxes=boxes,\n masks=mask_input,\n )\n\n # Predict masks\n low_res_masks, iou_predictions = self.model.mask_decoder(\n image_embeddings=self.features,\n image_pe=self.model.prompt_encoder.get_dense_pe(),\n sparse_prompt_embeddings=sparse_embeddings,\n dense_prompt_embeddings=dense_embeddings,\n multimask_output=multimask_output,\n )\n\n # Upscale the masks to the original image resolution\n masks = self.model.postprocess_masks(low_res_masks, self.input_size, self.original_size)\n\n if not return_logits:\n masks = masks > self.model.mask_threshold\n\n return masks, iou_predictions, low_res_masks\n\n def get_image_embedding(self) -> torch.Tensor:\n \"\"\"\n Returns the image embeddings for the currently set image, with\n shape 1xCxHxW, where C is the embedding dimension and (H,W) are\n the embedding spatial dimension of SAM (typically C=256, H=W=64).\n \"\"\"\n if not self.is_image_set:\n raise RuntimeError(\n \"An image must be set with .set_image(...) to generate an embedding.\"\n )\n assert self.features is not None, \"Features must exist if an image has been set.\"\n return self.features\n\n @property\n def device(self) -> torch.device:\n return self.model.device\n\n def reset_image(self) -> None:\n \"\"\"Resets the currently set image.\"\"\"\n self.is_image_set = False\n self.features = None\n self.orig_h = None\n self.orig_w = None\n self.input_h = None\n self.input_w = None\n" }, { "path": "segment_anything/utils/__init__.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n" }, { "path": "segment_anything/utils/amg.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport numpy as np\nimport torch\n\nimport math\nfrom copy import deepcopy\nfrom itertools import product\nfrom typing import Any, Dict, Generator, ItemsView, List, Tuple\n\n\nclass MaskData:\n \"\"\"\n A structure for storing masks and their related data in batched format.\n Implements basic filtering and concatenation.\n \"\"\"\n\n def __init__(self, **kwargs) -> None:\n for v in kwargs.values():\n assert isinstance(\n v, (list, np.ndarray, torch.Tensor)\n ), \"MaskData only supports list, numpy arrays, and torch tensors.\"\n self._stats = dict(**kwargs)\n\n def __setitem__(self, key: str, item: Any) -> None:\n assert isinstance(\n item, (list, np.ndarray, torch.Tensor)\n ), \"MaskData only supports list, numpy arrays, and torch tensors.\"\n self._stats[key] = item\n\n def __delitem__(self, key: str) -> None:\n del self._stats[key]\n\n def __getitem__(self, key: str) -> Any:\n return self._stats[key]\n\n def items(self) -> ItemsView[str, Any]:\n return self._stats.items()\n\n def filter(self, keep: torch.Tensor) -> None:\n for k, v in self._stats.items():\n if v is None:\n self._stats[k] = None\n elif isinstance(v, torch.Tensor):\n self._stats[k] = v[torch.as_tensor(keep, device=v.device)]\n elif isinstance(v, np.ndarray):\n self._stats[k] = v[keep.detach().cpu().numpy()]\n elif isinstance(v, list) and keep.dtype == torch.bool:\n self._stats[k] = [a for i, a in enumerate(v) if keep[i]]\n elif isinstance(v, list):\n self._stats[k] = [v[i] for i in keep]\n else:\n raise TypeError(f\"MaskData key {k} has an unsupported type {type(v)}.\")\n\n def cat(self, new_stats: \"MaskData\") -> None:\n for k, v in new_stats.items():\n if k not in self._stats or self._stats[k] is None:\n self._stats[k] = deepcopy(v)\n elif isinstance(v, torch.Tensor):\n self._stats[k] = torch.cat([self._stats[k], v], dim=0)\n elif isinstance(v, np.ndarray):\n self._stats[k] = np.concatenate([self._stats[k], v], axis=0)\n elif isinstance(v, list):\n self._stats[k] = self._stats[k] + deepcopy(v)\n else:\n raise TypeError(f\"MaskData key {k} has an unsupported type {type(v)}.\")\n\n def to_numpy(self) -> None:\n for k, v in self._stats.items():\n if isinstance(v, torch.Tensor):\n self._stats[k] = v.detach().cpu().numpy()\n\n\ndef is_box_near_crop_edge(\n boxes: torch.Tensor, crop_box: List[int], orig_box: List[int], atol: float = 20.0\n) -> torch.Tensor:\n \"\"\"Filter masks at the edge of a crop, but not at the edge of the original image.\"\"\"\n crop_box_torch = torch.as_tensor(crop_box, dtype=torch.float, device=boxes.device)\n orig_box_torch = torch.as_tensor(orig_box, dtype=torch.float, device=boxes.device)\n boxes = uncrop_boxes_xyxy(boxes, crop_box).float()\n near_crop_edge = torch.isclose(boxes, crop_box_torch[None, :], atol=atol, rtol=0)\n near_image_edge = torch.isclose(boxes, orig_box_torch[None, :], atol=atol, rtol=0)\n near_crop_edge = torch.logical_and(near_crop_edge, ~near_image_edge)\n return torch.any(near_crop_edge, dim=1)\n\n\ndef box_xyxy_to_xywh(box_xyxy: torch.Tensor) -> torch.Tensor:\n box_xywh = deepcopy(box_xyxy)\n box_xywh[2] = box_xywh[2] - box_xywh[0]\n box_xywh[3] = box_xywh[3] - box_xywh[1]\n return box_xywh\n\n\ndef batch_iterator(batch_size: int, *args) -> Generator[List[Any], None, None]:\n assert len(args) > 0 and all(\n len(a) == len(args[0]) for a in args\n ), \"Batched iteration must have inputs of all the same size.\"\n n_batches = len(args[0]) // batch_size + int(len(args[0]) % batch_size != 0)\n for b in range(n_batches):\n yield [arg[b * batch_size : (b + 1) * batch_size] for arg in args]\n\n\ndef mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:\n \"\"\"\n Encodes masks to an uncompressed RLE, in the format expected by\n pycoco tools.\n \"\"\"\n # Put in fortran order and flatten h,w\n b, h, w = tensor.shape\n tensor = tensor.permute(0, 2, 1).flatten(1)\n\n # Compute change indices\n diff = tensor[:, 1:] ^ tensor[:, :-1]\n change_indices = diff.nonzero()\n\n # Encode run length\n out = []\n for i in range(b):\n cur_idxs = change_indices[change_indices[:, 0] == i, 1]\n cur_idxs = torch.cat(\n [\n torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),\n cur_idxs + 1,\n torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),\n ]\n )\n btw_idxs = cur_idxs[1:] - cur_idxs[:-1]\n counts = [] if tensor[i, 0] == 0 else [0]\n counts.extend(btw_idxs.detach().cpu().tolist())\n out.append({\"size\": [h, w], \"counts\": counts})\n return out\n\n\ndef rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:\n \"\"\"Compute a binary mask from an uncompressed RLE.\"\"\"\n h, w = rle[\"size\"]\n mask = np.empty(h * w, dtype=bool)\n idx = 0\n parity = False\n for count in rle[\"counts\"]:\n mask[idx : idx + count] = parity\n idx += count\n parity ^= True\n mask = mask.reshape(w, h)\n return mask.transpose() # Put in C order\n\n\ndef area_from_rle(rle: Dict[str, Any]) -> int:\n return sum(rle[\"counts\"][1::2])\n\n\ndef calculate_stability_score(\n masks: torch.Tensor, mask_threshold: float, threshold_offset: float\n) -> torch.Tensor:\n \"\"\"\n Computes the stability score for a batch of masks. The stability\n score is the IoU between the binary masks obtained by thresholding\n the predicted mask logits at high and low values.\n \"\"\"\n # One mask is always contained inside the other.\n # Save memory by preventing unnecesary cast to torch.int64\n intersections = (\n (masks > (mask_threshold + threshold_offset))\n .sum(-1, dtype=torch.int16)\n .sum(-1, dtype=torch.int32)\n )\n unions = (\n (masks > (mask_threshold - threshold_offset))\n .sum(-1, dtype=torch.int16)\n .sum(-1, dtype=torch.int32)\n )\n return intersections / unions\n\n\ndef build_point_grid(n_per_side: int) -> np.ndarray:\n \"\"\"Generates a 2D grid of points evenly spaced in [0,1]x[0,1].\"\"\"\n offset = 1 / (2 * n_per_side)\n points_one_side = np.linspace(offset, 1 - offset, n_per_side)\n points_x = np.tile(points_one_side[None, :], (n_per_side, 1))\n points_y = np.tile(points_one_side[:, None], (1, n_per_side))\n points = np.stack([points_x, points_y], axis=-1).reshape(-1, 2)\n return points\n\n\ndef build_all_layer_point_grids(\n n_per_side: int, n_layers: int, scale_per_layer: int\n) -> List[np.ndarray]:\n \"\"\"Generates point grids for all crop layers.\"\"\"\n points_by_layer = []\n for i in range(n_layers + 1):\n n_points = int(n_per_side / (scale_per_layer**i))\n points_by_layer.append(build_point_grid(n_points))\n return points_by_layer\n\n\ndef generate_crop_boxes(\n im_size: Tuple[int, ...], n_layers: int, overlap_ratio: float\n) -> Tuple[List[List[int]], List[int]]:\n \"\"\"\n Generates a list of crop boxes of different sizes. Each layer\n has (2**i)**2 boxes for the ith layer.\n \"\"\"\n crop_boxes, layer_idxs = [], []\n im_h, im_w = im_size\n short_side = min(im_h, im_w)\n\n # Original image\n crop_boxes.append([0, 0, im_w, im_h])\n layer_idxs.append(0)\n\n def crop_len(orig_len, n_crops, overlap):\n return int(math.ceil((overlap * (n_crops - 1) + orig_len) / n_crops))\n\n for i_layer in range(n_layers):\n n_crops_per_side = 2 ** (i_layer + 1)\n overlap = int(overlap_ratio * short_side * (2 / n_crops_per_side))\n\n crop_w = crop_len(im_w, n_crops_per_side, overlap)\n crop_h = crop_len(im_h, n_crops_per_side, overlap)\n\n crop_box_x0 = [int((crop_w - overlap) * i) for i in range(n_crops_per_side)]\n crop_box_y0 = [int((crop_h - overlap) * i) for i in range(n_crops_per_side)]\n\n # Crops in XYWH format\n for x0, y0 in product(crop_box_x0, crop_box_y0):\n box = [x0, y0, min(x0 + crop_w, im_w), min(y0 + crop_h, im_h)]\n crop_boxes.append(box)\n layer_idxs.append(i_layer + 1)\n\n return crop_boxes, layer_idxs\n\n\ndef uncrop_boxes_xyxy(boxes: torch.Tensor, crop_box: List[int]) -> torch.Tensor:\n x0, y0, _, _ = crop_box\n offset = torch.tensor([[x0, y0, x0, y0]], device=boxes.device)\n # Check if boxes has a channel dimension\n if len(boxes.shape) == 3:\n offset = offset.unsqueeze(1)\n return boxes + offset\n\n\ndef uncrop_points(points: torch.Tensor, crop_box: List[int]) -> torch.Tensor:\n x0, y0, _, _ = crop_box\n offset = torch.tensor([[x0, y0]], device=points.device)\n # Check if points has a channel dimension\n if len(points.shape) == 3:\n offset = offset.unsqueeze(1)\n return points + offset\n\n\ndef uncrop_masks(\n masks: torch.Tensor, crop_box: List[int], orig_h: int, orig_w: int\n) -> torch.Tensor:\n x0, y0, x1, y1 = crop_box\n if x0 == 0 and y0 == 0 and x1 == orig_w and y1 == orig_h:\n return masks\n # Coordinate transform masks\n pad_x, pad_y = orig_w - (x1 - x0), orig_h - (y1 - y0)\n pad = (x0, pad_x - x0, y0, pad_y - y0)\n return torch.nn.functional.pad(masks, pad, value=0)\n\n\ndef remove_small_regions(\n mask: np.ndarray, area_thresh: float, mode: str\n) -> Tuple[np.ndarray, bool]:\n \"\"\"\n Removes small disconnected regions and holes in a mask. Returns the\n mask and an indicator of if the mask has been modified.\n \"\"\"\n import cv2 # type: ignore\n\n assert mode in [\"holes\", \"islands\"]\n correct_holes = mode == \"holes\"\n working_mask = (correct_holes ^ mask).astype(np.uint8)\n n_labels, regions, stats, _ = cv2.connectedComponentsWithStats(working_mask, 8)\n sizes = stats[:, -1][1:] # Row 0 is background label\n small_regions = [i + 1 for i, s in enumerate(sizes) if s < area_thresh]\n if len(small_regions) == 0:\n return mask, False\n fill_labels = [0] + small_regions\n if not correct_holes:\n fill_labels = [i for i in range(n_labels) if i not in fill_labels]\n # If every region is below threshold, keep largest\n if len(fill_labels) == 0:\n fill_labels = [int(np.argmax(sizes)) + 1]\n mask = np.isin(regions, fill_labels)\n return mask, True\n\n\ndef coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:\n from pycocotools import mask as mask_utils # type: ignore\n\n h, w = uncompressed_rle[\"size\"]\n rle = mask_utils.frPyObjects(uncompressed_rle, h, w)\n rle[\"counts\"] = rle[\"counts\"].decode(\"utf-8\") # Necessary to serialize with json\n return rle\n\n\ndef batched_mask_to_box(masks: torch.Tensor) -> torch.Tensor:\n \"\"\"\n Calculates boxes in XYXY format around masks. Return [0,0,0,0] for\n an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x...x4.\n \"\"\"\n # torch.max below raises an error on empty inputs, just skip in this case\n if torch.numel(masks) == 0:\n return torch.zeros(*masks.shape[:-2], 4, device=masks.device)\n\n # Normalize shape to CxHxW\n shape = masks.shape\n h, w = shape[-2:]\n if len(shape) > 2:\n masks = masks.flatten(0, -3)\n else:\n masks = masks.unsqueeze(0)\n\n # Get top and bottom edges\n in_height, _ = torch.max(masks, dim=-1)\n in_height_coords = in_height * torch.arange(h, device=in_height.device)[None, :]\n bottom_edges, _ = torch.max(in_height_coords, dim=-1)\n in_height_coords = in_height_coords + h * (~in_height)\n top_edges, _ = torch.min(in_height_coords, dim=-1)\n\n # Get left and right edges\n in_width, _ = torch.max(masks, dim=-2)\n in_width_coords = in_width * torch.arange(w, device=in_width.device)[None, :]\n right_edges, _ = torch.max(in_width_coords, dim=-1)\n in_width_coords = in_width_coords + w * (~in_width)\n left_edges, _ = torch.min(in_width_coords, dim=-1)\n\n # If the mask is empty the right edge will be to the left of the left edge.\n # Replace these boxes with [0, 0, 0, 0]\n empty_filter = (right_edges < left_edges) | (bottom_edges < top_edges)\n out = torch.stack([left_edges, top_edges, right_edges, bottom_edges], dim=-1)\n out = out * (~empty_filter).unsqueeze(-1)\n\n # Return to original shape\n if len(shape) > 2:\n out = out.reshape(*shape[:-2], 4)\n else:\n out = out[0]\n\n return out\n" }, { "path": "segment_anything/utils/onnx.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport torch\nimport torch.nn as nn\nfrom torch.nn import functional as F\n\nfrom typing import Tuple\n\nfrom ..modeling import Sam\nfrom .amg import calculate_stability_score\n\n\nclass SamOnnxModel(nn.Module):\n \"\"\"\n This model should not be called directly, but is used in ONNX export.\n It combines the prompt encoder, mask decoder, and mask postprocessing of Sam,\n with some functions modified to enable model tracing. Also supports extra\n options controlling what information. See the ONNX export script for details.\n \"\"\"\n\n def __init__(\n self,\n model: Sam,\n return_single_mask: bool,\n use_stability_score: bool = False,\n return_extra_metrics: bool = False,\n ) -> None:\n super().__init__()\n self.mask_decoder = model.mask_decoder\n self.model = model\n self.img_size = model.image_encoder.img_size\n self.return_single_mask = return_single_mask\n self.use_stability_score = use_stability_score\n self.stability_score_offset = 1.0\n self.return_extra_metrics = return_extra_metrics\n\n @staticmethod\n def resize_longest_image_size(\n input_image_size: torch.Tensor, longest_side: int\n ) -> torch.Tensor:\n input_image_size = input_image_size.to(torch.float32)\n scale = longest_side / torch.max(input_image_size)\n transformed_size = scale * input_image_size\n transformed_size = torch.floor(transformed_size + 0.5).to(torch.int64)\n return transformed_size\n\n def _embed_points(self, point_coords: torch.Tensor, point_labels: torch.Tensor) -> torch.Tensor:\n point_coords = point_coords + 0.5\n point_coords = point_coords / self.img_size\n point_embedding = self.model.prompt_encoder.pe_layer._pe_encoding(point_coords)\n point_labels = point_labels.unsqueeze(-1).expand_as(point_embedding)\n\n point_embedding = point_embedding * (point_labels != -1)\n point_embedding = point_embedding + self.model.prompt_encoder.not_a_point_embed.weight * (\n point_labels == -1\n )\n\n for i in range(self.model.prompt_encoder.num_point_embeddings):\n point_embedding = point_embedding + self.model.prompt_encoder.point_embeddings[\n i\n ].weight * (point_labels == i)\n\n return point_embedding\n\n def _embed_masks(self, input_mask: torch.Tensor, has_mask_input: torch.Tensor) -> torch.Tensor:\n mask_embedding = has_mask_input * self.model.prompt_encoder.mask_downscaling(input_mask)\n mask_embedding = mask_embedding + (\n 1 - has_mask_input\n ) * self.model.prompt_encoder.no_mask_embed.weight.reshape(1, -1, 1, 1)\n return mask_embedding\n\n def mask_postprocessing(self, masks: torch.Tensor, orig_im_size: torch.Tensor) -> torch.Tensor:\n masks = F.interpolate(\n masks,\n size=(self.img_size, self.img_size),\n mode=\"bilinear\",\n align_corners=False,\n )\n\n prepadded_size = self.resize_longest_image_size(orig_im_size, self.img_size)\n masks = masks[..., : int(prepadded_size[0]), : int(prepadded_size[1])]\n\n orig_im_size = orig_im_size.to(torch.int64)\n h, w = orig_im_size[0], orig_im_size[1]\n masks = F.interpolate(masks, size=(h, w), mode=\"bilinear\", align_corners=False)\n return masks\n\n def select_masks(\n self, masks: torch.Tensor, iou_preds: torch.Tensor, num_points: int\n ) -> Tuple[torch.Tensor, torch.Tensor]:\n # Determine if we should return the multiclick mask or not from the number of points.\n # The reweighting is used to avoid control flow.\n score_reweight = torch.tensor(\n [[1000] + [0] * (self.model.mask_decoder.num_mask_tokens - 1)]\n ).to(iou_preds.device)\n score = iou_preds + (num_points - 2.5) * score_reweight\n best_idx = torch.argmax(score, dim=1)\n masks = masks[torch.arange(masks.shape[0]), best_idx, :, :].unsqueeze(1)\n iou_preds = iou_preds[torch.arange(masks.shape[0]), best_idx].unsqueeze(1)\n\n return masks, iou_preds\n\n @torch.no_grad()\n def forward(\n self,\n image_embeddings: torch.Tensor,\n point_coords: torch.Tensor,\n point_labels: torch.Tensor,\n mask_input: torch.Tensor,\n has_mask_input: torch.Tensor,\n orig_im_size: torch.Tensor,\n ):\n sparse_embedding = self._embed_points(point_coords, point_labels)\n dense_embedding = self._embed_masks(mask_input, has_mask_input)\n\n masks, scores = self.model.mask_decoder.predict_masks(\n image_embeddings=image_embeddings,\n image_pe=self.model.prompt_encoder.get_dense_pe(),\n sparse_prompt_embeddings=sparse_embedding,\n dense_prompt_embeddings=dense_embedding,\n )\n\n if self.use_stability_score:\n scores = calculate_stability_score(\n masks, self.model.mask_threshold, self.stability_score_offset\n )\n\n if self.return_single_mask:\n masks, scores = self.select_masks(masks, scores, point_coords.shape[1])\n\n upscaled_masks = self.mask_postprocessing(masks, orig_im_size)\n\n if self.return_extra_metrics:\n stability_scores = calculate_stability_score(\n upscaled_masks, self.model.mask_threshold, self.stability_score_offset\n )\n areas = (upscaled_masks > self.model.mask_threshold).sum(-1).sum(-1)\n return upscaled_masks, scores, stability_scores, areas, masks\n\n return upscaled_masks, scores, masks\n" }, { "path": "segment_anything/utils/transforms.py", "content": "# Copyright (c) Meta Platforms, Inc. and affiliates.\n# All rights reserved.\n\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport numpy as np\nimport torch\nfrom torch.nn import functional as F\nfrom torchvision.transforms.functional import resize, to_pil_image # type: ignore\n\nfrom copy import deepcopy\nfrom typing import Tuple\n\n\nclass ResizeLongestSide:\n \"\"\"\n Resizes images to longest side 'target_length', as well as provides\n methods for resizing coordinates and boxes. Provides methods for\n transforming both numpy array and batched torch tensors.\n \"\"\"\n\n def __init__(self, target_length: int) -> None:\n self.target_length = target_length\n\n def apply_image(self, image: np.ndarray) -> np.ndarray:\n \"\"\"\n Expects a numpy array with shape HxWxC in uint8 format.\n \"\"\"\n target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)\n return np.array(resize(to_pil_image(image), target_size))\n\n def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:\n \"\"\"\n Expects a numpy array of length 2 in the final dimension. Requires the\n original image size in (H, W) format.\n \"\"\"\n old_h, old_w = original_size\n new_h, new_w = self.get_preprocess_shape(\n original_size[0], original_size[1], self.target_length\n )\n coords = deepcopy(coords).astype(float)\n coords[..., 0] = coords[..., 0] * (new_w / old_w)\n coords[..., 1] = coords[..., 1] * (new_h / old_h)\n return coords\n\n def apply_boxes(self, boxes: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:\n \"\"\"\n Expects a numpy array shape Bx4. Requires the original image size\n in (H, W) format.\n \"\"\"\n boxes = self.apply_coords(boxes.reshape(-1, 2, 2), original_size)\n return boxes.reshape(-1, 4)\n\n def apply_image_torch(self, image: torch.Tensor) -> torch.Tensor:\n \"\"\"\n Expects batched images with shape BxCxHxW and float format. This\n transformation may not exactly match apply_image. apply_image is\n the transformation expected by the model.\n \"\"\"\n # Expects an image in BCHW format. May not exactly match apply_image.\n target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)\n return F.interpolate(\n image, target_size, mode=\"bilinear\", align_corners=False, antialias=True\n )\n\n def apply_coords_torch(\n self, coords: torch.Tensor, original_size: Tuple[int, ...]\n ) -> torch.Tensor:\n \"\"\"\n Expects a torch tensor with length 2 in the last dimension. Requires the\n original image size in (H, W) format.\n \"\"\"\n old_h, old_w = original_size\n new_h, new_w = self.get_preprocess_shape(\n original_size[0], original_size[1], self.target_length\n )\n coords = deepcopy(coords).to(torch.float)\n coords[..., 0] = coords[..., 0] * (new_w / old_w)\n coords[..., 1] = coords[..., 1] * (new_h / old_h)\n return coords\n\n def apply_boxes_torch(\n self, boxes: torch.Tensor, original_size: Tuple[int, ...]\n ) -> torch.Tensor:\n \"\"\"\n Expects a torch tensor with shape Bx4. Requires the original image\n size in (H, W) format.\n \"\"\"\n boxes = self.apply_coords_torch(boxes.reshape(-1, 2, 2), original_size)\n return boxes.reshape(-1, 4)\n\n @staticmethod\n def get_preprocess_shape(oldh: int, oldw: int, long_side_length: int) -> Tuple[int, int]:\n \"\"\"\n Compute the output size given input size and target long side length.\n \"\"\"\n scale = long_side_length * 1.0 / max(oldh, oldw)\n newh, neww = oldh * scale, oldw * scale\n neww = int(neww + 0.5)\n newh = int(newh + 0.5)\n return (newh, neww)\n" } ]