code/model_vis_tools/extractor_sd.py

import itertools
from contextlib import ExitStack
import torch
from mask2former.data.datasets.register_ade20k_panoptic import ADE20K_150_CATEGORIES
from PIL import Image
import numpy as np
import torch.nn.functional as F
from detectron2.config import instantiate
from detectron2.data import MetadataCatalog
from detectron2.data import detection_utils as utils
from detectron2.config import LazyCall as L
from detectron2.data import transforms as T
from detectron2.data.datasets.builtin_meta import COCO_CATEGORIES
from detectron2.evaluation import inference_context
from detectron2.utils.env import seed_all_rng
from detectron2.utils.visualizer import ColorMode, Visualizer, random_color
from detectron2.utils.logger import setup_logger

from odise import model_zoo
from odise.checkpoint import ODISECheckpointer
from odise.config import instantiate_odise
from odise.data import get_openseg_labels
from odise.modeling.wrapper import OpenPanopticInference
from third_party.ODISE.odise.config.instantiate import instantiate_odise_backbone
from third_party.utils.utils_correspondence import resize

import faiss


COCO_THING_CLASSES = [
    label
    for idx, label in enumerate(get_openseg_labels("coco_panoptic", True))
    if COCO_CATEGORIES[idx]["isthing"] == 1
]
COCO_THING_COLORS = [c["color"] for c in COCO_CATEGORIES if c["isthing"] == 1]
COCO_STUFF_CLASSES = [
    label
    for idx, label in enumerate(get_openseg_labels("coco_panoptic", True))
    if COCO_CATEGORIES[idx]["isthing"] == 0
]
COCO_STUFF_COLORS = [c["color"] for c in COCO_CATEGORIES if c["isthing"] == 0]

ADE_THING_CLASSES = [
    label
    for idx, label in enumerate(get_openseg_labels("ade20k_150", True))
    if ADE20K_150_CATEGORIES[idx]["isthing"] == 1
]
ADE_THING_COLORS = [c["color"] for c in ADE20K_150_CATEGORIES if c["isthing"] == 1]
ADE_STUFF_CLASSES = [
    label
    for idx, label in enumerate(get_openseg_labels("ade20k_150", True))
    if ADE20K_150_CATEGORIES[idx]["isthing"] == 0
]
ADE_STUFF_COLORS = [c["color"] for c in ADE20K_150_CATEGORIES if c["isthing"] == 0]

LVIS_CLASSES = get_openseg_labels("lvis_1203", True)
# use beautiful coco colors
LVIS_COLORS = list(
    itertools.islice(itertools.cycle([c["color"] for c in COCO_CATEGORIES]), len(LVIS_CLASSES))
)


class StableDiffusionSeg(object):
    def __init__(self, model, metadata, aug, instance_mode=ColorMode.IMAGE):
        """
        Args:
            model (nn.Module):
            metadata (MetadataCatalog): image metadata.
            instance_mode (ColorMode):
            parallel (bool): whether to run the model in different processes from visualization.
                Useful since the visualization logic can be slow.
        """
        self.model = model
        self.metadata = metadata
        self.aug = aug
        self.cpu_device = torch.device("cpu")
        self.instance_mode = instance_mode

    def get_features(self, original_image, caption=None, pca=None):
        """
        Args:
            original_image (np.ndarray): an image of shape (H, W, C) (in BGR order).

        Returns:
            features (dict):
                the output of the model for one image only.
        """
        height, width = original_image.shape[:2]
        aug_input = T.AugInput(original_image, sem_seg=None)

        self.aug(aug_input)
        image = aug_input.image

        image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1))
        inputs = {"image": image, "height": height, "width": width}

        if caption is not None:
            features = self.model.get_features([inputs],caption,pca=pca)
        else:
            features = self.model.get_features([inputs],pca=pca)
        return features
    
    def predict(self, original_image):
        """
        Args:
            original_image (np.ndarray): an image of shape (H, W, C) (in BGR order).

        Returns:
            predictions (dict):
                the output of the model for one image only.
                See :doc:`/tutorials/models` for details about the format.
        """
        height, width = original_image.shape[:2]
        aug_input = T.AugInput(original_image, sem_seg=None)
        self.aug(aug_input)
        image = aug_input.image
        image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1))

        inputs = {"image": image, "height": height, "width": width}
        predictions = self.model([inputs])[0]
        return predictions

def build_demo_classes_and_metadata(vocab, label_list):
    extra_classes = []

    if vocab:
        for words in vocab.split(";"):
            extra_classes.append([word.strip() for word in words.split(",")])
    extra_colors = [random_color(rgb=True, maximum=1) for _ in range(len(extra_classes))]

    demo_thing_classes = extra_classes
    demo_stuff_classes = []
    demo_thing_colors = extra_colors
    demo_stuff_colors = []

    if "COCO" in label_list:
        demo_thing_classes += COCO_THING_CLASSES
        demo_stuff_classes += COCO_STUFF_CLASSES
        demo_thing_colors += COCO_THING_COLORS
        demo_stuff_colors += COCO_STUFF_COLORS
    if "ADE" in label_list:
        demo_thing_classes += ADE_THING_CLASSES
        demo_stuff_classes += ADE_STUFF_CLASSES
        demo_thing_colors += ADE_THING_COLORS
        demo_stuff_colors += ADE_STUFF_COLORS
    if "LVIS" in label_list:
        demo_thing_classes += LVIS_CLASSES
        demo_thing_colors += LVIS_COLORS

    MetadataCatalog.pop("odise_demo_metadata", None)
    demo_metadata = MetadataCatalog.get("odise_demo_metadata")
    demo_metadata.thing_classes = [c[0] for c in demo_thing_classes]
    demo_metadata.stuff_classes = [
        *demo_metadata.thing_classes,
        *[c[0] for c in demo_stuff_classes],
    ]
    demo_metadata.thing_colors = demo_thing_colors
    demo_metadata.stuff_colors = demo_thing_colors + demo_stuff_colors
    demo_metadata.stuff_dataset_id_to_contiguous_id = {
        idx: idx for idx in range(len(demo_metadata.stuff_classes))
    }
    demo_metadata.thing_dataset_id_to_contiguous_id = {
        idx: idx for idx in range(len(demo_metadata.thing_classes))
    }

    demo_classes = demo_thing_classes + demo_stuff_classes

    return demo_classes, demo_metadata

import sys


def load_model(config_path="Panoptic/odise_label_coco_50e.py", seed=42, diffusion_ver="v1-3", image_size=1024, num_timesteps=0, block_indices=(2,5,8,11), decoder_only=False, encoder_only=False, resblock_only=False):

    cfg = model_zoo.get_config(config_path, trained=True)
    cfg.model.backbone.feature_extractor.init_checkpoint = "sd://"+diffusion_ver
    cfg.model.backbone.feature_extractor.steps = (num_timesteps,)
    cfg.model.backbone.feature_extractor.unet_block_indices = block_indices
    cfg.model.backbone.feature_extractor.encoder_only = encoder_only
    cfg.model.backbone.feature_extractor.decoder_only = decoder_only
    cfg.model.backbone.feature_extractor.resblock_only = resblock_only
    cfg.model.overlap_threshold = 0
    cfg.dataloader.test.mapper.augmentations=[
            L(T.ResizeShortestEdge)(short_edge_length=image_size, sample_style="choice", max_size=2560),
        ]
    dataset_cfg = cfg.dataloader.test
    aug = instantiate(dataset_cfg.mapper).augmentations
    model = instantiate_odise(cfg.model)
    model.to(cfg.train.device)

    ODISECheckpointer(model).load(cfg.train.init_checkpoint)
    return model, aug


def load_sd_backbone(config_path="Panoptic/odise_label_coco_50e.py", seed=42, diffusion_ver="v1-3", image_size=1024, num_timesteps=0, block_indices=(2,5,8,11), decoder_only=False, encoder_only=False, resblock_only=False):

    cfg = model_zoo.get_config(config_path, trained=True)
    cfg.model.backbone.feature_extractor.init_checkpoint = "sd://"+diffusion_ver
    cfg.model.backbone.feature_extractor.steps = (num_timesteps,)
    cfg.model.backbone.feature_extractor.unet_block_indices = block_indices
    cfg.model.backbone.feature_extractor.encoder_only = encoder_only
    cfg.model.backbone.feature_extractor.decoder_only = decoder_only
    cfg.model.backbone.feature_extractor.resblock_only = resblock_only
    cfg.model.overlap_threshold = 0
    model = instantiate_odise_backbone(cfg.model)
    odise_backbone_ckpt = torch.load("third_party/ODISE/ckpts/odise_backbone_weights.pth", map_location="cpu")['model']
    missing_keys, unexpected_keys = model.load_state_dict(odise_backbone_ckpt, strict=False)
    model.to(cfg.train.device)

    return model


def inference(model, aug, image, vocab, label_list):

    demo_classes, demo_metadata = build_demo_classes_and_metadata(vocab, label_list)
    with ExitStack() as stack:
        inference_model = OpenPanopticInference(
            model=model,
            labels=demo_classes,
            metadata=demo_metadata,
            semantic_on=False,
            instance_on=False,
            panoptic_on=True,
        )
        stack.enter_context(inference_context(inference_model))
        stack.enter_context(torch.no_grad())

        demo = StableDiffusionSeg(inference_model, demo_metadata, aug)
        pred = demo.predict(np.array(image))
        return (pred, demo_classes)
    
def get_features(model, aug, image, vocab, label_list, caption=None, pca=False):
    
    demo_classes, demo_metadata = build_demo_classes_and_metadata(vocab, label_list)
    with ExitStack() as stack:
        inference_model = OpenPanopticInference(
            model=model,
            labels=demo_classes,
            metadata=demo_metadata,
            semantic_on=False,
            instance_on=False,
            panoptic_on=True,
        )
        stack.enter_context(inference_context(inference_model))
        stack.enter_context(torch.no_grad())

        demo = StableDiffusionSeg(inference_model, demo_metadata, aug)
        if caption is not None:
            features = demo.get_features(np.array(image), caption, pca=pca)
        else:
            features = demo.get_features(np.array(image), pca=pca)
        return features


def pca_process(features):
    # Get the feature tensors
    size_s5=features['s5'].shape[-1]
    size_s4=features['s4'].shape[-1]
    size_s3=features['s3'].shape[-1]

    s5 = features['s5'].reshape(features['s5'].shape[0], features['s5'].shape[1], -1)
    s4 = features['s4'].reshape(features['s4'].shape[0], features['s4'].shape[1], -1)
    s3 = features['s3'].reshape(features['s3'].shape[0], features['s3'].shape[1], -1)

    # Define the target dimensions
    target_dims = {'s5': 128, 's4': 128, 's3': 128}

    # Apply PCA to each tensor using Faiss CPU
    for name, tensor in zip(['s5', 's4', 's3'], [s5, s4, s3]):
        target_dim = target_dims[name]

        # Transpose the tensor so that the last dimension is the number of features
        tensor = tensor.permute(0, 2, 1)

        # # Norm the tensor
        # tensor = tensor / tensor.norm(dim=-1, keepdim=True)

        # Initialize a Faiss PCA object
        pca = faiss.PCAMatrix(tensor.shape[-1], target_dim)

        # Train the PCA object
        pca.train(tensor[0].cpu().numpy())

        # Apply PCA to the data
        transformed_tensor_np = pca.apply(tensor[0].cpu().numpy())

        # Convert the transformed data back to a tensor
        transformed_tensor = torch.tensor(transformed_tensor_np, device=tensor.device).unsqueeze(0)

        # Store the transformed tensor in the features dictionary
        features[name] = transformed_tensor

    # Reshape the tensors back to their original shapes
    features['s5'] = features['s5'].permute(0, 2, 1).reshape(features['s5'].shape[0], -1, size_s5, size_s5)
    features['s4'] = features['s4'].permute(0, 2, 1).reshape(features['s4'].shape[0], -1, size_s4, size_s4)
    features['s3'] = features['s3'].permute(0, 2, 1).reshape(features['s3'].shape[0], -1, size_s3, size_s3)
    # Upsample s5 spatially by a factor of 2
    upsampled_s5 = torch.nn.functional.interpolate(features['s5'], scale_factor=2, mode='bilinear', align_corners=False)

    # Concatenate upsampled_s5 and s4 to create a new s5
    features['s5'] = torch.cat((upsampled_s5, features['s4']), dim=1)

    # Set s3 as the new s4
    features['s4'] = features['s3']

    # Remove s3 from the features dictionary
    del features['s3']
    
    return features
    

def process_features_and_mask(model, aug, image, category=None, input_text=None, mask=True, pca=False, raw=False):

    input_image = image
    caption = input_text
    vocab = ""
    label_list = ["COCO"]
    category_convert_dict={
        'aeroplane':'airplane',
        'motorbike':'motorcycle',
        'pottedplant':'potted plant',
        'tvmonitor':'tv',
    }
    if type(category) is not list and category in category_convert_dict:
        category=category_convert_dict[category]
    elif type(category) is list:
        category=[category_convert_dict[cat] if cat in category_convert_dict else cat for cat in category]
    features = get_features(model, aug, input_image, vocab, label_list, caption, pca=(pca or raw))

    if pca:
        features = pca_process(features)
    if raw:
        return features
    
    features_gether_s4_s5 = torch.cat([features['s4'], F.interpolate(features['s5'], size=(features['s4'].shape[-2:]), mode='bilinear')], dim=1)
    
    if mask:
        (pred,classes) =inference(model, aug, input_image, vocab, label_list)
        seg_map=pred['panoptic_seg'][0]
        target_mask_id = []
        for item in pred['panoptic_seg'][1]:
            item['category_name']=classes[item['category_id']]
            if category in item['category_name']:
                target_mask_id.append(item['id'])
        resized_seg_map_s4 = F.interpolate(seg_map.unsqueeze(0).unsqueeze(0).float(), 
                                    size=(features['s4'].shape[-2:]), mode='nearest')
        # to do adjust size
        binary_seg_map = torch.zeros_like(resized_seg_map_s4)
        for i in target_mask_id:
            binary_seg_map += (resized_seg_map_s4 == i).float()
        if len(target_mask_id) == 0 or binary_seg_map.sum() < 6:
            binary_seg_map = torch.ones_like(resized_seg_map_s4)
        features_gether_s4_s5 = features_gether_s4_s5 * binary_seg_map
        # set where mask is 0 to inf
        features_gether_s4_s5[(binary_seg_map == 0).repeat(1,features_gether_s4_s5.shape[1],1,1)] = -1

    return features_gether_s4_s5

def get_mask(model, aug, image, category=None, input_text=None):
    model.backbone.feature_extractor.decoder_only = False
    model.backbone.feature_extractor.encoder_only = False
    model.backbone.feature_extractor.resblock_only = False
    input_image = image
    caption = input_text
    vocab = ""
    label_list = ["COCO"]
    category_convert_dict={
        'aeroplane':'airplane',
        'motorbike':'motorcycle',
        'pottedplant':'potted plant',
        'tvmonitor':'tv',
    }
    if type(category) is not list and category in category_convert_dict:
        category=category_convert_dict[category]
    elif type(category) is list:
        category=[category_convert_dict[cat] if cat in category_convert_dict else cat for cat in category]

    (pred,classes) =inference(model, aug, input_image, vocab, label_list)
    seg_map=pred['panoptic_seg'][0]
    target_mask_id = []
    for item in pred['panoptic_seg'][1]:
        item['category_name']=classes[item['category_id']]
        if type(category) is list:
            for cat in category:
                if cat in item['category_name']:
                    target_mask_id.append(item['id'])
        else:
            if category in item['category_name']:
                target_mask_id.append(item['id'])
    resized_seg_map_s4 = seg_map.float()
    binary_seg_map = torch.zeros_like(resized_seg_map_s4)
    for i in target_mask_id:
        binary_seg_map += (resized_seg_map_s4 == i).float()
    if len(target_mask_id) == 0 or binary_seg_map.sum() < 6:
        binary_seg_map = torch.ones_like(resized_seg_map_s4)

    return binary_seg_map

if __name__ == "__main__":
    image_path = sys.argv[1]
    try:
        input_text = sys.argv[2]
    except:
        input_text = None

    model, aug = load_model()
    img_size = 960
    image = Image.open(image_path).convert('RGB')
    image = resize(image, img_size, resize=True, to_pil=True)

    features = process_features_and_mask(model, aug, image, category=input_text, pca=False, raw=True)
    features = features['s4'] # save the features of layer 5
    
    # save the features
    np.save(image_path[:-4]+'.npy', features.cpu().numpy())