cat_seg/config.py

# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates.
from detectron2.config import CfgNode as CN


def add_cat_seg_config(cfg):
    """
    Add config for MASK_FORMER.
    """
    # data config
    # select the dataset mapper
    cfg.INPUT.DATASET_MAPPER_NAME = "mask_former_semantic"

    cfg.DATASETS.VAL_ALL = ("coco_2017_val_all_stuff_sem_seg",)

    # Color augmentation
    cfg.INPUT.COLOR_AUG_SSD = False
    # We retry random cropping until no single category in semantic segmentation GT occupies more
    # than `SINGLE_CATEGORY_MAX_AREA` part of the crop.
    cfg.INPUT.CROP.SINGLE_CATEGORY_MAX_AREA = 1.0
    # Pad image and segmentation GT in dataset mapper.
    cfg.INPUT.SIZE_DIVISIBILITY = -1

    # solver config
    # weight decay on embedding
    cfg.SOLVER.WEIGHT_DECAY_EMBED = 0.0
    # optimizer
    cfg.SOLVER.OPTIMIZER = "ADAMW"
    cfg.SOLVER.BACKBONE_MULTIPLIER = 0.1

    # mask_former model config
    cfg.MODEL.MASK_FORMER = CN()

    # Sometimes `backbone.size_divisibility` is set to 0 for some backbone (e.g. ResNet)
    # you can use this config to override
    cfg.MODEL.MASK_FORMER.SIZE_DIVISIBILITY = 32

    # swin transformer backbone
    cfg.MODEL.SWIN = CN()
    cfg.MODEL.SWIN.PRETRAIN_IMG_SIZE = 224
    cfg.MODEL.SWIN.PATCH_SIZE = 4
    cfg.MODEL.SWIN.EMBED_DIM = 96
    cfg.MODEL.SWIN.DEPTHS = [2, 2, 6, 2]
    cfg.MODEL.SWIN.NUM_HEADS = [3, 6, 12, 24]
    cfg.MODEL.SWIN.WINDOW_SIZE = 7
    cfg.MODEL.SWIN.MLP_RATIO = 4.0
    cfg.MODEL.SWIN.QKV_BIAS = True
    cfg.MODEL.SWIN.QK_SCALE = None
    cfg.MODEL.SWIN.DROP_RATE = 0.0
    cfg.MODEL.SWIN.ATTN_DROP_RATE = 0.0
    cfg.MODEL.SWIN.DROP_PATH_RATE = 0.3
    cfg.MODEL.SWIN.APE = False
    cfg.MODEL.SWIN.PATCH_NORM = True
    cfg.MODEL.SWIN.OUT_FEATURES = ["res2", "res3", "res4", "res5"]

    # zero shot config
    cfg.MODEL.SEM_SEG_HEAD.TRAIN_CLASS_JSON = "datasets/ADE20K_2021_17_01/ADE20K_847.json"
    cfg.MODEL.SEM_SEG_HEAD.TEST_CLASS_JSON = "datasets/ADE20K_2021_17_01/ADE20K_847.json"
    cfg.MODEL.SEM_SEG_HEAD.TRAIN_CLASS_INDEXES = "datasets/coco/coco_stuff/split/seen_indexes.json"
    cfg.MODEL.SEM_SEG_HEAD.TEST_CLASS_INDEXES = "datasets/coco/coco_stuff/split/unseen_indexes.json"

    cfg.MODEL.SEM_SEG_HEAD.CLIP_PRETRAINED = "ViT-B/16"

    cfg.MODEL.PROMPT_ENSEMBLE = False
    cfg.MODEL.PROMPT_ENSEMBLE_TYPE = "single"

    cfg.MODEL.CLIP_PIXEL_MEAN = [122.7709383, 116.7460125, 104.09373615]
    cfg.MODEL.CLIP_PIXEL_STD = [68.5005327, 66.6321579, 70.3231630]
    # three styles for clip classification, crop, mask, cropmask

    cfg.MODEL.SEM_SEG_HEAD.TEXT_AFFINITY_DIM = 512
    cfg.MODEL.SEM_SEG_HEAD.TEXT_AFFINITY_PROJ_DIM = 128
    cfg.MODEL.SEM_SEG_HEAD.APPEARANCE_AFFINITY_DIM = 512
    cfg.MODEL.SEM_SEG_HEAD.APPEARANCE_AFFINITY_PROJ_DIM = 128

    cfg.MODEL.SEM_SEG_HEAD.DECODER_DIMS = [64, 32]
    cfg.MODEL.SEM_SEG_HEAD.DECODER_AFFINITY_DIMS = [256, 128]
    cfg.MODEL.SEM_SEG_HEAD.DECODER_AFFINITY_PROJ_DIMS = [32, 16]

    cfg.MODEL.SEM_SEG_HEAD.NUM_LAYERS = 4
    cfg.MODEL.SEM_SEG_HEAD.NUM_HEADS = 4
    cfg.MODEL.SEM_SEG_HEAD.HIDDEN_DIMS = 128
    cfg.MODEL.SEM_SEG_HEAD.POOLING_SIZES = [6, 6]
    cfg.MODEL.SEM_SEG_HEAD.FEATURE_RESOLUTION = [24, 24]
    cfg.MODEL.SEM_SEG_HEAD.WINDOW_SIZES = 12
    cfg.MODEL.SEM_SEG_HEAD.ATTENTION_TYPE = "linear"

    cfg.MODEL.SEM_SEG_HEAD.PROMPT_DEPTH = 0
    cfg.MODEL.SEM_SEG_HEAD.PROMPT_LENGTH = 0
    cfg.SOLVER.CLIP_MULTIPLIER = 0.01

    cfg.MODEL.SEM_SEG_HEAD.CLIP_FINETUNE = "attention"
    cfg.TEST.SLIDING_WINDOW = False