Spaces:
Runtime error
Runtime error
File size: 7,126 Bytes
2171e8f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 |
# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates.
from detectron2.config import CfgNode as CN
__all__ = ["add_common_config", "add_oneformer_config", "add_swin_config",
"add_dinat_config", "add_convnext_config"]
def add_common_config(cfg):
"""
Add config for common configuration
"""
# data config
# select the dataset mapper
cfg.INPUT.DATASET_MAPPER_NAME = "oneformer_unified"
# 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
cfg.INPUT.TASK_SEQ_LEN = 77
cfg.INPUT.MAX_SEQ_LEN = 77
cfg.INPUT.TASK_PROB = CN()
cfg.INPUT.TASK_PROB.SEMANTIC = 0.33
cfg.INPUT.TASK_PROB.INSTANCE = 0.66
# test dataset
cfg.DATASETS.TEST_PANOPTIC = ("",)
cfg.DATASETS.TEST_INSTANCE = ("",)
cfg.DATASETS.TEST_SEMANTIC = ("",)
# solver config
# weight decay on embedding
cfg.SOLVER.WEIGHT_DECAY_EMBED = 0.0
# optimizer
cfg.SOLVER.OPTIMIZER = "ADAMW"
cfg.SOLVER.BACKBONE_MULTIPLIER = 0.1
# wandb
cfg.WANDB = CN()
cfg.WANDB.PROJECT = "OneFormer"
cfg.WANDB.NAME = None
cfg.MODEL.IS_TRAIN = True
cfg.MODEL.IS_DEMO = False
# text encoder config
cfg.MODEL.TEXT_ENCODER = CN()
cfg.MODEL.TEXT_ENCODER.WIDTH = 256
cfg.MODEL.TEXT_ENCODER.CONTEXT_LENGTH = 77
cfg.MODEL.TEXT_ENCODER.NUM_LAYERS = 12
cfg.MODEL.TEXT_ENCODER.VOCAB_SIZE = 49408
cfg.MODEL.TEXT_ENCODER.PROJ_NUM_LAYERS = 2
cfg.MODEL.TEXT_ENCODER.N_CTX = 16
# oneformer inference config
cfg.MODEL.TEST = CN()
cfg.MODEL.TEST.SEMANTIC_ON = True
cfg.MODEL.TEST.INSTANCE_ON = False
cfg.MODEL.TEST.PANOPTIC_ON = False
cfg.MODEL.TEST.DETECTION_ON = False
cfg.MODEL.TEST.OBJECT_MASK_THRESHOLD = 0.0
cfg.MODEL.TEST.OVERLAP_THRESHOLD = 0.0
cfg.MODEL.TEST.SEM_SEG_POSTPROCESSING_BEFORE_INFERENCE = False
cfg.MODEL.TEST.TASK = "panoptic"
# TEST AUG Slide
cfg.TEST.AUG.IS_SLIDE = False
cfg.TEST.AUG.CROP_SIZE = (640, 640)
cfg.TEST.AUG.STRIDE = (426, 426)
cfg.TEST.AUG.SCALE = (2048, 640)
cfg.TEST.AUG.SETR_MULTI_SCALE = True
cfg.TEST.AUG.KEEP_RATIO = True
cfg.TEST.AUG.SIZE_DIVISOR = 32
# pixel decoder config
cfg.MODEL.SEM_SEG_HEAD.MASK_DIM = 256
# adding transformer in pixel decoder
cfg.MODEL.SEM_SEG_HEAD.TRANSFORMER_ENC_LAYERS = 0
# pixel decoder
cfg.MODEL.SEM_SEG_HEAD.PIXEL_DECODER_NAME = "BasePixelDecoder"
cfg.MODEL.SEM_SEG_HEAD.SEM_EMBED_DIM = 256
cfg.MODEL.SEM_SEG_HEAD.INST_EMBED_DIM = 256
# LSJ aug
cfg.INPUT.IMAGE_SIZE = 1024
cfg.INPUT.MIN_SCALE = 0.1
cfg.INPUT.MAX_SCALE = 2.0
# MSDeformAttn encoder configs
cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES = ["res3", "res4", "res5"]
cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_N_POINTS = 4
cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_N_HEADS = 8
def add_oneformer_config(cfg):
"""
Add config for ONE_FORMER.
"""
# oneformer model config
cfg.MODEL.ONE_FORMER = CN()
# loss
cfg.MODEL.ONE_FORMER.DEEP_SUPERVISION = True
cfg.MODEL.ONE_FORMER.NO_OBJECT_WEIGHT = 0.1
cfg.MODEL.ONE_FORMER.CLASS_WEIGHT = 1.0
cfg.MODEL.ONE_FORMER.DICE_WEIGHT = 1.0
cfg.MODEL.ONE_FORMER.MASK_WEIGHT = 20.0
cfg.MODEL.ONE_FORMER.CONTRASTIVE_WEIGHT = 0.5
cfg.MODEL.ONE_FORMER.CONTRASTIVE_TEMPERATURE = 0.07
# transformer config
cfg.MODEL.ONE_FORMER.NHEADS = 8
cfg.MODEL.ONE_FORMER.DROPOUT = 0.1
cfg.MODEL.ONE_FORMER.DIM_FEEDFORWARD = 2048
cfg.MODEL.ONE_FORMER.ENC_LAYERS = 0
cfg.MODEL.ONE_FORMER.CLASS_DEC_LAYERS = 2
cfg.MODEL.ONE_FORMER.DEC_LAYERS = 6
cfg.MODEL.ONE_FORMER.PRE_NORM = False
cfg.MODEL.ONE_FORMER.HIDDEN_DIM = 256
cfg.MODEL.ONE_FORMER.NUM_OBJECT_QUERIES = 120
cfg.MODEL.ONE_FORMER.NUM_OBJECT_CTX = 16
cfg.MODEL.ONE_FORMER.USE_TASK_NORM = True
cfg.MODEL.ONE_FORMER.TRANSFORMER_IN_FEATURE = "res5"
cfg.MODEL.ONE_FORMER.ENFORCE_INPUT_PROJ = False
# Sometimes `backbone.size_divisibility` is set to 0 for some backbone (e.g. ResNet)
# you can use this config to override
cfg.MODEL.ONE_FORMER.SIZE_DIVISIBILITY = 32
# transformer module
cfg.MODEL.ONE_FORMER.TRANSFORMER_DECODER_NAME = "ContrastiveMultiScaleMaskedTransformerDecoder"
# point loss configs
# Number of points sampled during training for a mask point head.
cfg.MODEL.ONE_FORMER.TRAIN_NUM_POINTS = 112 * 112
# Oversampling parameter for PointRend point sampling during training. Parameter `k` in the
# original paper.
cfg.MODEL.ONE_FORMER.OVERSAMPLE_RATIO = 3.0
# Importance sampling parameter for PointRend point sampling during training. Parametr `beta` in
# the original paper.
cfg.MODEL.ONE_FORMER.IMPORTANCE_SAMPLE_RATIO = 0.75
def add_swin_config(cfg):
"""
Add config forSWIN Backbone.
"""
# 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"]
cfg.MODEL.SWIN.USE_CHECKPOINT = False
def add_dinat_config(cfg):
"""
Add config for NAT Backbone.
"""
# DINAT transformer backbone
cfg.MODEL.DiNAT = CN()
cfg.MODEL.DiNAT.DEPTHS = [3, 4, 18, 5]
cfg.MODEL.DiNAT.OUT_FEATURES = ["res2", "res3", "res4", "res5"]
cfg.MODEL.DiNAT.EMBED_DIM = 64
cfg.MODEL.DiNAT.MLP_RATIO = 3.0
cfg.MODEL.DiNAT.NUM_HEADS = [2, 4, 8, 16]
cfg.MODEL.DiNAT.DROP_PATH_RATE = 0.2
cfg.MODEL.DiNAT.KERNEL_SIZE = 7
cfg.MODEL.DiNAT.DILATIONS = [[1, 16, 1], [1, 4, 1, 8], [1, 2, 1, 3, 1, 4], [1, 2, 1, 2, 1]]
cfg.MODEL.DiNAT.OUT_INDICES = (0, 1, 2, 3)
cfg.MODEL.DiNAT.QKV_BIAS = True
cfg.MODEL.DiNAT.QK_SCALE = None
cfg.MODEL.DiNAT.DROP_RATE = 0
cfg.MODEL.DiNAT.ATTN_DROP_RATE = 0.
cfg.MODEL.DiNAT.IN_PATCH_SIZE = 4
def add_convnext_config(cfg):
"""
Add config for ConvNeXt Backbone.
"""
# swin transformer backbone
cfg.MODEL.CONVNEXT = CN()
cfg.MODEL.CONVNEXT.IN_CHANNELS = 3
cfg.MODEL.CONVNEXT.DEPTHS = [3, 3, 27, 3]
cfg.MODEL.CONVNEXT.DIMS = [192, 384, 768, 1536]
cfg.MODEL.CONVNEXT.DROP_PATH_RATE = 0.4
cfg.MODEL.CONVNEXT.LSIT = 1.0
cfg.MODEL.CONVNEXT.OUT_INDICES = [0, 1, 2, 3]
cfg.MODEL.CONVNEXT.OUT_FEATURES = ["res2", "res3", "res4", "res5"] |