def add_fbnet_v2_default_configs(_C):
_C.MODEL.FBNET_V2 = CN()
_C.MODEL.FBNET_V2.ARCH = "default"
_C.MODEL.FBNET_V2.ARCH_DEF = []
# number of channels input to trunk
_C.MODEL.FBNET_V2.STEM_IN_CHANNELS = 3
_C.MODEL.FBNET_V2.SCALE_FACTOR = 1.0
# the output channels will be divisible by WIDTH_DIVISOR
_C.MODEL.FBNET_V2.WIDTH_DIVISOR = 1
# normalization configs
# name of norm such as "bn", "sync_bn", "gn"
_C.MODEL.FBNET_V2.NORM = "bn"
# for advanced use case that requries extra arguments, passing a list of
# dict such as [{"num_groups": 8}, {"momentum": 0.1}] (merged in given order).
# Note that string written it in .yaml will be evaluated by yacs, thus this
# node will become normal python object.
# https://github.com/rbgirshick/yacs/blob/master/yacs/config.py#L410
_C.MODEL.FBNET_V2.NORM_ARGS = []
_C.MODEL.VT_FPN = CN()
_C.MODEL.VT_FPN.IN_FEATURES = ["res2", "res3", "res4", "res5"]
_C.MODEL.VT_FPN.OUT_CHANNELS = 256
_C.MODEL.VT_FPN.LAYERS = 3
_C.MODEL.VT_FPN.TOKEN_LS = [16, 16, 8, 8]
_C.MODEL.VT_FPN.TOKEN_C = 1024
_C.MODEL.VT_FPN.HEADS = 16
_C.MODEL.VT_FPN.MIN_GROUP_PLANES = 64
_C.MODEL.VT_FPN.NORM = "BN"
_C.MODEL.VT_FPN.POS_HWS = []
_C.MODEL.VT_FPN.POS_N_DOWNSAMPLE = []
def test_pit_model(self):
cfg = CN()
cfg.MODEL = CN()
add_pit_backbone_config(cfg)
build_model = BACKBONE_REGISTRY.get("pit_d2go_model_wrapper")
pit_models = {
"pit_ti_ours": 160,
"pit_ti": 224,
"pit_s_ours_v1": 256,
"pit_s": 224,
}
pit_model_weights = {
"pit_ti_ours":
"manifold://mobile_vision_workflows/tree/workflows/kyungminkim/20210515/deit_[model]pit_scalable_distilled_[bs]128_[mcfg]pit_ti_ours_.HImkjNCpJI/checkpoint_best.pth",
"pit_ti":
"manifold://mobile_vision_workflows/tree/workflows/kyungminkim/20210515/deit_[model]pit_scalable_distilled_[bs]128_[mcfg]pit_ti_.QJeFNUfYOD/checkpoint_best.pth",
"pit_s_ours_v1":
"manifold://mobile_vision_workflows/tree/workflows/kyungminkim/20210515/deit_[model]pit_scalable_distilled_[bs]64_[mcfg]pit_s_ours_v1_.LXdwyBDaNY/checkpoint_best.pth",
"pit_s":
"manifold://mobile_vision_workflows/tree/workflows/kyungminkim/20210515/deit_[model]pit_scalable_distilled_[bs]128_[mcfg]pit_s_.zReQLPOuJe/checkpoint_best.pth",
}
for model_name, org_size in pit_models.items():
print("model_name", model_name)
cfg.MODEL.PIT.MODEL_CONFIG = f"manifold://mobile_vision_workflows/tree/workflows/wbc/deit/model_cfgs/{model_name}.json"
cfg.MODEL.PIT.WEIGHTS = pit_model_weights[model_name]
cfg.MODEL.PIT.DILATED = True
model = build_model(cfg, None)
model.eval()
for input_size_h in [org_size, 192, 224, 256, 320]:
for input_size_w in [org_size, 192, 224, 256, 320]:
x = torch.rand(1, 3, input_size_h, input_size_w)
y = model(x)
print(f"x.shape: {x.shape}, y.shape: {y.shape}")
def add_subclass_configs(cfg):
_C = cfg
_C.MODEL.SUBCLASS = CN()
_C.MODEL.SUBCLASS.SUBCLASS_ON = False
_C.MODEL.SUBCLASS.NUM_SUBCLASSES = 0 # must be set
_C.MODEL.SUBCLASS.NUM_LAYERS = 1
_C.MODEL.SUBCLASS.SUBCLASS_ID_FETCHER = "SubclassFetcher" # ABC, must be set
def create_cfg_from_cli_args(args, default_cfg):
"""
Instead of loading from defaults.py, this binary only includes necessary
configs building from scratch, and overrides them from args. There're two
levels of config:
_C: the config system used by this binary, which is a sub-set of training
config, override by configurable_cfg. It can also be override by
args.opts for convinience.
configurable_cfg: common configs that user should explicitly specify
in the args.
"""
_C = CN()
_C.INPUT = default_cfg.INPUT
_C.DATASETS = default_cfg.DATASETS
_C.DATALOADER = default_cfg.DATALOADER
_C.TEST = default_cfg.TEST
if hasattr(default_cfg, "D2GO_DATA"):
_C.D2GO_DATA = default_cfg.D2GO_DATA
if hasattr(default_cfg, "TENSORBOARD"):
_C.TENSORBOARD = default_cfg.TENSORBOARD
# NOTE configs below might not be necessary, but must add to make code work
_C.MODEL = CN()
_C.MODEL.META_ARCHITECTURE = default_cfg.MODEL.META_ARCHITECTURE
_C.MODEL.MASK_ON = default_cfg.MODEL.MASK_ON
_C.MODEL.KEYPOINT_ON = default_cfg.MODEL.KEYPOINT_ON
_C.MODEL.LOAD_PROPOSALS = default_cfg.MODEL.LOAD_PROPOSALS
assert _C.MODEL.LOAD_PROPOSALS is False, "caffe2 model doesn't support"
_C.OUTPUT_DIR = args.output_dir
configurable_cfg = [
"DATASETS.TEST",
args.datasets,
"INPUT.MIN_SIZE_TEST",
args.min_size,
"INPUT.MAX_SIZE_TEST",
args.max_size,
]
cfg = _C.clone()
cfg.merge_from_list(configurable_cfg)
cfg.merge_from_list(args.opts)
return cfg
def add_subclass_configs(cfg):
_C = cfg
_C.MODEL.SUBCLASS = CN()
_C.MODEL.SUBCLASS.SUBCLASS_ON = False
_C.MODEL.SUBCLASS.NUM_SUBCLASSES = 0 # must be set
_C.MODEL.SUBCLASS.NUM_LAYERS = 1
_C.MODEL.SUBCLASS.SUBCLASS_ID_FETCHER = "SubclassFetcher" # ABC, must be set
_C.MODEL.SUBCLASS.SUBCLASS_MAPPING = [
] # subclass mapping from model output to annotation
def add_bifpn_default_configs(_C):
_C.MODEL.BIFPN = CN()
_C.MODEL.BIFPN.DEPTH_MULTIPLIER = 1
_C.MODEL.BIFPN.SCALE_FACTOR = 1
_C.MODEL.BIFPN.WIDTH_DIVISOR = 8
_C.MODEL.BIFPN.NORM = "bn"
_C.MODEL.BIFPN.NORM_ARGS = []
_C.MODEL.BIFPN.TOP_BLOCK_BEFORE_FPN = False
def add_kmeans_anchors_cfg(_C):
_C.MODEL.KMEANS_ANCHORS = CN()
_C.MODEL.KMEANS_ANCHORS.KMEANS_ANCHORS_ON = False
_C.MODEL.KMEANS_ANCHORS.NUM_CLUSTERS = 0
_C.MODEL.KMEANS_ANCHORS.NUM_TRAINING_IMG = 0
_C.MODEL.KMEANS_ANCHORS.DATASETS = ()
_C.MODEL.ANCHOR_GENERATOR.OFFSET = 0.0
_C.MODEL.KMEANS_ANCHORS.RNG_SEED = 3
return _C
def add_model_ema_configs(_C):
from d2go.config import CfgNode as CN
_C.MODEL_EMA = CN()
_C.MODEL_EMA.ENABLED = False
_C.MODEL_EMA.DECAY = 0.999
# use the same as MODEL.DEVICE when empty
_C.MODEL_EMA.DEVICE = ""
# When True, loading the ema weight to the model when eval_only=True in build_model()
_C.MODEL_EMA.USE_EMA_WEIGHTS_FOR_EVAL_ONLY = False
def get_default_cfg():
"""
Override `get_default_cfg` for adding non common config.
"""
from detectron2.config import get_cfg as get_d2_cfg
cfg = get_d2_cfg()
cfg = CN(cfg) # upgrade from D2's CfgNode to D2Go's CfgNode
cfg.SOLVER.AUTO_SCALING_METHODS = ["default_scale_d2_configs"]
return cfg
def get_default_cfg():
_C = super(GeneralizedRCNNRunner,
GeneralizedRCNNRunner).get_default_cfg()
_C.EXPORT_CAFFE2 = CN()
_C.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT = False
_C.RCNN_PREPARE_FOR_EXPORT = "default_rcnn_prepare_for_export"
_C.RCNN_PREPARE_FOR_QUANT = "default_rcnn_prepare_for_quant"
_C.RCNN_PREPARE_FOR_QUANT_CONVERT = "default_rcnn_prepare_for_quant_convert"
return _C
def add_tensorboard_default_configs(_C):
_C.TENSORBOARD = CN()
# Output from dataloader will be written to tensorboard at this frequency
_C.TENSORBOARD.TRAIN_LOADER_VIS_WRITE_PERIOD = 20
# This controls max number of images over all batches, be considerate when
# increasing this number because it takes disk space and slows down the training
_C.TENSORBOARD.TRAIN_LOADER_VIS_MAX_IMAGES = 16
# Max number of images per dataset to visualize in tensorboard during evaluation
_C.TENSORBOARD.TEST_VIS_MAX_IMAGES = 16
# TENSORBOARD.LOG_DIR will be determined solely by OUTPUT_DIR
_C.register_deprecated_key("TENSORBOARD.LOG_DIR")
def add_d2go_data_default_configs(_C):
_C.D2GO_DATA = CN()
# Config for "detectron2go.data.extended_coco.extended_coco_load"
_C.D2GO_DATA.DATASETS = CN()
# List of class names to use when loading the data, this applies to train
# and test separately. Default value means using all classes, otherwise it'll create
# new json file containing only given categories.
_C.D2GO_DATA.DATASETS.TRAIN_CATEGORIES = ()
_C.D2GO_DATA.DATASETS.TEST_CATEGORIES = ()
# Register a list of COCO datasets in config
_C.D2GO_DATA.DATASETS.COCO_INJECTION = CN()
_C.D2GO_DATA.DATASETS.COCO_INJECTION.NAMES = []
_C.D2GO_DATA.DATASETS.COCO_INJECTION.IM_DIRS = []
_C.D2GO_DATA.DATASETS.COCO_INJECTION.JSON_FILES = []
# On-the-fly register a list of datasets located under detectron2go/datasets
# by specifying the filename (without .py).
_C.D2GO_DATA.DATASETS.DYNAMIC_DATASETS = []
# TODO: potentially add this config
# # List of extra keys in annotation, the item will be forwarded by
# # extended_coco_load.
# _C.D2GO_DATA.DATASETS.ANNOTATION_FIELDS_TO_FORWARD = ()
# Config for D2GoDatasetMapper
_C.D2GO_DATA.MAPPER = CN()
# dataset mapper name
_C.D2GO_DATA.MAPPER.NAME = "D2GoDatasetMapper"
# When enabled, image item from json dataset doesn't need to have width/hegiht,
# they will be backfilled once image is loaded. This may cause issue when
# width/hegiht is acutally been used by extended_coco_load, eg. grouping
# by aspect ratio.
_C.D2GO_DATA.MAPPER.BACKFILL_SIZE = False
_C.D2GO_DATA.MAPPER.RETRY = 3
_C.D2GO_DATA.MAPPER.CATCH_EXCEPTION = True
_C.D2GO_DATA.AUG_OPS = CN()
# List of transforms that are represented by string. Each string starts with
# a registered name in TRANSFORM_OP_REGISTRY, optionally followed by a string
# argument (separated by "::") which can be used for initializing the
# transform object. See build_transform_gen for the detail.
# Some examples are:
# example 1: RandomFlipOp
# example 2: RandomFlipOp::{}
# example 3: RandomFlipOp::{"prob":0.5}
# example 4: RandomBrightnessOp::{"intensity_min":1.0, "intensity_max":2.0}
# NOTE: search "example repr:" in fbcode for examples.
_C.D2GO_DATA.AUG_OPS.TRAIN = ["ResizeShortestEdgeOp", "RandomFlipOp"]
_C.D2GO_DATA.AUG_OPS.TEST = ["ResizeShortestEdgeOp"]
_C.D2GO_DATA.TEST = CN()
# Evaluate on the first specified number of images for each datset during
# testing, default value 0 means using all images.
# NOTE: See maybe_subsample_n_images for details.
_C.D2GO_DATA.TEST.MAX_IMAGES = 0
_C.D2GO_DATA.TEST.SUBSET_SAMPLING = "frontmost" # one of {"frontmost", "random"}
return _C
def test_deit_model(self):
cfg = CN()
cfg.MODEL = CN()
add_deit_backbone_config(cfg)
build_model = BACKBONE_REGISTRY.get("deit_d2go_model_wrapper")
deit_models = {
"8X-7-RM_4": 170,
"DeiT-Tiny": 224,
"DeiT-Small": 224,
"32X-1-RM_2": 221,
"8X-7": 160,
"32X-1": 256,
}
deit_model_weights = {
"8X-7-RM_4":
"manifold://mobile_vision_workflows/tree/workflows/kyungminkim/20210511/deit_[model]deit_scaling_distill_[bs]128_[mcfg]8X-7-RM_4_.OIXarYpbZw/checkpoint_best.pth",
"DeiT-Tiny":
"manifold://mobile_vision_workflows/tree/workflows/cl114/DeiT-official-ckpt/deit_tiny_distilled_patch16_224-b40b3cf7.pth",
"DeiT-Small":
"manifold://mobile_vision_workflows/tree/workflows/cl114/DeiT-official-ckpt/deit_small_distilled_patch16_224-649709d9.pth",
"32X-1-RM_2":
"manifold://mobile_vision_workflows/tree/workflows/kyungminkim/20210511/deit_[model]deit_scaling_distill_[bs]64_[mcfg]32X-1-RM_2_.xusuFyNMdD/checkpoint_best.pth",
"8X-7":
"manifold://mobile_vision_workflows/tree/workflows/cl114/scaled_best/8X-7.pth",
"32X-1":
"manifold://mobile_vision_workflows/tree/workflows/cl114/scaled_best/32X-1.pth",
}
for model_name, org_size in deit_models.items():
print("model_name", model_name)
cfg.MODEL.DEIT.MODEL_CONFIG = f"manifold://mobile_vision_workflows/tree/workflows/wbc/deit/model_cfgs/{model_name}.json"
cfg.MODEL.DEIT.WEIGHTS = deit_model_weights[model_name]
model = build_model(cfg, None)
model.eval()
for input_size_h in [org_size, 192, 224, 256, 320]:
for input_size_w in [org_size, 192, 224, 256, 320]:
x = torch.rand(1, 3, input_size_h, input_size_w)
y = model(x)
print(f"x.shape: {x.shape}, y.shape: {y.shape}")
def add_detr_config(cfg):
"""
Add config for DETR.
"""
cfg.MODEL.DETR = CN()
cfg.MODEL.DETR.NAME = "DETR"
cfg.MODEL.DETR.NUM_CLASSES = 80
# simple backbone
cfg.MODEL.BACKBONE.SIMPLE = False
cfg.MODEL.BACKBONE.STRIDE = 1
cfg.MODEL.BACKBONE.CHANNEL = 0
# FBNet
cfg.MODEL.FBNET_V2.OUT_FEATURES = ["trunk3"]
# For Segmentation
cfg.MODEL.DETR.FROZEN_WEIGHTS = ""
# LOSS
cfg.MODEL.DETR.DEFORMABLE = False
cfg.MODEL.DETR.USE_FOCAL_LOSS = False
cfg.MODEL.DETR.CENTERED_POSITION_ENCODIND = False
cfg.MODEL.DETR.CLS_WEIGHT = 1.0
cfg.MODEL.DETR.NUM_FEATURE_LEVELS = 4
cfg.MODEL.DETR.GIOU_WEIGHT = 2.0
cfg.MODEL.DETR.L1_WEIGHT = 5.0
cfg.MODEL.DETR.DEEP_SUPERVISION = True
cfg.MODEL.DETR.NO_OBJECT_WEIGHT = 0.1
cfg.MODEL.DETR.WITH_BOX_REFINE = False
cfg.MODEL.DETR.TWO_STAGE = False
cfg.MODEL.DETR.DECODER_BLOCK_GRAD = True
# TRANSFORMER
cfg.MODEL.DETR.NHEADS = 8
cfg.MODEL.DETR.DROPOUT = 0.1
cfg.MODEL.DETR.DIM_FEEDFORWARD = 2048
cfg.MODEL.DETR.ENC_LAYERS = 6
cfg.MODEL.DETR.DEC_LAYERS = 6
cfg.MODEL.DETR.BBOX_EMBED_NUM_LAYERS = 3
cfg.MODEL.DETR.PRE_NORM = False
cfg.MODEL.DETR.HIDDEN_DIM = 256
cfg.MODEL.DETR.NUM_OBJECT_QUERIES = 100
# solver
cfg.SOLVER.OPTIMIZER = "ADAMW"
cfg.SOLVER.BACKBONE_MULTIPLIER = 0.1
# tgt & embeddings
cfg.MODEL.DETR.LEARNABLE_TGT = False
def add_fcos_configs(cfg):
cfg.MODEL.FCOS = CN()
# the number of foreground classes.
cfg.MODEL.FCOS.NUM_CLASSES = 80
cfg.MODEL.FCOS.IN_FEATURES = ["p3", "p4", "p5", "p6", "p7"]
cfg.MODEL.FCOS.NUM_CONVS = 4
cfg.MODEL.FCOS.HEAD_NORM = "GN"
# inference parameters
cfg.MODEL.FCOS.SCORE_THRESH_TEST = 0.04
cfg.MODEL.FCOS.TOPK_CANDIDATES_TEST = 1000
cfg.MODEL.FCOS.NMS_THRESH_TEST = 0.6
# Focal loss parameters
cfg.MODEL.FCOS.FOCAL_LOSS_ALPHA = 0.25
cfg.MODEL.FCOS.FOCAL_LOSS_GAMMA = 2.0
def add_fbnet_default_configs(_C):
""" FBNet options and default values
"""
_C.MODEL.FBNET = CN()
_C.MODEL.FBNET.ARCH = "default"
# custom arch
_C.MODEL.FBNET.ARCH_DEF = ""
_C.MODEL.FBNET.BN_TYPE = "bn"
_C.MODEL.FBNET.NUM_GROUPS = 32 # for gn usage only
_C.MODEL.FBNET.SCALE_FACTOR = 1.0
# the output channels will be divisible by WIDTH_DIVISOR
_C.MODEL.FBNET.WIDTH_DIVISOR = 1
_C.MODEL.FBNET.DW_CONV_SKIP_BN = True
_C.MODEL.FBNET.DW_CONV_SKIP_RELU = True
# > 0 scale, == 0 skip, < 0 same dimension
_C.MODEL.FBNET.DET_HEAD_LAST_SCALE = 1.0
_C.MODEL.FBNET.DET_HEAD_BLOCKS = []
# overwrite the stride for the head, 0 to use original value
_C.MODEL.FBNET.DET_HEAD_STRIDE = 0
# > 0 scale, == 0 skip, < 0 same dimension
_C.MODEL.FBNET.KPTS_HEAD_LAST_SCALE = 0.0
_C.MODEL.FBNET.KPTS_HEAD_BLOCKS = []
# overwrite the stride for the head, 0 to use original value
_C.MODEL.FBNET.KPTS_HEAD_STRIDE = 0
# > 0 scale, == 0 skip, < 0 same dimension
_C.MODEL.FBNET.MASK_HEAD_LAST_SCALE = 0.0
_C.MODEL.FBNET.MASK_HEAD_BLOCKS = []
# overwrite the stride for the head, 0 to use original value
_C.MODEL.FBNET.MASK_HEAD_STRIDE = 0
# 0 to use all blocks defined in arch_def
_C.MODEL.FBNET.RPN_HEAD_BLOCKS = 0
_C.MODEL.FBNET.RPN_BN_TYPE = ""
# number of channels input to trunk
_C.MODEL.FBNET.STEM_IN_CHANNELS = 3
def add_subclass_configs(cfg):
_C = cfg
_C.MODEL.SUBCLASS = CN()
_C.MODEL.SUBCLASS.SUBCLASS_ON = False
_C.MODEL.SUBCLASS.NUM_SUBCLASSES = 0 # must be set
_C.MODEL.SUBCLASS.NUM_LAYERS = 1
def get_default_cfg():
_C = super(GeneralizedRCNNRunner,
GeneralizedRCNNRunner).get_default_cfg()
_C.EXPORT_CAFFE2 = CN()
_C.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT = False
return _C
def add_abnormal_checker_configs(_C):
_C.ABNORMAL_CHECKER = CN()
# check and log the iteration with bad losses if enabled
_C.ABNORMAL_CHECKER.ENABLED = False
def get_default_cfg(self):
_C = super().get_default_cfg()
add_detr_config(_C)
_C.MODEL.DETR = CN(_C.MODEL.DETR)
return _C
