GLIP-BLIP-Object-Detection-VQA / configs /glip_Swin_T_O365_GoldG.yaml
Pinwheel's picture
HF Demo
128757a
MODEL:
META_ARCHITECTURE: "GeneralizedVLRCNN"
WEIGHT: "swin_tiny_patch4_window7_224.pth"
RPN_ONLY: True
RPN_ARCHITECTURE: "VLDYHEAD"
BACKBONE:
CONV_BODY: "SWINT-FPN-RETINANET"
OUT_CHANNELS: 256
FREEZE_CONV_BODY_AT: -1
LANGUAGE_BACKBONE:
FREEZE: False
MODEL_TYPE: "bert-base-uncased" # "roberta-base", "clip"
MASK_SPECIAL: False
RPN:
USE_FPN: True
ANCHOR_SIZES: (64, 128, 256, 512, 1024)
ANCHOR_STRIDE: (8, 16, 32, 64, 128)
ASPECT_RATIOS: (1.0,)
SCALES_PER_OCTAVE: 1
DYHEAD:
CHANNELS: 256
NUM_CONVS: 6
USE_GN: True
USE_DYRELU: True
USE_DFCONV: True
USE_DYFUSE: True
TOPK: 9 # topk for selecting candidate positive samples from each level
SCORE_AGG: "MEAN"
LOG_SCALE: 0.0
FUSE_CONFIG:
EARLY_FUSE_ON: True
TYPE: "MHA-B"
USE_CLASSIFICATION_LOSS: False
USE_TOKEN_LOSS: False
USE_CONTRASTIVE_ALIGN_LOSS: False
CONTRASTIVE_HIDDEN_DIM: 64
USE_DOT_PRODUCT_TOKEN_LOSS: True
USE_FUSED_FEATURES_DOT_PRODUCT: True
USE_LAYER_SCALE: True
CLAMP_MIN_FOR_UNDERFLOW: True
CLAMP_MAX_FOR_OVERFLOW: True
CLAMP_BERTATTN_MIN_FOR_UNDERFLOW: True
CLAMP_BERTATTN_MAX_FOR_OVERFLOW: True
CLAMP_DOT_PRODUCT: True
USE_CHECKPOINT: True
TEST:
DURING_TRAINING: False
IMS_PER_BATCH: 64
# use for grounding model
DATASETS:
TRAIN: ("object365_dt_train", "mixed_train_no_coco", "flickr30k_train", )
TEST: ("coco_2017_val", )
DISABLE_SHUFFLE: False
ADD_DET_PROMPT: False
RANDOM_SAMPLE_NEG: 85
CONTROL_PROB: (0.0, 0.0, 0.5, 0.0)
SEPARATION_TOKENS: ". "
INPUT:
PIXEL_MEAN: [ 103.530, 116.280, 123.675 ]
PIXEL_STD: [ 57.375, 57.120, 58.395 ]
MIN_SIZE_TRAIN: 800
MAX_SIZE_TRAIN: 1333
MIN_SIZE_TEST: 800
MAX_SIZE_TEST: 1333
AUGMENT:
MULT_MIN_SIZE_TRAIN: (480,560,640,720,800)
DATALOADER:
SIZE_DIVISIBILITY: 32
SOLVER:
OPTIMIZER: ADAMW
BASE_LR: 0.0001
LANG_LR: 0.00001
WEIGHT_DECAY: 0.0001
STEPS: (0.67, 0.89)
MAX_EPOCH: 30
IMS_PER_BATCH: 64
WARMUP_ITERS: 2000
WARMUP_FACTOR: 0.001
USE_AMP: True
MODEL_EMA: 0.999
FIND_UNUSED_PARAMETERS: False
CLIP_GRADIENTS:
ENABLED: True
CLIP_TYPE: "full_model"
CLIP_VALUE: 1.0
NORM_TYPE: 2.0