arsity
commented
2 hours ago Config file is as follows, where I modify GPUS, IMAGES_PER_GPU and WORKERS.
AUTO_RESUME: True
DATA_DIR: ''
GPUS: [0]
LOG_DIR: log
OUTPUT_DIR: output
PRINT_FREQ: 25
VERBOSE: False
MULTIPROCESSING_DISTRIBUTED: True
CUDNN:
BENCHMARK: True
DETERMINISTIC: False
ENABLED: True
DATASET:
DATASET: exlpose_kpt
DATA_FORMAT: png
FLIP: 0.5
INPUT_SIZE: 512
OUTPUT_SIZE: 128
MAX_NUM_PEOPLE: 30
MAX_ROTATION: 30
MAX_SCALE: 1.5
SCALE_TYPE: 'short'
MAX_TRANSLATE: 40
MIN_SCALE: 0.75
NUM_JOINTS: 14
ROOT: '../datasets/ExLPose'
TEST: all
TRAIN: train
OFFSET_RADIUS: 4
SIGMA: 2.0
CENTER_SIGMA: 4.0
BG_WEIGHT: 0.1
ADJUST_ELLA: True
LOSS:
WITH_HEATMAPS_LOSS: True
HEATMAPS_LOSS_FACTOR: 1.0
WITH_OFFSETS_LOSS: True
OFFSETS_LOSS_FACTOR: 0.03
WITH_AE_LOSS: True
AE_LOSS_TYPE: exp
PUSH_LOSS_FACTOR: 0.001
PULL_LOSS_FACTOR: 0.001
MODEL:
SPEC:
FINAL_CONV_KERNEL: 1
PRETRAINED_LAYERS: ['*']
STAGES:
NUM_STAGES: 3
NUM_MODULES:
- 1
- 4
- 3
NUM_BRANCHES:
- 2
- 3
- 4
BLOCK:
- BASIC
- BASIC
- BASIC
NUM_BLOCKS:
- [4, 4]
- [4, 4, 4]
- [4, 4, 4, 4]
NUM_CHANNELS:
- [32, 64]
- [32, 64, 128]
- [32, 64, 128, 256]
FUSE_METHOD:
- SUM
- SUM
- SUM
HEAD_HEATMAP:
BLOCK: BASIC
NUM_BLOCKS: 1
NUM_CHANNELS: 32
DILATION_RATE: 1
HEAD_OFFSET:
BLOCK: ADAPTIVE
NUM_BLOCKS: 2
NUM_CHANNELS_PERKPT: 15
DILATION_RATE: 1
INIT_WEIGHTS: True
NAME: hrnet_main
NUM_JOINTS: 14
PRETRAINED_MAIN: 'model/imagenet/hrnet_w32-36af842e.pth'
PRETRAINED_COMP: 'model/imagenet/hrnet_w32-36af842e.pth'
TAG_PER_JOINT: True
TEST:
FLIP_TEST: True
IMAGES_PER_GPU: 1
MODEL_FILE: ''
SCALE_FACTOR: [1]
NMS_THRE: 0.15
NMS_NUM_THRE: 8
KEYPOINT_THRESHOLD: 0.01
ADJUST_THRESHOLD: 0.05
MAX_ABSORB_DISTANCE: 75
GUASSIAN_KERNEL: 6
DECREASE: 0.9
DETECTION_THRESHOLD: 0.1
TRAIN:
BEGIN_EPOCH: 0
CHECKPOINT: ''
END_EPOCH: 10
GAMMA1: 0.99
GAMMA2: 0.0
IMAGES_PER_GPU: 10
LR: 0.001
LR_FACTOR: 0.1
LR_STEP: [90, 120]
MOMENTUM: 0.9
NESTEROV: False
OPTIMIZER: adam
RESUME: False
SHUFFLE: True
WD: 0.0001
MAX_NUM_CENTERS: 15
TEACHER_THRESHOLD: 0.9
STAGE: 'KA'
WORKERS: 8
In KA stage training,
lunsupterm is always zero in log. When I used this final checkpoint after training, the performance is abnormal.