Skip to content

Latest commit

 

History

History
99 lines (89 loc) · 3.89 KB

experiment.md

File metadata and controls

99 lines (89 loc) · 3.89 KB
Raw

vgg with BN dataset=oxfordhand

train/resume (训练中evaluate是单gpu)

one gpu:
CUDA_VISIBLE_DEVICES="2" python train.py --config-file configs/vgg_bn_ssd300_hand.yaml
tow gpu：
export NGPUS=2
CUDA_VISIBLE_DEVICES="2,3" python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --config-file configs/vgg_bn_ssd300_hand.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000

evaluate (只支持单gpu)

CUDA_VISIBLE_DEVICES="2" python test.py --config-file configs/vgg_bn_ssd300_hand.yaml TEST.BN_FUSE True

mAP:77.64

demo

CUDA_VISIBLE_DEVICES="2" python demo.py --config-file configs/vgg_bn_ssd300_hand.yaml --ckpt /path_to/model_002500.pth --dataset_type oxfordhand --score_threshold 0.4

mobile_netv2 dataset=voc

train/resume

export NGPUS=2
CUDA_VISIBLE_DEVICES="2,3" python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --config-file configs/mobile_v2_ssd_voc0712.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000

mAP:model_final-->70.66

evaluate (只支持单gpu)

CUDA_VISIBLE_DEVICES="2" python test.py --config-file configs/mobile_v2_ssd_voc0712.yaml TEST.BN_FUSE True

mobile_netv2 dataset=hand prune=normal_sparse (稀疏化需要轮数多一些训练，才会容易剪枝)

train/resume

one gpu:
CUDA_VISIBLE_DEVICES="3" python train.py --config-file configs/mobile_v2_ssd_hand_normal_sparse.yaml
two:
export NGPUS=2
CUDA_VISIBLE_DEVICES="2,3" python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --config-file configs/mobile_v2_ssd_hand_normal_sparse.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000

normal prune

CUDA_VISIBLE_DEVICES="3" python prune.py --config-file configs/mobile_v2_ssd_hand_normal_sparse.yaml --regular 0 --percent 0.1 --quick 0 --model model_final.pth

mobile_netv2 dataset=hand prune=shortcut_sparse

train/resume

one gpu:
CUDA_VISIBLE_DEVICES="2" python train.py --config-file configs/mobile_v2_ssd_hand_shortcut_sparse.yaml
two:
export NGPUS=2
CUDA_VISIBLE_DEVICES="2,3" python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --config-file configs/mobile_v2_ssd_hand_shortcut_sparse.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000

shortcut prune

CUDA_VISIBLE_DEVICES="3" python prune.py --config-file configs/mobile_v2_ssd_hand_shortcut_sparse.yaml --percent 0.2 --quick 0 --model model_final.pth

voc(vgg with BN) use_07_metric=False

train/resume

one_gpu:
CUDA_VISIBLE_DEVICES="2" python train.py --config-file configs/vgg_bn_ssd300_voc0712.yaml
two_gpu:
export NGPUS=2
CUDA_VISIBLE_DEVICES="2,3" python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --config-file configs/vgg_bn_ssd300_voc0712.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000
four_gpu:
export NGPUS=4
python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --config-file configs/vgg_bn_ssd300_voc0712.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000

evaluate(one gpu)

CUDA_VISIBLE_DEVICES="2" python test.py --config-file configs/vgg_bn_ssd300_voc0712.yaml

mAP:79.01

voc(vgg fpga) use_07_metric=False

train/resume

one_gpu:
CUDA_VISIBLE_DEVICES="2" python train.py --config-file configs/vgg_ssd300_voc0712_fpga.yaml
two_gpu:
export NGPUS=2
CUDA_VISIBLE_DEVICES="2,3" python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --eval_step -1 --config-file configs/vgg_ssd300_voc0712_fpga.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000 
four_gpu:
export NGPUS=4
python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --config-file configs/vgg_ssd300_voc0712_fpga.yaml SOLVER.WARMUP_FACTOR 0.03333 SOLVER.WARMUP_ITERS 1000

evaluate(one gpu)

CUDA_VISIBLE_DEVICES="2" python test.py --config-file configs/vgg_ssd300_voc0712_fpga.yaml

mAP:77.99