Skip to content

Latest commit

 

History

History
53 lines (41 loc) · 1.85 KB

OpenVINO.md

File metadata and controls

53 lines (41 loc) · 1.85 KB
Raw

Train Instance-Segmentation Models from Open Model Zoo

instance-segmentation-security-0010

Description

$ python3 tools/train.py \
    --dataset coco2017 \
    --max_image_size 800 1333 \
    --bs 16 \
    --bs_per_gpu 1 \
    --lr 0.02 \
    --max_iter 180000 \
    --drop_lr 120000 160000 \
    --model segmentoly.rcnn.model_zoo.resnet_panet_mask_rcnn.ResNeXt101PANetMaskRCNN \
    --load_backbone data/pretrained_models/converted/imagenet/detectron/resnext101.pth

NOTE: This model is trained on 4 P100.

instance-segmentation-security-0083

Description

$ python3 tools/train.py \
    --dataset coco2017 \
    --max_image_size 480 640 \
    --bs 16 \
    --bs_per_gpu 4 \
    --lr 0.02 \
    --max_iter 90000 \
    --drop_lr 60000 80000 \
    --model segmentoly.rcnn.model_zoo.resnet_fpn_mask_rcnn.ResNeXt152FPNMaskRCNN \
    --load_backbone data/pretrained_models/converted/imagenet/detectron/resnext152.pth

NOTE: This model is trained on 2 P100.

instance-segmentation-security-0050

Description

To train from ImageNet weights, run the command below:

$ python3 tools/train_0050.py

Find the fine-tuning sample in tools/finetune_0050.py.

NOTE: Download checkpoints via tools/download_pretrained_weights.py first. Before running, make necessary changes to learning rate, batch size, number of training steps or other training parameters there.