CUDA OOM with small gts per image

kye9216789 commented 4 years ago

Hello.

I am training Mask R-CNN for my own dataset using same format MS COCO Dataset. The only difference is the number of classes is 2 (means single class). I only want find a single object in one image, so I set number of gts is just 1. Image size is (600,600) and size_divizer is 32, so True image size is 608. I often encounter CUDA OOM when I use resnext 32x4 backbone.

I read https://github.com/open-mmlab/mmdetection/issues/188, and found that main factor for gpu memory is number of gt, and aspect ratio.

But I have no problem in both factors. I encounter OOM when I use images_per_gpu=2, and 1 is ok.

I modified mask head size from 28 to 128 because my object is very thin, and I needed high resolution in mask head and mask RoIExtractor. RoIExtractor out size is 64.

I don`t understand why it happens. My machine has two of GTX1080ti, with graphic memory 11G Below is my config file.


model = dict(
    type='MaskRCNN',
    pretrained='open-mmlab://resnext101_32x4d',
    backbone=dict(
        type='ResNeXt',
        depth=101,
        groups=32,
        base_width=4,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=-1,
        style='pytorch'),
    neck=dict(
        type='FPN',
        in_channels=[256, 512, 1024, 2048],
        out_channels=256,
        num_outs=5),
    rpn_head=dict(
        type='RPNHead',
        in_channels=256,
        feat_channels=256,
        anchor_scales=[8],
        anchor_ratios=[0.5, 1.0, 2.0],
        anchor_strides=[4, 8, 16, 32, 64],
        target_means=[.0, .0, .0, .0],
        target_stds=[1.0, 1.0, 1.0, 1.0],
        loss_cls=dict(
            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
    bbox_roi_extractor=dict(
        type='SingleRoIExtractor',
        roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
        out_channels=256,
        featmap_strides=[4, 8, 16, 32]),
    bbox_head=dict(
        type='SharedFCBBoxHead',
        num_fcs=2,
        in_channels=256,
        fc_out_channels=1024,
        roi_feat_size=7,
        num_classes=2,
        target_means=[0., 0., 0., 0.],
        target_stds=[0.1, 0.1, 0.2, 0.2],
        reg_class_agnostic=False,
        loss_cls=dict(
            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
        loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
    mask_roi_extractor=dict(
        type='SingleRoIExtractor',
        roi_layer=dict(type='RoIAlign', out_size=64, sample_num=2),
        out_channels=256,
        featmap_strides=[4]),
    mask_head=dict(
        type='FCNMaskHead',
        num_convs=4,
        in_channels=256,
        conv_out_channels=256,
        num_classes=2,
        loss_mask=dict(
            type='CrossEntropyLoss', use_mask=True, loss_weight=1.0)))
# model training and testing settings
train_cfg = dict(
    rpn=dict(
        assigner=dict(
            type='MaxIoUAssigner',
            pos_iou_thr=0.7,
            neg_iou_thr=0.3,
            min_pos_iou=0.3,
            ignore_iof_thr=-1),
        sampler=dict(
            type='RandomSampler',
            num=256,
            pos_fraction=0.5,
            neg_pos_ub=-1,
            add_gt_as_proposals=False),
        allowed_border=0,
        pos_weight=-1,
        debug=False),
    rpn_proposal=dict(
        nms_across_levels=False,
        nms_pre=2000,
        nms_post=2000,
        max_num=2000,
        nms_thr=0.7,
        min_bbox_size=0),
    rcnn=dict(
        assigner=dict(
            type='MaxIoUAssigner',
            pos_iou_thr=0.5,
            neg_iou_thr=0.5,
            min_pos_iou=0.5,
            ignore_iof_thr=-1),
        sampler=dict(
            type='RandomSampler',
            num=512,
            pos_fraction=0.25,
            neg_pos_ub=-1,
            add_gt_as_proposals=True),
        mask_size=128,
        pos_weight=-1,
        debug=False))
test_cfg = dict(
    rpn=dict(
        nms_across_levels=False,
        nms_pre=1000,
        nms_post=1000,
        max_num=1000,
        nms_thr=0.7,
        min_bbox_size=0),
    rcnn=dict(
        score_thr=0.05,
        nms=dict(type='nms', iou_thr=0.5),
        max_per_img=100,
        mask_thr_binary=0.5))```

ZwwWayne commented 4 years ago

The ResNext_x32x4d is already a very big backbone, it is almost the biggest backbone can be used by 1080Ti with Mask R-CNN.
Enlarge the output size of Mask RoI Features will also enlarge the memory usage significantly since mask head is very costly.

kye9216789 commented 4 years ago

The ResNext_x32x4d is already a very big backbone, it is almost the biggest backbone can be used by 1080Ti with Mask R-CNN.

Enlarge the output size of Mask RoI Features will also enlarge the memory usage significantly since mask head is very costly.

Thanks for your answer.

open-mmlab / mmdetection

CUDA OOM with small gts per image #2051