Discrepancy in AP50 while Re-implementing MetaRCNN

[mani2002]

I encountered an issue while re-implementing Metarcnn. When finetuning for all classes, I observed a difference of up to 6% in the AP50 values for novel classes. Initially, I utilized 2 GPUs for the implementation. Due to limited GPU access, I adjusted the batch size to 2 and increased the iterations to 36,000 to manage the base training process efficiently.

Here are base training configurations for split1:

"img_norm_cfg = dict(
mean=[103.53, 116.28, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False)
train_multi_pipelines = dict(
query=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(type='Resize', img_scale=(1000, 600), keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
],
support=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='GenerateMask', target_size=(224, 224)),
dict(type='RandomFlip', flip_ratio=0.0),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
])
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1000, 600),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
]
data_root = 'data/VOCdevkit/'
data = dict(
samples_per_gpu=2,
workers_per_gpu=2,
train=dict(
type='NWayKShotDataset',
num_support_ways=15,
num_support_shots=1,
one_support_shot_per_image=True,
num_used_support_shots=200,
save_dataset=False,
dataset=dict(
type='FewShotVOCDataset',
ann_cfg=[
dict(
type='ann_file',
ann_file=
'data/VOCdevkit/VOC2007/ImageSets/Main/trainval.txt'),
dict(
type='ann_file',
ann_file=
'data/VOCdevkit/VOC2012/ImageSets/Main/trainval.txt')
],
img_prefix='data/VOCdevkit/',
multi_pipelines=dict(
query=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(
type='Resize', img_scale=(1000, 600), keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='DefaultFormatBundle'),
dict(
type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
],
support=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='GenerateMask', target_size=(224, 224)),
dict(type='RandomFlip', flip_ratio=0.0),
dict(type='DefaultFormatBundle'),
dict(
type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]),
classes='BASE_CLASSES_SPLIT1',
use_difficult=True,
instance_wise=False,
dataset_name='query_dataset'),
support_dataset=dict(
type='FewShotVOCDataset',
ann_cfg=[
dict(
type='ann_file',
ann_file=
'data/VOCdevkit/VOC2007/ImageSets/Main/trainval.txt'),
dict(
type='ann_file',
ann_file=
'data/VOCdevkit/VOC2012/ImageSets/Main/trainval.txt')
],
img_prefix='data/VOCdevkit/',
multi_pipelines=dict(
query=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(
type='Resize', img_scale=(1000, 600), keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='DefaultFormatBundle'),
dict(
type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
],
support=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='GenerateMask', target_size=(224, 224)),
dict(type='RandomFlip', flip_ratio=0.0),
dict(type='DefaultFormatBundle'),
dict(
type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]),
classes='BASE_CLASSES_SPLIT1',
use_difficult=False,
instance_wise=False,
dataset_name='support_dataset')),
val=dict(
type='FewShotVOCDataset',
ann_cfg=[
dict(
type='ann_file',
ann_file='data/VOCdevkit/VOC2007/ImageSets/Main/test.txt')
],
img_prefix='data/VOCdevkit/',
pipeline=[
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1000, 600),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
],
classes='BASE_CLASSES_SPLIT1'),
test=dict(
type='FewShotVOCDataset',
ann_cfg=[
dict(
type='ann_file',
ann_file='data/VOCdevkit/VOC2007/ImageSets/Main/test.txt')
],
img_prefix='data/VOCdevkit/',
pipeline=[
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1000, 600),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
],
test_mode=True,
classes='BASE_CLASSES_SPLIT1'),
model_init=dict(
copy_from_train_dataset=True,
samples_per_gpu=16,
workers_per_gpu=1,
type='FewShotVOCDataset',
ann_cfg=None,
img_prefix='data/VOCdevkit/',
pipeline=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(
type='Normalize',
mean=[103.53, 116.28, 123.675],
std=[1.0, 1.0, 1.0],
to_rgb=False),
dict(type='GenerateMask', target_size=(224, 224)),
dict(type='RandomFlip', flip_ratio=0.0),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
],
use_difficult=False,
instance_wise=True,
classes='BASE_CLASSES_SPLIT1',
dataset_name='model_init_dataset'))
evaluation = dict(interval=12000, metric='mAP')
optimizer = dict(type='SGD', lr=0.005, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(grad_clip=None)
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=100,
warmup_ratio=0.001,
step=[16000])
runner = dict(type='IterBasedRunner', max_iters=36000)
norm_cfg = dict(type='BN', requires_grad=False)
pretrained = 'open-mmlab://detectron2/resnet101_caffe'
model = dict(
type='MetaRCNN',
pretrained='open-mmlab://detectron2/resnet101_caffe',
backbone=dict(
type='ResNetWithMetaConv',
depth=101,
num_stages=3,
strides=(1, 2, 2),
dilations=(1, 1, 1),
out_indices=(2, ),
frozen_stages=2,
norm_cfg=dict(type='BN', requires_grad=False),
norm_eval=True,
style='caffe'),
rpn_head=dict(
type='RPNHead',
in_channels=1024,
feat_channels=512,
anchor_generator=dict(
type='AnchorGenerator',
scales=[2, 4, 8, 16, 32],
ratios=[0.5, 1.0, 2.0],
scale_major=False,
strides=[16]),
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0.0, 0.0, 0.0, 0.0],
target_stds=[1.0, 1.0, 1.0, 1.0]),
loss_cls=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
roi_head=dict(
type='MetaRCNNRoIHead',
shared_head=dict(
type='MetaRCNNResLayer',
pretrained='open-mmlab://detectron2/resnet101_caffe',
depth=50,
stage=3,
stride=2,
dilation=1,
style='caffe',
norm_cfg=dict(type='BN', requires_grad=False),
norm_eval=True),
bbox_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0),
out_channels=1024,
featmap_strides=[16]),
bbox_head=dict(
type='MetaBBoxHead',
with_avg_pool=False,
roi_feat_size=1,
in_channels=2048,
num_classes=15,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0.0, 0.0, 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', loss_weight=1.0),
num_meta_classes=15,
meta_cls_in_channels=2048,
with_meta_cls_loss=True,
loss_meta=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
aggregation_layer=dict(
type='AggregationLayer',
aggregator_cfgs=[
dict(
type='DotProductAggregator',
in_channels=2048,
with_fc=False)
])),
train_cfg=dict(
rpn=dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.7,
neg_iou_thr=0.3,
min_pos_iou=0.3,
match_low_quality=True,
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_pre=12000,
max_per_img=2000,
nms=dict(type='nms', iou_threshold=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,
match_low_quality=False,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=128,
pos_fraction=0.25,
neg_pos_ub=-1,
add_gt_as_proposals=True),
pos_weight=-1,
debug=False)),
test_cfg=dict(
rpn=dict(
nms_pre=6000,
max_per_img=300,
nms=dict(type='nms', iou_threshold=0.7),
min_bbox_size=0),
rcnn=dict(
score_thr=0.05,
nms=dict(type='nms', iou_threshold=0.3),
max_per_img=100)))
checkpoint_config = dict(interval=12000)
log_config = dict(interval=50, hooks=[dict(type='TextLoggerHook')])
custom_hooks = [dict(type='NumClassCheckHook')]
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
resume_from = None
workflow = [('train', 1)]
use_infinite_sampler = True
seed = 42
work_dir = './work_dirs/meta-rcnn_r101_c4_8xb4_voc-split1_base-training'
gpu_ids = range(0, 2)"

Could you please support in resolving this issue?