Closed luohao123 closed 1 year ago
Hi, you are on the model zoo page now. If you are on version 1.x, you can refer to the dataset documentation here. Please check your MMPose version and document version and read through our guide before training in MMPose.
@LareinaM I stilll didn't found any example configs to reuse......
Hi @luohao123 , currently the training configs are not migrated from 0.x . Would you like tu create a PR to support it? Just refer to the configs of coco-wholebody.
@Tau-J I'd like to but I am not very famalliar how to start. If you guys can have a wholebody halpe config, I'll help write a 26 one. Currently mmpose really lack some 26 kyepoints model..... (In real world, we need hand, face but not whole133body it's tooooo heavy).
Please consider add at least halp wholebody support since it was supported before.
Thanks for your feedback, we'll consider to support 26-kpt models in RTMPose. For halpe, I suggest having a look on the config of coco-wholebody, you can easily write a halpe version by modifying number of keypoints and annotation files
@Tau-J Would help take a look did I missed anythign at this configuration?
# base dataset settings
dataset_type = 'HalpePoseDataset'
data_mode = 'topdown'
data_root = 'data/halpe/'
# data loaders
train_dataloader = dict(
batch_size=64,
num_workers=10,
persistent_workers=True,
sampler=dict(type='DefaultSampler', shuffle=True),
dataset=dict(
type=dataset_type,
data_root=data_root,
data_mode=data_mode,
ann_file='annotations/halpe_train_v1.json',
data_prefix=dict(img='hico_20160224_det/images/train2015'),
pipeline=train_pipeline,
))
val_dataloader = dict(
batch_size=32,
num_workers=10,
persistent_workers=True,
drop_last=False,
sampler=dict(type='DefaultSampler', shuffle=False, round_up=False),
dataset=dict(
type=dataset_type,
data_root=data_root,
data_mode=data_mode,
ann_file='annotations/halpe_val_v1.json',
bbox_file='person_detection_results/COCO_val2017_detections_AP_H_56_person.json',
data_prefix=dict(img='val2017/'),
test_mode=True,
pipeline=val_pipeline,
))
test_dataloader = val_dataloader
# hooks
default_hooks = dict(
checkpoint=dict(
save_best='halpe-wholebody/AP', rule='greater', max_keep_ckpts=1))
custom_hooks = [
dict(
type='EMAHook',
ema_type='ExpMomentumEMA',
momentum=0.0002,
update_buffers=True,
priority=49),
dict(
type='mmdet.PipelineSwitchHook',
switch_epoch=max_epochs - stage2_num_epochs,
switch_pipeline=train_pipeline_stage2)
]
# evaluators
val_evaluator = dict(
type='CocoWholeBodyMetric',
ann_file=data_root + 'person_detection_results/COCO_val2017_detections_AP_H_56_person.json',
use_area=False,
iou_type='keypoints_crowd',
prefix='crowdpose')
test_evaluator = val_evaluator
where should I insert a KeypointConverter
Prerequisite
Environment
[Bug] Didn't found any Halpe dataset config......
Reproduces the problem - code sample
[Bug] Didn't found any Halpe dataset config......
Reproduces the problem - command or script
[Bug] Didn't found any Halpe dataset config......
Reproduces the problem - error message
[Bug] Didn't found any Halpe dataset config......
Additional information
[Bug] Didn't found any Halpe dataset config......
Please give some practical suggestions train on Halpe...