in distill, how print teacher model accuracy of data

[Manningchan]

No description provided.

[wilxy]

Thanks for your issue.
You can set the type of val_cfg to mmrazor.SingleTeacherDistillValLoop like this config.
Then you can get the accuracy of the teacher model during training like this log:

{"student.accuracy/top1": 29.884000778198242, "student.accuracy/top5": 55.50400161743164, "step": 1}
{"teacher.accuracy/top1": 73.62600708007812, "teacher.accuracy/top5": 91.59000396728516, "step": 1}

[Manningchan]

Thanks for your issue. You can set the type of val_cfg to mmrazor.SingleTeacherDistillValLoop like this config. Then you can get the accuracy of the teacher model during training like this log:

{"student.accuracy/top1": 29.884000778198242, "student.accuracy/top5": 55.50400161743164, "step": 1}
{"teacher.accuracy/top1": 73.62600708007812, "teacher.accuracy/top5": 91.59000396728516, "step": 1}

thanks for your reply, but i used mmseg，if i add val_cfg，i still can not get the teacher accuracy，and if i add student_recorders and teacher_recorders, it will occurs TypeError: GeneralDistill: SingleTeacherDistiller: init() got an unexpected keyword argument 'student_recorders'. i can not find ‘ConfigurableDistiller‘，

[wilxy]

Please provide your training config and the output of pip list | grep "mmcv\|mmrazor\|^torch".

[Manningchan]

grep "mmcv|mmrazor|^torch
my download mmrazor is master and i see your provide is dev-1.x, can i modify some codes to print the teacher model accuracy in master branch? the config and output of mmcv,mmrazor,torch is below, the model is a mask semantic segmentation model defined by myself

config:
algorithm = dict(
type='GeneralDistill',
architecture=dict(
type='MMSegArchitecture',
model=student,
),
distiller=dict(
type='SingleTeacherDistiller',
teacher=teacher,
teacher_trainable=False,
#student_recorders=dict(
# conv_seg=dict(type='',source='decode_head.conv_seg')),
#teacher_recorders=dict(
# conv_seg=dict(type='',source='decode_head.conv_seg')),
components=[
dict(
student_module='decode_head.conv_seg1',
teacher_module='decode_head.conv_seg1',
losses=[
dict(
type='ChannelWiseDivergence_cls',
name='loss_cwd_logits_cls',
tau=1,
loss_weight=5,
)
]),
dict(
student_module='decode_head.conv_seg2',
teacher_module='decode_head.conv_seg2',
losses=[
dict(
type='ChannelWiseDivergence_mask',
name='loss_cwd_logits_mask',
tau=1,
loss_weight=5,
)
])
]),
)

environment

mmcv-full 1.6.0
mmrazor 0.3.1
torch 1.12.1+cu113
torchaudio 0.12.1+cu113
torchvision 0.13.1+cu113

[wilxy]

Your MMRazor version is 0.3.1, which does not support the recorder mechanism and SingleTeacherDistillValLoop.
Please update the version to 1.0.0rc1 by using the dev-1.x branch.

[wilxy]

Please reopen it if there are any problems in the future.