search

open-mmlab / mmpretrain

OpenMMLab Pre-training Toolbox and Benchmark
https://mmpretrain.readthedocs.io/en/latest/
Apache License 2.0
3.38k stars 1.05k forks source link

The epoch number of ResNet and SeResNet are different #344

Closed J-JunChen closed 3 years ago

J-JunChen commented 3 years ago

Hey, author! I found the epoch number of ResNet and SeResNet are different in resnet50_b32x8_imagenet.py and seresnet50_b32x8_imagenet.py.

the following code in resnet50_b32x8_imagenet.py is:

_base_=[
...
    '../_base_/schedules/imagenet_bs256.py'
...
]

it has 100 epochs , however in seresnet50_b32x8_imagenet.py:

_base_=[
...
    '../_base_/schedules/imagenet_bs256_140e.py',
...
]

it has 140 epochs. The two models with different epoch number can be compared equally?

mzr1996 commented 3 years ago

Different networks need different epochs to converge. And usually, we only concern with the final performance of the model.