search

decisionforce / TPN

[CVPR 2020] Temporal Pyramid Network for Action Recognition
https://decisionforce.github.io/TPN/
Apache License 2.0
394 stars 55 forks source link

About tsm101_tpn pretrained model on sth-sth-v2 dataset #19

Closed JR-Wang closed 4 years ago

JR-Wang commented 4 years ago

Hello, may I ask you for your tsm101_tpn pretrained model on sth-sth-v2 dataset? By the way, I am now change config_files/sthv2/tsm_tpn.py as follow to train a tsm101_tpn model on sth-sth-v2-dataset. model = dict( type='TSN2D', backbone=dict( type='ResNet', pretrained='modelzoo://resnet101', depth=101, nsegments=8, out_indices=(2, 3), tsm=True, bn_eval=False, partial_bn=False), As I use 2 gpus, I set my learning rate as 0.0025. However, the top1 precision remains around 56 after 45 epochs training, shall I decay my learning rate? Looking forward to your reply and many thanks.

limbo0000 commented 4 years ago

Hi @JR-Wang,

Yes. you can follow the same learning rate scheduler in our config, which decays lr at 75, 125 epoch respectively.