search

chenyilun95 / tf-cpn

Cascaded Pyramid Network for Multi-Person Pose Estimation (CVPR 2018)
MIT License
792 stars 197 forks source link

something about continue_train #50

Closed chenpengf0223 closed 5 years ago

chenpengf0223 commented 5 years ago

Hi, if I want to finetune the 350epoch-snapshot.ckpt model on my own keypoints data, if I need to set the cfg variable "continue_train" to True?

ttdd11 commented 5 years ago

@chenpengf0223 Did you ever fine tune the model with your own keypoints?

chenpengf0223 commented 5 years ago

@ttdd11 yes.

ttdd11 commented 5 years ago

@chenpengf0223 So you loaded the 350 and tuned it with more joints/different joints or just different data? What were your training parameters and how long did you run it? Was the end result pretty good? Just wondering if I should bother tuning or restart the training with my own data (since it takes so long).

chenpengf0223 commented 5 years ago

@ttdd11 I tuned the 350 ckpt with different joints. My training parameters are: lr = 1e-3, lr_gamma = 0.5, lr_dec_epoch = 60, optimizer = 'adam', batch_size = 28, weight_decay = 1e-5. I finally use the model after 50 epochs. I think the training depends on your data, if it is a different distribution with coco and has a large amount, you should train from scratch, otherwise, you can fine tune the model with it.