search

lmbxmu / FilterSketch

Pytorch implementation of our paper accepted by IEEE TNNLS, 2021 -- Filter Sketch for Network Pruning
https://arxiv.org/abs/2001.08514
52 stars 7 forks source link

How to resume training from the last model_epoch.pt #1

Open atrah22 opened 4 years ago

atrah22 commented 4 years ago

Hello, I am trying to resume the training from epoch=19 as the last saved model is model_19.pt. However, the training accuracy significantly drops compared to the actual accuracy of model_19.pt I am pruning resnet50v1 model.

I am using remote server and my submitted jobs terminates in 24 hours. I hence need to start the job again from the previously trained model.

BRs, atul

lmbxmu commented 4 years ago

Truly apologize. I haven't notice this since I did not receive any email towards this issue.

Can your provide me more about your training loggers and configurations ?

Besides, my WeChat (lmbcxy) is suggested to be added (if you are using it) so that we can have more convenient communications.

atrah22 commented 4 years ago

Hello, I ended up not pursuing this project anymore. I reproduced ABCPruner github project (https://github.com/lmbxmu/abcpruner) and had no problem running the code. Is ABC Pruner better than FilterSketch in terms of accuracy?

lmbxmu commented 4 years ago

Yes. But the results are not deterministic since the structure search is not fixed.

lmbxmu commented 4 years ago

As I can remember, I did not provide the resuming function in this project. I guess that you finish it by yourself and something is wrong in your codes. If necessary, let me know and I can do this for you.