search

sairin1202 / Semantic-Aware-Attention-Based-Deep-Object-Co-segmentation

Semantic Aware Attention Based Deep Object Co-segmentation
68 stars 17 forks source link

Training is very slow #6

Open sidgairo18 opened 5 years ago

sidgairo18 commented 5 years ago

The training is very slow. Even the evaluate part takes hours to finish.

How much time did it take to train the model?

sairin1202 commented 5 years ago

To accelerate training step, you can fix the parameters in the backbone model or input the features instead of the images. If you have two or more GPUs, you can finish training the model within one day.

sidgairo18 commented 5 years ago

Thank you for your reply. I am currently using a Batch-Size of 20 with 4 GPUs and training the model end to end.

I have 500K pairs for training. It takes 1 day for a single epoch with batch size 20.