UX-Decoder/Semantic-SAM

Training questions

alexcbb opened this issue · 5 comments

alexcbb commented

Hello, thank you for your work !

I have some technical questions regarding your training of Semantic-SAM. On how many GPUs did you perform the training and which type of GPU (A100, V100, others) ?
Did you use some advanced parallelism method such as ZeRO or FSDP ? Or did you only use the classical DataDistributedParallel ?

Next to that, did you implement an Automatic Mask Generator script like original work or not at all ?

Thank you in advance for your answers !

fujianhai commented

+1

Leon1207 commented

+1

FengLi-ust commented

Hi, we use 64V100 for training. We did not use the advanced parallelism method. We will release the automatic mask generator later. Thank you.

alexcbb commented

I saw that you released a first version of the demo for automatic mask generation, thank you ! Do you plan on make it possible to choose parameter for the filtering ?

praeclarumjj3 commented

Hi, we use 64V100 for training. We did not use the advanced parallelism method. We will release the automatic mask generator later. Thank you.

Hi @FengLi-ust, thanks for the great work! Could you also share information about how many hours you train on 64 V100 GPUs?