Hello, I am trying to reproduce your experiments from scratch on the SemanticKitti. I am running the code on a node with 8x titan. But the performance of mine is lower than yours about 4 points. I have tested your pre-trained models, the performance is higer than mine. I also use the arch.yaml and data.yaml from your pre-trained models and I find the performance is similar to config/arch and config/labels.
test on your pre-trained models: darknet21 (noknn, 47.4) and darknet53 (noknn, 50.4)
test on mine: darknet21 (noknn, 43.7) and darknet53 (noknn, 44.8)
what do you think the main difference between your training and mine?
can you share the training log with me?
abhigoku10
commented
4 years ago
Hello, I am trying to reproduce your experiments from scratch on the SemanticKitti. I am running the code on a node with 8x titan. But the performance of mine is lower than yours about 4 points. I have tested your pre-trained models, the performance is higer than mine. I also use the arch.yaml and data.yaml from your pre-trained models and I find the performance is similar to config/arch and config/labels.
test on your pre-trained models: darknet21 (noknn, 47.4) and darknet53 (noknn, 50.4) test on mine: darknet21 (noknn, 43.7) and darknet53 (noknn, 44.8)
what do you think the main difference between your training and mine? can you share the training log with me?