princeton-vl/CornerNet-Lite

The accuracy and speed of CornerNet-Squeeze

jerrywgz opened this issue · 3 comments

jerrywgz commented

I have trained CornerNet-Squeeze on P40 4gpu with default config and met some problems:

  1. The training process costed around 10 days and I got mAP=33.6, but the released mAP
    is 34.4. Is there any trick in training?
  2. The test speed is 17iter/s which is much slower than yolov3. I test them on same environment.
jerrywgz commented

Besides, I found you mentioned that CornerNet-Squeeze costed 2 weeks to train. Which gpu did you use in training?

heilaw commented
  1. We trained the network with 4 1080Tis. We did not use any tricks.
  2. Our experiments were done on machines with 1080Tis. We don't have access to P40 GPUs so I am not sure why slower inference speed was observed on P40. If you don't mind, can you share more details about your setup, such as the version of your Python, PyTorch and hardwares? This may help us understand the issue better.

BTW, Merry Christmas!

rebotnix commented

@jerrywgz sounds for me like an cuda issue. Please provide Python, PyTorch and CUDA, Cudnn version and I can compare it to ours. 2 weeks is definitely too long for the training process.