Closed encore-zhou closed 4 years ago
Hi, I am able to reproduce the results on my machines. However, I do observe some unstable behaviors. With some new experiments, I think the following parameters may give you a more stable performance.
Use batch size 16, and use the following learning rate setting: BASE_LEARNING_RATE = 0.01 or 0.001 FLAGS_LR_DECAY_STEPS = '100, 160, 220, 260, 360' FLAGS_LR_DECAY_RATES = '0.1, 0.1, 0.1, 0.1, 1'
I am able to achieve around 60 mAP@0.25 and 39 mAP@0.5.
What are the results you are getting?
Thank you so much. Let me try again with the new configs.
Close it for now. Feel free to reopen it.
Hi, I try to training the H3DNet on SUN-RGBD dataset(using the default configs), but my reproduced performance is lower than the paper reported. Are there others tricks that I missed?