thinklab-sjtu / dcl_retinanet_tensorflow Goto Github PK

我尝试按照您的描述去复现 DCL ,用DOTA数据集
图片路径为：/home/cdzk/datasets/DOTA/DOTA1.0
代码路径为: /home/cdzk/DCL
python=3.7
cuda=11.1
tensorflow=1.15.1
可是最后只显示了参数量，但没有训练结果？GPU利用率也为0，所以想请问一下，这是否训练成功了呢？

Hi, I see you calculate Flops and Params in your paper. I find some codes on the Internet but cant't run those. Could you please share codes about this? Thank you very much!

Hello and thanks for your hard work!

I want to train the models with a single gpu. Could you please give me some analogy between multi_gpu training and single_gpu configurations?

Looking on this configuration i have the following hyperparameters i need to change

SAVE_WEIGHTS_INTE = 27000*4 
LR = 5e-4
BATCH_SIZE = 1
MAX_ITERATION = SAVE_WEIGHTS_INTE*20

Could you please give some insights as to what values should those hyperparameters change into? Are there any other hyperparameters i need to change to support single gpu training?

Thank you!

How about the inference speed of DCL?

In the "densely_coded_label.py", the decoding process is a "for" loop. So for dense anchors, does this way of the "for" loop has a great impact on the inference speed?