Open ZhiyuanChen opened 4 years ago
jianjiandandande
commented
4 years ago More recently, I've tried to reproduce siamrpn++ and make small changes to it, but the results of the reproduction are slightly different from the original paper. So I don't know what to do. And sometimes the total number of training data is different, I don't know why@ZhiyuanChen
htwdls
commented
4 years ago hi, i want to know where to download 'alexnet-bn.pth'? there is only alexnet.pth in the Google Drive.
WhXl
commented
2 years ago 首先感谢你开源的工作！
我在训练的过程中遇到了一些问题
我使用命令python -m torch.distributed.launch --nproc_per_node=1 --master_port=2333 ./tools/train.py --cfg ./experiments/siamrpn_r50_l234_dwxcorr_8gpu/config.yaml训练coco数据集中的val2017;
（1）我想知道训练日志中的
[2022-01-25 16:16:53,492-rk0-train.py#249] Epoch: [1][2420/21428] lr: 0.001000 batch_time: 0.669450 (0.665431) data_time: 0.000064 (0.000074) cls_loss: 0.091744 (0.071709) loc_loss: 0.279706 (0.232551) total_loss: 0.427391 (0.350770) [2022-01-25 16:16:53,493-rk0-log_helper.py#105] Progress: 2420 / 428560 [0%], Speed: 0.665 s/iter, ETA 3:06:46 (D:H:M)
中的Epoch:1中的21428是是如何计算出来的？因为我的coco数据集中的val2017只有5000张图像，剪切的crop511/val2017为73,562张剪切图像！
（2）我想知道如果训练自己的数据集，单个类别，目标占比较小，在制作数据集方面你有什么建议吗？
（3）你是否考虑过将模型部署在jetson 开发板上，使用tensorrt，是否有这方面的开源打算？
logs.txt
xumengmeng-97
commented
2 years ago 你好，请问直接利用siamrpn_mobilev2_l234_dwxcorr.yaml然后train.py能复现预训练模型的精度出来嘛，我这边复现出来精度很差，不清楚什么原因
fujetson
commented
2 years ago Hey, where can i find the code for the distractors priority and online learning? i cant seem to find where the distractors score are calculated. Thanks !
What's new?
We will be using PySOT Discussions for answering all questions (how to use the code), and leave issues for all working staff (how to edite the code) only. Please do NOT create any issue which is neither bug report nor a work in progress. This also means we will be converting all future, past and present "discussions", questions in issues to discussions. The
useageandpaperlabel is now deprecated, and will be removed in the near future. Enhancement proposals should be discussed in discussion first and move to issues should it be accepted. Enjoy~什么是新的？
我们将使用PySOT Discussions来回答所有的问题（如何使用代码），并让issues只承担与工作有关的事宜（如何修改代码）。请不要创建任何并非bug报告或者工作的议题。这也代表着我们会将所有未来、过去以及现在的在issues当中的讨论和问题转换到Discussions当中。
useage和paper标签现在被停用，并将在之后被移除。改进提案应当现在Discussions中讨论，当其被接受后再转换到issues当中。享受吧~