Repository navigation
Train from scratch #14
Description
Activity
Hi @haildceu1,
Sorry for the late reply. The model checkpoints were saved after training for 30 epochs. It's normal for the model to perform poorly during the initial epochs. The performance typically starts improving after the 5th epoch.
我的训练之后TP=0，一个正样本都没有，是因为数据标注的原因吗？还是说另有原因？Excuse me, after my training, TP=0, there is no positive sample, is it because of the data labeling? Or is there another reason?
这个解决了，就是轮数少了，还有就是那个推理的csv的帧数问题帮忙看一下
@lzalzalza 你是多少轮次之后会好点
@lzalzalza 你是多少轮次之后会好点
将近20轮好点了，还有就是您做过推理吗，推理出来的帧坐标都是按顺序排的吗，为什么我最后倒数第六帧会跑到最后一帧后面。。。
你好，我这里自己标注的数据集还是tp是0，但是我跑v1缺效果好，跑v3效果就是这样，这是怎么回事呢@lzalzalza 嗨！朋友，请问你训练了多久，我训练2个epoch花费了4天？用的4090，你也是训练了很长时间吗？
@lzalzalzaHi! Friend, how long have you been training, I spent 4 days training 2 epochs? With 4090, have you also been training for a long time?
there are alot of training optimisations you can make with the train.py module. I'm using a 5090 on vast.ai doing an epoch in around 30 mins. Also downscaling the dataset to 522x288 (as that's the resolution that the model sees) can help alot with the size of the dataset.
May I ask how many epochs did you train to obtain such high metrics results? I was using your checkpoints to train tennis dataset and it only uses 2 epochs to reach max accuracy, but when training from scratch the first few eopchs showed poor results.