pysot中比较有价值的issue

2019年11月1日 下午4:33

Pysot中比较有价值的issue

• 训练的问题 · Issue #135 · STVIR/pysot · GitHub
  • 为什么8块和16块的GPU在训练的参数不同？
  • 为什么针对不同测试数据集，训练的参数不同？
• 能解释下对一张图像shift意味着什么吗？
  • 至于central bias 原因，是因为 0-padding引起的，导致网络认为受 0-padding 的 feature 都是对应没有目标的区域 （如果没有 shift 的话），于是网络就倾向于认为 中心的feature是对应有目标的 （因为中心的feature不会被0-padding影响，或收到的影响相对边缘的feature弱）
• 想请教一下：COCO2017数据集预处理的图像仿射变换后和train2017.json的truth box坐标的关系
• Why the final hot-map shape is 25x25? · Issue #203 · STVIR/pysot · GitHub
• Discussion of Central Bias in Tracking with Deepnet · Issue #176 · STVIR/pysot · GitHub
• 想请教一下：COCO2017数据集预处理的图像仿射变换后和train2017.json的truth box坐标的关系？ 急，在线等~ · Issue #163 · STVIR/pysot · GitHub
  • SiameseFC tracker
• A problem about TRAIN.EXEMPLAR_SIZE and TRAIN.SEARCH_SIZE · Issue #164 · STVIR/pysot · GitHub
  • RuntimeError: sizes must be non-negative
• checkpoint_e10.pth · Issue #80 · STVIR/pysot · GitHub
  • why after checkpoint_e10.pth, it load model and train?
    • There are two stages in SiamRPN++ training. The first 10 epochs only trains head(RPN/MASK). The other 10 epochs fine tunes backbone. So there are some difference in optimizer.
  • how do you fine-tuning in detail?
    • All of our experiments are using the same datasets, Youtube+VID+cooc+imageNetDet.
    • https://github.com/STVIR/pysot/blob/master/experiments/siamrpn_alex_dwxcorr_16gpu/config.yaml#L8
    • That is an example of finetuning, where you can define when and what to fine-tune.
  • As for SiamMask, currently pysot do not support training yet.
  • why Fine-tune sometimes won’t bring performance gain ？
    • The most gain from fine-tune is that regression becomes better after fine-tune. However, training data only contains few categories compares to ImageNet. The robustness will be a little worse after fine-tuning.
  • If I want to continue train at checkpoint_e9.pth
    • config.yaml
      • START_EPOCH: 10
      • RESUME: snapshot/checkpoint_e10.pth
• batchsize=16 · Issue #156 · STVIR/pysot · GitHub
  • when using batchsize=32, the memory (11G) may not be enough from epoch 11. When fine-tuning the backbone network from epoch 11, each GPU uses about 8G of memory (2GPUs, batchsize=24. r50-l234)
• training · Issue #42 · STVIR/pysot · GitHub
• Train on pretrained baseline models · Issue #37 · STVIR/pysot · GitHub
  • In config.yaml, at the same level of RESUME, add a PRETRAINED: “path_to_pretain/model” will use a pretained model.
• Question about negative pair training · Issue #152 · STVIR/pysot · GitHub
• some questions on test and training · Issue #62 · STVIR/pysot · GitHub
  • 16 GPUs takes about 12hrs.
  • The siamrpn_alex_dwxcorr is used for inference. siamrpn_alex_dwxcorr_16gpu is used for training
• Discussion of design of penalty · Issue #158 · STVIR/pysot · GitHub
  • Yeah, it’s tricky. We mention it in the first paper(SiamRPN). We penalize scores whose scale or aspect ratio varies too much compared to the previous.
  • But unfortunately, it is a hyper-parameter. Currently we only perform hyper-parameter searching on it. We haven’t found a way to calculate it. The same as window_influence.
• problem about modify the EXEMPLAR_SIZE？ · Issue #167 · STVIR/pysot · GitHub
  • I try to train the model in my own dataset. And I have change the Base Size and Exemplar Size. however it is strange that the loc_loss is 0 through the all train process. Why is it?
  • Maybe there is no positive examples (No anchor whose iou with gt > threshold). I think there is something wrong about your image size and anchor setting.
• train siamrpn_alex failed · Issue #85 · STVIR/pysot · GitHub
  • If you want a ImageNet pretrained model to reimplement our SiamRPN’s performance, use TRAIN.md . If you want to finetune from our pretrained models (our tracking pretrained model) on other dataset, use the link provided in ModelZoo.
  • If only 4 gpu to train, should I change learning rate corresponding？
    • You don’t need to change the learning rate, the default reduce is sum.
• Resume bug： ValueError: loaded state dict contains a parameter group that doesn’t match the size of optimizer’s group · Issue #92 · STVIR/pysot · GitHub
• AttributeError: ‘NoneType’ object has no attribute ‘shape’ · Issue #72 · STVIR/pysot · GitHub
  • NoneType means image not exists. Check your image path.
• train siamrpn_alex failed · Issue #85 · STVIR/pysot · GitHub
  • If you want a ImageNet pretrained model to reimplement our SiamRPN’s performance, use TRAIN.md . If you want to finetune from our pretrained models (our tracking pretrained model) on other dataset, use the link provided in ModelZoo.
• 框回归后处理问题 · Issue #228 · STVIR/pysot · GitHub
  • tracker里生成的anchor是以255255的中心点为参考的，所以回归后也是bbox也是相对于255255中心点的坐标。