Paul Zhang Yixing

Different result compared to sklearn.metric.auc calculated

4

in main.py, function, voc_ap(); I tried to add the following lines of codes there `auc_ap = metrics.auc(rec, prec) print(ap, auc_ap)` However, I found the result is hugely different.

I tried to use python api to quantize the SSD model. I want to make use of plugin, so I copy several files and make a plugin. Following the idea...

AdaRound Example

1

I tried to run the AdaRound Example provided in the sample code. I used the ImageNet validation data downloaded from tf slim However I only achieved accuracy 0.22 if I...

Just open an issue and wanna discuss about it. Now I have encountered this phenomenon. I have a bunch of data from daily communication, well, as you may be realized,...

I tired to use darkon influence to find the influential example for wrong predicted examples. However, one example took about 45 mins to complete. Is it that slow?

Can Hyperparameter Tuning done on local server?

1

## Proposal Summary Just to be curious about the fact if the hyperparameter can be done on the local server? ## Motivation checked the documents, but not much discussions on...

Segmentation Fault after 2 line integration

4

Thank you for helping us making ClearML better! ## Describe the bug After adding clearml and clearml.task, I run the code as per-normal, and it happens to be like ...

I tried to download the whole file, and downloaded CIFAR100 dataset. while I try to run the script: python3 trainer.py --dataset CIFAR100 --batch_size 25 --num_dis_conv 6 --gan_type wgan-gp I waited...

我仿照tensorflow-addons的GIOU写了一个CIOU的 用于VGG16SSD 参考了一下您的代码 有几个问题， 在train.py script 里面， 您为什么在81行， CIOU的计算方式，每个都乘以1.0呢？是为了保持都变成float么 第二个是， 我之前的VGG16SSD用的是smooth L1, 看起来这个loss没问题 但是我一换成DIOU/CIOU， 只要一有距离项目， 整个training很快就爆掉了 因为loss很快就增大了，不知道您有没有遇到过这种情况， 但是单独排除又没有问题 那个clip_by_value是用来防止loss太大爆掉的 我也觉得很奇怪 不知道怎么回事 ` def calculate_ciou(self, b1, b2): ''' Arguments: target (nD...

please noted that :scipy.misc.imsave is deprecated! imsave is deprecated in SciPy 1.0.0, and will be removed in 1.2.0. Use imageio.imwrite instead.