chenf99

@yatengLG 同时还有一句 Finally we note that α, the weight assigned to the rare class， 这是不是更加说明1−α 才应该赋给background

我明白您那样写的意思，代码很简洁明了，我仔细看了论文，也看了很多别人的解读，α=0.25是赋予正样本的，这样来看在加大背景类的权重，解释是因为衰减部分(1-pt)**γ把作为easy examples的背景类的loss减小得太多了。另外，请问您用focal loss训练的SSD效果比原来的更好吗，我最近在想办法改进SSD的效果。

我用focal loss训练出来的SSD效果要差很多，我是先在VOC数据集上训练好SSD后再在别的数据集上微调，用focal loss达不到之前用OHEM的效果。我想再问一下您除了调大正则化之外还有改别的超参数吗，比如学习率，focal loss的α和γ，以及初始化方式等，谢谢！

@naoto0804 I'm looking forward to your reply, thank you!

Thank you for your reply, I tried it again and I found `322583407.jpg` does exist: ``` cross-domain-detection/datasets$ ls -lat clipart/JPEGImages | grep 322583407.jpg -rw-r--r-- 1 chenfan chenfan 32095 Mar 11...

I used `stat` to check what's wrong and I got this: ``` cross-domain-detection/datasets$ stat dt_pl_clipart/JPEGImages File: 'dt_pl_clipart/JPEGImages' -> './clipart/JPEGImages' Size: 20 Blocks: 0 IO Block: 4096 symbolic link Device: 811h/2065d...

In tensorflow_src/SCN.py, `with open(embedding_file, 'rb') as f: embeddings = pickle.load(f,encoding="bytes")` the 'encoding' parameter is only valid for Python 3, so running this file will get an error Should we run...