tangchen2

PNet cls_loss问题

1

作者您好，我看了一下关于cls_ohem部分的代码 label_prob相当于对gt_label为1的pos sample 取第二个cls_prob，对其余label取第一个cls_prob 然后 -tf.log(label_prob) 对于pos sample而言没错，但对于neg sample而言，其应该是 -tf.log(1 - label_prob) 但代码中没有实现 -tf.log(1- label_prob)的部分 下面的valid_inds 是将pos sample和neg sample都筛选出来的 所以和loss相乘的时候 并没有将neg sample的 -tf.log(label_prob)消除，所以我认为这里的loss计算会有一些问题 还望大家讨论一下

when the network process the attention match rnn, in the first loop, it actually process question - passage attention. In the function 'attention_rnn' when it calls attention, the outputs_ is...

Variable G_conv_mnist/fully_connected/weights already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope? Originally defined at: File "C:\Users\tc\Anaconda3\envs\keras-gpu\lib\site-packages\tensorflow\python\framework\ops.py", line 1718, in __init__ self._traceback = self._graph._extract_stack() # pylint: disable=protected-access...

卷积BN融合

2

您好请教一下，ResNetV2这种BN-Relu-Conv的结构，是不是不能把所有的conv和bn都融合，比如channel selection之前的bn，他前一个conv应该是上一个shortcut两个支路相加的结果，这种情况就不能融合对嘛？

set num threads

1

if i didnt't use "set_num_threads", hnsw index would use default threads values which i printed(num_threads) were 96, why the default num_threads were 96? is it set according to the system?

bad recall performance

9

i build hnsw index ( cosine distance ) based on 1000w data, however, the recall performance is bad, some query data (cosine distance very small ) not be recalled, while...

Excuse me, i have a question about embeddings, in ur code, u train the word embedding matrix without using the pre-trained embedding, right ?

About Triplet prediction Accuracy

12

I have implemented the code in Pytorch, and trained it with the whole dataset. However the triplet prediction accuracy seems much lower than the paper which reached 81%. My question...

Nice work first! Is it possible for ASLFeat to train on a custom dataset which only consists of image pairs (without camera.txt or other info)? I find ASLFeat loss calculation...