sillytheif

fail to froze the graph

1

Hi, I want to froze the model to conduct unittest. When I run the command "g2p-seq2seq --model_dir model_folder_bre --freeze" There exisits the bug: AssertionError: transformer/parallel_0_5/transformer/body/decoder/layer_0/self_attention/multihead_attention/dot_product_attention/Softmax is not in graph. I...

proble about how to concentrate on center part

4

哲东大佬好。感谢这篇论文对reid align问题的启示，小星星送上 orz 看这篇论文，我有一个小疑惑是：如何保证STN起始 align的时候是，是从center part 开始。我看文中提到是把scale 参数设置成0.8，其他设置成0，是不是就是 把grid的预测网络，那6个参数的 bias初始化成 0.8和0呢？ best wishes

Your work is very interesting. I want to obtain this dataset and conduct my experiments. Could you provide a download link?

At cifar10 0.2 sym noise rate setting, cann't repeat the results reported at paper.

3

The cifar10 0.2 noise rate, I can only get about 91.5% acc useing the code, and I can't find the reason

疑似存在代码实现错误

1

co-train.py中： ptl = batch_prob_l * batch_labels_l + (1 - batch_prob_l) * pl[0] 中pl[0]仅是batch内第一个伪标签，这里应该是和batch内所有的进行累加。 希望能得到您的回复。

about the forward loss function problem

1

Hi, I have some problem about your Forward loss function: def forward_loss(output, target, trans_mat, index = None): # l_{forward}(y, h(x)) = l_{ce}(y, h(x) @ T) outputs = F.softmax(output, dim=1) if...

I conduct many experiments on TVsum dataset，and I find that I cann't get an stable results for each domain. Take VT domain as an example, top-5 mAP varies from 45%~79%...