zengqunzhao/ma-net

[TIP'21] Learning Deep Global Multi-scale and Local Attention Features for Facial Expression Recognition in the Wild

ma-net's Issues

CAER-S

你好，我最近看了您的论文，想请教一下你在用CEAR-S做训练时是直接将原图放进去还是把人脸部分提取出来做的？

Why the num_classes in fc_1 is 12666?

Looking forward to your reply.

在沒有預訓練的情況下對 RAF-DB 進行評估

您好，我在使用您提供的代碼及步驟進行實驗時，在 RAF-DB 上能夠如論文得到 88% 以上的準確率，但在不使用預訓練模型時無法重現論文中 TABLE II 86.34% 的效果，請問您在沒有預訓練的情況下訓練時有做什麼改動嗎？謝謝！

关于训练后模型的使用问题

您好，我在加载了您的预训练模型并且自己训练后，得到了多个checkpoint，如果我想使用我自己训练得到的checkpoint去对新的图片进行输出和标注，需要怎么做呢

Some questions about the difference of your model and your provided code:

Hello Mr.Zhao, thanks for your work in facial expression recognition. When I try to read your paper and understand the code that you provided, I met with this issues:
in Fig.2, the feature map was divided by channel into four groups. However, the left stage's first block(called block 1) enter one 3 x 3 conv, and then it is concatenated with block2. And in the right stage, the block 1 was concatenated in the last. in other words, it is not absolutely symmetry. However, in your code, we find that the block 1 was the first added in both stage, which means it isn't the same as the paper told. Can you explain it clearer?

model = torch.load('./checkpoint/Pretrained_on_MSCeleb.pth.tar') error

when I run the code :checkpoint = torch.load('./checkpoint/Pretrained_on_MSCeleb.pth.tar')
then I got:
Traceback (most recent call last):
File "/home/hyj/桌面/master_projects/MA-Net-main/just_test.py", line 5, in
model = torch.load('./checkpoint/Pretrained_on_MSCeleb.pth.tar')
File "/home/hyj/anaconda3/envs/hyj_env/lib/python3.8/site-packages/torch/serialization.py", line 608, in load
return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
File "/home/hyj/anaconda3/envs/hyj_env/lib/python3.8/site-packages/torch/serialization.py", line 787, in _legacy_load
result = unpickler.load()
AttributeError: Can't get attribute 'RecorderMeter' on <module 'main' from '/home/hyj/桌面/master_projects/MA-Net-main/just_test.py'>

Test on an image

Excellent work. But please, I am new in this field. So I wonder if I can use main.py to test facial expression on any face from other datasets.
Thanks in advance.

数据集的问题

你好，请问能提供一份你使用的AffectNet、Sfew数据集和FED-RO数据集吗？非常感谢！！

how can I generate the heatmap-liked pictures?

in your article,I notice you use the heatmap-liked pictures to show the where the net is focusing.
it's cool, however, when I tried to use pytorch-grad-cam to do this,I found out that the net has 2 fc layers(different from the nomal net like resnet50) and i failed to generate the pictures.
Can you tell me your method?

Data preprocessing

Hello, thank you for your kind words. I would like to know if it is necessary to use RetinaFace for face image detection and alignment on each dataset before running the program. I look forward to your reply.

pre-training model is damaged

Hello, the contents of the compressed package of pre-training model that I downloaded through Google_drive is damaged. Do you have any other way to download it? Looking for your reply, thank you.

Has anything changed of hyper-pararmeter on the Affectnet dataset?

Hello, it is a nice work.
But I can't reproduce the results of paper in Affectnet using the setting of this repo.
Could you tell me the mainly changed setting?
Ps . Imbalanced Sample is using。。。

预训练模型