ricvolpi / adversarial-feature-augmentation Goto Github PK

Thanks for your nice paper and works.
I have some questions about contents of paper and implementation.

1. I don`t understand that why to use Feature generator that generates features that are similar to original source features instead of source features.

2. How can I draw t-SNE plots after make features using your code(with train_feature_extractor function)
   Can you tell me where`s the reference codes or yours?

thanks!

when I train the generator of step 1. I find that the discriminator is powerful .the g_loss will be 1 and the d_loss will be 0. How could I do

It's a nice work of DIFA.
Recently, I need to evaluate my own DA model on NYUD cropped datasets. And I follow the description of Tzeng[1] and tried to crop object in RGB and HHA images. However, the number of images in each category is not match to the discription of NYUD in [1] and experiments part of DIFA. Could you give me some suggests about the NYUD?

Many thanks.
[1]Tzeng, Eric, Judy Hoffman, Kate Saenko and Trevor Darrell. “Adversarial Discriminative Domain Adaptation.” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017): 2962-2971.

Thank you for your work!
I'd like to know if it's possible to apply this setting to other domain adaptation datasets, say Office-31 or ImageCLEF. Given that the training images are larger (256x256 etc.), how should I implement those feature extractor, generator etc.

Many thanks.

Hi there, I came across your work and was truly impressed. However could your update a Pytorch version of this work? Cause it is so hard for me to understand the TF 1.3.0 version and the new python version do not support the TF 1.3.0 anymore.

I trained a encoder $E_s$ and classifier $C$ on mnist with the final accuracy about 98%, and I computed the $APs$ using the method in the paper.

1. The $APs$ of the features obtained using $E_s$ is nearly 60000 which is the size of training data.
2. Besides, I trained a feature generator $S$ using $E_s$ and CGAN. After finishing the training, I feed the $S$ with 60000 noises. However the $APs$ of the features generated is about 1800.

That is to say the above results is in contradiction with the results in the paper. But I have not found any bugs in my code, I want to learn that have you ever got such results.

Thanks!