yanbeic / semi-memory Goto Github PK

Thanks for sharing the code.
In the paper, the memory module is no longer needed during test. Why? why not use the predicted results of memory module?

Hi, have you ever run your proposed model on the cifar10 dataset with fewer labeled data, eg 25 per class? I run the experiment but the test accuracy is only 10%, which is the same as the random guess.

BTW, I run the code on cifar10 with 400 labeled data per class, the test accuracy is 86.48%, which is about 1.5% lower than that in the original paper. So the running environment is ok.

Hi, thanks for your code, it's elegant, and I learned a lot from it,
I have some questions when I read your paper,

1. I noticed that you do a lot of data argumentation when training, and I wonder how much this impacts the performance in semi-supervised learning?
2. In my research field, I can not do data argument for samples, and I just have a few like one or five samples per class, I wonder the keys and values define in memory could learn the semi-supervised, and how could we guarantee the memory updated with just very few labeled samples? think about this, in extra situation, we just have one sample, and I update the keys and values with this only sample, please asking your advice may this work?

Thank you.
Best wishes.

Hi and thanks for the great work!
I noticed after running the train_svhn_semi.sh script that test accuracy is always around zero percent.
The same problem happens again if I train with the supervised setting (this time prediction is always 5 for some reason).
In both case training reaches 98/99% accuracy after only 2 epochs which seems a bit strange.
I'm using tensorflow 1.9.0 and python 2.7.6

Hi,v@yanbeic Thank for the code.

While running CIFAR10 bash script I am facing the following issue:

model_checkpoint_path = ckpt.model_checkpoint_path
AttributeError: 'NoneType' object has no attribute 'model_checkpoint_path'

Could you help me out?

Hi

Thanks for the wonderful work. I found it to be a great read and very easy to understand.

(1) I am wondering what loss function did you use for the key value updates?
based on eqn. (3) it seems that mean square error loss has been utilized like
loss_k_j = sum_i=1^n_j (k_j - x_i)^2
Is there any particular reason that 1/(n_j + 1) has been selected instead of n_j ?

(2) I am also wondering if the MND and ME loss was simply defined for the unlabeled portion of the data, how does the performance degrade ?

(3) lastly what happens if instead of updating the key and values after every epoch, we simply average out the intermediate representations and the softmax of the labeled data to re-define the key and value pairs ?

(4) Any plans to release the code in Pytorch ? I am not very familiar with tensorflow but would like to understand your method more by studying the code.

any clarifications will be helpful.

Thanks
Devraj

Thanks for sharing.

I tried python3.5 to convert the images into tfrecords.
But seems the codes are not supporting python3.5.
And it works well under the condition of python2.7,
I suggest add the python version in the tutorial setting.

Thanks a lot.