8000 Ask for some help · Issue #1 · wqingzex/tensorflow-ram · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Ask for some help #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
hhhmoan opened this issue Jul 31, 2016 · 15 comments
Open

Ask for some help #1

hhhmoan opened this issue Jul 31, 2016 · 15 comments

Comments

@hhhmoan
Copy link
hhhmoan commented Jul 31, 2016

thanks for release your code. now i just want to re-implement the RAM in tensorflow. but it always get bad result. can you code achieve the accuary release by the paper “Recurrent models of Visual attention”?

@wqingzex
Copy link
Owner
wqingzex commented Aug 1, 2016

No, I didn't, maybe because I didn't use rnn.

@hhhmoan
Copy link
Author
hhhmoan commented Aug 1, 2016

thanks for your answer,but why you use your grad_reinforce function but not just
tf.graident((l_mean-l_sample)**2/balabala * (R-tf.reduce_mean(R))). Is there difference?
and in your ram.py the line 294, there really need a nagetive?

@wqingzex
Copy link
Owner
wqingzex commented Aug 2, 2016

@hhhmoan the method is called "Policy gradient methods", one implementation of attention model, you can read this for more details, another implementation can be found in here

maybe, I was wrong, if so ,please tell me

@hhhmoan
Copy link
Author
hhhmoan commented Aug 15, 2016

when you train your model in translate clutter mnist, how much image do you use? and your gradient method is right,but in tensorflow, maybe you code can be more easier. and in tensorflow , the randomnormal function also have gradients,maybe you should use stop_gradient. and really thank you release your code .

@wqingzex
Copy link
Owner

@hhhmoan my colleague share the training data with me, ~5k per digit.

do you mean tf.random_normal function? although the function also gradients, but i think only the trainable variables need to be compute, so if you have some good ideas, please share it with me, also including the stop_gradient function

@hhhmoan
Copy link
Author
hhhmoan commented Aug 16, 2016
< 8000 /tbody>

yes, Maybe you are right .And,could your release the code in translate clutter mnist,i want to try it by myself.If I've been a little abrupt, then I sincerely apologize. thank you

@wqingzex
Copy link
Owner

I test these codes on dataset of translate clutter mnist, but with size 40 * 40, and also original mnist dataset, it's ok. If something is wrong, please tell me.

By the way, I convert the original dataset file to images per digit

@hhhmoan
Copy link
Author
hhhmoan commented Aug 16, 2016

maybe you should test the code in translate clutter mnist 100_100 with depth 3 and glimpse 12_12,because my code also can make ok in small image, but when in 100*100,it's fail,if you need data,i can share with you

@wqingzex
Copy link
Owner

really? if true, it's my fault. please share the dataset with me, I will try it.

@wqingzex
Copy link
Owner
wqingzex commented Aug 16, 2016

did you try different parameters? maybe it's not the right configuration. what's your problem

@hhhmoan
Copy link
Author
hhhmoan commented Aug 16, 2016

let me check my data,and share you latter

@hhhmoan
Copy link
Author
hhhmoan commented Aug 16, 2016

the num of image is really too much, so i'd like to give you the script to get them. https://github.com/deepmind/mnist-cluttered if you finfish your code, please tell me . and you can see my easy code in https://github.com/hhhmoan/RAM-tensorflow

@hhhmoan
Copy link
Author
hhhmoan commented Aug 27, 2016

have you finish this experiment

@wqingzex
Copy link
Owner
wqingzex commented Sep 1, 2016

I am not familiar with lua, so I do not know how to generate a image with 3 channels, did you try decrease the glimpse? maybe 8 would be ok?

@hhhmoan
Copy link
Author
hhhmoan commented Sep 3, 2016

maybe you can get the grey image and then use python cv2 script to transfer to 3 channel,and i think decrease the number of glimpse is no use, follow the paper's experiment may be better. Thank you for taking time to discuss with me

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants
0