bbdamodaran / deepjdot Goto Github PK

Currently I am using my ground truth labels of the target dataset for searching the optimal loss weights.
In your code they are called sloss, tloss, jdot_alpha.

Are there other simple methods (heuristics) for finding the optimal hyperparameters, such that the target accuracy is maximized, but without using the target labels.

Dear Prof. Bharath Bhushan Damodaran,
In the paper,in page 9 ,about the embedding function g,the author show that there is a fully-connected layer behind the convolutional layers .Whether it need a pooling layer between these layers?
What 's the pool_size?
Thank you very much!
Best regards.

In dnn.py, there is an error when trying to import Merge. I believe this can be removed:
from keras.layers import Merge, merge

Not important, I am just curious about your code style in the Deepjdot class.

Why do you assign functions to the class with the following structure:

class Deepjdot(object):
    ....
    def foo(baz):
         do_sth(baz)
    self.foo = foo

Instead of:

class Deepjdot(object):
    ....
    def foo(self, baz):
         do_sth(baz)

Has your implementation any advantages?

Hello.
Can you release the code for real datasets?such as MNIST or USPS.
Thank you very much.

According to PR #9, the learning rate was fixed. This issue is created as a reminder that we have to experiment on the demo and SVHN data to see if high accuracy is still obtained with the learning rate change.
After someone experiments and see high accuracy then this issue should be closed.

I want to improve accuracy or loss of target dataset. So I would want to ask a few questions that might affect the accuracy.

1. If I want to increase accuracy of target, should I train with a varied source domain? E.g. to increase accuracy for MNIST (for SVHN to MNIST adaptation), should I augment SVHN to include variations like a grayscale version of SVHN, different colored SVHN, etc?
2. What are the example source and target datasets that would achieve high target accuracy? Give an idea if you haven't experiment before.
3. What are the example source and target datasets that would achieve low target accuracy? Give an idea if you haven't experiment before.

Recently, I read your ECCV2018 paper. It's an awesome paper. However, when I read this repository, I had some questions about the loss. When you fixed gamma to optimize the embedding function (g) and classifier (f), you used the cross-entropy loss. When you fixed g and f to optimize the gamma, you used the L2 loss. However, in your paper, you used the same function which was Lt. May you tell me why the Lt is different in these two steps?

Hi! Thank you for the procedure, I recently applied your program to solve some practical problems. But I encountered some difficulties. In your jdot_align model, you use self.model.train_on_batch([data], [np.vstack((ys, l_dummy)), g_dummy]) to train, but I encountered the problem of overfitting. I learned by checking the information that the function train_on_batch doesn't have Callback functions, So I can't save the best model. Do you have any other good solutions? thanks!

Hello,

Hello,I want to transfer from MNIST to USPS.
How to resize MNIST or USPS in keras?
Thank you.

@bbdamodaran I want to do regression from 0 to 1. (Not 0 or 1), how do I modify the code to do so? Suppose I already have the data available in the deepjdot_demo by replacing target values with 0.9 or 0.1 instead of 1 or 0. One of the changes I can do to the demo is to predict how close the sample is to the center.
My real use case is to predict how much someone opens their mouth or how much their open their eyes (0 = completely closed, 1 = completely opened) with source domain as image from day lighting and target domain as image from night lighting (and maybe from different person face).

If deepjdot doesn't work with regression, what are some of the other approaches I can use?