kochbj / deep-learning-for-causal-inference Goto Github PK

Hi,

When i am running first tutorial of TarNet on colad, got the training loss = nan.

In Tutorial 1, there is the following saying:

While you should experiment with different learning rates, I recommend having a conservative (smaller) learning rate because we really want our estimator to be unbiased.

Here, "unbiased" means that no selection bias or that the estimator converges around a global minimum?

Hi:

Thanks for your amazing tutorial first. I want to know where is the dragonnet tutorial you mentioned?

Thanks!

Hi,

I have read the tutorial and it is really detailed and helpful! Thanks a lot!

But I got a question and hope you can help me. So in dragonnet implementation in tutorial 3, we estimate our target with 2 method, one is ATE and another is TARREG_CATE. So my question is

1. Is TARREG_CATE actually the ATE if we take mean of it? If it is, when we try to get the better estimation of ATE, should we choose ATE or TARREG_CATE?

Hi, this is a great tutorial! Thank you for sharing.

I have a question about implementing Dragonnet with a large dataset (in my case 200k subjects). Since to calculate loss it needs to construct a large matrix (200k x 200k) in float32 dtype, that cannot fit into memory. Do you have any suggestions?

Thanks

In the implementation, na and nb is not used.

def pdist2sq(A, B):
    #helper for PEHEnn
    #calculates squared euclidean distance between rows of two matrices  
    #https://gist.github.com/mbsariyildiz/34cdc26afb630e8cae079048eef91865
    # squared norms of each row in A and B
    na = tf.reduce_sum(tf.square(A), 1)
    nb = tf.reduce_sum(tf.square(B), 1)    
    # na as a row and nb as a column vectors
    na = tf.reshape(na, [-1, 1])
    nb = tf.reshape(nb, [1, -1])
    # return pairwise euclidean difference matrix
    D=tf.reduce_sum((tf.expand_dims(A, 1)-tf.expand_dims(B, 0))**2,2) 
    return D

In Tutorial 1, it is calculated according to the following equation:

No matter t equals 0 or 1, the value is opposite to what it is supposed to be.

Hi,
for short tutorial 1 I still get training loss == nan

Hi,

Very usefull and great tutorials! Any plans to add a multi treatments tutorial with tarnet or dragonnet ?

Thanks !