Vectorize to Apply Function to 3d Array

Question

I am trying to apply a function to 3d torch tensor while the function is applied to 2d tensor which is read through the axis 1 of the 3d torch tensor.

For example, I have a torch tensor of the shape (51, 128, 20100) (a variable with name autoencode_logprob) and the function(rawid2sentence) runs on the input of the shape (51, 20100).

Right now I wrote the code to run with naive for loop, looping one by one with range(128).

However, it’s too slow. Following is the code part that matters.

autoencode_logprobs is the 3d tensor and I need to apply rawids2sentence function along its second axis. Any help to vectorize it?

for i in range(128):
    output_sent = self.dictionary.rawids2sentence(
        autoencode_logprobs[:, i].max(1)[
            1].data.cpu().numpy(),
        oov_dicts[i],
    )
    output_sent_encoding = ifst_model.encode([output_sent])

With only 128 iterations I'd focus the speed of the task itself, not on eliminating the iterations. — hpaulj
– hpaulj, Commented May 7, 2019 at 19:36
There are millions of batches to process the same task. Eliminating iteration is the basic and fundamental in deep learning. — Beverlie
– Beverlie, Commented May 7, 2019 at 19:47

Wasi Ahmad · Accepted Answer · 2019-05-08 00:55:19Z

1

Since I do not know what rawids2sentence or encode function does, I can help you with to do the max operation.

In the following statement,

autoencode_logprobs[:, i].max(1)[1]

You identify the index of the maximum values along dim=1 for each 51 x 20100 tensor. So, the output is a vector of size 51.

You can perform the same operation in your full tensor of shape 51 x 128 x 20100 and get the output as 128 x 51 tensor.

autoencode_logprobs.transpose(0, 1).max(2)[1] # 128 x 51

So, if your rawids2sentence or encode methods can tackle batch inputs, the above change should work for you without any loop.

answered May 8, 2019 at 0:55

Wasi Ahmad

38.1k34 gold badges129 silver badges169 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Vectorize to Apply Function to 3d Array

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related