Google Professional Machine Learning Engineer Exam - Topic 5 Question 80 Discussion

Actual exam question for Google's Professional Machine Learning Engineer exam

Question #: 80
Topic #: 5

[All Professional Machine Learning Engineer Questions]

You recently developed a deep learning model using Keras, and now you are experimenting with different training strategies. First, you trained the model using a single GPU, but the training process was too slow. Next, you distributed the training across 4 GPUs using tf.distribute.MirroredStrategy (with no other changes), but you did not observe a decrease in training time. What should you do?

ADistribute the dataset with tf.distribute.Strategy.experimental_distribute_dataset

BCreate a custom training loop.

CUse a TPU with tf.distribute.TPUStrategy.

DIncrease the batch size.

Show Suggested Answer

Suggested Answer: D

by Chauncey at Jun 21, 2024, 09:30 AM

Limited Time Offer

25%

Off

Get Premium Professional Machine Learning Engineer Questions as Interactive Web-Based Practice Test or PDF

Contribute your Thoughts:

Submit Cancel

Cherilyn

3 months ago

TPUs are cool, but not always necessary for speed.

upvoted 0 times

...

Denise

3 months ago

Definitely try distributing the dataset first.

upvoted 0 times

...

Nadine

4 months ago

I disagree, a custom training loop might be more effective.

upvoted 0 times

...

Britt

4 months ago

Surprised that 4 GPUs didn't help at all!

upvoted 0 times

...

Myra

4 months ago

Increasing the batch size can help speed things up!

upvoted 0 times

...

Anastacia

4 months ago

Using a TPU sounds interesting, but I think it might be overkill for this situation. I’d lean towards option A or D instead.

upvoted 0 times

...

Mary

5 months ago

I practiced a similar question, and I feel like a custom training loop could give more control, but I’m not confident that’s necessary here.

upvoted 0 times

...

Luz

5 months ago

I'm not entirely sure, but I think increasing the batch size could help utilize the GPUs better. That might be option D?

upvoted 0 times

...

Corazon

5 months ago

I remember reading that distributing the dataset can help with performance, so maybe option A is the right choice?

upvoted 0 times

...

Viva

5 months ago

This seems like a good opportunity to try out a custom training loop. I've heard that can really help optimize the training process, especially when working with multiple GPUs. I'm going to go with option B.

upvoted 0 times

...

Wynell

5 months ago

I'm a bit confused here. Distributing the dataset seems like a logical step, but the question says that didn't work. Maybe I should consider creating a custom training loop or using a TPU instead?

upvoted 0 times

...

Marge

5 months ago

Okay, let's think this through. If the training time didn't improve with 4 GPUs, then the issue might not be with the hardware. A custom training loop could help optimize the process, so I'm leaning towards option B.

upvoted 0 times

...

Raina

6 months ago

Hmm, this is a tricky one. I'm not sure why the training time didn't decrease with the 4 GPUs. Maybe I need to look into how to properly distribute the dataset.

upvoted 0 times

...

Joanne

6 months ago

Ah, I see now. Routers are layer 3 devices that use protocols to learn paths, not layer 2 switching. And they definitely aren't layer 4 devices. I've got my answer.

upvoted 0 times

...

Clorinda

6 months ago

Hmm, I'm a little unsure about this one. I'll need to think through the different IP addressing scenarios carefully.

upvoted 0 times

...

Halina

6 months ago

I'm a bit unsure about this one. I know enterprise analysis is important, but I'm not totally clear on the specific document that would outline the governances. I'll have to think this through carefully.

upvoted 0 times

...

Sharmaine

11 months ago

Ah, the age-old dilemma of training a deep learning model - GPUs, TPUs, and batch sizes, oh my! I say we just throw the whole thing in the microwave and see what happens. *chuckles*

upvoted 0 times

...

Rebbecca

11 months ago

Well, this is a tough one. I'm leaning towards option A - distributing the dataset with tf.distribute.Strategy.experimental_distribute_dataset. Seems like the most straightforward approach to me.

upvoted 0 times

Trinidad

10 months ago

Let's give it a try and see if it makes a difference.

upvoted 0 times

...

Carol

10 months ago

I agree, distributing the dataset might be the key to improving training time.

upvoted 0 times

...

Verona

10 months ago

I think option A is a good choice. It could help speed up the training process.

upvoted 0 times

...

Nada

11 months ago

Oh, I bet option D is the way to go! Increasing the batch size might just do the trick. After all, who needs GPUs when you have big batches, right? *wink*

upvoted 0 times

Curt

10 months ago

User 3: I agree, let's give it a shot and see if it makes a difference.

upvoted 0 times

...

Dorathy

10 months ago

User 2: Yeah, that could be a good solution to try out.

upvoted 0 times

...

Lourdes

10 months ago

User 1: I think increasing the batch size might help speed up the training process.

upvoted 0 times

...

Micheal

11 months ago

Interesting question. I think option C looks promising - using a TPU with tf.distribute.TPUStrategy could really speed up the training process.

upvoted 0 times

Leoma

9 months ago

C) Use a TPU with tf.distribute.TPUStrategy.

upvoted 0 times

...

Chi

10 months ago

B) Create a custom training loop.

upvoted 0 times

...

Dominque

10 months ago

A) Distribute the dataset with tf.distribute.Strategy.experimental_distribute_dataset

upvoted 0 times

...

Omega

11 months ago

Hmm, I would say option B. Creating a custom training loop can help you fine-tune the distribution of the training process and potentially improve the performance.

upvoted 0 times

...