Random seed management in MLOps - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why set a random seed in machine learning pipelines?

What is the main reason to set a random seed when running machine learning experiments?

A To reduce the size of the training dataset

B To speed up the training process

C To increase the randomness of the model initialization

D To ensure the results are reproducible across runs

💡 Hint

Think about why you want to get the same results every time you run your code.

Setting a random seed fixes the starting point for random number generation, making experiments reproducible.

💻 Command Output

intermediate

Output of setting random seed in Python's random module

What is the output of the following Python code snippet?

MLOps

import random
random.seed(42)
print([random.randint(1, 10) for _ in range(3)])

A [7, 1, 3]

B [10, 1, 4]

C [2, 1, 5]

D [2, 10, 7]

💡 Hint

Run the code or recall the sequence generated by seed 42 in Python's random module.

With seed 42, the first three random integers between 1 and 10 are 7, 1, and 3.

🔀 Workflow

advanced

Proper random seed management in a multi-step ML pipeline

In a multi-step machine learning pipeline involving data shuffling, model initialization, and data augmentation, which approach best ensures reproducibility?

A Set a global random seed and pass it explicitly to each step's random generator

B Set different random seeds for each step independently without coordination

C Set the same random seed once at the start of the pipeline and use the default random generators throughout

D Do not set any random seed and rely on system randomness

💡 Hint

Consider how to control randomness in each step to get consistent results.

Passing a global seed explicitly to each step's random generator ensures all randomness is controlled and reproducible.

❓ Troubleshoot

advanced

Why does setting random seed not produce reproducible results in PyTorch training?

You set torch.manual_seed(123) before training your model, but results differ between runs. What is the most likely cause?

A You need to set the seed after training starts

B You forgot to set the seed for CUDA operations with torch.cuda.manual_seed_all

C PyTorch does not support random seeds for reproducibility

D The seed value 123 is too small to produce reproducible results

💡 Hint

Think about GPU operations and their randomness control.

CUDA operations have their own random seed control; forgetting to set it causes non-reproducible results.

✅ Best Practice

expert

Best practice for random seed management in distributed training

In distributed training across multiple machines and GPUs, what is the best practice to manage random seeds to ensure reproducibility?

A Use the same fixed seed on all processes and rely on synchronization

B Use different seeds on each process to increase randomness

C Set a base seed and derive unique seeds per process using their rank

D Do not set seeds and let each process generate random numbers independently

💡 Hint

Consider how to balance reproducibility and independent randomness per process.

Deriving unique seeds per process from a base seed ensures reproducibility while avoiding identical random sequences across processes.