MLA-C01 · Question #225
A company runs a neural network model and retrains the model when the performance degrades. The company uses a training job that uses Amazon SageMaker AI distributed data parallelism (DDP). The traini
The correct answer is D. Increase the number of instances.. The job already uses Amazon SageMaker Distributed Data Parallelism (DDP), which splits training data across multiple instances and synchronizes gradients between them. The most direct way to reduce training time with DDP is to add more instances - each instance processes a smalle
Question
A company runs a neural network model and retrains the model when the performance degrades. The company uses a training job that uses Amazon SageMaker AI distributed data parallelism (DDP). The training job takes several hours to run. The company wants to decrease the required time for the training job. Which solution will meet this requirement?
Options
- AIncrease the number of epochs.
- BIncrease the number of neurons in the hidden layers.
- CIncrease the number of layers.
- DIncrease the number of instances.
How the community answered
(41 responses)- A2% (1)
- B15% (6)
- C5% (2)
- D78% (32)
Explanation
The job already uses Amazon SageMaker Distributed Data Parallelism (DDP), which splits training data across multiple instances and synchronizes gradients between them. The most direct way to reduce training time with DDP is to add more instances - each instance processes a smaller shard of data per step, so the overall number of steps per epoch decreases proportionally, reducing wall-clock time. Options A, B, and C (more epochs, more neurons, more layers) all increase the amount of computation the model must perform, which would make training take longer, not shorter. Only scaling out the instance count leverages the DDP architecture already in place.
Topics
Community Discussion
No community discussion yet for this question.