Your team is building a convolutional neural network (CNN)-based architecture from scratch.
The preliminary experiments running on your on-premises CPU-only infrastructure were encouraging, but have slow convergence.
You have been asked to speed up model training to reduce time-to-market.
You want to experiment with virtual machines (VMs) on Google Cloud to leverage more powerful hardware.
Your code does not include any manual device placement and has not been wrapped in Estimator model-level abstraction.
Which environment should you train your model on?
Click on the arrows to vote for the correct answer
A. B. C. D.A.
Based on the given scenario, the team is building a convolutional neural network (CNN) from scratch, and preliminary experiments running on their on-premises CPU-only infrastructure were encouraging, but slow in convergence. To speed up model training, they want to experiment with virtual machines (VMs) on Google Cloud to leverage more powerful hardware.
In this case, the best option is to use a Deep Learning VM with an n1-standard-2 machine and 1 GPU with all libraries pre-installed, which is answer C. Here's why:
Option A suggests using a VM on Compute Engine and one TPU with all dependencies installed manually. However, TPUs are specifically designed for Google's TensorFlow framework, and since the team's code does not include any manual device placement, it is unlikely that it is already optimized for TPUs. As a result, the team would need to invest additional time to modify the code to take advantage of TPUs, which might not be worth it in this case.
Option B suggests using a VM on Compute Engine and 8 GPUs with all dependencies installed manually. While multiple GPUs can speed up training time, it may not be the most efficient solution for the team's current situation. Additionally, manually installing all dependencies can be time-consuming and prone to errors, which can result in further delays.
Option D suggests using a Deep Learning VM with more powerful CPU e2-highcpu-16 machines with all libraries pre-installed. However, since the team's model is a CNN, it is likely that the GPU's parallel processing power will be more beneficial than the additional CPU cores.
Option C, a Deep Learning VM with an n1-standard-2 machine and 1 GPU with all libraries pre-installed, is the best choice. This configuration allows the team to leverage the power of a GPU without the need to manually install dependencies. Moreover, since the n1-standard-2 machine has 2 vCPUs and 7.5 GB of memory, it provides a balance between processing power and cost.
In summary, option C - a Deep Learning VM with an n1-standard-2 machine and 1 GPU with all libraries pre-installed - is the best choice for the team's scenario.