I’m working on a deep learning model with a training dataset containing 2500 observations. I’m uncertain about the optimal batch size to use for training to achieve the best balance between performance and efficiency.

Here are my main considerations:

Avoiding memory issues (I’m using a GPU with limited resources).
Ensuring good stability in the model’s convergence.
Keeping training times reasonable.
Could anyone suggest a suitable batch size to start with for a dataset of this size? Also, are there any best practices or strategies to test and determine the ideal batch size based on available hardware resources?

Thanks in advance for your advice!

I haven’t tried a specific batch size yet. I’m looking for guidance to understand which batch size would be most appropriate for my dataset of 2500 observations, considering factors like performance, efficiency, and hardware limitations.