2
$\begingroup$

This is a conceptual question regarding model training (e.g. CNNs). I have since solved the issue that raised this question, but I was still curious.

Preliminaries: In the typical training setting, we have $N$ training examples, which we batch to use with mini-batch SGD (or similar optimization). One run through all training examples is a single epoch. Let's say you plan to run $M$ epochs.

Now the question: During the training it's generally good practice to shuffle the data such that the mini-batches are not the same during each epoch. If one is using the mini-batches to update gradients, can one instead train over a "super-epoch" by pre-shuffling the data and feeding it $N\cdot M$ training examples? That is, perform $M$ random shuffles (without replacement) and concatenate to generate $N \cdot M$ examples. Then train for only a single epoch.

Is the only downside of this (say, for TensorFlow) that you cannot check the progress of accuracy/etc. of your validation set following each epoch? I know that is a pretty big downside, and I'm not advocating for this method...was just curious if there was anything else I was missing in my understanding.

$\endgroup$
1
  • $\begingroup$ For more clarity-- my query stemmed from a deep metric network context. I have many images (each belonging to one of many classes) to compare and there are obviously a huge number of combinations one can choose for training using the N-pairs loss method. Ultimately the real question I was aiming for was: Is it problematic to generate, say, 10M batches by sampling (so each batch is very likely to be unique given the sample space) and just run one huge epoch over those batches? Are there benefits to be gained by sticking with a smaller set of batches and shuffling them appropriately? $\endgroup$ – blawney_dfci Sep 14 '19 at 1:29
3
$\begingroup$

As long as you’re using mini-batching for both, you’ve described the same procedure in two different ways. Dividing the $M \times N$ shuffled samples into $M$ epochs is the same number of updates as all $M\times N$ updates concatenated together. If you further require that each of the $N$ samples appear exactly once before being used again, your method just a different description of the ordinary method of several epochs. (You can do "special" things include common early-stopping methods which check the model quality against a validation set on a regular basis, so you might not achieve all $M$ epochs before a termination condition is satisfied.)

Checking model statistics at the end of an epoch is a common way to see how well the model is doing, but it's purely a matter of social practice. You can report validation statistics every $k\ge 1$ mini-batches if you want, with the understanding that you'll be doing more computation because you're evaluating the model more frequently than once per epoch.

$\endgroup$
1
$\begingroup$

The entire idea of mini batches is two fold

  1. Faster Epoch Processing Time

A lower epoch processing time helps you evaluate the progress of your model, and make necessary adjustments etc

  1. A more Generalized Model

Since your model is forced to re-evaluate gradients for different data each time, the final model after convergence will be robust

$\endgroup$
1
  • $\begingroup$ The question was not about the reasoning for mini-batch. $\endgroup$ – blawney_dfci Sep 14 '19 at 1:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.