0
$\begingroup$

I have a small datset (530 images) trained on a simple CNN called AquaSight. This is the architecture. architecture of aquaSight

I had an underfitting problem, 75% accuracy and 0.6 loss. How can I solve the underfitting problem ? If I do Data Augmentation, will there be any improvement ?

$\endgroup$
1
  • $\begingroup$ with 21M parameters your problem might be overfitting... $\endgroup$
    – lcrmorin
    Jul 27 at 23:27
1
$\begingroup$

some questions will help give better answers:

  1. When you say underfitting, I assume you mean that the low accuracy is on the train set, correct? I'm asking also because with that amount of parameters for such a small training set I would be far more concerned with overfitting

  2. 530 images is very small dataset, I would consider going with a pretrained architecture and possibly finetuning. CNNs trained from scratch isn't a very good solution for such data in many cases. Also, when you consider your results bad, are you comparing to something? Do you have a benchmark for what kind of accuracy you should expect on this task?

  3. As a general practice, if I encounter underfitting I normally remove Dropout and put it back once I have good learning on train set and start to overfit. If you attach the loss curve we could see what's going on better

  4. Data augmentation is usually more of a solution to overfitting, allowing the model to generalize better

$\endgroup$

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