如何在此类机器学习任务中修复 batch_size 和 epoch

Question

I train a ResNet50 network.
I have a dataset with 1500 images, I fix epochs = 100 and batch_size = 16, I find that accuracy reaches 0.8 from the 10th epoch and continues to increase until 0.95. Now, I add images for this dataset, it becomes 15 000 and with epochs = 100 and batch_size = 16. I observed that accuracy does not exceed 0.6 . I can't understand reasons?

Answer 1

The accuracy of a deep learning model mainly depends on the nature of the dataset and number of samples within the dataset. The more samples, the more epochs you would need. However, batch size has greater effect on the speed of training a model rather than accuracy.

In your case, keeping epochs constant (100), it's understandable that dataset with 15000 samples will have less accuracy then a dataset with 1500 samples. If you want to achieve the desired accuracy on larger dataset, then you'd need to train your model for more epochs, let say try 200 and see the results.

Another thing you need to keep in mind that how you are calculating the accuracy? Is it training accuracy or accuracy calculated using unseen test set? It totally depends on the diversity of your dataset.

In short, for better training results, you need to make your dataset diverse with enough samples to get a generalized model. For better accuracy results, you need to keep training your model for enough epochs unless you see improvements in the accuracy. Hope these information help you !!!