I have been actively working on image classification using deep learning for last some months. Now I want to work on Glaucoma classification problem, but there is no large dataset available, but there are still some smaller datasets. I want to know from you that would it be wise if I merge one more dataset and train my model based on merged one? Cause, in that case, I would get much data at a time for training and evaluating my model.
It is a good idea to use the different datasets to test and evaluate the model but for training 1 dataset will be enough if it is relevant to your project.
It depends on how alike the data is. If the images are of different sizes then it'll cause trouble for the model. Also if its simple binary then it wont be a problem but if you have different classes of the disease then you have to be sure there is no un balanced distribution!
I would recommend the use of multi-source Domain Adaptation/Domain Generalization for efficiently using the available labels from the related but distinct domains. Zhang et al. has a great compilation of such methods:
https://www.researchgate.net/profile/Jing_Zhang218/publication/316921517_Transfer_Learning_for_Cross-Dataset_Recognition_A_Survey/links/595e001ba6fdcc6d8562a7b6/Transfer-Learning-for-Cross-Dataset-Recognition-A-Survey.pdf
Soumya Snigdha Kundu, Hugo Neves Oliveira thank you both for your valuable ideas. I will try to follow what you have recommended.
For merging two datasets you need to be sure regarding distribution of the data in both and features.
Moreover, for deep learning models if you need large dataset for learning you can try data augmentation.
Please refer following link for reference:
https://nanonets.com/blog/data-augmentation-how-to-use-deep-learning-when-you-have-limited-data-part-2/
In my opinion, it may depend on how different the datasets are. Is ther anything that sets them apart? WIll anyone be able to tell if you merged them? Is the merging going to have an effect that individual use of the datasets wouldn't?
Newbie's advice here.
It is a good idea to use the different datasets to test and evaluate the model but for training 1 dataset will be enough if it is relevant to your project.
Ritika Lohiya Dilruba Tariq Jinia Mohammed Ashikur Rahman Thank you, everybody, for your valuable suggestions. Unfortunately training with one dataset and evaluating my model with different datasets gave me very bad results because the samples from the different dataset were different in characteristics though they were for solving the same problem. So, when I try to evaluate, I get poorer results. But I guess this will be okay if the samples from different datasets look closer, unfortunately in my case, they did not look so close. But anyway, now I am working on class weights during training my model with different datasets in order to avoid biases of the unbalanced data. But a new problem has arisen. For the larger dataset, my model is getting correctly fitted, but for smaller datasets, I am getting high overfitting though I have taken almost all necessary steps to reduce overfitting except cross-validation. I think I will also need to consider offline data augmentation and give both cases a comparison.
Yes, it is not strange (it is custom among data analysts). Your objective is to test performance and validity of your algorithm. Most the publicly available database for testing models validations are already merged together.
Dear Md. Muhtadir Rahman
Yes, you can do that. Check this paper
Article Novel Feature Selection and Voting Classifier Algorithms for...
If the datasets are similar then they can be merged, otherwise not, because a single recognition algorithm will not be able to work efficiently on two visually different datasets.
- Badges
- Science topic