Training with ~3000 images

[Anne]

Hi all,

I want to train my own model, but creating the data for training and testing takes a lot of time. Is it possible to know if the net will work by training it with around 3000 images, and around 500 for testing ? Or is it compulsory to have like 10k+ images in order to have the training working?
Thank you for sharing !

[Hossein Hasanpour]

Your options as far as I know are as follows: 
1.find a pre trained model that was trained on data similar  to yours, and then fine-tune it with your current dataset.

2.if there is no pretrained model in frist place, but there are datasets that are similar to your dataset, first train on that dataset and then use the pre-trained model for fine-tuning. 

3.use data augmentation to increase the size of your dataset. ( you can at least double or quadruple it )

and for the data, you really need as much data as you can gather, when you are dealing with deeplearning methods, or you have a very high chance of overfitting.

[Anne]

Thank you for your answer !
Do you know any pretrained model using images as label? From what I saw it's not very common...

[Hossein Hasanpour]

Unfortunately I dont know any.