I have just started the course and completed dogscats example. Just for fun, i have tried with different datasets:
Cricket and baseketball.
Initially i have taken 10 photos each in training data and 5 in validation dataset. Rest details are same as class. My accuracy comes to ~46%. I than tried with more photos in training set (~30), accuracy improved to ~53. Still not good. Further i have increased set to 50 and add some photos to valid set as well. Now my accuracy reduced to 52%.
Just one more thing, i changed the learning rate as well. For different learning rate my accuracy varied. The best i got is 70% with learning rate .003.
Any idea what should i try further to improve the accuracy?
Make sure the image dimensions are consistent. Images downloaded from google will vary in their dimensions. You can try plotting a histogram of the width and height as Jeremy mentions in the second lesson. It’ll give you a good idea about the distribution of image dimensions. You want most of your images to be similar in dimension since the kernel tends to work with square sections of the image. So it might be neglecting other parts of the image which have meaningful information.
Also, it never hurts to download more images. Here is a really nice tool that lets you automatically download google images (I found it in one of the other threads on the forum, can’t remember which one).
I used it to download 100 images of 10 Harry Potter characters. Here is a blog post I wrote about it, or you can check out the Jupyter notebook.
Note: if you automatically download google images, you will have to either have to look at them and clean them up, or you have to accept that some of them will be just random & unrelated. Your topic should have plenty of images though, so this shouldn’t be a major drawback.