#5 Working on Tensorflow Image Classification with Transfer Learning - Training

#machinelearning #machinelearningproject #kaggle #tensorflow

Most machine learning models for image classification are trained on square images, which means that if you try to give them an image that is not square, then you need to either crop it to be a square, or you distort it, or you have to add black bars onto them.
Most times cropping is ok, but what if your subject is at the edges of your images? Then it becomes a bit more tricky.
I am using MobileNet pre-trained model with transfer learning using a dataset of images I gathered from different datasets on Kaggle and off the internet. The model does great with images from my validation dataset, but it does terrible when I evaluate the model against the images taken from my own Raspberry PI camera.
One of the reasons is because the foxes, birds and cats in my garden don't stay in the center of the frame. They seem to sit on the edges, which means that when TF crops the image to make it square, it loses the subject that am trying to identify.
Even when I tried to use my own Pi images dataset as the validation dataset, the results were terrible.

✅ Recommended Courses
(Includes Amazon Affiliate Links)
==================================

✅ Recommended Playlists
==================================

✅ My Favorite Tech
(Includes Amazon Affiliate Links)
==================================

✅ My RTX 3070 and Ryzen 5900 Machine Learning PC Build
(Includes Amazon Affiliate Links)
=======================================================

✅ MY EQUIPMENT FOR YOUTUBE
(Includes Amazon Affiliate Links)
==================================

As an Amazon Associate, I earn a small commission from qualifying purchases on the amazon links above. It costs you nothing but helps me in keeping the content coming.

✅ Article

TBD

✅ BUSINESS ENQUIRIES

✅ SUPPORT THE CHANNEL or get MENTORING

✅ FOLLOW ME

✅ Timestamps: