Train an Image Classifier
Part of the Data Science Roadmap
In short
This is where it all comes together: your own photos, a pretrained network, augmentation and honest evaluation. Finish it and you have a real deep learning project for your portfolio.
The practice project for this phase — train a CNN (from scratch or via transfer learning) to classify images into categories, and document your architecture choices and results.
Think of it like this
Sorting a pile of pet photos
Someone hands you a shoebox of cat and dog photos to sort into two albums. At first you check carefully, then you get fast and confident. You'd still double-check the blurry ones. Your model learns the same way, and you test it on photos it has never seen.
Watch it happen
Project build plan
1/5Gather a few hundred images per class from a public dataset or your own phone. Remove duplicates.
The big ideas
What you need to know
Folder per class
data/train/cats, data/train/dogs and so on. Keras reads labels straight from folder names.
Augmentation
Random flips and rotations act like extra photos and cut overfitting on small datasets.
Show the confusion matrix
Which classes get mixed up tells you more than one accuracy number.
Ship a demo
Wrap the saved model in a small Gradio or Streamlit app so anyone can try it.
See it in code
Walk through the code
1import keras2from keras import layers3 4train = keras.utils.image_dataset_from_directory(5 "data/train", image_size=(160, 160), batch_size=32)6val = keras.utils.image_dataset_from_directory(7 "data/val", image_size=(160, 160), batch_size=32)8 9augment = keras.Sequential([10 layers.RandomFlip("horizontal"),11 layers.RandomRotation(0.1),12])13train = train.map(lambda x, y: (augment(x, training=True), y))1/4Read training images and get labels from the sub-folder names.
The data pipeline is the trickiest part to get right. Plug these datasets into the transfer learning model from the earlier lesson.
Save yourself the pain
Common mistakes
Oops
Letting near-duplicate photos land in both train and test folders.
Do this instead
Remove duplicates and split by source or photo session, so the test score is honest.
Oops
Augmenting the validation and test sets too.
Do this instead
Augment only training data. Evaluate on clean, untouched images.
Your mission
First milestone: a working baseline
Quick check
Why apply data augmentation only to the training set?
Go deeper
- 1Choose a manageable image dataset (not enormous) to keep training time reasonable.
- 2Start with transfer learning before attempting to train a CNN fully from scratch.
- 3Document your model's accuracy, training curves, and a few example predictions (right and wrong).
Stuck on this topic? Ask an Insider
Get 1:1 guidance from people who've walked this exact path — on the InsideEdge app.
