Transfer Learning on a Small Custom Dataset
Fine-tuning pre-trained CNNs (ResNet / VGG / AlexNet) on a small image set: how to get a useful model when labelled data is scarce and expensive.
The business question
Most businesses do not have hundreds of thousands of labelled images, and labelling them costs time and money. Transfer learning reuses a model already trained on millions of images and adapts it to a small custom dataset, here a subset of Fashion-MNIST.
What I built
- Loading and preprocessing a custom CSV image dataset
- Replacing the final classifier layer of a pre-trained model (ResNet, AlexNet or VGG)
- Fine-tuning selected layers and comparing the architectures on accuracy and training time
Where this applies in a business
Sorting product photos, spotting visible defects in goods, or reading simple document types, when only a few hundred labelled examples are available.
What it shows
How to reach a usable result with a fraction of the data and compute, which in a business means a fraction of the cost.
Full Stack Data Scientist & Business Systems Developer · Finance background · ML, Automation, Django
View full portfolio →
More projects
Brijwasi Sweets — Company Website & Recommendation System
A new company website and an AI product-recommendation system for a food & beverages brand, aimed at larger orders and more repeat customers.
Custom Business SoftwareClient Portal & GST Billing System
The system BeeProton AI runs on: client portal, GST-compliant invoicing, payments, milestones, support tickets and encrypted client communication, in one place.
Web & App DevelopmentAnvi Foundation — Website for a Special-Education NGO
A ten-page website for an NGO in Navi Mumbai, built around one goal: helping parents book an assessment. Includes UPI and bank donations with no gateway fees.
Have a project in mind? Get a free quote in 24 hours.
Transparent, affordable pricing in INR. Fixed-price packages or hourly — you choose.