Issue: 8: Forget Accuracy, Try from scratch

[kyanmahajan]

Objective

In this task, participants are expected to design and train a Convolutional Neural Network (CNN) from scratch for image classification.

The focus of this issue is to help you understand how CNNs work internally, without relying on pretrained models or fine-tuning.

---

Task Description

• Implement a custom CNN architecture using a deep learning framework of your choice (PyTorch / TensorFlow).
• Train the network from random initialization.
• Perform image classification on the provided dataset.

Pretrained models are NOT allowed
(Do not use ResNet, EfficientNet, VGG, etc.)

---

🔗 Connecting with Previous Issues

Participants are strongly recommended to reuse the data loaders created in previous issues.

This helps in:

• Maintaining modular and reusable code
• Understanding how different components of an ML pipeline connect
• Building a clean end-to-end training setup

---

Expected Components

Your solution should include:

• A CNN model implemented from scratch
• Convolutional, pooling, and fully connected layers
• A training loop with loss calculation and optimization
• Basic evaluation on validation/test data

---

📁 Where to Work

• All work should be done inside the participants/ folder
• You may use a new notebook/script or extend an existing one

[Aaryan-Degama]

Claim
If competitive needs claiming

[kyanmahajan]

Competitive does not need claiming, you can do it