Digit Doctor -- The Broken CNN (Medium Difficulty)

"The previous developer built a CNN for digit recognition. It compiles, it trains... but it is basically guessing randomly. Can you fix it?"

---

The Story

A junior developer was tasked with building a Convolutional Neural Network (CNN) to classify handwritten digits using the MNIST dataset. They wrote the full pipeline -- data loading, preprocessing, model building, training, and evaluation.

The code runs without any errors. It compiles fine. It trains.

But the accuracy? Suspiciously low. That is essentially random guessing across 10 digit classes (0-9).

Something is fundamentally wrong with the pipeline.

Your mission: Diagnose the issues, fix the bugs, and optimize the model to achieve maximum accuracy.

---

Objective

1. Read and understand the existing code in notebook.ipynb
2. Find and fix the bugs that make the model perform like a random guesser
3. Optimize the model to achieve the highest accuracy possible
4. Save your model and submit via Pull Request

Scoring Checkpoints

Your submission is automatically evaluated when you create a Pull Request:

Test Case	Threshold	Points	Description
TC-2	Test Accuracy >= 75%	+11 pts	Pipeline is fixed and working
TC-3	Test Accuracy >= 85%	+9 pts	Model architecture is solid
TC-4	Test Accuracy >= 90%	+7 pts	Well-optimized model
TC-5	Unseen Data >= 80%	+9 pts	Model generalizes well
TC-6	No Overfitting	+4 pts	Train-test gap <= 10%

Maximum score: 40 points (cumulative -- earn points for each threshold you clear)

Overfitting Check

Your model is evaluated on a hidden, unseen portion of the MNIST dataset. If the gap between your test accuracy and unseen accuracy exceeds 10%, you will not receive the No Overfitting bonus.

---

What You Get

DevSphere-ML-Medium/
├── notebook.ipynb            <-- THE BROKEN NOTEBOOK (fix this!)
├── requirements.txt          <-- Python dependencies
└── README.md                 <-- You are here

The evaluation is handled automatically -- you do not need to run any test scripts.

---

Rules

You MAY:

• Fix bugs in the notebook
• Change the model architecture (layers, filters, activations)
• Tune hyperparameters (learning rate, epochs, batch size, optimizer)
• Modify preprocessing (normalization, reshaping)
• Add regularization (Dropout, BatchNormalization, etc.)

You may NOT:

• Change the dataset -- must use keras.datasets.mnist
• Use pretrained or external models
• Change the model save filename -- must save as model.h5
• Hardcode predictions or cheat the evaluation

---

How to Submit

Step 1: Fork this repository

Click the Fork button on GitHub to create your own copy.

Step 2: Clone your fork

git clone https://github.com/YOUR-USERNAME/DevSphere-ML-Medium.git
cd DevSphere-ML-Medium

Step 3: Install dependencies

pip install -r requirements.txt

Step 4: Open and fix the notebook

jupyter notebook notebook.ipynb

Read the code carefully -- look for the HINT and BUG comments. Fix the bugs, optimize the model, and make sure it saves as model.h5.

Alternatively, open the notebook directly in Google Colab.

Step 5: Commit your changes

Commit message should be in the format - "Your_Roll_Number_CNN model"
Ex: "LCI2025001_CNN_model"

git add .
git commit -m "Your_Roll_Number_CNN model"
git push origin main

Step 6: Create a Pull Request

Go to the original repository and create a Pull Request from your fork.

Step 7: Get your score

The evaluation runs automatically on your PR. Your score and detailed feedback will appear as a comment on your Pull Request within a few minutes.

Tip: Every time you push a new commit to your PR branch, the evaluation re-runs automatically. Iterate until you are satisfied with your score.

---

Hints

The notebook contains BUG and HINT comments at suspicious locations. Read them carefully.

Stuck? Click for general guidance

Think about:

1. Data preprocessing -- What range should pixel values be in?
2. Architecture -- What makes a CNN different from a plain neural network?
3. Activation functions -- Can a network learn complex patterns without non-linearity?
4. Output layer -- How does a model output probabilities for classification?
5. Loss function -- Is the loss function designed for classification?
6. Optimizer -- Some optimizers converge much faster than others
7. Training -- Is the model training long enough?

---

Sample Output (what success looks like)

When your PR is evaluated, you will see a comment like this:

Digit Doctor -- Evaluation Report

Test Results (6/7 passed)

| TC-2 | Accuracy >= 75%  | Pass | +11 |
| TC-3 | Accuracy >= 85%  | Pass | +9 |
| TC-4 | Accuracy >= 90%  | Pass | +7 |
| TC-5 | Unseen >= 80%    | Pass | +9 |
| TC-6 | No Overfitting   | Pass | +4 |

Score: 40/40
Status: Outstanding -- You are a true Digit Doctor!

---

Leaderboard Criteria

Participants are ranked by:

1. Total Score (primary)
2. Test Accuracy (tiebreaker)
3. Submission Time (earlier is better)

note that points regarding this might fluctuate after proper reviews from the members at the end of the event, but for the time being this repo contributes 40 pts to your score for completion

---

Built for DevSphere by GDG IIIT Lucknow