Day 19/100: #100DaysOfMLCode 📊 Paper: YOLO v1Task: Dataset Pipeline & Label EncodingThe tricky part of YOLO isn't just the model; it's the data preparation.Standard datasets give you a list of boxes [x, y, w, h]. YOLO needs a 3D tensor [7, 7, 30]. Implemented 👇

Day 16/100: #100DaysOfMLCode Project 3 Starts: YOLO v1 (You Only Look Once).Leaving image classification to tackle Object Detection.📖 Reading Joseph Redmon's paper.The genius of YOLO: Instead of sliding windows or region proposals, it reframes detection as a single 👇

✅ Paper 2/10: ResNet Complete!Project 2 of #100DaysOfMLCode is done. Implemented ResNet-18 from scratch and trained it on CIFAR-10 to ~88% accuracy.But I didn't just code it; I tested the theory. 🧪 Thread on my Ablation Study 👇

Day 14/100: #100DaysOfMLCode Paper: ResNet (2015) Task: Ablation Study (ResNet vs PlainNet)I challenged the residual learning hypothesis by training a "broken" ResNet (PlainNet-18) without skip connections. Results (10 Epochs): 🔵 ResNet-18: 87.98% (Faster convergence) 👇

Day 13/100: #100DaysOfMLCode Paper: ResNet (2015) Task: First Training RunTrained ResNet-18 on CIFAR-10 for 10 epochs. Hardware: Apple M2 Pro (MPS) - ~15 mins. Results: Test Accuracy: 87.98% Overfitting: None (Train 89% vs Test 88%)The skip connections are working their

Day 12/100: #100DaysOfMLCode Paper: ResNet (2015) Task: Data Pipeline & Augmentation Successfully loaded CIFAR-10. 🖼️ Input Pipeline: Random Crop (32x32) Random Horizontal Flip Normalization (mean/std) Batch Shape: [4, 3, 32, 32] ✅ The augmentation pipeline is live.

Day 11/100: #100DaysOfMLCode Paper: ResNet (He et al., 2015) Task: Full AssemblySuccessfully assembled ResNet-18. The logic handles dynamic downsampling through 4 layers of Residual Blocks. 📊 Specs: Layers: 18 deep Parameters: 11,173,962 🔥

Day 10/100: #100DaysOfMLCode Paper: ResNet (He et al., 2015) Task: Implementing the Residual BlockSuccessfully built the core unit of ResNet: The BasicBlock. Technical Win:Implemented the conditional skip connection. If dimensions match: Identity mapping (x). If dimensions

Day 9/100: #100DaysOfMLCode Project 2 Starts: ResNet (2015). Leaving the 5-layer LeNet behind to tackle deep networks (18+ layers). Reading Kaiming He's paper on Deep Residual Learning.The Core Insight: The "Degradation Problem" isn't overfitting. It's optimization difficulty.

✅ Paper 1/10: LeNet-5 Complete! I just wrapped up the first project of my #100DaysOfMLCode challenge. Reproduced LeCun's 1998 classic from scratch using PyTorch. 🏆 Final Result: 99.00% Accuracy on MNIST. ⏱️ Time: 8 Days. A thread on what I learned 👇 🧵

Day 7/100: #100DaysOfMLCode Paper: LeNet-5 Task: Engineering & MLOps Graduated from Jupyter Notebooks to modular Python scripts. 🎓 Refactored the codebase into: 📄 model.py (Architecture) 📄 data_setup.py (ETL) 📄 train.py (Training Engine)

Day 6/100: #100DaysOfMLCode Paper: LeNet-5 Task: Visualization & Error Analysis Milestone Reached: 99.00% Test Accuracy on MNIST! Why isn't it 100%? Visualized the misclassified samples. As shown below, the errors are mostly on ambiguous digits (e.g., a '5' written like a '6').

Day 5/100: #100DaysOfMLCode 🚀Paper: LeNet-5 (1998) Task: Full Training & Validation Cycle🏆 Final Results (5 Epochs): Training Loss: 0.03 Test Accuracy: 98.7% Implemented test_loop with torch.no_grad() to validate performance.

Day 4/100: #100DaysOfMLCode Paper: LeNet-5 (1998)Task: The Training LoopThe engine is ON. Implemented the full backpropagation cycle: Forward → Loss → Backward → Optimizer Step.First Epoch Results: Start Loss: 2.30 (Random guessing) End Loss: 0.18 (Significant learning)

Day 3/100: #100DaysOfMLCode Paper: LeNet-5 (1998)Status: Architecture Implemented via PyTorch.Successfully translated LeCun's paper into code. Tech Specs: Input: 32x32 (padded/resized from MNIST). Activation: Tanh (Historical accuracy). Pooling: Average Pooling.

Day 2/100 of #100DaysOfMLCode Paper: LeNet-5 Focus: Data Pipeline & Preprocessing Successfully built the PyTorch DataLoader for MNIST. Key Detail: The original LeNet-5 paper requires 32x32 input, but MNIST is 28x28.