MLP

MLP implementation from scratch with scalar autograd, trained on the Iris dataset.

Run

pip install -r requirements.txt

• Open mlp_iris.ipynb and run all cells from the repo root so mlp_lib is on the path.

Project Structure

• mlp_lib/: minimal autograd + MLP implementation
  • tensor.py: scalar Tensor with backward pass
  • mlp.py: Neuron/Layer/MLP
  • loss.py: stable log-sum-exp cross-entropy on logits
  • optimizer.py: Adam optimizer
• mlp_iris.ipynb: training + evaluation notebook
• training_curves.png, confusion_matrix.png: sample outputs

Model

• Input: 4 features (Iris)
• Hidden 1: 16 neurons, tanh
• Hidden 2: 8 neurons, tanh
• Output: 3 logits (softmax handled in the loss)

Data

• StandardScaler is fit on the training split and applied to both train and test.

Results

~95-100% accuracy on the test set after 100 epochs.

Authors

• Marc GUILLEMOT
• Emre ULUSOY
• Rayan DRISSI
• Yanis MARTIN