Deep Learning Projects

A PyTorch notebook implementing multi-label image classification on the Pascal VOC 2012 dataset (20 object categories, one or more labels per image).

Contents

image_classification.ipynb:

• Downloads Pascal VOC 2012 via torchvision and builds a multi-label dataset from the object-detection annotations.
• Uses an ImageNet-pretrained EfficientNet-B3 backbone with a replaced multi-label classification head.
• Training setup: mixed precision (torch.amp), AdamW optimizer with differential learning rates, OneCycleLR schedule, and label-smoothed binary cross-entropy loss.
• Evaluation via mean average precision (mAP), with helper code for training curves, per-class average precision, prediction visualization, and a label co-occurrence heatmap.
• Includes an optional (disabled by default) ResNet-50 baseline for comparison.

Status

The saved notebook shows the data pipeline, model, and training loop defined and partially exercised (data loading and model construction ran successfully: 5,717 training / 5,823 validation images, 10.7M trainable parameters), but the training-loop cell in the saved output ends in a NameError before completing an epoch. No trained-model metrics (mAP, loss curves) are present in the saved outputs, so none are reported here. Running the notebook top to bottom in a fresh kernel should resolve the variable-ordering issue.

Requirements

Python 3 with torch, torchvision, scikit-learn, matplotlib, numpy, and Pillow. Written for a GPU-backed Colab runtime.