ece595

Repository for MS Autonomy ECE59500CV

Project Archives: https://drive.google.com/drive/folders/1UUPBdLMmMBIj3a8dw_dFwxBuVDol618I?usp=drive_link

Final Project: ConvNeXt Critique and CIFAR-10 Implementation

The final project is the main artifact in this repository. It combines a literature-style critique of three recent CVPR papers with a PyTorch implementation and ablation study of a ConvNeXt-Tiny-inspired image classifier on CIFAR-10.

The project studies:

• ConvNeXt: A ConvNet for the 2020s (CVPR 2022) - the implementation focus, used to study how transformer-era design choices translate back into convolutional networks.
• UniAD: Planning-oriented Autonomous Driving (CVPR 2023) - reviewed as a system-level example of planning-oriented perception, prediction, and control.
• Reasoning in Visual Navigation of End-to-End Trained Agents (CVPR 2025) - reviewed for its dynamical-systems perspective on learned visual navigation policies.

Final Project Implementation

The implemented portion focuses on ConvNeXt. It builds a ConvNeXt-Tiny-like model from scratch in PyTorch and compares it against a ResNet-18 baseline on CIFAR-10. The code exposes the major architectural choices as command-line flags so they can be evaluated independently:

• Depthwise convolution kernel size: 3x3 vs 7x7
• Normalization: LayerNorm vs BatchNorm
• MLP expansion ratio: 2, 4, or 6
• Input stem: ConvNeXt-style 4x4 patchify stem vs ResNet-style 7x7 + max-pool stem
• Baseline model: torchvision ResNet-18

Key files:

paper/
  train_cifar10.py        # CIFAR-10 training/evaluation entrypoint
  models/convnext.py      # ConvNeXt-Tiny-like model and ablation knobs
  run_exps.sh             # Experiment sweep driver
  plot_results.py         # Log aggregation and plot generation
  requirements.txt        # Final project Python dependencies
  LaTeX/                  # CVPR-style paper source

Architecture and experiment flow:

flowchart LR
    A["CIFAR-10 images"] --> B["Augmentation + normalization"]
    B --> C["Stem ablation: patchify or ResNet-style"]
    C --> D["ConvNeXt stages: depths 3, 3, 9, 3"]
    D --> E["Global average pool + LayerNorm"]
    E --> F["10-class classifier"]
    F --> G["CSV logs, checkpoints, plots"]

Loading

Final Project Results

The final report and presentation show that ConvNeXt's ImageNet-scale design choices do not transfer uniformly to low-resolution CIFAR-10. ResNet-18 remains the strongest baseline, while ConvNeXt benefits most from a more traditional ResNet-style input stem.

Configuration	Best Validation Accuracy
ResNet-18	84.95%
ConvNeXt-Tiny, patchify stem, k=7, LayerNorm, MLP x4	72.93%
ConvNeXt-Tiny, patchify stem, k=3, LayerNorm, MLP x4	73.18%
ConvNeXt-Tiny, patchify stem, k=7, BatchNorm, MLP x4	70.56%
ConvNeXt-Tiny, patchify stem, k=7, LayerNorm, MLP x2	71.42%
ConvNeXt-Tiny, patchify stem, k=7, LayerNorm, MLP x6	71.44%
ConvNeXt-Tiny, ResNet-style stem, k=7, LayerNorm, MLP x4	79.90%

Main takeaways:

• ConvNeXt's modernization ideas are useful, but scale and image resolution matter.
• The large 7x7 depthwise kernel is not clearly beneficial on CIFAR-10; 3x3 performs slightly better in this setup.
• LayerNorm is more stable than BatchNorm inside the ConvNeXt blocks tested here.
• The original MLP expansion ratio of 4 is the best among the tested ratios.
• Replacing the patchify stem with a ResNet-style stem substantially improves ConvNeXt on CIFAR-10, suggesting early low-level feature extraction is especially important for small images.

Running the Final Project

From the paper/ directory:

pip install -r requirements.txt

python train_cifar10.py \
  --model convnext_tiny \
  --data-dir ./data \
  --epochs 50 \
  --batch-size 128 \
  --kernel-size 7 \
  --norm-layer ln \
  --mlp-ratio 4 \
  --stem-type patchify

To run the included ablation sweep:

bash run_exps.sh

Training writes CSV logs to paper/results/ and checkpoints to paper/checkpoints/. The plotting helper reads the logged results and generates comparison figures.

Other Repository Contents

HW1: Reverse-Mode Autodiff and Toy MLPs

hw1/ implements a small reverse-mode automatic differentiation system and uses it to train simple MLPs on toy Boolean tasks.

Workflow:

flowchart LR
    A["Toy datasets: XOR and two-bit adder"] --> B["Train/test splits"]
    B --> C["Packed MLP parameters"]
    C --> D["Reverse-mode autodiff tape"]
    D --> E["Gradient descent"]
    E --> F["Loss and accuracy demos"]

Loading

Key pieces:

• reverse_mode.py implements the cobundle/tape-based reverse-mode autodiff primitives.
• mlp.py defines vector helpers, sigmoid layers, L2 loss, parameter packing, and MLP evaluation.
• datasets.py provides XOR and two-bit-adder datasets.
• splits.py, evaluate.py, and train_loop.py provide data splitting, metrics, and gradient descent.
• train_mlp_xor_demo.py and train_mlp_adder_demo.py run experiments across small model sizes, splits, learning rates, and iteration counts.
• submission/, submission/v2/, and submission/tmp/ preserve packaged submission versions, including zipped turn-ins.

Run:

cd hw1
bash run

HW2: SmallNet Image Classifier

hw2/ contains a compact CNN image classifier for a four-class image dataset: Egyptian cat, banana, African elephant, and mountain bike. The model is intentionally lightweight and VGG-like, with BatchNorm, ReLU activations, max-pooling, a small residual connection in the first block, and an adaptive-pooling classifier head.

Model flow:

flowchart LR
    A["Input image: 3 x 128 x 128"] --> B["Stem: Conv + BatchNorm + ReLU"]
    B --> C["ConvBlock 1: residual + max-pool"]
    C --> D["ConvBlock 2: max-pool"]
    D --> E["ConvBlock 3: max-pool"]
    E --> F["Adaptive average pool + dropout"]
    F --> G["4-class linear head"]

Loading

Key pieces:

• models/small_net.py defines the SmallNet architecture.
• utils/dataset.py implements the dataset loader, class mapping, resizing/cropping, augmentation, and normalization.
• train.py, test.py, inference.py, and trace.py cover training, evaluation, sample inference, and shape/parameter tracing.
• arch.txt documents the model blueprint.
• artifacts/ includes the trained model checkpoint, metrics, confusion matrix, inference grid, and test report.
• submission/tmp/ preserves the packaged submission copy.

Recorded evaluation artifacts report:

• Top-1 accuracy: 70.08%
• Top-2 accuracy: 84.58%
• Test samples: 1200

Demo outputs already included in the repository:

Run:

cd hw2
bash run

The script installs dependencies, downloads/unzips the Purdue homework dataset if needed, prints the model blueprint, trains the model, evaluates it, and writes outputs into hw2/artifacts/.

Root-Level Files

• .gitignore excludes Python caches, environments, generated datasets, and large training artifacts such as hw2/data/, h2-data.zip, paper/data/, and paper/checkpoints/.
• LICENSE contains the Apache License 2.0.