LeafSR

GAN-based 4x super-resolution for plant leaf images.

LeafSR packages a Kaggle/coursework experiment into a reproducible PyTorch project. It trains an ESRGAN-lite conditional GAN that predicts a high-resolution residual over a bicubic upsampled leaf image.

Project site: https://briskam.github.io/leafsr/

The project site includes real samples from the dataset and predictions generated from the trained checkpoint.

Results

Validation was measured with mean absolute error on held-out training images.

Method	Validation MAE
Bicubic baseline	18.20
LeafSR ESRGAN-lite checkpoint	16.89

The original run used a Tesla T4, mixed precision, EMA-smoothed generator weights, top-k checkpoint averaging, and 8-way test-time augmentation.

Highlights

• ESRGAN-lite generator with Residual-in-Residual Dense Blocks.
• Conditional discriminator over bicubic image + target/generated HR image.
• Combined Charbonnier, residual, perceptual, FFT, and adversarial losses.
• CUDA AMP support, cosine schedule, gradient clipping, EMA checkpointing, and early stopping.
• Submission/inference path with test-time augmentation and optional top-k checkpoint ensemble.

Repository Layout

configs/default.yaml        Training and inference configuration
src/leafsr/                 Reusable Python package
scripts/train.py            Train and validate the model
scripts/infer.py            Generate image outputs or Kaggle submission CSV
scripts/make_comparison.py  Build LR/bicubic/prediction/HR comparison grids
notebooks/                  Original notebook archive
tests/                      Lightweight smoke tests

Data Layout

The dataset is not committed. Place it like this, or pass custom paths in the config:

data/
  train_Low_Resolution/
    image_001.png
  train_High_Resolution/
    image_001.png
  test_Low_Resolution/
    test_001.png
  sample_submission.csv
  vgg19_weights.pth

Low-resolution and high-resolution training images must share filenames. The default scale factor is 4.

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Prepare Data

python scripts/prepare_data.py plant-leaves-super-resolution-challenge.zip --out-dir data

Train

python scripts/train.py --config configs/default.yaml

Useful overrides:

python scripts/train.py \
  --config configs/default.yaml \
  --data-root /path/to/dataset \
  --out-dir runs/leafsr

Inference

Generate enhanced PNG files:

python scripts/infer.py \
  --checkpoint runs/leafsr/best_leafsr.pt \
  --input-dir data/test_Low_Resolution \
  --output-dir runs/leafsr/predictions

Generate a Kaggle-style submission CSV:

python scripts/infer.py \
  --checkpoint runs/leafsr/best_leafsr.pt \
  --input-dir data/test_Low_Resolution \
  --sample-submission data/sample_submission.csv \
  --submission runs/leafsr/submission.csv

Visual Comparisons

python scripts/make_comparison.py \
  --checkpoint runs/leafsr/best_leafsr.pt \
  --lr-dir data/train_Low_Resolution \
  --hr-dir data/train_High_Resolution \
  --output assets/comparison_grid.png

The deployed frontend uses three real paired samples and precomputed predictions from the trained checkpoint:

Sample	Bicubic MAE	LeafSR MAE
agrivision_train_0000	22.21	21.37
agrivision_train_0005	14.96	13.86
agrivision_train_0010	20.21	18.85

Docker

docker build -t leafsr .
docker run --rm -v "$PWD/data:/app/data" -v "$PWD/runs:/app/runs" leafsr

Resume Bullet

Built LeafSR, a PyTorch super-resolution pipeline for plant leaf imagery using an ESRGAN-lite conditional GAN, VGG perceptual loss, EMA checkpointing, mixed precision training, and 8-way test-time augmentation, improving validation MAE from 18.20 bicubic baseline to 16.89.

Notes

The original dataset and VGG weights are expected to come from the competition/course environment and are intentionally excluded from git.