| --- |
| license: mit |
| tags: |
| - pytorch |
| - image-classification |
| - gzsl |
| - agriculture |
| - weeds |
| - crops |
| - mobilenetv2 |
| - resnet18 |
| - squeezenet |
| - shufflenetv2 |
| - squeeze-and-excitation |
| - depthwise-separable-convolution |
| - weed-identification |
| --- |
| |
| # GZSL Weeds Identification: Lightweight Classifier Weights |
|
|
| This repository hosts the PyTorch checkpoints used in our generalized zero‑shot learning (GZSL) pipeline for weed identification in agricultural imagery. |
| Backbones were fine‑tuned on **CropAndWeed** and evaluated for cross‑dataset generalization to **Plant Phenotyping** and a self‑collected, real‑field dataset. |
|
|
| ## Available models |
|
|
| | File name | Architecture / variant | |
| |-----------|------------------------| |
| | `mobilenet.pt` | MobileNetV2 (ImageNet stem, width 1.0) | |
| | `resnet18.pt` | ResNet‑18 | |
| | `squeezenet.pt` | SqueezeNet 1.1 | |
| | `shufflenet.pt` | ShuffleNet V2 (baseline) | |
| | `shufflenet_squeeze_excitation.pt` | ShuffleNet V2 + Squeeze‑and‑Excitation (SE) | |
| | `shufflenet_sep_conv.pt` | ShuffleNet V2 + Depthwise Separable Convolution (SC) | |
| | `shufflenet_sep_conv_squeeze_excitation.pt` | ShuffleNet V2 + SC + SE | |
|
|
| ## Getting started |
|
|
| All inference scripts, data loaders and architecture definitions live in the companion GitHub repository: |
| https://github.com/SyArsRa/WeedZSL.git |
|
|
|
|
| The quick‑start guide there walks through: |
|
|
| 1. Instantiating the desired backbone (for example, MobileNetV2 or ShuffleNetV2 + SE) |
| 2. Loading the matching `.pt` file from this weights hub |
| 3. Running single‑image or batch inference |
| 4. Fine‑tuning on a custom dataset if needed |
|
|
| ## License |
|
|
| Weights and code are released under the MIT license for research and non‑commercial use. |
| See `LICENSE` for details or contact the maintainers for alternative licensing. |
|
|
| ## Citation |
|
|
| These checkpoints support a study currently submitted to a conference. |
| Please cite the forthcoming paper or contact the authors for an interim reference. |
|
|
|
|
