Description
This is the repository for the code in Jia et al 2023. We design a SwinMin model for mineral photo image recognition in this paper, which embeds convolution information into the Transformer sequences and fuses multi-scale features with the proposed dynamic feature fusion module to exploit multi-scale contexts more effectively. SwinMin is based on Swin Transformer, and we borrow the code from its repository. can be used to easily to recognize 45 categories mineral.
Requirement
torch
timm
numpy
Usage
you can run this code to train SwinMin in your dataset:
python -m torch.distributed.launch --nproc_per_node 1 --master_port 12345 main.py --resume your-path-of-the-pretrained-Swin-Tiny --data-path your-path-of-the-dataset
Pretrained model in mineral dataset
| name | resolution | acc@1 | acc@5 | #params | model |
|---|---|---|---|---|---|
| SwinMin | 224*224 | 92.86% | 98.75% | 32.67M | link |
Dataset
── dataset_name
| ├── train
| | ├── class_1
| | | ├── 1_1_images
| | | ├── 1_2_images
| | | ├── .....
| | ├── class_2
| | | ├── 2_1_images
| | | ├── 2_2_images
| | | ├── .....
| | ├── .....
| | ├── class_X
| ├── val
| | ├── class_1
| | | ├── 1_1_images
| | | ├── 1_2_images
| | | ├── .....
| | ├── class_2
| | | ├── 2_1_images
| | | ├── 2_2_images
| | | ├── .....
| | ├── .....
| | ├── class_X