nailongnaiwa

KO | EN

gitlite — search

#python #hacktoberfest #react #ai #go #machine-learning #rust #windows #awesome #awesome-list #security #mcp #ios #openai

(★ 21)

识别奶龙与奶蛙

nailongnaiwa 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Nailong / Naiwa Image Classifier

This is a small PyTorch image classification project for distinguishing `nailong` and `naiwa` images. It includes dataset preparation, CNN training, classic image-feature baselines, a local web UI, and algorithm comparison reports.

## Highlights

- Train a lightweight CNN classifier.
- Fine-tune a pretrained torchvision classifier head.
- Preprocess Naiwa watermark regions.
- Build balanced train / test / generalization splits.
- Upload an image in a local web UI and choose the prediction algorithm, including an optional cloud multimodal LLM.
- Compare CNN predictions with classic image algorithms.
- Show image distribution charts in the UI.
- Export accuracy rankings, confusion matrices, and per-image predictions.

## Project Structure

## Install

```powershell
pip install -r requirements.txt
```

## Train

```powershell
python train_nailong_naiwa.py --cpu --epochs 12 --repeats 3 --batch-size 16 --image-size 128 --print-every 3 --out models/nailong_naiwa_balanced_cnn.pt
```

The training script rebuilds the balanced experiment dataset, creates splits, trains the CNN, and saves a checkpoint under `models/`.

## Fine-tune a Pretrained Model

```powershell
python train_pretrained_classifier.py --cpu --arch resnet18 --epochs 10 --batch-size 16 --image-size 160 --freeze-backbone --out models/nailong_naiwa_resnet18_head.pt
```

This loads an ImageNet-pretrained backbone, replaces the final classification head with a two-class Nailong/Naiwa head, and fine-tunes it on `nailong_naiwa_splits/train`. The saved `.pt` file can be evaluated and compared with the existing CNN models:

```powershell
python compare_nailong_algorithms.py --data nailong_naiwa_splits/test --split-dir nailong_naiwa_splits --models-dir models --out-dir reports --cpu
```

For a stricter comparison against training from scratch with the same architecture, add `--random-init`.

### Try a Vision Transformer

To upgrade the CNN experiment to a Transformer-style image classifier, use torchvision's ImageNet-pretrained ViT-B/16 backbone:

```powershell
python train_pretrained_classifier.py --cpu --arch vit_b_16 --epochs 10 --batch-size 4 --freeze-backbone --out models/nailong_naiwa_vit_b_16_head.pt
```

`vit_b_16` uses fixed 224 x 224 position embeddings in torchvision, so the training script defaults to `--image-size 224` for this architecture. After training, the ViT checkpoint can be used by the existing evaluator, comparison report, and local prediction UI:

```powershell
python evaluate_nailong_naiwa.py --cpu --model models/nailong_naiwa_vit_b_16_head.pt --data nailong_naiwa_splits/test
python compare_nailong_algorithms.py --data nailong_naiwa_splits/test --split-dir nailong_naiwa_splits --models-dir models --out-dir reports --cpu
python predict_nailong_naiwa_ui.py --host 127.0.0.1 --port 8770 --cpu --model models/nailong_naiwa_vit_b_16_head.pt
```

## Run the UI

```powershell
python predict_nailong_naiwa_ui.py --host 127.0.0.1 --port 8770 --cpu
```

Open:

```text
http://127.0.0.1:8770
```

The UI supports:

- `CNN deep model`
- `Cloud multimodal LLM`
- `Compare all algorithms`
- `Color mean prototype`
- `RGB histogram prototype`
- `Thumbnail kNN`
- `Edge histogram prototype`

It also displays train / test / generalization image distribution and confidence hints.

To enable the cloud multimodal LLM option, set an OpenAI-compatible API key before starting the UI:

```powershell
$env:OPENAI_API_KEY="your_api_key"
$env:OPENAI_MODEL="gpt-4o-mini"
python predict_nailong_naiwa_ui.py --host 127.0.0.1 --port 8770 --cpu
```

For another OpenAI-compatible provider, also set `OPENAI_BASE_URL`, for example `https://example.com/v1`. Without `OPENAI_API_KEY`, the CNN and classic algorithms still work, and the LLM option will show a configuration error.

## Compare Algorithms

```powershell
python compare_nailong_algorithms.py --data nailong_naiwa_splits/test --split-dir nailong_naiwa_splits --models-dir models --out-dir reports --cpu
```

Generated outputs:

- `reports/algorithm_comparison.md`
- `reports/algorithm_comparison.json`
- `reports/algorithm_predictions.csv`

Current test-set ranking:

| Algorithm | Accuracy |
| --- | ---: |
| CNN balanced model | 90.24% |
| Thumbnail kNN | 87.80% |
| RGB histogram prototype | 85.37% |
| Color mean prototype | 78.05% |
| Edge histogram prototype | 58.54% |
| ResNet18 pretrained head | 48.78% |

## Dataset Distribution

```powershell
python visualize_dataset_distribution.py --split-dir nailong_naiwa_splits
```

Example split:

```text
train:          nailong=60, naiwa=65
test:           nailong=20, naiwa=21
generalization: nailong=20, naiwa=21
```

## Notes

This is an educational experiment, not a production classifier. The dataset is small, the labels are limited to two classes, and predictions on unrelated images should be manually reviewed.

nailongnaiwa

nailongnaiwa 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

nailongnaiwa

nailongnaiwa 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README