|
--- |
|
license: mit |
|
--- |
|
|
|
|
|
FLUX Detection model with sklearn API using wavelets and UMAP embeddings and then K-Nearest neighbors for classification. |
|
The process is as follows: DWT -> UMAP -> KNN. The discrete wavelet transform converts an image to its wavelet representation. |
|
The model was trained on around 300 FLUX images and 300 photographs from Unsplash but a new dataset can be used instead of FLUX. |
|
Accuracy varies depending on the dataset but can be anywhere from 80%-90% depending on the dataset you use. |
|
|
|
Methods include fit, predict, score, and predict_proba. First, load an image using PIL (Pillow) and then store using an array. Load the class using joblib and then predict. |
|
The model is already pretrained but can be trained again using fit. The images have to be resized to 512x512 before prediction, otherwise the model will give an error. |
|
|
|
```python |
|
model = joblib.load("flux_classifier.pkl") |
|
images = [Image.open("image.jpeg").resize((512, 512))] |
|
predictions = model.predict(images) |
|
``` |
