benchmarks/rsscn7.py

import torch
import numpy as np
import kagglehub
import glob
import os
import pandas as pd
from PIL import Image
from mapminer import models
from tqdm import tqdm
from sklearn.preprocessing import LabelEncoder

from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import accuracy_score


def load_models():
    dinov3 = models.DINOv3(architecture="vit-l-sat", pretrained=True).cuda()
    _ = dinov3.eval()

    lejepa = models.DINOv3(architecture="vit-l-sat", pretrained=True).cuda()

    weight_path = "lejepa-l.pt"
    state_dict = torch.load(weight_path)

    from collections import OrderedDict

    new_state_dict = OrderedDict()
    for k, v in state_dict.items():
        if k.startswith('encoder.model.'):
            new_key = k.replace('encoder.model.', 'model.')
            new_state_dict[new_key] = v
        else:
            new_state_dict[k] = v

    state_dict = new_state_dict
    lejepa.load_state_dict(state_dict,strict=False)

    _ = lejepa.eval()
    return dinov3, lejepa

def load_dataset():
    path  = kagglehub.dataset_download("nifulislam/rsscn7-dataset")
    print("Path to dataset files:", path)
    return path

def get_embedding(model, img_path, device="cuda"):
    img = Image.open(img_path).convert("RGB").resize((256, 256))

    # to tensor
    arr = np.asarray(img).transpose(2,0,1)
    x = torch.tensor(arr, dtype=torch.float32)[None].to(device)

    # normalize with model's preprocess
    x = model.normalize(x)

    # forward pass
    with torch.no_grad():
        out = model.model.forward_features(x)
        emb = out['x_norm_clstoken'].squeeze().cpu().numpy()
    return emb


def evaluate():
    path = load_dataset()
    root = "/root/.cache/kagglehub/datasets/nifulislam/rsscn7-dataset/versions/2/gamma-correct"
    classes = sorted(os.listdir(root))
    files = []
    for cls in classes:
        imgs = glob.glob(f"{root}/{cls}/*.jpg")
        for f in imgs:
            files.append((f, cls))

    df = pd.DataFrame(files,columns=['path','class'])
    files = df
    files = files.sample(frac=1.0,random_state=42)

    dinov3, lejepa = load_models()
    X_dino_ = []
    X_jepa_ = []
    y_ = []

    for index in tqdm(range(len(files))):
        path = files.iloc[index]['path']
        cls = files.iloc[index]['class']
        emb_d = get_embedding(dinov3, path)
        emb_j = get_embedding(lejepa, path)

        X_dino_.append(emb_d)
        X_jepa_.append(emb_j)
        y_.append(cls)

    X_dino = np.array(X_dino_)
    X_jepa = np.array(X_jepa_)

    # encode labels
    le = LabelEncoder()
    y_enc = le.fit_transform(y_)

    kf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)

    accs_dino = []
    accs_jepa = []

    for train_idx, test_idx in kf.split(X_dino, y_enc):

        Xtr_d, Xte_d = X_dino[train_idx], X_dino[test_idx]
        Xtr_j, Xte_j = X_jepa[train_idx], X_jepa[test_idx]
        ytr, yte = y_enc[train_idx], y_enc[test_idx]

        clf_dino = LogisticRegression(max_iter=8000, n_jobs=-1)
        clf_jepa = LogisticRegression(max_iter=8000, n_jobs=-1)

        clf_dino.fit(Xtr_d, ytr)
        clf_jepa.fit(Xtr_j, ytr)

        pred_dino = clf_dino.predict(Xte_d)
        pred_jepa = clf_jepa.predict(Xte_j)

        accs_dino.append(accuracy_score(yte, pred_dino))
        accs_jepa.append(accuracy_score(yte, pred_jepa))

    print("DINOv3 K-fold Accuracy:", round(np.mean(accs_dino)*100,1))
    print("LeJEPA K-fold Accuracy:", round(np.mean(accs_jepa)*100,1))