app.py

import streamlit as st
import numpy as np
from scipy.sparse import csr_matrix
import joblib
from sklearn.metrics import f1_score
from sklearn.decomposition import TruncatedSVD
from tqdm import tqdm
import tempfile
import os

st.set_page_config(page_title="RLF Model Tester", layout="wide")
st.title("Random Label Forest (RLF) - Test Dataset Evaluation")

# ----------------------- Parsing Function ----------------------- #
def parse_rcv1_sparse_file(filename):
    with open(filename, 'r') as f:
        header = f.readline()
        num_samples, num_features, num_labels = map(int, header.strip().split())
        
        feat_data, feat_rows, feat_cols = [], [], []
        label_data, label_rows, label_cols = [], [], []

        for row_idx, line in enumerate(f):
            line = line.strip()
            if not line:
                continue
            label_part, feature_part = line.split(' ', 1)
            labels = list(map(int, label_part.split(',')))
            features = feature_part.strip().split()

            for feat in features:
                fid, fval = feat.split(':')
                feat_rows.append(int(fid))
                feat_cols.append(row_idx)
                feat_data.append(float(fval))

            for lid in labels:
                label_rows.append(lid)
                label_cols.append(row_idx)
                label_data.append(1.0)

    X = csr_matrix((feat_data, (feat_rows, feat_cols)), shape=(num_features, num_samples), dtype=np.float32)
    Y = csr_matrix((label_data, (label_rows, label_cols)), shape=(num_labels, num_samples), dtype=np.uint8)
    return X.transpose(), Y.transpose()

# ----------------------- Prediction Functions ----------------------- #
def predict_scores_one_tree(x, tree):
    kmeans1 = tree['level1_kmeans']
    level2_dict = tree['level2_kmeans']
    tree_model = tree['model_tree']

    l1_cluster = kmeans1.predict(x.reshape(1, -1))[0]

    if l1_cluster not in level2_dict:
        return {}

    kmeans2 = level2_dict[l1_cluster]
    l2_cluster = kmeans2.predict(x.reshape(1, -1))[0]

    node_key = (l1_cluster, l2_cluster)
    if node_key not in tree_model:
        return {}

    classifiers = tree_model[node_key]['classifiers']
    label_ids = tree_model[node_key]['label_ids']

    preds = {}
    for clf, lid in zip(classifiers, label_ids):
        score = clf.decision_function(x.reshape(1, -1))[0]
        preds[lid] = score
    return preds

def predict_ensemble(X_test_reduced, ensemble):
    all_preds = []
    for i in range(X_test_reduced.shape[0]):
        instance_scores = {}
        x = X_test_reduced[i]
        for tree in ensemble:
            preds = predict_scores_one_tree(x, tree)
            for lid, score in preds.items():
                if lid not in instance_scores:
                    instance_scores[lid] = []
                instance_scores[lid].append(score)
        avg_scores = {lid: np.mean(scores) for lid, scores in instance_scores.items()}
        all_preds.append(avg_scores)
    return all_preds

def get_topk(pred_scores, k):
    top_labels = []
    for scores in pred_scores:
        if len(scores) == 0:
            top_labels.append([])
            continue
        sorted_labels = sorted(scores.items(), key=lambda x: -x[1])
        top_labels.append([lid for lid, _ in sorted_labels[:k]])
    return top_labels

def precision_at_k(preds_topk, Y_true, k):
    hits = 0
    total = len(preds_topk)
    for i, pred_labels in enumerate(preds_topk):
        true_labels = set(Y_true[i].nonzero()[1])
        hits += len(set(pred_labels) & true_labels)
    return hits / (total * k)

def evaluate_rforest(ensemble, X_test, Y_test, dim=300):
    svd = TruncatedSVD(n_components=dim, random_state=42)
    X_test_reduced = svd.fit_transform(X_test)

    pred_scores = predict_ensemble(X_test_reduced, ensemble)

    preds_at_1 = get_topk(pred_scores, k=1)
    preds_at_3 = get_topk(pred_scores, k=3)

    p1 = precision_at_k(preds_at_1, Y_test, k=1)
    p3 = precision_at_k(preds_at_3, Y_test, k=3)

    y_true = Y_test.toarray()
    y_pred_bin = np.zeros_like(y_true)
    for i, pred in enumerate(get_topk(pred_scores, k=5)):
        y_pred_bin[i, pred] = 1

    macro_f1 = f1_score(y_true, y_pred_bin, average='macro', zero_division=0)

    return p1, p3, macro_f1

# ----------------------- Load Pretrained Model ----------------------- #
@st.cache_resource
def load_model():
    return joblib.load("random_label_forest_model.pkl")

ensemble_model = load_model()

# ----------------------- User Interface ----------------------- #
uploaded_file = st.file_uploader("Upload Test Dataset (.txt in RCV1 format)", type=["txt"])

if uploaded_file:
    with tempfile.NamedTemporaryFile(delete=False, suffix=".txt") as tmp:
        tmp.write(uploaded_file.getvalue())
        test_file_path = tmp.name

    if st.button("Test Model"):
        st.info("Parsing dataset and evaluating. Please wait...")
        X_test, Y_test = parse_rcv1_sparse_file(test_file_path)
        p1, p3, macro_f1 = evaluate_rforest(ensemble_model, X_test, Y_test)

        st.success("Evaluation Completed!")
        st.metric("Precision@1", f"{p1:.4f}")
        st.metric("Precision@3", f"{p3:.4f}")
        st.metric("Macro F1 Score", f"{macro_f1:.4f}")