Added titles to some figures

experiments/searchspace_and_greedy_optimality.py

@@ -16,82 +16,83 @@
 scoreset = [0, 1]
 
 if __name__ == "__main__":
-    dataset = "thorax"
-    df = pd.read_csv(f"../../data/{dataset}.csv")
-    X = df.iloc[:, 1:].to_numpy()
-    y = df.iloc[:, 0].to_numpy()
-    wpos = y.mean()
-
-    opt, meas = "BACC", "BACC"
-    loss = get_loss(opt, wpos)
-    perf = lambda y, pred: -get_loss(meas, wpos)(y, pred)
-
-
-    def fit_predict(scores_):
-        features, scores = from_scorevec(scores_)
-        clf = _ClassifierAtK(features, scores=list(scores),
-                             initial_feature_thresholds=np.full_like(features, .5),
-                             threshold_optimizer=None)
-        clf.fit(X, y)
-        y_prob = clf.predict_proba(X)[:, 1]
-
-        return scores_, perf(y, y_prob)
-        # return accuracy_score(y, y_prob>.5)
-
-
-    G = nx.Graph()
-    node = dict()
-
-    with Pool(12) as p:
-        for scores, metric in tqdm(p.imap(fit_predict, product(scoreset, repeat=X.shape[1])),
-                                   total=len(scoreset) ** X.shape[1]):
-
-            id_ = np.count_nonzero(scores), metric
-            node[tuple(scores)] = id_
-
-            for index in np.nonzero(scores)[0]:
-                # get previous nodes by removing one non-zero feature
-                from_ = np.array(scores)
-                from_[index] = 0
-
-                G.add_edge(node[tuple(from_)], id_)
-
-    pos = {p: p for p in G.nodes}
-
-    sns.set(font_scale=1.5, rc={'text.usetex': True})
-    sns.set_style("whitegrid")
-    plt.rc('font', **{'family': 'serif'})
-    fig, ax = plt.subplots()
-    fig.set_size_inches(13, 4)
-    ax.set_ylabel("Balanced Accuracy in sample")
-    ax.set_xlabel("")
-
-    print("drawing nodes")
-    nx.draw_networkx_nodes(G, pos, node_color="#000000", node_size=10, ax=ax, alpha=.1)
-    print("drawing edges")
-    nx.draw_networkx_edges(G, pos, edge_color="#000000", ax=ax, width=0.5, node_size=10, alpha=.1)
-
-    # highlight cascade
-    print("fitting cascade")
-    psl = ProbabilisticScoringList(score_set=set(scoreset) - {0},
-                                   stage_loss=loss,
-                                   lookahead=LOOKAHEAD).fit(X, y)
-    cascade = [(i, perf(y, clf.predict_proba(X)[:, 1]))
-               for i, clf in enumerate(psl.stage_clfs)]
-
-    G = nx.Graph()
-    for u, v in zip(cascade, cascade[1:]):
-        G.add_edge(u, v)
-    print("drawing cascade")
-    nx.draw_networkx_nodes(G, pos, node_color="#1f78b4", node_size=50, ax=ax)
-    nx.draw_networkx_edges(G, pos, edge_color="#1f78b4", ax=ax, width=1, node_size=10)
-
-    # plt.box(False)
-    ax.tick_params(left=True, bottom=True, labelleft=True, labelbottom=True)
-    ax.xaxis.set_major_locator(MultipleLocator(1))
-    ax.set_xlim(-0.2, X.shape[1] + 0.2)
-
-    print("generating file")
-    code = f"local{opt}_globalSUM_metr{meas}"
-    fig.savefig(f"../../fig/{dataset}_{code}.pdf")
-    # fig.savefig(f"../../fig/{dataset}_{code}.png", dpi=300)
+    for dataset in ["thorax", "covid"]:
+        df = pd.read_csv(f"../data/{dataset}.csv")
+        X = df.iloc[:, 1:].to_numpy()
+        y = df.iloc[:, 0].to_numpy()
+        wpos = y.mean()
+
+        opt, meas = "BACC", "BACC"
+        loss = get_loss(opt, wpos)
+        perf = lambda y, pred: -get_loss(meas, wpos)(y, pred)
+
+
+        def fit_predict(scores_):
+            features, scores = from_scorevec(scores_)
+            clf = _ClassifierAtK(features, scores=list(scores),
+                                 initial_feature_thresholds=np.full_like(features, .5),
+                                 threshold_optimizer=None)
+            clf.fit(X, y)
+            y_prob = clf.predict_proba(X)[:, 1]
+
+            return scores_, perf(y, y_prob)
+            # return accuracy_score(y, y_prob>.5)
+
+
+        G = nx.Graph()
+        node = dict()
+
+        with Pool(12) as p:
+            for scores, metric in tqdm(p.imap(fit_predict, product(scoreset, repeat=X.shape[1])),
+                                       total=len(scoreset) ** X.shape[1]):
+
+                id_ = np.count_nonzero(scores), metric
+                node[tuple(scores)] = id_
+
+                for index in np.nonzero(scores)[0]:
+                    # get previous nodes by removing one non-zero feature
+                    from_ = np.array(scores)
+                    from_[index] = 0
+
+                    G.add_edge(node[tuple(from_)], id_)
+
+        pos = {p: p for p in G.nodes}
+
+        sns.set(font_scale=1.5, rc={'text.usetex': True})
+        sns.set_style("whitegrid")
+        plt.rc('font', **{'family': 'serif'})
+        fig, ax = plt.subplots()
+        fig.set_size_inches(13, 4)
+        ax.set_ylabel("Balanced Accuracy in sample")
+        ax.set_xlabel("Model Complexity")
+
+        print("drawing nodes")
+        nx.draw_networkx_nodes(G, pos, node_color="#000000", node_size=10, ax=ax, alpha=.1)
+        print("drawing edges")
+        nx.draw_networkx_edges(G, pos, edge_color="#000000", ax=ax, width=0.5, node_size=10, alpha=.1)
+
+        # highlight cascade
+        print("fitting cascade")
+        psl = ProbabilisticScoringList(score_set=set(scoreset) - {0},
+                                       stage_loss=loss,
+                                       lookahead=LOOKAHEAD).fit(X, y)
+        cascade = [(i, perf(y, clf.predict_proba(X)[:, 1]))
+                   for i, clf in enumerate(psl.stage_clfs)]
+
+        G = nx.Graph()
+        for u, v in zip(cascade, cascade[1:]):
+            G.add_edge(u, v)
+        print("drawing cascade")
+        nx.draw_networkx_nodes(G, pos, node_color="#1f78b4", node_size=50, ax=ax)
+        nx.draw_networkx_edges(G, pos, edge_color="#1f78b4", ax=ax, width=1, node_size=10)
+
+        # plt.box(False)
+        ax.tick_params(left=True, bottom=True, labelleft=True, labelbottom=True)
+        ax.xaxis.set_major_locator(MultipleLocator(1))
+        ax.set_xlim(-0.2, X.shape[1] + 0.2)
+
+        print("generating file")
+        code = f"local{opt}_globalSUM_metr{meas}"
+        fig.suptitle(dataset.title())
+        fig.savefig(f"../fig/{dataset}_{code}.pdf", bbox_inches="tight")
+        # fig.savefig(f"../../fig/{dataset}_{code}.png", dpi=300)