Plot progress of obj scores across generations

protosignet/plot_results.py

@@ -1,9 +1,11 @@
-import ast
 from pathlib import Path
 
+import matplotlib as mpl
+import matplotlib.pyplot as plt
 import numpy as np
-import pandas as pd
-from natsort import natsorted
+import seaborn as sns
+
+from protosignet.util import eval_pareto, tag_objectives
 
 CUSTOM_PALETTE = ["#648FFF", "#2ECC71", "#8069EC", "#EA822C", "#D143A4", "#F1C40F", "#34495E"]
 
@@ -29,35 +31,64 @@
 }
 
 
-def plot_figure_1d(data_dp, save_dp):
+def plot_figure_1d(data_dp, fig_fp):
     """Generate scatterplot of obj 1 (simplicity) vs obj 2 (performance) over all runs/repeats.
 
     Args:
         data_dp (str): absolute path to data directory
-        save_dp (str): absolute path to save directory
+        fig_fp (str): absolute path for saving generated figure
     """
-    obj_scores = np.empty((0, 5))
-    for i, csv_fp in enumerate(natsorted(Path(data_dp).glob("*.csv"))):
-        df_rep = pd.read_csv(csv_fp)
-        os_rep = df_rep["objective"].values
-        for j in range(len(os_rep)):
-            os_gen = np.array(ast.literal_eval(os_rep[j]))
-            os_gen[:, 0] = os_gen[:, 0]
-            os_gen[:, 1] = os_gen[:, 1]
-            csv_i = i * np.ones(os_gen.shape[0])
-            gen_j = j * np.ones(os_gen.shape[0])
-            pop_k = np.arange(os_gen.shape[0])
-            address = np.column_stack((csv_i, gen_j, pop_k))
-            os_ijk = np.concatenate((os_gen, address), axis=1)
-            obj_scores = np.vstack((obj_scores, os_ijk))
-    df = pd.DataFrame(data=np.array(obj_scores), columns=["obj1", "obj2", "rep_i", "gen_j", "pop_k"])
-    print(df.shape)
+    df = tag_objectives(data_dp)
+    df_gen_001 = df.loc[df["gen_j"] == 0]
+    df_gen_010 = df.loc[df["gen_j"] == 9]
+    df_gen_100 = df.loc[df["gen_j"] == 99]
+    df_top = df.iloc[df.groupby("obj1")["obj2"].idxmax().values].copy()
+    df_top["is_pareto"] = eval_pareto(df_top[["obj1", "obj2"]].to_numpy())
+    df_pareto = df_top.loc[df_top["is_pareto"] == 1]
+    # print(df_pareto)
+    with plt.style.context(("seaborn-v0_8-whitegrid", CUSTOM_STYLE)):
+        fig, ax = plt.subplots(figsize=(24, 20))
+        sns.scatterplot(data=df_gen_001, x="obj1", y="obj2", edgecolor="#212121", facecolor="#2ECC71", alpha=0.8, linewidth=2, s=600)
+        sns.scatterplot(data=df_gen_010, x="obj1", y="obj2", edgecolor="#212121", facecolor="#F1C40F", alpha=0.8, linewidth=2, s=600)
+        sns.scatterplot(data=df_gen_100, x="obj1", y="obj2", edgecolor="#212121", facecolor="#EA822C", alpha=0.8, linewidth=2, s=600)
+        sns.scatterplot(data=df_pareto, x="obj1", y="obj2", edgecolor="#212121", facecolor="#D143A4", alpha=1.0, linewidth=2, s=600)
+        handles = [
+            mpl.lines.Line2D([], [], color="#2ECC71", marker="o", markersize=8, linewidth=0),
+            mpl.lines.Line2D([], [], color="#F1C40F", marker="o", markersize=8, linewidth=0),
+            mpl.lines.Line2D([], [], color="#EA822C", marker="o", markersize=8, linewidth=0),
+            mpl.lines.Line2D([], [], color="#D143A4", marker="o", markersize=8, linewidth=0),
+        ]
+        group_labels = ["Gen 1", "Gen 10", "Gen 100", "Best (Pareto)"]
+        ax.legend(
+            handles,
+            group_labels,
+            loc="best",
+            markerscale=4,
+            frameon=True,
+            shadow=False,
+            handletextpad=0.4,
+            borderpad=0.2,
+            labelspacing=0.2,
+            handlelength=1,
+        )
+        ax.set_xlabel("Simplicity")
+        ax.set_ylabel("Performance")
+        ax.xaxis.set_ticks(np.arange(0, 1.1, 0.2))
+        ax.set_xlim(-0.1, 1.1)
+        ax.yaxis.set_ticks(np.arange(0, 1.1, 0.2))
+        ax.set_ylim(-0.1, 1.1)
+        fig.tight_layout()
+        fig.canvas.draw()
+        fig.savefig(fig_fp, pad_inches=0.3, dpi=200, bbox_inches="tight", transparent=False)
+    plt.close("all")
 
 
 def main():
     data_dp = Path("/home/phuong/data/protosignet/dual_fm/data/")
     save_dp = Path("/home/phuong/data/protosignet/dual_fm/figs/")
-    plot_figure_1d(data_dp, save_dp)
+    save_dp.mkdir(parents=True, exist_ok=True)
+    fig_fp = save_dp / "fig_1d.png"
+    plot_figure_1d(data_dp, fig_fp)
 
 
 if __name__ == "__main__":

protosignet/util.py

@@ -1,4 +1,9 @@
+import ast
+from pathlib import Path
+
 import numpy as np
+import pandas as pd
+from natsort import natsorted
 
 
 def calc_hypervolume2D(pf_obj, ref):
@@ -18,3 +23,65 @@ def calc_hypervolume2D(pf_obj, ref):
     df2 = np.abs(pf_obj[:, 1] - ref[1])  # rectangle heights
     hv = (df1 * df2).sum()
     return hv
+
+
+def tag_objectives(data_dp):
+    """Tag every objective score set with an address (repeat index, generation index, population index).
+
+    Args:
+        data_dp (str): absolute path to data directory
+
+    Returns:
+        df (DataFrame): reorganized data with columns ["obj1", "obj2", "rep_i", "gen_j", "pop_k"]
+    """
+    obj_scores = np.empty((0, 5))
+    for i, csv_fp in enumerate(natsorted(Path(data_dp).glob("*.csv"))):
+        df_rep = pd.read_csv(csv_fp)
+        os_rep = df_rep["objective"].values
+        for j in range(len(os_rep)):
+            os_gen = np.array(ast.literal_eval(os_rep[j]))
+            rep_i = i * np.ones(os_gen.shape[0])
+            gen_j = j * np.ones(os_gen.shape[0])
+            pop_k = np.arange(os_gen.shape[0])
+            address = np.column_stack((rep_i, gen_j, pop_k))
+            os_ijk = np.concatenate((os_gen, address), axis=1)
+            obj_scores = np.vstack((obj_scores, os_ijk))
+    df = pd.DataFrame(data=np.array(obj_scores), columns=["obj1", "obj2", "rep_i", "gen_j", "pop_k"])
+    return df
+
+
+def dominates(p_obj, q_obj):
+    """Evaluates whether individual p dominates individual q.
+
+    Individual p dominates individual q if p is no worse than q in all objectives and p is
+    strictly better than q in at least one objective.
+
+    Args:
+        p_obj (1D array-like): array of j objective scores corresponding to individual p
+        q_obj (1D array-like): array of j objective scores corresponding to individual q
+
+    Returns:
+        True if p dominates q else False
+    """
+    return np.all(p_obj >= q_obj) and np.any(p_obj > q_obj)
+
+
+def eval_pareto(objectives):
+    pop_idx = range(len(objectives))
+    dom_count = [0 for i in pop_idx]
+    is_pareto = [0 for i in pop_idx]
+    for p in pop_idx:
+        for q in pop_idx:
+            if dominates(objectives[q], objectives[p]):
+                dom_count[p] += 1
+        if dom_count[p] == 0:
+            is_pareto[p] = 1
+    return np.array(is_pareto)
+
+
+def fetch_indiv(csv_fp, gen_j, pop_k):
+    df = pd.read_csv(Path(csv_fp))
+    pop_rep = df["population"].values
+    pop_gen = np.array(ast.literal_eval(pop_rep[int(gen_j)]))
+    indiv = pop_gen[int(pop_k)]
+    return indiv