plot_improvement works OK, could use fine tuning

plot_improvement.py

@@ -1,79 +1,50 @@
 import argparse
-from numpy import load, ndarray
+import numpy as np
 import os
-
-parser = argparse.ArgumentParser()
-parser.add_argument("-f", "--filepath", required=True, help="Specify the file path to the agent.", type=str)
-args = parser.parse_args()
-
-filepath = args.filepath
-npdata = load(filepath)
-
-evaluations = ndarray.tolist(npdata['results'])
-# print(evaluations)
-sorted_evals = []
-for eval in evaluations:
-    sorted_evals.append(sorted(eval))
-
-# Now I have a sorted list.
-# Now just pop the first and last elements of each eval
-for eval in sorted_evals:
-    eval.pop(0)
-    eval.pop()
-
-print()
-# print(sorted_evals)
-
-# Now that I have my sorted evaluations, I can calculate the mean episode reward for each eval
-mean_eval_rewards = []
-for eval in sorted_evals:
-    mean_eval_rewards.append(sum(eval) / len(eval))
-    
-# Now I should have a list with the mean evaluation reward with the highest and lowest score tossed out
-print(mean_eval_rewards)
-print("num evals: " + str(len(mean_eval_rewards)))
-
-# I'm dealing with a 2D array. Each element contains an array of ten data points
-# The number of elements is going to vary for each training run
-# The number of evaluation episodes will be constant, 10.
-# I need to convert to a regular list first
-# I could iterate over each element
-
-agent_dirs = []
+from matplotlib import pyplot as plt
+
+def calc_stats(filepath):
+    # load the numpy file
+    data = np.load(filepath)["results"]
+    # sort the arrays and delete the first and last elements
+    data = np.sort(data, axis=1)
+    data = np.delete(data, -1, axis=1)
+    data = np.delete(data, 0, axis=1)
+    avg = round(np.mean(data), 2)
+    std = round(np.std(data), 2)
+    return avg, std
+
+# parser = argparse.ArgumentParser()
+# parser.add_argument("-f", "--filepath", required=True, help="Specify the file path to the agent.", type=str)
+# args = parser.parse_args()
+
+# Get the file paths and store in list.
+# For each file path, I want to calculate the mean reward. This would be the mean reward for the training run over all evaluations.
+# For each file path, append the mean reward to an averages list
+# Plot the averages!
+
+filepaths = []
 for d in os.listdir("agents/"):
     if "dqn_v2" in d:
-        agent_dirs.append(d)
-# Now I have a list of dirs with the evals.
-# Iterate over the dirs, append the file path, load the evals, calculate the average score of the eval, then return a list with averages
-eval_list = []
-for d in agent_dirs:
-    path = "agents/" + d + "/evaluations.npz"
-    evals = ndarray.tolist(load(path)["results"])
-    eval_list.append(evals)
-# for i in eval_list:
-#     print(i)
-#     print()
-
-def remove_outliers(evals: list) -> list:
-    trimmed = []
-    for eval in evals:
-        eval.sort()
-        eval.pop(0)
-        eval.pop()
-        trimmed.append(eval)
-    return trimmed
+        path = "agents/" + d + "/evaluations.npz"
+        filepaths.append(path)
+
+means = []
+stds = []
+for path in filepaths:
+    avg, std = calc_stats(path)
+    means.append(avg)
+    stds.append(std)
+
+runs = []
+for i in range(len(filepaths)):
+    runs.append(i + 1)
+plt.xlabel("training runs")
+plt.ylabel("score")
+plt.bar(runs, means)
+plt.bar(runs, stds)
+plt.legend(["Mean evaluation score", "Standard deviation"])
+plt.title("Average Evaluation Score and Standard Deviation\nAdjusted for Outliers\nAgent: dqn_v2")
+plt.show()
 
-avgs = [[]]
-index = 0
-for i in eval_list:
-    avgs.append(i)
-    for j in i:
-        j.sort()
-        j.pop()
-        j.pop(0)
-        avgs[index].append(sum(j) / len(j))
-    index += 1
-    
-print(avgs)
 
-