Improved plot visualisation.

- Fixed problem showing the accuracy as if it was error rate
- Added loss in additional axis
- Added option to export the plot

DenseNet/plot_results.py

@@ -3,18 +3,35 @@
 import numpy as np
 
 
-def plot_cifar10():
+def plot_cifar10(save=True):
 
-    with open("experiment_log_cifar10.json", "r") as f:
+    with open("./log/experiment_log_cifar10.json", "r") as f:
         d = json.load(f)
 
-    train_accuracy = 100 * (1 - np.array(d["train_loss"])[:, 1])
-    test_accuracy = 100 * (1 - np.array(d["test_loss"])[:, 1])
+    train_accuracy = 100 * (np.array(d["train_loss"])[:, 1])
+    test_accuracy = 100 * (np.array(d["test_loss"])[:, 1])
 
-    plt.plot(train_accuracy, color="tomato", linewidth=2)
-    plt.plot(test_accuracy, color="steelblue", linewidth=2)
+    fig = plt.figure()
+    ax1 = fig.add_subplot(111)
+    ax1.set_ylabel('Accuracy')
+    ax1.plot(train_accuracy, color="tomato", linewidth=2, label='train_acc')
+    ax1.plot(test_accuracy, color="steelblue", linewidth=2, label='test_acc')
+    ax1.legend(loc=0)
+
+    train_loss = np.array(d["train_loss"])[:, 0]
+    test_loss = np.array(d["test_loss"])[:, 0]
+
+    ax2 = ax1.twinx()
+    ax2.set_ylabel('Loss')
+    ax2.plot(train_loss, '--', color="tomato", linewidth=2, label='train_loss')
+    ax2.plot(test_loss, '--', color="steelblue", linewidth=2, label='test_loss')
+    ax2.legend(loc=1)
+
+    ax1.grid(True)
+
+    if save:
+        fig.savefig('./figures/plot_cifar10.svg')
 
-    plt.grid()
     plt.show()
     plt.clf()
     plt.close()