)
print("Now we evaluate on the test data")
results = model.evaluate(test_ds)
print("test loss:", results[0], ", test acc:", results[1])
# Now we generate the confusion matrix and loss graph
# Confusion matrix
# It's jank because we have a dataset of tuples, and there's not a great way to separate the halves without iterating twice.
# However, iterating twice will not work because our dataset is shuffled on each iteration.
test_y_hat = []
test_y = []
# This is actually very slow
for e in test_ds:
test_y_hat.extend(list(np.argmax(model.predict(e[0]), axis=1)))
test_y.extend(list(np.argmax(e[1].numpy(), axis=1)))
#test_y_hat = np.ndarray.flatten(np.array(test_y_hat))
#test_y = np.ndarray.flatten(np.array(test_y))
# I've checked both of these calls, they work correctly
confusion = tf.math.confusion_matrix(test_y, test_y_hat)
#plot_confusion_matrix(confusion)
save_confusion_matrix(confusion)
# Loss curve
#plot_loss_curve(history)
save_loss_curve(history)
)
with open(results_csv_path, "a") as f:
f.write("{},{},{},{},{}\n".format(distance, val_results[0],
val_results[1], test_results[0],
test_results[1]))
print("Calculate the confusion matrix")
total_confusion = None
f = None
for e in test_ds.unbatch().batch(50000).prefetch(5):
confusion = tf.math.confusion_matrix(np.argmax(e[1].numpy(),
axis=1),
np.argmax(model.predict(e[0]),
axis=1),
num_classes=RANGE)
if total_confusion == None:
total_confusion = confusion
else:
total_confusion = total_confusion + confusion
save_confusion_matrix(confusion,
path="confusion_distance-{}".format(distance))
save_loss_curve(history)
end_time = time.time()
with open(results_csv_path, "a") as f:
f.write("total time seconds: {}\n".format(end_time - start_time))
