def custom_demo(use_old_data=True):
draw = Draw()
if use_old_data:
data = pickle.load("custom_data.p")
if not use_old_data:
img = np.concatenate(([1], np.ravel(draw.get_img())))
label = raw_input("What did you draw? ")
num_features = np.size(img) - 1 # Subtract 1 for the bias term
features = [img]
labels = [label]
data = Data(features=features, labels=labels, theta=np.zeros((1, 785)),
num_features=num_features, num_labels=1, num_examples=1,
alpha=0.01, epsilon=0.1, label_set=[label])
_gradient_ascent(data)
try:
while True:
img = np.concatenate(([1], np.ravel(draw.get_img())))
prediction = _predict(img, data)
print 'I think you drew a', data.label_set[prediction]
label = raw_input("What did you draw? ")
data.features.append(img)
data.labels.append(label)
data.num_examples += 1
if label not in data.label_set:
data.label_set.append(label)
data.num_labels += 1
new_row = np.zeros((1, data.num_features+1)) #New label means we must add a row to theta
data.theta = np.vstack((data.theta, new_row))
_gradient_ascent(data)
except KeyboardInterrupt:
save = raw_input("Save data? ").lower()
if save == 'y' or save == 'yes':
pickle.dump(data, "custom_data.p")
