def learning(train_set, dataset_path, lbl_train, neural_nets, nn_for_learn, indexes, debug=False):
if nn_for_learn[NET_12]:
if debug:
print("First network learning")
for i in range(0, indexes[NET_12]):
if debug:
print(i)
all_images, all_labels, coordinates = prepare_images.prepare(dataset_path + train_set[i].decode('utf8'),
lbl_train[i], debug=debug)
if debug:
print("Image prepared")
images = all_images[all_labels == 1]
labels = np.ones(images.shape[0])
neg_size = int(labels.shape[0] * NEGATIVE_MULTIPLIER)
neg_indexes = np.random.choice(np.arange(all_images.shape[0] * all_images.shape[1]),
neg_size, replace=False)
neg_indexes = np.unravel_index(neg_indexes, (all_images.shape[0], all_images.shape[1]))
neg_labels = all_labels[neg_indexes]
neg_images = all_images[neg_indexes]
images = np.concatenate((images, neg_images))
labels = np.concatenate((labels, neg_labels))
if debug:
print("images.shape, labels.shape")
print(images.shape, labels.shape)
neural_nets[NET_12].learning(dataset=convert48to12(images), labels=labels, debug_print=debug, n_epochs=5)
if nn_for_learn[NET_24]:
if debug:
print("Second network learning")
for i in range(indexes[NET_24]):
all_images, all_labels, coordinates = prepare_images.prepare(dataset_path + train_set[i].decode('utf8'),
lbl_train[i], debug=debug)
predicted_labels = neural_nets[NET_12].predict(all_images)
images = all_images[predicted_labels == 1]
labels = all_labels[predicted_labels == 1]
nn_for_learn[NET_24].learning(dataset=convert48to24(images), labels=labels, debug_print=debug, n_epochs=10)
if nn_for_learn[NET_48]:
if debug:
print("Second network learning")
for i in range(indexes[NET_48]):
all_images, all_labels, coordinates = prepare_images.prepare(dataset_path + train_set[i].decode('utf8'),
lbl_train[i], debug=debug)
predicted_labels = neural_nets[NET_12].predict(all_images)
images = all_images[predicted_labels == 1]
predicted_labels = neural_nets[NET_24].predict(images)
images = all_images[predicted_labels == 1]
labels = all_labels[predicted_labels == 1]
nn_for_learn[NET_48].learning(dataset=convert48to24(images), labels=labels, debug_print=debug, n_epochs=10)
def testing_results(neural_nets, nn_params, nn_for_test, test_set_without_negatives, numbers_of_test_imgs=10):
first_net, second_net = neural_nets[:2]
first_net.load_params(nn_params[0])
if nn_for_test[1]:
second_net.load_params(nn_params[1])
print("Testing...")
test_img = np.array(list(test_set_without_negatives.keys())[:numbers_of_test_imgs])
lbl_test = np.array([test_set_without_negatives.get(key).get_coordinates() for key in test_img])
tp_all = np.zeros(numbers_of_test_imgs)
tn_all = np.zeros(numbers_of_test_imgs)
fp_all = np.zeros(numbers_of_test_imgs)
fn_all = np.zeros(numbers_of_test_imgs)
f1_score_all = np.zeros(numbers_of_test_imgs)
tn_percent_all = np.zeros(numbers_of_test_imgs)
for i in range(test_img.shape[0]):
imgs, lbls, coords = prepare_images.prepare(DATASET_PATH + test_img[i].decode('utf8'), lbl_test[i])
y_pred = np.zeros_like(lbls)
for j in range(imgs.shape[0]):
# TODO добавить nms в цепочку
y_pred[j] = first_net.predict(nn.convert48to24(imgs[j]))
if nn_for_test[1]:
tmp = imgs[y_pred == 1]
y_pred[y_pred == 1] = second_net.predict(tmp)
tmp = lbls - y_pred
tp = np.sum((y_pred == 1) & (lbls == 1))
tn = np.sum((y_pred == 0) & (lbls == 0))
fp = np.sum(tmp == -1)
fn = np.sum(tmp == 1)
f1_score = 2 * tp / (2 * tp + fn + fp)
tp_all[i] = tp
tn_all[i] = tn
fp_all[i] = fp
fn_all[i] = fn
f1_score_all[i] = f1_score
print(" f1 score = {}, true positive = {}, true negative = {}, false positive = {}, false negative = {}"
.format(f1_score, tp, tn, fp, fn))
tn_percent = tn / (tn + fp) * 100
tn_percent_all[i] = tn_percent
print("True negative percent from all negatives = {}".format(tn_percent))
tmp = np.arange(lbls.shape[0] * lbls.shape[1]).reshape(lbls.shape)
tmp = tmp[y_pred == 1]
rects = [coords.get(key, None) for key in tmp]
prepare_images.save_img_with_rectangles(DATASET_PATH, test_img[i].decode('utf8'), rects)
# prepare_images.show_rectangles(dataset_path + test_img[i].decode('utf8'), rects, show_type='opencv')
# rects = prepare_images.nms(rects, 0.3)
# prepare_images.show_rectangles(dataset_path + test_img[i].decode('utf8'), rects, show_type='opencv')
print("f1 score = {}, true positive = {}, true negative = {}, false positive = {}, false negative = {}"
.format(f1_score_all.mean(), tp_all.mean(), tn_all.mean(), fp_all.mean(), fn_all.mean()))
print("True negative percent from all negatives = {}".format(tn_percent_all.mean()))
return tp_all, tn_all, fp_all, fn_all, f1_score_all, tn_percent_all
