def predict_ml_yolo(img_path, model, lang_db, dictionary, k_closest=1):
img = read_img(img_path)
boxes = load_image_boxes_yolo(img, dictionary, img_path)
if len(boxes) == 0: return []
pred_vects = model.predict_on_batch(
boxes) # Return predictions on the boxes, predictions 300d vector
predictions = lang_db.closest_labels(
pred_vects,
k_closest=k_closest) # Return closest label to each vector box
dict_counter = {}
preds = []
for k in range(k_closest):
for i in range(len(predictions)):
preds.append(predictions[i][k])
for pred in preds:
if pred in dict_counter:
dict_counter[pred] += 1
else:
dict_counter[pred] = 1
sorted_dict = sorted(dict_counter.items(),
key=operator.itemgetter(1),
reverse=True)
return [x[0] for x in sorted_dict[:25]]
def load_val_data(self, lang_db):
val_file_set = osp.join(self._devkit_path, 'ImageSets/DET/val_bts.txt')
val_folder_path = osp.join(self._devkit_path, 'Data', 'DET', 'val')
with open(val_file_set) as f:
images_data = [line.split() for line in f]
while 1:
for img_data in images_data:
img_path = osp.join(val_folder_path,
img_data[0] + self._image_ext[0])
x = read_img(img_path)
word_vec = lang_db.word_vector(img_data[1])
yield np.expand_dims(x, axis=0), np.expand_dims(word_vec,
axis=0)
def load_data(imdb, lang_db, batch_size=32):
total_batch = int(imdb.num_images / batch_size)
index = [ix for ix in range(imdb.num_images)]
while 1:
shuffle(index)
i = 0
for _ in range(total_batch):
X_train = []
y_train = []
for j in range(batch_size):
x_path, y = imdb.get_image_data_clean(index[i], lang_db)
x = read_img(x_path)
X_train.append(x)
y_train.append(y)
i = i + 1
yield np.array(X_train,
dtype=np.float32), np.array(y_train,
dtype=np.float32)
def predict_ml_random(img_path, model, lang_db, extra):
# Generate boxes for image
img = read_img(img_path, target_size=(299, 299))
boxes = generate_n_boxes(img, pool=extra)
pred_vects = model.predict_on_batch(
boxes) # Return predictions on the boxes, predictions 300d vector
predictions = lang_db.closest_labels(
pred_vects, k_closest=1) # Return closest label to each vector box
dict_counter = {}
for pred in predictions:
pred = pred[0]
# Retrieve closest word in semantic space
if pred in dict_counter:
dict_counter[pred] += 1
else:
dict_counter[pred] = 1
sorted_dict = sorted(dict_counter.items(),
key=operator.itemgetter(1),
reverse=True)
return [x[0] for x in sorted_dict[:20]]
def predict_image_singlelabel(img_path, model, lang_db, extra=None):
img = read_img(img_path, target_size=(299, 299))
img = np.expand_dims(img, axis=0)
pred_vector = model.predict(img)
predictions = lang_db.closest_labels(pred_vector, k_closest=20)
return predictions
