def cifar_generator():
"""
X_data : (data_size, 32, 32, 3)
y_label : (data_size, 100)
batch_size : batch size
"""
(X_train, y_train), (X_test, y_test) = c100.load_data() # Load Data
y_train, y_test = prepare_output_data(y_train, y_test) # prepare y_label
data_size = len(X_train)
batch_iter_per_epoch = int(data_size / batch_size)
while True:
shuffle_idx = perm(np.arange(len(X_train)))
print('\n')
print("*" * 30)
print("Data Size : {}".format(data_size))
print("Batch Size : {}".format(batch_size))
print("Batch iterations per Epoch : {}".format(batch_iter_per_epoch))
print("*" * 30)
print(shuffle_idx[0:10])
for b in range(batch_iter_per_epoch):
batch_features = np.zeros((batch_size, inp_w, inp_h, inp_c))
batch_labels = np.zeros((batch_size, 100))
for b_i, i in enumerate(
range(b * batch_size, b * batch_size + batch_size)):
# choose random index in features
batch_features[b_i] = preprocess(X_train[shuffle_idx[i]],
color_type='RGB')
batch_labels[b_i] = y_train[shuffle_idx[i]]
yield batch_features, batch_labels
print("Done Generator")
def _batch(chunk, is_test=False):
"""
Takes a chunk of parsed annotations
returns value for placeholders of net's
input & loss layer correspond to this chunk
chunk : ['006098.jpg', [375, 500, [['boat', 92, 74, 292, 178], ['bird', 239, 88, 276, 133], ['bird', 93, 100, 142, 140]]]]
"""
S, B = cfg.cell_size, cfg.boxes_per_cell
C, labels = cfg.num_classes, cfg.classes_name
# preprocess
jpg = chunk[0]
w, h, allobj_ = chunk[1]
allobj = deepcopy(allobj_)
if not is_test:
path = os.path.join(cfg.imageset_location, jpg)
else:
path = os.path.join(cfg.test_imageset_location, jpg)
img = preprocess(path, allobj)
# Calculate regression target
cellx = 1. * w / S
celly = 1. * h / S
for obj in allobj:
centerx = .5 * (obj[1] + obj[3]) #xmin, xmax
centery = .5 * (obj[2] + obj[4]) #ymin, ymax
cx = centerx / cellx
cy = centery / celly
if cx >= S or cy >= S: return None, None
obj[3] = float(obj[3] - obj[1]) / w
obj[4] = float(obj[4] - obj[2]) / h
obj[3] = np.sqrt(obj[3])
obj[4] = np.sqrt(obj[4])
obj[1] = cx - np.floor(cx) # centerx
obj[2] = cy - np.floor(cy) # centery
obj += [int(np.floor(cy) * S + np.floor(cx))]
# show(im, allobj, S, w, h, cellx, celly) # unit test
# Calculate placeholders' values
probs = np.zeros([S * S, C])
confs = np.zeros([S * S, B])
coord = np.zeros([S * S, B, 4])
proid = np.zeros([S * S, C])
prear = np.zeros([S * S, 4])
for obj in allobj:
# print(type(obj), obj) # <class 'list'> ['horse', 0.024000000000000021, 0.48952095808383245, 0.92303846073714613, 0.85995404416970578, 24]
probs[obj[5], :] = [0.] * C
probs[obj[5], labels.index(obj[0])] = 1.
proid[obj[5], :] = [1] * C
coord[obj[5], :, :] = [obj[1:5]] * B
prear[obj[5], 0] = obj[1] - obj[3]**2 * .5 * S # xleft
prear[obj[5], 1] = obj[2] - obj[4]**2 * .5 * S # yup
prear[obj[5], 2] = obj[1] + obj[3]**2 * .5 * S # xright
prear[obj[5], 3] = obj[2] + obj[4]**2 * .5 * S # ybot
confs[obj[5], :] = [1.] * B
# Finalise the placeholders' values
upleft = np.expand_dims(prear[:, 0:2], 1)
botright = np.expand_dims(prear[:, 2:4], 1)
wh = botright - upleft
area = wh[:, :, 0] * wh[:, :, 1]
upleft = np.concatenate([upleft] * B, 1)
botright = np.concatenate([botright] * B, 1)
areas = np.concatenate([area] * B, 1)
# value for placeholder at input layer
inp_feed_val = img
# value for placeholder at loss layer
loss_feed_val = {
'probs': probs,
'confs': confs,
'coord': coord,
'proid': proid,
'areas': areas,
'upleft': upleft,
'botright': botright
}
return inp_feed_val, loss_feed_val
frameNum = -1
items = []
cap = cv2.VideoCapture(video_path)
try:
while True:
ret, frame = cap.read()
frameNum += 1
if frameNum > maxFrame or not ret:
break
if frameNum % frameSteps != 0:
continue
print("FrameNum : {}".format(frameNum))
img = preprocess(frame)
batch = np.expand_dims(img, axis=0)
net_out = model.predict(batch)
out_img, objects, is_object = post_progress(net_out[0],
im=frame,
is_save=False,
threshold=test_threshold)
if is_object:
for each_object in objects:
items.append(
DB_Item(videoId, frameNum, each_object[0], each_object[1],
each_object[2], each_object[3], each_object[4],
each_object[5]))
if len(items) >= 10:
# # In[7]:
# imagePath = './test/my_testset/001618.jpg'
# imagePath = './test/my_testset/000892.jpg'
# imagePath = './test/my_testset/000906.jpg'
# imagePath = './test/my_testset/000467.jpg'
# imagePath = './test/my_testset/000386.jpg'
# imagePath = './test/my_testset/many_person.jpg'
imagePath = './test/my_testset/person.jpg'
# imagePath = './test/my_testset/person_car3.jpg'
# imagePath = './test/my_testset/person.jpg'
# imagePath = './test/my_testset/car1.jpg'
image = cv2.imread(imagePath)
print("1", image.shape)
image = preprocess(image)
batch = np.expand_dims(image, axis=0)
print("4", batch.shape)
# # In[8]:
# out = model.predict(batch)
# print("5", out.shape)
# # In[9]:
# out_img = post_progress(out[0], im=image, is_save=False, threshold=0.1)
# print("6", out_img.shape)
# out_img = cv2.cvtColor(out_img, cv2.COLOR_BGR2RGB)
# plt.imshow(out_img)