def __init__(self, model_def_file, image_height, image_width, raw_scale, gpu_mode):
logging.info('Loading net and associated files...')
if not gpu_mode:
os.environ['CUDA_VISIBLE_DEVICES'] = ""
self.low_level = [27, 32, 50, 56]#, 61, 62, 63, 64
self.mid_level = [0, 6, 7, 8, 9, 11, 12, 13, 17, 20, 21, 22, 23, 24, 25, 26, 28, 29, 30, 33, 35, 36, 37, 38, 39, 41, 42, 43, 44, 45, 46, 47, 48, 49, 51, 52, 53, 54, 55, 57, 58, 59, 60]
self.high_level = [1, 2, 3, 4, 5, 10, 14, 15, 16, 18, 19, 31, 34, 40]
self.net = hiarBayesGoogLeNet.build(image_height, image_width, 3, [len(self.low_level), len(self.mid_level), len(self.high_level)], weights="None")
self.net.load_weights(model_def_file)
self.labels = ['accessoryHeadphone', 'personalLess15', 'personalLess30', 'personalLess45', 'personalLess60', 'personalLarger60', 'carryingBabyBuggy', 'carryingBackpack', 'hairBald', 'footwearBoots', 'lowerBodyCapri', 'carryingOther', 'carryingShoppingTro', 'carryingUmbrella', 'lowerBodyCasual', 'upperBodyCasual', 'personalFemale', 'carryingFolder', 'lowerBodyFormal', 'upperBodyFormal', 'accessoryHairBand', 'accessoryHat', 'lowerBodyHotPants', 'upperBodyJacket', 'lowerBodyJeans', 'accessoryKerchief', 'footwearLeatherShoes', 'upperBodyLogo', 'hairLong', 'lowerBodyLongSkirt', 'upperBodyLongSleeve', 'lowerBodyPlaid', 'lowerBodyThinStripes', 'carryingLuggageCase', 'personalMale', 'carryingMessengerBag', 'accessoryMuffler', 'accessoryNothing', 'carryingNothing', 'upperBodyNoSleeve', 'upperBodyPlaid', 'carryingPlasticBags', 'footwearSandals', 'footwearShoes', 'hairShort', 'lowerBodyShorts', 'upperBodyShortSleeve', 'lowerBodyShortSkirt', 'footwearSneakers', 'footwearStocking', 'upperBodyThinStripes', 'upperBodySuit', 'carryingSuitcase', 'lowerBodySuits', 'accessorySunglasses', 'upperBodySweater', 'upperBodyThickStripes', 'lowerBodyTrousers', 'upperBodyTshirt', 'upperBodyOther', 'upperBodyVNeck', 'footwear', 'hair', 'lowerbody', 'upperbody']
[len(low_level), len(mid_level),
len(high_level)])
loss_func = 'binary_crossentropy' #weighted_categorical_crossentropy(alpha)
loss_weights = None
metrics = ['accuracy']
elif args.model == "hiarGoogLeNet_low":
model = hiarGoogLeNet_low.build(
image_width, image_height, 3,
[len(low_level), len(mid_level),
len(high_level)])
loss_func = 'binary_crossentropy' #weighted_categorical_crossentropy(alpha)
loss_weights = None
metrics = ['accuracy']
elif args.model == "hiarBayesGoogLeNet":
model = hiarBayesGoogLeNet.build(
image_height, image_width, 3,
[len(low_level), len(mid_level),
len(high_level)])
loss_func = 'binary_crossentropy' #bayes_binary_crossentropy(alpha, y_train)#weighted_categorical_crossentropy(alpha)
loss_func = weighted_binary_crossentropy(alpha)
loss_weights = None
metrics = ['accuracy']
elif args.model == "hiarBayesInception_v4":
model = hiarBayesInception_v4(
image_width, image_height, 3,
[len(low_level), len(mid_level),
len(high_level)])
loss_func = 'binary_crossentropy' #bayes_binary_crossentropy(alpha, y_train)#weighted_categorical_crossentropy(alpha)
loss_func = weighted_binary_crossentropy(alpha)
loss_weights = None
metrics = ['accuracy']
metrics = [weighted_acc]
def train(observation, action, save_name, batch_size_p=32, nb_epoch_p=50, ma_as_reward=True):
#print(int(save_name[save_name.rindex('_')+1:]))
heavy_augmentation = True
if heavy_augmentation:
datagen = ImageDataGenerator(
featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=45,
width_shift_range=0.25,
height_shift_range=0.25,
horizontal_flip=True,
vertical_flip=False,
zoom_range=0.5,
channel_shift_range=0.5,
fill_mode='nearest')
else:
datagen = ImageDataGenerator(
featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.125,
height_shift_range=0.125,
horizontal_flip=True,
vertical_flip=False,
fill_mode='nearest')
image_width = 160
image_height = 75
class_num = action.shape[0]
filename = r"../results/PETA.csv"
data = np.array(pd.read_csv(filename))[:, 1:]
length = len(data)
data_x = np.zeros((length, image_width, image_height, 3))
data_y = np.zeros((length, class_num))
low_level = []
mid_level = []
high_level = []
for i in range(action.shape[0]):
if action[i] == 0:
low_level.append(i)
elif action[i] == 1:
mid_level.append(i)
elif action[i] == 2:
high_level.append(i)
else:
print("ERROR ACTION!!!")
print(low_level)
print(mid_level)
print(high_level)
labels_list_file = r"/home/anhaoran/data/pedestrian_attributes_PETA/PETA/labels.txt"
labels_list_data = open(labels_list_file)
lines = labels_list_data.readlines()
tmp_list = []
for line in lines:
dat = line.split()
#print(attr)
tmp_list.append(dat[1])
#attributes_list = list(np.array(tmp_list))
attributes_list = list(np.array(tmp_list)[[0,1,2,3,4,5,6,7,8,9,27,32,40,50,56]])
print("-------------------------------------------------------------------------")
print("***Low level attributes: ")
for i in low_level:
print(attributes_list[i], end=", ")
print("\n***Medium level attributes: ")
for i in mid_level:
print(attributes_list[i], end=", ")
print("\n***High level attributes: ")
for i in high_level:
print(attributes_list[i], end=", ")
print()
print("-------------------------------------------------------------------------")
if np.shape(low_level)[0] == 0 or np.shape(mid_level)[0] == 0 or np.shape(high_level)[0] == 0:
reward = -1
return action, reward, reward>=0.9 or int(save_name[save_name.rindex('_')+1:]) >= 10
for i in range(length):
#img = image.load_img(path + m)
img = image.load_img(data[i, 0], target_size=(image_width, image_height, 3))
data_x[i] = image.img_to_array(img)
#data_y[i] = np.array(data[i, 1:1+class_num], dtype="float32")
data_y[i] = np.array(data[i, [1,2,3,4,5,6,7,8,9,10,28,33,41,51,57]], dtype="float32")
data_y = data_y[:, list(np.hstack((low_level, mid_level, high_level)))]
X_train = data_x[:9500]
X_test = data_x[9500:11400]
y_train = data_y[:9500]#, len(low_level)+len(mid_level):
y_test = data_y[9500:11400]#, len(low_level)+len(mid_level):
XX = data_x[11400:]
yy = data_y[11400:]
print("The shape of the X_train is: ", X_train.shape)
print("The shape of the y_train is: ", y_train.shape)
print("The shape of the X_test is: ", X_test.shape)
print("The shape of the y_test is: ", y_test.shape)
model = hiarBayesGoogLeNet.build(image_width, image_height, 3, [len(low_level), len(mid_level), len(high_level)])
model.compile(loss='binary_crossentropy', optimizer='adam', loss_weights=None, metrics=['accuracy'])
batch_size = batch_size_p
nb_epoch = nb_epoch_p
train_generator = datagen.flow(X_train, y_train, batch_size=batch_size)
val_generator = datagen.flow(X_test, y_test, batch_size=batch_size)
model.fit_generator(train_generator,
steps_per_epoch = int(X_train.shape[0] / batch_size),
epochs = nb_epoch,
validation_data = val_generator,
validation_steps = int(X_test.shape[0] / batch_size))
#model.save_weights('../models/imagenet_models/' + save_name + '_final_model.h5')
predictions_prob = model.predict(XX)
predictions = np.array(predictions_prob >= 0.5, dtype="float64")
label = yy
reward = mA(predictions, label)
return action, reward, True