def test_guided_backprop():
# placeholder for input image
image = tf.placeholder(tf.float32, shape=[None, None, None, 3])
# initialize input dataflow
# change '.png' to other image types if other types of images are used
input_im = ImageFromFile('.png', data_dir=IMPATH,
num_channel=3, shuffle=False)
# batch size has to be one
input_im.set_batch_size(1)
# initialize guided back propagation class
# use VGG19 as an example
# images will be rescaled to smallest side = 224 is is_rescale=True
model = GuideBackPro(vis_model=VGG19_FCN(is_load=False,
is_rescale=True))
# get op to compute guided back propagation map
# final output respect to input image
back_pro_op = model.get_visualization(image)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
im = input_im.next_batch()[0]
guided_backpro, label, o_im =\
sess.run([back_pro_op, model.pre_label,
model.input_im],
feed_dict={image: im})
print(label)
tf.reset_default_graph()
def train():
FLAGS = get_args()
style_name = os.path.splitext(FLAGS.styleim)[0]
style_im = scipy.misc.imread('../data/{}'.format(FLAGS.styleim))
style_im = [imagetool.resize_image_with_smallest_side(style_im, 512)]
style_shape = [style_im[0].shape[0], style_im[0].shape[1]]
train_data = ImageFromFile(ext_name='.jpg',
data_dir=DATA_PATH,
num_channel=3,
shuffle=True,
batch_dict_name=['im'],
pf=imagetool.im_normalize)
train_data.setup(epoch_val=0, batch_size=FLAGS.batch)
test_im = scipy.misc.imread('../data/cat.png')
test_im = [test_im]
train_model = FastStyle(content_size=256,
style_size=style_shape,
c_channel=3,
s_channel=3,
vgg_path=VGG_PATH,
s_weight=FLAGS.style,
c_weight=FLAGS.content,
tv_weight=FLAGS.tv)
train_model.create_train_model()
generate_model = FastStyle(c_channel=3)
generate_model.create_generate_model()
writer = tf.summary.FileWriter(SAVE_PATH)
saver = tf.train.Saver(var_list=tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, scope='style_net'))
sessconfig = tf.ConfigProto()
sessconfig.gpu_options.allow_growth = True
with tf.Session(config=sessconfig) as sess:
sess.run(tf.global_variables_initializer(),
feed_dict={train_model.style_image: style_im})
writer.add_graph(sess.graph)
# 40000 steps
for i in range(400):
train_model.train(sess,
train_data,
num_iteration=100,
summary_writer=writer)
generate_model.generate(sess, test_im, summary_writer=writer)
saver.save(sess, '{}{}_step_{}'.format(SAVE_PATH, style_name, i))
writer.close()
def get_predict_config(FLAGS):
dataset_test = ImageFromFile(FLAGS.type,
data_dir=config_path.test_data_dir,
shuffle=False,
resize=224,
num_channel=NUM_CHANNEL)
# dataset_test = ImageLabelFromFolder('.jpg',
# data_dir = config_path.data_dir,
# num_class = FLAGS.nclass,
# resize = 224,
# num_channel = NUM_CHANNEL)
prediction_list = [
# PredictionScalar(['pre_label'], ['label']),
# PredictionMeanScalar('accuracy/result', 'test_accuracy'),
PredictionMat('classmap/result', ['test']),
PredictionOverlay(['classmap/result', 'image'], ['map', 'image'],
color=True, merge_im=True),
PredictionImage(['image'], ['image'], color=True, merge_im=True)]
return PridectConfig(
dataflow=dataset_test,
model=VGGCAM(num_class=FLAGS.nclass, inspect_class=FLAGS.label,
is_load=True, pre_train_path=config_path.vgg_dir),
model_name=FLAGS.model,
predictions=prediction_list,
batch_size=FLAGS.bsize,
default_dirs=config_path)
def test_gradcam():
# merge several output images in one large image
merge_im = 1
grid_size = np.ceil(merge_im**0.5).astype(int)
# class label for Grad-CAM generation
# 355 llama 543 dumbbell 605 iPod 515 hat 99 groose 283 tiger cat
# 282 tabby cat 233 border collie 242 boxer
# class_id = [355, 543, 605, 515]
class_id = [283, 242]
# initialize Grad-CAM
# using VGG19
gcam = ClassifyGradCAM(vis_model=VGG19_FCN(is_load=False, is_rescale=True))
gbackprob = GuideBackPro(
vis_model=VGG19_FCN(is_load=False, is_rescale=True))
# placeholder for input image
image = tf.placeholder(tf.float32, shape=[None, None, None, 3])
# create VGG19 model
gcam.create_model(image)
gcam.setup_graph()
# generate class map and prediction label ops
map_op = gcam.get_visualization(class_id=class_id)
label_op = gcam.pre_label
back_pro_op = gbackprob.get_visualization(image)
# initialize input dataflow
# change '.png' to other image types if other types of images are used
input_im = ImageFromFile('.png',
data_dir=IMPATH,
num_channel=3,
shuffle=False)
input_im.set_batch_size(1)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
cnt = 0
merge_cnt = 0
o_im_list = []
im = input_im.next_batch()[0]
gcam_map, b_map, label, o_im =\
sess.run([map_op, back_pro_op, label_op, gcam.input_im],
feed_dict={image: im})
print(label)
o_im_list.extend(o_im)
for idx, cid, cmap in zip(count(), gcam_map[1], gcam_map[0]):
overlay_im = image_overlay(cmap, o_im)
weight_im = image_weight_mask(b_map[0], cmap)
try:
weight_im_list[idx].append(weight_im)
overlay_im_list[idx].append(overlay_im)
except NameError:
gcam_class_id = gcam_map[1]
weight_im_list = [[] for i in range(len(gcam_class_id))]
overlay_im_list = [[] for i in range(len(gcam_class_id))]
weight_im_list[idx].append(weight_im)
overlay_im_list[idx].append(overlay_im)
tf.reset_default_graph()
parser.add_argument('--type',
default='.jpg',
type=str,
help='image file extension')
return parser.parse_args()
if __name__ == '__main__':
FLAGS = get_args()
model = GoogleNet(is_load=True, pre_train_path=conf.PARA_DIR)
image = tf.placeholder(tf.float32, shape=[None, None, None, 3])
test_data = ImageFromFile(FLAGS.type,
data_dir=conf.DATA_DIR,
num_channel=3)
display_data(test_data, 'test_data')
word_dict = get_word_list('../data/imageNetLabel.txt')
model.create_model([image, 1])
test_op = tf.nn.top_k(tf.nn.softmax(model.layer['output']),
k=5,
sorted=True)
input_op = model.layer['input']
writer = tf.summary.FileWriter(conf.SAVE_DIR)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
def get_config(FLAGS):
# data for training
dataset_train = ImageLabelFromFolder(FLAGS.type,
data_dir=config_path.data_dir,
num_class=FLAGS.nclass,
resize=224,
num_channel=NUM_CHANNEL)
# Print image class name and label
# print(dataset_train.label_dict)
# Since the aim of training is visulization of class map, all the images
# are used for training. Using the training set as validation set is just
# for checking whether the training works correctly.
dataset_val = ImageLabelFromFolder(FLAGS.type,
data_dir=config_path.data_dir,
num_class=FLAGS.nclass,
resize=224,
num_channel=NUM_CHANNEL)
# Check accuracy during training using training set
inference_list_validation = InferScalars('accuracy/result',
'test_accuracy')
training_callbacks = [
ModelSaver(periodic=100),
TrainSummary(key='train', periodic=50),
FeedInferenceBatch(dataset_val, batch_count=10, periodic=100,
inferencers=inference_list_validation),
CheckScalar(['accuracy/result', 'loss/result'], periodic=10)]
inspect_class = None
if FLAGS.label > 0:
inspect_class = FLAGS.label
# Image use for inference the class acitivation map during training
dataset_test = ImageFromFile(FLAGS.type,
data_dir=config_path.infer_data_dir,
shuffle=False,
resize=224,
num_channel=NUM_CHANNEL)
# Check class acitivation map during training
inference_list_test = [
InferOverlay(['classmap/result', 'image'], ['map', 'image'],
color=True),
InferImages('classmap/result', 'map', color=True)]
training_callbacks += FeedInference(dataset_test, periodic=50,
infer_batch_size=1,
inferencers=inference_list_test),
return TrainConfig(
dataflow=dataset_train,
model=VGGCAM(num_class=FLAGS.nclass,
inspect_class=inspect_class,
learning_rate=0.001, is_load=True,
pre_train_path=config_path.vgg_dir),
monitors=TFSummaryWriter(),
callbacks=training_callbacks,
batch_size=FLAGS.bsize,
max_epoch=25,
summary_periodic=50,
default_dirs=config_path)
parser.add_argument('-cid', '--class_id', type=int, default=None,
help='Assign class id! Default = None')
parser.add_argument('-t', '--top', type=int, default=1,
help='前幾大激活值得導向反向傳播，預設為1。')
return parser.parse_args()
if __name__ == '__main__':
FLAGS = get_parse()
# placeholder for input image
image = tf.placeholder(tf.float32, shape=[None, None, None, 3])
# initialize input dataflow
# change '.png' to other image types if other types of images are used
input_im = ImageFromFile(FLAGS.imtype, data_dir = config.im_path,
num_channel=3, shuffle=False)
# batch size has to be one
input_im.set_batch_size(1)
# initialize guided back propagation class
# use VGG19 as an example
# images will be rescaled to smallest side = 224 if is_rescale=True
#class_id != None:會返回辨認為指定class的導向反向傳播的圖片
#class_id = None:則會告知Top5(imagnet預設為1000類)的分類結果與對應的機率，並透過導向反向傳播解釋。
model = GuideBackPro(vis_model=VGG19_FCN(is_load=True,
pre_train_path = config.vgg_path,
is_rescale=True)
, class_id = FLAGS.class_id
, top = FLAGS.top)
# get op to compute guided back propagation map
image = tf.placeholder(tf.float32, shape=[None, None, None, 3])
# create VGG19 model
gcam.create_model(image)
gcam.setup_graph()
# generate class map and prediction label ops
map_op = gcam.get_visualization(class_id=class_id)
label_op = gcam.pre_label
back_pro_op = gbackprob.get_visualization(image)
# initialize input dataflow
# change '.png' to other image types if other types of images are used
input_im = ImageFromFile('.png',
data_dir=IM_PATH,
num_channel=3,
shuffle=False)
input_im.set_batch_size(1)
writer = tf.summary.FileWriter(SAVE_DIR)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
writer.add_graph(sess.graph)
cnt = 0
merge_cnt = 0
# weight_im_list = [[] for i in range(len(class_id))]
o_im_list = []
while input_im.epochs_completed < 1:
im = input_im.next_batch()[0]
def get_config(FLAGS):
# data for training
dataset_train = ImageLabelFromFolder(FLAGS.type,
data_dir=config_path.data_dir,
num_class=FLAGS.nclass,
resize=224,
num_channel=NUM_CHANNEL)
# Print image class name and label
# print(dataset_train.label_dict)
# Since the aim of training is visulization of class map, all the images
# are used for training. Using the training set as validation set is just
# for checking whether the training works correctly.
dataset_val = ImageLabelFromFolder(FLAGS.type,
data_dir=config_path.val_data_dir,
num_class=FLAGS.nclass,
resize=224,
num_channel=NUM_CHANNEL)
# Check accuracy during training using training set
inference_list_validation = InferScalars('accuracy/result',
'test_accuracy')
training_callbacks = [
ModelSaver(periodic=100
), #每100個training step就儲存model到config.checkpoint_dir裡面。
TrainSummary(
key='train', periodic=50
), #每50個training step就會儲存訓練過程的準確率、loss等資訊到summary，方便之後用tensorboard查看
FeedInferenceBatch(
dataset_val,
batch_count=10,
periodic=
100, #每100個training step就透過所有dataset_val(同dataset_train)去測試準確率與lost。
inferencers=inference_list_validation),
CheckScalar(['accuracy/result', 'loss/result'], periodic=10)
] #每10個training step就顯示準確率與lost。
inspect_class = None
if FLAGS.label > 0:
#訓練時指定inspect_class(如：imagenet的282是tiger cat)，紀錄訓練過程中不同epochs的CAM變化。
inspect_class = FLAGS.label
# Image use for inference the class acitivation map during training
dataset_test = ImageFromFile(FLAGS.infer_type,
data_dir=config_path.infer_data_dir,
shuffle=False,
resize=224,
num_channel=NUM_CHANNEL)
# Check class acitivation map during training
inference_list_test = [
InferOverlay(['classmap/result', 'image'], ['map', 'image'],
color=True),
InferImages('classmap/result', 'map', color=True)
]
training_callbacks += FeedInference(
dataset_test,
periodic=
50, #每50個training step就測試infer image，並將測試結果儲存到config.infer_dir裡面。
infer_batch_size=1,
inferencers=inference_list_test),
return TrainConfig(dataflow=dataset_train,
model=VGGCAM(num_class=FLAGS.nclass,
inspect_class=inspect_class,
learning_rate=0.001,
is_load=True,
pre_train_path=config_path.vgg_dir),
monitors=TFSummaryWriter(),
callbacks=training_callbacks,
batch_size=FLAGS.bsize,
max_epoch=25,
summary_periodic=50,
default_dirs=config_path)
parser.add_argument('--id', type=int, default=None,
help='feature map id')
return parser.parse_args()
def im_scale(im):
return uim.im_rescale(im, [IM_SIZE, IM_SIZE])
if __name__ == '__main__':
FLAGS = get_parse()
input_im = ImageFromFile(FLAGS.imtype,
data_dir=config.im_path,
num_channel=3,
shuffle=False,
pf=im_scale,
)
input_im.set_batch_size(1)
print('size', input_im.size())
vizmodel = DeconvBaseVGG19(config.vgg_path,
feat_key=FLAGS.feat,
pick_feat=FLAGS.id)
vizmap = vizmodel.layers['deconvim'] #
print('vizmap', vizmap)
feat_op = vizmodel.feats # 4D Tensor, Dim is [N, H, W, C], depending on FLAGS.feat
max_act_op = vizmodel.max_act # 1D Tensor, the cur_feats_pick max value, depending on FLAGS.feat
act_size = vizmodel.receptive_size[FLAGS.feat]
act_scale = vizmodel.stride[FLAGS.feat]
default='.jpg',
type=str,
help='image file extension')
return parser.parse_args()
if __name__ == '__main__':
FLAGS = get_args()
model = GoogleNet(is_load=True, pre_train_path=conf.PARA_DIR)
image = tf.placeholder(tf.float32, shape=[None, None, None, 3])
test_data = ImageFromFile(
# ed: FLAGS.type이 argparse라서 ipython에서 실행시키기 힘드므로 아래와 같이 한다
# FLAGS.type,
'.jpg',
data_dir=conf.DATA_DIR,
num_channel=3)
display_data(test_data, 'test_data')
word_dict = get_word_list('../data/imageNetLabel.txt')
model.create_model([image, 1])
test_op = tf.nn.top_k(tf.nn.softmax(model.layer['output']),
k=5,
sorted=True)
input_op = model.layer['input']