def _comp_feature_importance_weight(self, class_id):
if not isinstance(class_id, list):
class_id = [class_id]
with tf.name_scope('feature_weight'):
self._feature_w_list = []
for idx, cid in enumerate(class_id):
# 將輸出轉為指定類別為1，其餘為0，shape為(nclass, 1)的one hot 形式
one_hot = tf.sparse_to_dense([[cid, 0]], [self._nclass, 1],
1.0)
# _out_act: 每個類別的激活值，reshape to (1, nclass)
out_act = tf.reshape(self._out_act, [1, self._nclass])
#class_act(1, 1) = _out_act(1, nclass) * one_hot(nclass, 1)，代表僅保留該類別的激活值
class_act = tf.matmul(out_act,
one_hot,
name='class_act_{}'.format(idx))
#該類別的激活值(class_act)對特徵圖的每個像素(_conv_out)求導，con_out是conv5_4的卷積結果(最後一層卷積特徵圖)
feature_grad = tf.gradients(class_act,
self._conv_out,
name='grad_{}'.format(idx))
#將feature_grad刪除第0維度，shape會和con_out的shape一樣，con_out中每張特徵圖的每個像素值，都有對應的導數(權值)。
feature_grad = tf.squeeze(tf.convert_to_tensor(feature_grad),
axis=0)
#feature_w: 對feature_grad做GAP，即針對特徵圖中每個像素值的導數(加權值)取平均，得一張特徵圖一個權值
feature_w = global_avg_pool(feature_grad,
name='feature_w_{}'.format(idx))
self._feature_w_list.append(feature_w)
def _create_model(self):
input_im = self.model_input[0]
keep_prob = self.model_input[1]
conv_out = self._create_conv(input_im)
init_b = tf.truncated_normal_initializer(stddev=0.01)
conv_cam = conv(conv_out,
3,
1024,
'conv_cam',
nl=tf.nn.relu,
wd=0.01,
init_b=init_b)
gap = global_avg_pool(conv_cam)
dropout_gap = dropout(gap, keep_prob, self.is_training)
with tf.variable_scope('cam'):
init = tf.truncated_normal_initializer(stddev=0.01)
#fc_w: shape is (1024, nclass)
fc_w = new_weights(
'weights',
1, [gap.get_shape().as_list()[-1], self._num_class],
initializer=init,
wd=0.01)
fc_cam = tf.matmul(dropout_gap, fc_w, name='output')
self.output = tf.identity(fc_cam, 'model_output')
self.prediction = tf.argmax(fc_cam, name='pre_label', axis=-1)
self.prediction_pro = tf.nn.softmax(fc_cam, name='pre_pro')
if self._inspect_class is not None:
with tf.name_scope('classmap'):
self.get_classmap(self._inspect_class, conv_cam, input_im)
def setup_graph(self):
self.input_im = self._vis_model.layer['input']
self._out_act = global_avg_pool(self._vis_model.layer['output'])
self._conv_out = self._vis_model.layer['conv_out']
self._nclass = self._out_act.shape.as_list()[-1]
self.pre_label = tf.nn.top_k(tf.nn.softmax(self._out_act),
k=5,
sorted=True)
def _create_model(self, image_a, image_b):
keep_prob = 1
self._vis_model.create_model([image_a, image_b, keep_prob])
self.input_im_a = self._vis_model.layer['input_a']
self.input_im_b = self._vis_model.layer['input_b']
self._out_act = global_avg_pool(self._vis_model.layer['output'])
self.pre_label = tf.nn.top_k(
tf.nn.softmax(self._out_act), k=5, sorted=True)
def setup_graph(self):
self.input_im = self._vis_model.layer['input']
# layer['output']: 是最後一個1*1 conv的輸出(命名為fc8)，其深度為nclass
# _out_act: 每個類別的激活值，用以代表每個類別的特徵圖
self._out_act = global_avg_pool(self._vis_model.layer['output'])
# layer['conv_out']: 是特徵抽取之捲基層的最後一層輸出(為'conv5_4'之輸出)
self._conv_out = self._vis_model.layer['conv_out']
self._nclass = self._out_act.shape.as_list()[-1]
#pre_label: 為所有類別的激活值中，前5大的激活值與對應的類別號碼，並從大致小排序
self.pre_label = tf.nn.top_k(tf.nn.softmax(self._out_act),
k=5,
sorted=True)
def _create_model(self):
inputs = self._input_dict['input']
conv_out = self._create_conv(inputs)
# init_b = tf.truncated_normal_initializer(stddev=0.01)
# conv_gap = conv(conv_out, 3, self._n_class, 'conv_gap',
# nl=tf.nn.relu, wd=0, init_b=init_b)
gap = global_avg_pool(conv_out)
self.layer['logits'] = gap
self.layer['feature'] = conv_out
self.layer['pred'] = tf.argmax(gap, name='pred', axis=-1)
self.layer['prob'] = tf.nn.softmax(gap, name='prob')
def _create_model(self, image):
keep_prob = 1
self._vis_model.create_model([image, keep_prob])
self.input_im = self._vis_model.layer['input']
#VGG19_FCN: 沒有全連接層，透過 4096個7*7*512 => 4096個1*1*4096 => 1000個1*1*4096 取代
#layer['output']: 一個四維的tensor
#_out_act: 為VGG19_FCN之輸出層(同樣命名為fc8)經過GAP的activate值 => 1000*1
self._out_act = global_avg_pool(self._vis_model.layer['output'])
#pre_label: 透過softmax將_out_act轉成機率型態，緊保留前k個機率最大的分類結果，並由大到小排序。
self.pre_label = tf.nn.top_k(tf.nn.softmax(self._out_act),
k=5,
sorted=True)
def _comp_feature_importance_weight(self, class_id):
if not isinstance(class_id, list):
class_id = [class_id]
with tf.name_scope('feature_weight'):
self._feature_w_list = []
for idx, cid in enumerate(class_id):
one_hot = tf.sparse_to_dense([[cid, 0]], [self._nclass, 1],
1.0)
out_act = tf.reshape(self._out_act, [1, self._nclass])
class_act = tf.matmul(out_act,
one_hot,
name='class_act_{}'.format(idx))
feature_grad = tf.gradients(class_act,
self._conv_out,
name='grad_{}'.format(idx))
feature_grad = tf.squeeze(tf.convert_to_tensor(feature_grad),
axis=0)
feature_w = global_avg_pool(feature_grad,
name='feature_w_{}'.format(idx))
self._feature_w_list.append(feature_w)
def _create_model(self):
with tf.name_scope('input'):
input_im = self.model_input[0]
keep_prob = self.model_input[1]
if self._is_rescale:
input_im =\
resize_tensor_image_with_smallest_side(input_im, 224)
self.layer['input'] = input_im
red, green, blue = tf.split(axis=3,
num_or_size_splits=3,
value=input_im)
input_bgr = tf.concat(axis=3,
values=[
blue - MEAN[0],
green - MEAN[1],
red - MEAN[2],
])
data_dict = {}
if self._is_load:
data_dict = np.load(self._pre_train_path, encoding='latin1').item()
inception5b = self._create_conv(input_bgr, data_dict)
gap = global_avg_pool(inception5b)
gap_dropout = dropout(gap, keep_prob, self.is_training)
fc1 = fc(gap_dropout, 1000, 'loss3_classifier', data_dict=data_dict)
self.layer['conv_out'] = inception5b
self.layer['output'] = fc1
self.layer['class_prob'] = tf.nn.softmax(fc1, name='class_prob')
self.layer['pre_prob'] = tf.reduce_max(self.layer['class_prob'],
axis=-1,
name='pre_prob')
self.layer['prediction'] = tf.argmax(self.layer['output'], axis=-1)
def _create_conv(inputs, data_dict):
arg_scope = tf.contrib.framework.arg_scope
with arg_scope([conv], trainable=False, data_dict=data_dict,
nl=tf.nn.relu):
conv1 = conv(inputs, 7, 64, name='conv1_7x7_s2', stride=2)
padding1 = tf.constant([[0, 0], [0, 1], [0, 1], [0, 0]])
conv1_pad = tf.pad(conv1, padding1, 'CONSTANT')
pool1 = max_pool(conv1_pad,
'pool1',
padding='VALID',
filter_size=3,
stride=2)
pool1_lrn = tf.nn.local_response_normalization(pool1,
depth_radius=2,
alpha=2e-05,
beta=0.75,
name='pool1_lrn')
conv2_reduce = conv(pool1_lrn, 1, 64, name='conv2_3x3_reduce')
conv2 = conv(conv2_reduce, 3, 192, name='conv2_3x3')
padding2 = tf.constant([[0, 0], [0, 1], [0, 1], [0, 0]])
conv2_pad = tf.pad(conv2, padding1, 'CONSTANT')
pool2 = max_pool(conv2_pad,
'pool2',
padding='VALID',
filter_size=3,
stride=2)
pool2_lrn = tf.nn.local_response_normalization(pool2,
depth_radius=2,
alpha=2e-05,
beta=0.75,
name='pool2_lrn')
with arg_scope([inception_layer], trainable=False, data_dict=data_dict):
inception3a = inception_layer(pool2_lrn,
64,
96,
128,
16,
32,
32,
name='inception_3a')
inception3b = inception_layer(inception3a,
128,
128,
192,
32,
96,
64,
name='inception_3b')
pool3 = max_pool(inception3b,
'pool3',
padding='SAME',
filter_size=3,
stride=2)
inception4a = inception_layer(pool3,
192,
96,
208,
16,
48,
64,
name='inception_4a')
inception4b = inception_layer(inception4a,
160,
112,
224,
24,
64,
64,
name='inception_4b')
inception4c = inception_layer(inception4b,
128,
128,
256,
24,
64,
64,
name='inception_4c')
inception4d = inception_layer(inception4c,
112,
144,
288,
32,
64,
64,
name='inception_4d')
inception4e = inception_layer(inception4d,
256,
160,
320,
32,
128,
128,
name='inception_4e')
pool4 = max_pool(inception4e,
'pool4',
padding='SAME',
filter_size=3,
stride=2)
inception5a = inception_layer(pool4,
256,
160,
320,
32,
128,
128,
name='inception_5a')
inception5b = inception_layer(inception5a,
384,
192,
384,
48,
128,
128,
name='inception_5b')
pool5 = global_avg_pool(inception5b, name='global_avg_pool')
return pool5