def inference(self, input):
reshaped_input = tf.reshape(
input, [-1, self.img_side, self.img_side, self.num_channels])
self.inception_model = InceptionV3(include_top=False,
weights='imagenet',
input_tensor=reshaped_input)
raw_inception_features = self.inception_model.output
pooled_inception_features = AveragePooling2D(
(8, 8), strides=(8, 8), name='avg_pool')(raw_inception_features)
self.inception_features = Flatten(
name='flatten')(pooled_inception_features)
with tf.variable_scope('softmax_linear'):
weights = variable_with_weight_decay(
'weights', [self.num_features],
stddev=1.0 / math.sqrt(float(self.num_features)),
wd=self.weight_decay)
logits = tf.matmul(self.inception_features,
tf.reshape(weights, [-1, 1]))
zeros = tf.zeros_like(logits)
logits_with_zeros = tf.concat([zeros, logits], 1)
self.weights = weights
return logits_with_zeros
def inference(self, input):
reshaped_input = tf.reshape(
input, [-1, self.img_side, self.img_side, self.num_channels])
if self.use_InceptionResNet:
self.inception_model = InceptionResNetV2(
include_top=False,
weights='imagenet',
input_tensor=reshaped_input)
else:
self.inception_model = InceptionV3(include_top=False,
weights='imagenet',
input_tensor=reshaped_input)
raw_inception_features = self.inception_model.output
pooled_inception_features = AveragePooling2D(
(8, 8), strides=(8, 8), name='avg_pool')(raw_inception_features)
self.inception_features = Flatten(
name='flatten')(pooled_inception_features)
with tf.variable_scope('softmax_linear'):
# if binary, the proper dimension of Log Reg's weight param
# should be input_dim * 1.
# Thus to find the logits, we need to append a zero column
# to get the proper input of the softmax layer
if self.num_classes == 2:
weights = variable_with_weight_decay(
'weights', [self.num_features],
stddev=1.0 / math.sqrt(float(self.num_features)),
wd=self.weight_decay)
logits = tf.matmul(self.inception_features,
tf.reshape(weights, [-1, 1]))
zeros = tf.reshape(tf.zeros_like(logits)[:, 0], [-1, 1])
logits = tf.concat([zeros, logits], 1)
# if multilabels, logits would be simply the product of
# weight and latent features
else:
weights = variable_with_weight_decay(
'weights', [self.num_features * self.num_classes],
stddev=1.0 / math.sqrt(float(self.num_features)),
wd=self.weight_decay)
logits = tf.matmul(self.inception_features,
tf.reshape(weights, [-1, self.num_classes]))
self.weights = weights
return logits