def inference(self, inputs):
# resnet
with tf.variable_scope('first_layer'):
kernel = tf.get_variable('weights', [7, 7, 3, 64],
initializer=xavier_initializer(
dtype=tf.float32),
dtype=tf.float32)
conv = tf.nn.conv2d(inputs, kernel, [1, 2, 2, 1],
padding='SAME',
name='conv')
batch_norm = self.__batch_norm_wrapper(conv, shape=[0, 1, 2, 3])
conv = tf.nn.elu(batch_norm, 'elu')
grid = self.__put_kernels_on_grid(kernel, (8, 8))
tf.summary.image('conv1/features', grid, max_outputs=1)
grid = self.__put_activations_on_grid(conv, (8, 8))
tf.summary.image('conv1/activations', grid, max_outputs=1)
inputs = tf.nn.max_pool(conv, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME',
name='pool')
if self.net == NetworkType.RESNET34:
inputs = self.__resnet34(inputs)
elif self.net == NetworkType.RESNET50:
inputs = self.__resnet50(inputs)
# classify regions and add final region adjustments
with tf.variable_scope('fully_connected'):
fc = tf.reduce_mean(inputs, [1, 2])
class_weights = tf.get_variable('class_weights',
[self.conv_feature_count,
self.num_classes],
initializer=xavier_initializer(
dtype=tf.float32),
dtype=tf.float32)
class_bias = tf.get_variable("class_bias", [
self.num_classes],
initializer=tf.constant_initializer(
0.1),
dtype=tf.float32)
class_score = tf.matmul(fc, class_weights)
class_score = tf.nn.bias_add(class_score, class_bias)
return class_score
评论列表
文章目录
