def my_small_regime(input, stages, filters, classes, dropout_rate, graph_model,
graph_param, graph_file_path, init_subsample,
training): #regular regime 기반
stage = 1
graph_data = gg.graph_generator(graph_model, graph_param, graph_file_path,
'conv' + str(stage) + '_' + graph_model)
input = build_stage2(input, filters, dropout_rate, training, graph_data,
'conv' + str(stage))
filters *= 2
for stage in range(2, stages + 1):
graph_data = gg.graph_generator(
graph_model, graph_param, graph_file_path,
'conv' + str(stage) + '_' + graph_model)
input = build_stage(input, filters, dropout_rate, training, graph_data,
'conv' + str(stage))
filters *= 2
with tf.variable_scope('classifier'):
input = conv_block2(input, 1, 1280, 1, dropout_rate, training,
'conv_block_classifier')
input = tf.layers.average_pooling2d(input,
pool_size=input.shape[1:3],
strides=[1, 1])
input = tf.layers.flatten(input)
input = tf.layers.dropout(input, rate=0.3, training=training)
input = tf.layers.dense(input, units=classes)
return input
def regular_regime(input, stages, filters, classes, dropout_rate, graph_model,
graph_param, graph_file_path, training):
with tf.variable_scope('conv1'):
input = tf.layers.separable_conv2d(input,
filters=int(filters / 2),
kernel_size=[3, 3],
strides=[2, 2],
padding='SAME')
input = tf.layers.batch_normalization(input, training=training)
for stage in range(2, stages + 1):
graph_data = gg.graph_generator(
graph_model, graph_param, graph_file_path,
'conv' + str(stage) + '_' + graph_model)
input = build_stage(input, filters, dropout_rate, training, graph_data,
'conv' + str(stage))
filters *= 2
with tf.variable_scope('classifier'):
input = conv_block(input, 1, 1280, 1, dropout_rate, training,
'conv_block_classifier')
input = tf.layers.average_pooling2d(input,
pool_size=input.shape[1:3],
strides=[1, 1])
input = tf.layers.flatten(input)
input = tf.layers.dense(input, units=classes)
input = tf.layers.dropout(input, rate=dropout_rate)
return input
