def inception_C(x, input_channel, name):
with tf.variable_scope(name) as scope:
pool_1 = avg_pool(x, [1, 2, 2, 1], 1, 'SAME',
'inception_C_pool_1') # 1 - avg_pool
conv_1_1 = conv_layer(pool_1, [1, 1, input_channel, 256], [256], 1,
'SAME', 'inception_C_conv_1_1') # 1 - 1x1
conv_2_1 = conv_layer(x, [1, 1, input_channel, 256], [256], 1, 'SAME',
'inception_C_conv_2_1') # 2 - 1x1
conv_3_1 = conv_layer(x, [1, 1, input_channel, 384], [384], 1, 'SAME',
'inception_C_conv_3_1') # 3 - 1x1
conv_3_2_1 = conv_layer(conv_3_1, [1, 3, 384, 256], [256], 1, 'SAME',
'inception_C_conv_3_2_1') # 3 - 1x3
conv_3_2_2 = conv_layer(conv_3_1, [3, 1, 384, 256], [256], 1, 'SAME',
'inception_C_conv_3_2_2') # 3 - 3x1
conv_4_1 = conv_layer(x, [1, 1, input_channel, 384], [384], 1, 'SAME',
'inception_C_conv_4_1') # 4 - 1x1
conv_4_2 = conv_layer(conv_4_1, [1, 3, 384, 448], [448], 1, 'SAME',
'inception_C_conv_4_2') # 4 - 1x3
conv_4_3 = conv_layer(conv_4_2, [3, 1, 448, 512], [512], 1, 'SAME',
'inception_C_conv_4_3') # 4 - 3x1
conv_4_4_1 = conv_layer(conv_4_3, [3, 1, 512, 256], [256], 1, 'SAME',
'inception_C_conv_4_4_1') # 4 - 3x1
conv_4_4_2 = conv_layer(conv_4_3, [1, 3, 512, 256], [256], 1, 'SAME',
'inception_C_conv_4_4_2') # 4 - 1x3
return tf.concat([
conv_1_1, conv_2_1, conv_3_2_1, conv_3_2_2, conv_4_4_1, conv_4_4_2
],
axis=3,
name='{}_concat'.format(name)) # 8x8x1536
def inception_B(x, input_channel, name):
with tf.variable_scope(name) as scope:
pool_1 = avg_pool(x, [1, 2, 2, 1], 1, 'SAME',
'inception_B_pool_1') # 1 - avg_pool
conv_1_1 = conv_layer(pool_1, [1, 1, input_channel, 128], [128], 1,
'SAME', 'inception_B_conv_1_1') # 1 - 1x1
conv_2_1 = conv_layer(x, [1, 1, input_channel, 384], [384], 1, 'SAME',
'inception_B_conv_2_1') # 2 - 1x1
conv_3_1 = conv_layer(x, [1, 1, input_channel, 192], [192], 1, 'SAME',
'inception_B_conv_3_1') # 3 - 1x1
conv_3_2 = conv_layer(conv_3_1, [1, 7, 192, 224], [224], 1, 'SAME',
'inception_B_conv_3_2') # 3 - 1x7
conv_3_3 = conv_layer(conv_3_2, [7, 1, 224, 256], [256], 1, 'SAME',
'inception_B_conv_3_3') # 3 - 7x1
conv_4_1 = conv_layer(x, [1, 1, input_channel, 192], [192], 1, 'SAME',
'inception_B_conv_4_1') # 4 - 1x1
conv_4_2 = conv_layer(conv_4_1, [1, 7, 192, 192], [192], 1, 'SAME',
'inception_B_conv_4_2') # 4 - 1x7
conv_4_3 = conv_layer(conv_4_2, [7, 1, 192, 224], [224], 1, 'SAME',
'inception_B_conv_4_3') # 4 - 7x1
conv_4_4 = conv_layer(conv_4_3, [1, 7, 224, 224], [224], 1, 'SAME',
'inception_B_conv_4_4') # 4 - 1x7
conv_4_5 = conv_layer(conv_4_4, [1, 7, 224, 256], [256], 1, 'SAME',
'inception_B_conv_4_5') # 4 - 7x1
return tf.concat([conv_1_1, conv_2_1, conv_3_3, conv_4_5],
axis=3,
name='{}_concat'.format(name)) # 17x17x1024
def __init__(self, cfg):
super(FCN16s, self).__init__()
self.backbone = build_backbone(cfg)
num_classes = cfg.MODEL.NUM_CLASSES
# fc1
self.fc1 = conv_layer(512, 4096, 7)
self.relu1 = nn.ReLU(inplace=True)
self.drop1 = nn.Dropout2d()
# fc2
self.fc2 = conv_layer(4096, 4096, 1)
self.relu2 = nn.ReLU(inplace=True)
self.drop2 = nn.Dropout2d()
self.score_fr = conv_layer(4096, num_classes, 1)
self.score_pool4 = conv_layer(512, num_classes, 1)
self.upscore2 = bilinear_upsampling(num_classes,
num_classes,
4,
stride=2,
bias=False)
self.upscore16 = bilinear_upsampling(num_classes,
num_classes,
32,
stride=16,
bias=False)
def inception_A(x, input_channel, name):
with tf.variable_scope(name) as scope:
pool_1 = avg_pool(x, [1, 2, 2, 1], 1, 'SAME',
'inception_A_pool_1') # 1 - avg_pool
conv_1_1 = conv_layer(pool_1, [1, 1, input_channel, 96], [96], 1,
'SAME', 'inception_A_conv_1_1') # 1 - 1x1
conv_2_1 = conv_layer(x, [1, 1, input_channel, 96], [96], 1, 'SAME',
'inception_A_conv_2_1') # 2 - 1x1
conv_3_1 = conv_layer(x, [1, 1, input_channel, 64], [64], 1, 'SAME',
'inception_A_conv_3_1') # 3 - 1x1
conv_3_2 = conv_layer(conv_3_1, [3, 3, 64, 96], [96], 1, 'SAME',
'inception_A_conv_3_2') # 3 - 3x3
conv_4_1 = conv_layer(x, [1, 1, input_channel, 64], [64], 1, 'SAME',
'inception_A_conv_4_1') # 4 - 1x1
conv_4_2 = conv_layer(conv_4_1, [3, 3, 64, 96], [96], 1, 'SAME',
'inception_conv_A_4_2') # 4 - 3x3
conv_4_3 = conv_layer(conv_4_2, [3, 3, 96, 96], [96], 1, 'SAME',
'inception_conv_A_4_3') # 4 - 3x3
return tf.concat([conv_1_1, conv_2_1, conv_3_2, conv_4_3],
axis=3,
name='{}_concat'.format(name)) # 35x35x384
def reduction_A(x, input_channel, name): # 35x35 -> 17x17
with tf.variable_scope(name) as scope:
pool_1 = max_pool(x, [1, 3, 3, 1], 2, 'VALID',
'reduction_A_pool_1') # 1 384
conv_2 = conv_layer(x, [3, 3, input_channel, 384], [384], 2, 'VALID',
'reduction_A_conv_2') # 2 - 3x3 384
conv_3_1 = conv_layer(x, [1, 1, input_channel, 192], [192], 1, 'SAME',
'reduction_A_conv_3_1') # 3 - 1x1
conv_3_2 = conv_layer(conv_3_1, [3, 3, 192, 224], [224], 1, 'SAME',
'reduction_A_conv_3_2') # 3 - 3x3
conv_3_3 = conv_layer(conv_3_2, [3, 3, 224, 256], [256], 2, 'VALID',
'reduction_A_conv_3_3') # 3 - 3x3 256
return tf.concat([pool_1, conv_2, conv_3_3],
axis=3,
name='{}_concat'.format(name)) # 17x17x1024
def __init__(self, cfg):
super(FCN32s, self).__init__()
self.backbone = build_backbone(cfg)
num_classes = cfg.MODEL.NUM_CLASSES
#TODO : to be continue, why fc here use the conv_layer?
self.fc1 = conv_layer(512, 4096, 7)
self.relu1 = nn.ReLU(inplace=True)
self.drop1 = nn.Dropout2d()
self.fc2 = conv_layer(4096, 4096, 1)
self.relu2 = nn.ReLU(inplace=True)
self.drop2 = nn.Dropout2d()
self.score_fr = conv_layer(4096, num_classes, 1)
self.upscore = bilinear_upsampling(num_classes,
num_classes,
64,
stride=32,
bias=False)
def inception_layer(x,
conv_1_size,
conv_3_reduce_size,
conv_3_size,
conv_5_reduce_size,
conv_5_size,
pool_proj_size,
name='inception'):
""" Create an Inception Layer """
with tf.variable_scope(name) as scope:
conv_1 = conv_layer(x,
filter_height=1,
filter_width=1,
num_filters=conv_1_size,
name='{}_1x1'.format(name))
conv_3_reduce = conv_layer(x,
filter_height=1,
filter_width=1,
num_filters=conv_3_reduce_size,
name='{}_3x3_reduce'.format(name))
conv_3 = conv_layer(conv_3_reduce,
filter_height=3,
filter_width=3,
num_filters=conv_3_size,
name='{}_3x3'.format(name))
conv_5_reduce = conv_layer(x,
filter_height=1,
filter_width=1,
num_filters=conv_5_reduce_size,
name='{}_5x5_reduce'.format(name))
conv_5 = conv_layer(conv_5_reduce,
filter_height=5,
filter_width=5,
num_filters=conv_5_size,
name='{}_5x5'.format(name))
pool = max_pool(x,
stride=1,
padding='SAME',
name='{}_pool'.format(name))
pool_proj = conv_layer(pool,
filter_height=1,
filter_width=1,
num_filters=pool_proj_size,
name='{}_pool_proj'.format(name))
return tf.concat([conv_1, conv_3, conv_5, pool_proj],
axis=3,
name='{}_concat'.format(name))
def reduction_B(x, input_channel, name): # 17x17 -> 8x8
with tf.variable_scope(name) as scope:
pool_1 = max_pool(x, [1, 3, 3, 1], 2, 'VALID',
'reduction_B_pool_1') # 1 1024
conv_2_1 = conv_layer(x, [1, 1, input_channel, 192], [192], 1, 'SAME',
'reduction_B_conv_2_1') # 3 - 1x1
conv_2_2 = conv_layer(conv_2_1, [3, 3, 192, 192], [192], 2, 'VALID',
'reduction_B_conv_2_2') # 2 - 3x3 192
conv_3_1 = conv_layer(x, [1, 1, input_channel, 256], [256], 1, 'SAME',
'reduction_B_conv_3_1') # 3 - 1x1
conv_3_2 = conv_layer(conv_3_1, [1, 7, 256, 256], [256], 1, 'SAME',
'reduction_B_conv_3_2') # 3 - 1x7
conv_3_3 = conv_layer(conv_3_2, [7, 1, 256, 320], [320], 1, 'SAME',
'reduction_B_conv_3_3') # 3 - 7x1
conv_3_4 = conv_layer(conv_3_3, [3, 3, 320, 320], [320], 2, 'VALID',
'reduction_B_conv_3_4') # 3 - 3x3 320
return tf.concat([pool_1, conv_2_2, conv_3_4],
axis=3,
name='{}_concat'.format(name)) # 8x8x1536
def dcnn(x, keep_prob, img_result):
# stem
conv_1 = conv_layer(x, [3, 3, 3, 32], [32], 2, 'VALID', 'conv_1')
conv_2 = conv_layer(conv_1, [3, 3, 32, 32], [32], 1, 'VALID', 'conv_2')
conv_3 = conv_layer(conv_2, [3, 3, 32, 64], [64], 1, 'SAME', 'conv_3')
pool_4 = max_pool(conv_3, [1, 3, 3, 1], 2, 'VALID', 'pool_4') # 1
conv_4 = conv_layer(conv_3, [3, 3, 64, 96], [96], 2, 'VALID',
'conv_4') # 2
concat_5 = tf.concat([pool_4, conv_4], axis=3,
name='concat_5') # 73x73x160
conv_5_1 = conv_layer(concat_5, [1, 1, 160, 64], [64], 1, 'SAME',
'conv_5_1') # 1
conv_6_1 = conv_layer(conv_5_1, [3, 3, 64, 96], [96], 1, 'VALID',
'conv_6_1')
conv_5_2 = conv_layer(concat_5, [1, 1, 160, 64], [64], 1, 'SAME',
'conv_5_2') # 2
conv_6_2 = conv_layer(conv_5_2, [7, 1, 64, 64], [64], 1, 'SAME',
'conv_6_2')
conv_7_2 = conv_layer(conv_6_2, [1, 7, 64, 64], [64], 1, 'SAME',
'conv_7_2')
conv_8_2 = conv_layer(conv_7_2, [3, 3, 64, 96], [96], 1, 'VALID',
'conv_8_2')
concat_9 = tf.concat([conv_6_1, conv_8_2], axis=3,
name='concat_9') # 71x71x192
conv_10 = conv_layer(concat_9, [3, 3, 192, 192], [192], 2, 'VALID',
'conv_10') # 1
pool_10 = max_pool(concat_9, [1, 2, 2, 1], 2, 'VALID', 'pool_10') # 1
concat_11 = tf.concat([conv_10, pool_10], axis=3,
name='concat_11') # 71x71x192
# print("stem shape : ",concat_11.shape)
inception_a_12 = inception_A(concat_11, 384, 'inception_a_12')
inception_a_13 = inception_A(inception_a_12, 384, 'inception_a_13')
inception_a_14 = inception_A(inception_a_13, 384, 'inception_a_14')
inception_a_15 = inception_A(inception_a_14, 384,
'inception_a_15') # 4번 반복
# print("inception A shape : ",inception_a_15.shape)
reduction_a_16 = reduction_A(inception_a_15, 384, 'reduction_a_16')
# print("reduction A shape : ",reduction_a_16.shape)
inception_b_17 = inception_B(reduction_a_16, 1024, 'inception_b_17')
inception_b_18 = inception_B(inception_b_17, 1024, 'inception_b_18')
inception_b_19 = inception_B(inception_b_18, 1024, 'inception_b_19')
inception_b_20 = inception_B(inception_b_19, 1024, 'inception_b_20')
inception_b_21 = inception_B(inception_b_20, 1024, 'inception_b_21')
inception_b_22 = inception_B(inception_b_21, 1024, 'inception_b_22')
inception_b_23 = inception_B(inception_b_22, 1024, 'inception_b_23')
# print("inception B shape : ",inception_b_23.shape)
reduction_b_24 = reduction_B(inception_b_23, 1024, 'reduction_b_24')
# print("reduction B shape : ",reduction_b_24.shape)
inception_c_25 = inception_C(reduction_b_24, 1536, 'inception_c_25')
inception_c_26 = inception_C(inception_c_25, 1536, 'inception_c_26')
inception_c_27 = inception_C(inception_c_26, 1536, 'inception_c_27')
# print("inception C shape : ",inception_c_27.shape)
pool_28 = avg_pool(inception_c_27, [1, 8, 8, 1], 1, 'VALID',
'avg_pool_28') # 1 - avg_pool
dropout = tf.nn.dropout(pool_28, keep_prob)
flatten = tf.reshape(dropout, [-1, 1 * 1 * 1536])
logits = fc_layer(flatten, 1536, img_result, 'fc_layer')
print("logits : ", logits.shape)
return tf.nn.softmax(logits)