def unet_decoder(features, params, mode):
training = mode == tf.estimator.ModeKeys.TRAIN
up6 = layers.upconv_concat_layer(features['conv_5'], features['conv_4'], params, num_filters=512,
kernel_size=2, strides=2, pad='valid', training=training, name='6')
conv_6 = layers.conv_pool_layer(up6, filters=512, params=params, kernel_size=3, training=training,
pool=False, pad='valid', name='6')
conv_6_1 = layers.conv_pool_layer(conv_6, filters=512, params=params, kernel_size=3, training=training,
pool=False, pad='valid', name='6_1')
# print('SHAPE up6: ', up6.shape)
# print('SHAPE conv_6: ', conv_6.shape)
# print('SHAPE conv_6_1: ', conv_6_1.shape)
up7 = layers.upconv_concat_layer(conv_6_1, features['conv_3'], params, num_filters=256,
kernel_size=2, strides=2, pad='valid', training=training, name='7')
conv_7 = layers.conv_pool_layer(up7, filters=256, params=params, kernel_size=3, training=training,
pool=False, pad='valid', name='7')
conv_7_1 = layers.conv_pool_layer(conv_7, filters=256, params=params, kernel_size=3, training=training,
pool=False, pad='valid', name='7_1')
# print('SHAPE up7: ', up7.shape)
# print('SHAPE conv_7: ', conv_7.shape)
# print('SHAPE conv_7_1: ', conv_7_1.shape)
up8 = layers.upconv_concat_layer(conv_7_1, features['conv_2'], params, num_filters=128,
kernel_size=2, strides=2, pad='valid', training=training, name='8')
conv_8 = layers.conv_pool_layer(up8, filters=128, params=params, kernel_size=3, training=training,
pool=False, pad='valid', name='8')
conv_8_1 = layers.conv_pool_layer(conv_8, filters=128, params=params, kernel_size=3, training=training,
pool=False, pad='valid', name='8_1')
# print('SHAPE up8: ', up8.shape)
# print('SHAPE conv_8: ', conv_8.shape)
# print('SHAPE conv_8_1: ', conv_8_1.shape)
up9 = layers.upconv_concat_layer(conv_8_1, features['conv_1'], params, num_filters=64,
kernel_size=2, strides=2, pad='valid', training=training, name='9')
conv_9 = layers.conv_pool_layer(up9, filters=64, params=params, kernel_size=3, training=training,
pool=False, pad='valid', name='9')
conv_9_1 = layers.conv_pool_layer(conv_9, filters=64, params=params, kernel_size=3, training=training,
pool=False, pad='valid', name='9_1')
# print('SHAPE up9: ', up9.shape)
# print('SHAPE conv_9: ', conv_9.shape)
# print('SHAPE conv_9_1: ', conv_9_1.shape)
# dropout = tf.layers.dropout(conv_9_1, rate=params['dropout_rate'], training=training, name='dropout')
# logits = tf.layers.conv2d(conv_9_1, params['num_classes'], (1, 1), activation=tf.nn.relu, padding='valid',
# kernel_initializer=tf.keras.initializers.GlorotUniform(),
# name='logits')
# print('LOGITS SHAPE: ', logits.shape)
return conv_9_1
def fcn8s_description(samples, labels, params, mode, config):
training = mode == tf.estimator.ModeKeys.TRAIN
evaluating = mode == tf.estimator.ModeKeys.EVAL
num_classes = params['num_classes']
# learning_rate = params['learning_rate']
# tf.identity(learning_rate, "learning_rate")
# tf.summary.scalar('learning_rate', learning_rate)
height, width, _ = samples[0].shape
# Base Network (VGG_16)
conv1_1 = layers.conv_pool_layer(bottom=samples,
filters=64,
params=params,
training=training,
name="1_1",
pool=False)
conv1_2, pool1 = layers.conv_pool_layer(bottom=conv1_1,
filters=64,
params=params,
training=training,
name="1_2")
# print("SHAPE Conv_1: ", pool1.shape)
conv2_1 = layers.conv_pool_layer(bottom=pool1,
filters=128,
params=params,
training=training,
name="2_1",
pool=False)
conv2_2, pool2 = layers.conv_pool_layer(bottom=conv2_1,
filters=128,
params=params,
training=training,
name="2_2")
# print("SHAPE Conv_2: ", pool2.shape)
conv3_1 = layers.conv_pool_layer(bottom=pool2,
filters=256,
params=params,
training=training,
name="3_1",
pool=False)
conv3_2 = layers.conv_pool_layer(bottom=conv3_1,
filters=256,
params=params,
training=training,
name="3_2",
pool=False)
conv3_3, pool3 = layers.conv_pool_layer(bottom=conv3_2,
filters=256,
params=params,
training=training,
name="3_3")
# print("SHAPE Conv_3: ", pool3.shape)
conv4_1 = layers.conv_pool_layer(bottom=pool3,
filters=512,
params=params,
training=training,
name="4_1",
pool=False)
conv4_2 = layers.conv_pool_layer(bottom=conv4_1,
filters=512,
params=params,
training=training,
name="4_2",
pool=False)
conv4_3, pool4 = layers.conv_pool_layer(bottom=conv4_2,
filters=512,
params=params,
training=training,
name="4_3")
# print("SHAPE Conv_4: ", pool4.shape)
conv5_1 = layers.conv_pool_layer(bottom=pool4,
filters=512,
params=params,
training=training,
name="5_1",
pool=False)
conv5_2 = layers.conv_pool_layer(bottom=conv5_1,
filters=512,
params=params,
training=training,
name="5_2",
pool=False)
conv5_3, pool5 = layers.conv_pool_layer(bottom=conv5_2,
filters=512,
params=params,
training=training,
name="5_3")
# print("SHAPE Conv_5: ", pool5.shape)
# Fully Convolutional part
fconv6 = layers.conv_pool_layer(bottom=pool5,
filters=4096,
kernel_size=7,
params=params,
training=training,
name="fc6",
pool=False)
if (training):
fconv6 = tf.compat.v1.layers.dropout(inputs=fconv6,
rate=params["dropout_rate"],
name="drop_6")
# print("SHAPE FConv_6: ", fconv6.shape)
fconv7 = layers.conv_pool_layer(bottom=fconv6,
filters=4096,
kernel_size=1,
params=params,
training=training,
name="fc7",
pool=False)
if training:
fconv7 = tf.compat.v1.layers.dropout(inputs=fconv7,
rate=params["dropout_rate"],
name="drop_7")
# print("SHAPE FConv_7: ", fconv7.shape)
score_layer = tf.compat.v1.layers.conv2d(inputs=fconv7,
filters=num_classes,
kernel_size=1,
padding="valid",
data_format="channels_last",
activation=None,
name="Score_Layer_FC_2")
if (training):
score_layer = tf.compat.v1.layers.dropout(inputs=score_layer,
rate=params["dropout_rate"],
name="drop_8")
up_score_1 = layers.up_conv_add_layer(score_layer,
pool4,
params=params,
kernel_size=4,
num_filters=num_classes,
strides=2,
pad="same",
name="1")
# print("SHAPE Up Score: ", up_score_1.shape)
up_score_2 = layers.up_conv_add_layer(up_score_1,
pool3,
params=params,
kernel_size=4,
num_filters=num_classes,
strides=2,
pad="same",
name="2")
# up_score_3 = layers.up_conv_add_layer(up_score_2, pool2, params=params, kernel_size=4,
# num_filters=num_classes, strides=2, pad="same", name="3")
logits = layers.up_conv_layer(up_score_2,
num_filters=num_classes,
kernel_size=8,
strides=8,
params=params,
out_size=height,
pad="same",
name="final")
# print("SHAPE Up Score Final: ", up_final.shape)
# up_final = tf.layers.conv2d(up_final, num_classes, (1, 1), name="output", activation=tf.nn.sigmoid, padding="same",
# kernel_initializer=tf.initializers.variance_scaling(scale=0.001, distribution="uniform"))
return logits
def fcn2s_description(samples, labels, params, mode, config):
training = mode == tf.estimator.ModeKeys.TRAIN
evaluating = mode == tf.estimator.ModeKeys.EVAL
num_classes = params['num_classes']
learning_rate = params['learning_rate']
# tf.identity(learning_rate, 'learning_rate')
# tf.summary.scalar('learning_rate', learning_rate)
height, width, _ = samples[0].shape
# print('SHAPE LABELS: ', labels.shape)
# print('SHAPE Input: ', samples.shape)
# labels_1hot = labels
# Base Network (VGG_16)
conv1_1 = layers.conv_pool_layer(bottom=samples,
filters=64,
params=params,
training=training,
name='1_1',
pool=False)
conv1_2, pool1 = layers.conv_pool_layer(bottom=conv1_1,
filters=64,
params=params,
training=training,
name='1_2')
# print('SHAPE Conv_1: ', pool1.shape)
conv2_1 = layers.conv_pool_layer(bottom=pool1,
filters=128,
params=params,
training=training,
name='2_1',
pool=False)
conv2_2, pool2 = layers.conv_pool_layer(bottom=conv2_1,
filters=128,
params=params,
training=training,
name='2_2')
# print('SHAPE Conv_2: ', pool2.shape)
conv3_1 = layers.conv_pool_layer(bottom=pool2,
filters=256,
params=params,
training=training,
name='3_1',
pool=False)
conv3_2 = layers.conv_pool_layer(bottom=conv3_1,
filters=256,
params=params,
training=training,
name='3_2',
pool=False)
conv3_3, pool3 = layers.conv_pool_layer(bottom=conv3_2,
filters=256,
params=params,
training=training,
name='3_3')
# print('SHAPE Conv_3: ', pool3.shape)
conv4_1 = layers.conv_pool_layer(bottom=pool3,
filters=512,
params=params,
training=training,
name='4_1',
pool=False)
conv4_2 = layers.conv_pool_layer(bottom=conv4_1,
filters=512,
params=params,
training=training,
name='4_2',
pool=False)
conv4_3, pool4 = layers.conv_pool_layer(bottom=conv4_2,
filters=512,
params=params,
training=training,
name='4_3')
# print('SHAPE Conv_4: ', pool4.shape)
conv5_1 = layers.conv_pool_layer(bottom=pool4,
filters=512,
params=params,
training=training,
name='5_1',
pool=False)
conv5_2 = layers.conv_pool_layer(bottom=conv5_1,
filters=512,
params=params,
training=training,
name='5_2',
pool=False)
conv5_3, pool5 = layers.conv_pool_layer(bottom=conv5_2,
filters=512,
params=params,
training=training,
name='5_3')
# print('SHAPE Conv_5: ', pool5.shape)
# Fully Convolutional part
fconv6 = layers.conv_pool_layer(bottom=pool5,
filters=4096,
kernel_size=7,
params=params,
training=training,
name='fc6',
pool=False)
if (training):
fconv6 = tf.compat.v1.layers.dropout(inputs=fconv6,
rate=params['dropout_rate'],
name='drop_6')
# print('SHAPE FConv_6: ', fconv6.shape)
fconv7 = layers.conv_pool_layer(bottom=fconv6,
filters=4096,
kernel_size=1,
params=params,
training=training,
name='fc7',
pool=False)
if (training):
fconv7 = tf.compat.v1.layers.dropout(inputs=fconv7,
rate=params['dropout_rate'],
name='drop_7')
# print('SHAPE FConv_7: ', fconv7.shape)
# fconv8 = tf.layers.conv2d(inputs=fconv7, filters=1000, kernel_size=1, padding='same',
# data_format='channels_last', activation=None, name='fc8')
score_layer = tf.compat.v1.layers.conv2d(inputs=fconv7,
filters=num_classes,
kernel_size=1,
padding='valid',
data_format='channels_last',
activation=None,
name='Score_Layer_FC_2')
if (training):
score_layer = tf.compat.v1.layers.dropout(inputs=score_layer,
rate=params['dropout_rate'],
name='drop_8')
up_score_1 = layers.up_conv_add_layer(score_layer,
pool4,
params=params,
kernel_size=4,
num_filters=num_classes,
strides=2,
pad='same',
name='1')
# print('SHAPE Up Score: ', up_score_1.shape)
up_score_2 = layers.up_conv_add_layer(up_score_1,
pool3,
params=params,
kernel_size=4,
num_filters=num_classes,
strides=2,
pad='same',
name='2')
up_score_3 = layers.up_conv_add_layer(up_score_2,
pool2,
params=params,
kernel_size=4,
num_filters=num_classes,
strides=2,
pad='same',
name='3')
up_score_4 = layers.up_conv_add_layer(up_score_3,
pool1,
params=params,
kernel_size=4,
num_filters=num_classes,
strides=2,
pad='same',
name='4')
logits = layers.up_conv_layer(up_score_4,
num_filters=num_classes,
kernel_size=8,
strides=2,
params=params,
out_size=height,
pad='same',
name='final')
# print('SHAPE Up Score Final: ', up_final.shape)
output = tf.compat.v1.layers.conv2d(
logits,
1, (1, 1),
name='output',
activation=tf.nn.sigmoid,
padding='same',
kernel_initializer=tf.compat.v1.initializers.variance_scaling(
scale=0.001, distribution='uniform'))
return logits
def unet_encoder(samples, params, mode, name_sufix=''):
training = mode == tf.estimator.ModeKeys.TRAIN
# TODO: Remove this from here. Put it in the description, before calling the encoder.
if 'fusion' in params:
if params['fusion'] == 'early':
total_channels = samples.get_shape().as_list()[3]
num_channels = round(total_channels / 2)
#total_channels = tf.shape(samples)[3]
#num_channels = tf.cast(tf.round(total_channels / 2), tf.int32)
samples = tf.compat.v1.layers.conv2d(samples, filters=num_channels, kernel_size=(1, 1), strides=1,
padding='valid', activation=tf.nn.relu,
kernel_initializer=tf.keras.initializers.GlorotUniform(),
name='time_fusion')
# print('SHAPE INPUT: ', samples.shape)
# TODO: review the whole implementation, the number of filters and all the parameters
conv_1 = layers.conv_pool_layer(bottom=samples, filters=64, params=params, training=training,
name='1_1' + name_sufix, pool=False, pad='valid')
conv_1_2, pool1 = layers.conv_pool_layer(bottom=conv_1, filters=64, params=params, training=training,
name='1_2' + name_sufix, pad='valid')
# print('SHAPE Conv_1: ', conv_1.shape)
# print('SHAPE Conv_1_2: ', conv_1_2.shape)
# print('SHAPE Pool_1: ', pool1.shape)
conv_2 = layers.conv_pool_layer(bottom=pool1, filters=128, params=params, training=training,
name='2_1' + name_sufix, pool=False, pad='valid')
conv_2_1, pool2 = layers.conv_pool_layer(bottom=conv_2, filters=128, params=params, training=training,
name='2_2' + name_sufix, pad='valid')
# print('SHAPE Conv_2: ', conv_2.shape)
# print('SHAPE Conv_2_1: ', conv_2_1.shape)
# print('SHAPE Pool_2: ', pool2.shape)
conv_3 = layers.conv_pool_layer(bottom=pool2, filters=256, params=params, training=training,
name='3_1' + name_sufix, pool=False, pad='valid')
conv_3_1, pool3 = layers.conv_pool_layer(bottom=conv_3, filters=256, params=params, training=training,
name='3_2' + name_sufix, pad='valid')
# print('SHAPE Conv_3: ', conv_3.shape)
# print('SHAPE Conv_3_1: ', conv_3_1.shape)
# print('SHAPE Pool_3: ', pool3.shape)
conv_4 = layers.conv_pool_layer(bottom=pool3, filters=512, params=params, training=training,
name='4_1' + name_sufix, pool=False, pad='valid')
conv_4_1, pool4 = layers.conv_pool_layer(bottom=conv_4, filters=512, params=params, training=training,
name='4_2' + name_sufix, pad='valid')
# print('SHAPE Conv_4: ', conv_4.shape)
# print('SHAPE Conv_4_1: ', conv_4_1.shape)
# print('SHAPE Pool_4: ', pool4.shape)
conv_5_1 = layers.conv_pool_layer(bottom=pool4, filters=1024, params=params, training=training,
name='5_1' + name_sufix, pool=False, pad='valid')
conv_5_2 = layers.conv_pool_layer(bottom=conv_5_1, filters=1024, params=params, training=training,
name='5_2' + name_sufix, pool=False, pad='valid')
# print('SHAPE Conv_5: ', conv_5_1.shape)
# print('SHAPE Conv_5_1: ', conv_5_1.shape)
# print('SHAPE Pool_5: ', conv_5_2.shape)
return {'conv_1': conv_1_2,
'conv_2': conv_2_1,
'conv_3': conv_3_1,
'conv_4': conv_4_1,
'conv_5': conv_5_2}
def fcn32s_description(samples, labels, params, mode, config):
training = mode == tf.estimator.ModeKeys.TRAIN
evaluating = mode == tf.estimator.ModeKeys.EVAL
num_classes = params['num_classes']
height, width, _ = samples[0].shape
# print("SHAPE LABELS: ", labels.shape)
# print("SHAPE Input: ", samples.shape)
# labels_1hot = labels
# Base Network (VGG_16)
conv1_1 = layers.conv_pool_layer(bottom=samples, filters=64, params=params, training=training, name="1_1",
pool=False)
conv1_2, pool1 = layers.conv_pool_layer(bottom=conv1_1, filters=64, params=params, training=training, name="1_2")
# print("SHAPE Conv_1: ", pool1.shape)
conv2_1 = layers.conv_pool_layer(bottom=pool1, filters=128, params=params, training=training, name="2_1",
pool=False)
conv2_1, pool2 = layers.conv_pool_layer(bottom=conv2_1, filters=128, params=params, training=training, name="2_2")
# print("SHAPE Conv_2: ", pool2.shape)
conv3_1 = layers.conv_pool_layer(bottom=pool2, filters=256, params=params, training=training, name="3_1",
pool=False)
conv3_2 = layers.conv_pool_layer(bottom=conv3_1, filters=256, params=params, training=training, name="3_2",
pool=False)
conv3_3, pool3 = layers.conv_pool_layer(bottom=conv3_2, filters=256, params=params, training=training, name="3_3")
# print("SHAPE Conv_3: ", pool3.shape)
conv4_1 = layers.conv_pool_layer(bottom=pool3, filters=512, params=params, training=training, name="4_1",
pool=False)
conv4_2 = layers.conv_pool_layer(bottom=conv4_1, filters=512, params=params, training=training, name="4_2",
pool=False)
conv4_3, pool4 = layers.conv_pool_layer(bottom=conv4_2, filters=512, params=params, training=training, name="4_3")
# print("SHAPE Conv_4: ", pool4.shape)
conv5_1 = layers.conv_pool_layer(bottom=pool4, filters=512, params=params, training=training, name="5_1",
pool=False)
conv5_2 = layers.conv_pool_layer(bottom=conv5_1, filters=512, params=params, training=training, name="5_2",
pool=False)
conv5_3, pool5 = layers.conv_pool_layer(bottom=conv5_2, filters=512, params=params, training=training, name="5_3")
# print("SHAPE Conv_5: ", pool5.shape)
# Fully Convolutional part
fconv6 = layers.conv_pool_layer(bottom=pool5, filters=4096, kernel_size=7, params=params, training=training,
pool=False, name="fc6")
if(training):
fconv6 = tf.compat.v1.layers.dropout(inputs=fconv6, rate=params['dropout_rate'], name="drop_6")
# print("SHAPE FConv_6: ", fconv6.shape)
fconv7 = layers.conv_pool_layer(bottom=fconv6, filters=4096, kernel_size=1, params=params, training=training,
pool=False, name="fc7")
if(training):
fconv7 = tf.compat.v1.layers.dropout(inputs=fconv7, rate=params['dropout_rate'], name="drop_7")
# print("SHAPE FConv_7: ", fconv7.shape)
score_layer = tf.compat.v1.layers.conv2d(inputs=fconv7, filters=num_classes, kernel_size=1, padding="same",
data_format="channels_last", activation=None, name="score_layer")
# print("SHAPE Score Layer: ", score_layer.shape)
logits = layers.up_conv_layer(score_layer, num_filters=num_classes, kernel_size=64, strides=32,
params=params, out_size=height, pad="same", name="uc")
return logits