def testInvalidStrides(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'strides'):
pooling_layers.max_pooling2d(images, 3, strides=(1, 2, 3))
with self.assertRaisesRegexp(ValueError, 'strides'):
pooling_layers.max_pooling2d(images, 3, strides=None)
def testInvalidPoolSize(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'pool_size'):
pooling_layers.max_pooling2d(images, (1, 2, 3), strides=2)
with self.assertRaisesRegexp(ValueError, 'pool_size'):
pooling_layers.max_pooling2d(images, None, strides=2)
def testInvalidPoolSize(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'pool_size'):
pooling_layers.max_pooling2d(images, (1, 2, 3), strides=2)
with self.assertRaisesRegexp(ValueError, 'pool_size'):
pooling_layers.max_pooling2d(images, None, strides=2)
def testInvalidDataFormat(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'data_format'):
pooling_layers.max_pooling2d(images,
3,
strides=2,
data_format='invalid')
def testInvalidStrides(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'strides'):
pooling_layers.max_pooling2d(images, 3, strides=(1, 2, 3))
with self.assertRaisesRegexp(ValueError, 'strides'):
pooling_layers.max_pooling2d(images, 3, strides=None)
def cnn_2d(images, is_training):
"""
Build the model for 2D-CNN.
Inputs:
-- images: Images placeholder
-- is_training: bool placeholder, training or not
Output:
-- Logits: Return the output of the model
"""
# Build the CNN model
l_conv1 = conv2d(images,
CONV1_FILTERS,
KERNEL_SIZE1,
strides=STRIDE_CONV1,
activation=relu,
name='Conv1')
l_maxpool1 = max_pooling2d(l_conv1,
POOL_SIZE1,
POOL_SIZE1,
padding='same',
name='Maxpool1')
l_conv2 = conv2d(l_maxpool1,
CONV2_FILTERS,
KERNEL_SIZE2,
strides=STRIDE_CONV2,
activation=relu,
name='Conv2')
l_maxpool2 = max_pooling2d(l_conv2,
POOL_SIZE2,
POOL_SIZE2,
padding='same',
name='Maxpool2')
l_flatten = flatten(l_maxpool2, scope='Flatten')
l_fc1 = dense(l_flatten, FC1, activation=relu, name='Fc1')
l_drop = dropout(l_fc1, DROP_RATE, training=is_training, name='Dropout')
l_fc2 = dense(l_drop, FC2, activation=relu, name='Fc2')
logits = dense(l_fc2, NUM_CLASSES, name='Output')
return logits
def _make_conv_layers(self):
self.image_input = tf.placeholder(
dtype=tf.float32,
shape=[None, self.HEIGHT, self.WIDTH, 3],
name="image_input")
layer = self.image_input
for p, pooling_config in enumerate(self.config.poolings, 1):
for c, conv_config in enumerate(pooling_config.convolutions, 1):
layer = conv2d(
inputs=layer,
filters=conv_config.n_filters,
kernel_size=conv_config.kernel_size,
activation=relu,
padding="SAME",
name=f"conv_{p}_{c}",
)
layer = max_pooling2d(inputs=layer,
pool_size=2,
strides=pooling_config.strides,
name=f"pool_{p}")
self.last_pool_layer = layer
def mpool(self,
k_height,
k_width,
d_height=2,
d_width=2,
mode='VALID',
input_layer=None,
num_channels_in=None):
"""Construct a max pooling layer."""
if input_layer is None:
input_layer = self.top_layer
else:
self.top_size = num_channels_in
name = 'mpool' + str(self.counts['mpool'])
self.counts['mpool'] += 1
pool = pooling_layers.max_pooling2d(
input_layer, [k_height, k_width], [d_height, d_width],
padding=mode,
data_format=self.channel_pos,
name=name)
self.top_layer = pool
return pool
def testInvalidDataFormat(self):
height, width = 7, 9
images = tf.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(
ValueError, 'data_format'):
pooling_layers.max_pooling2d(images, 3, strides=2, data_format='invalid')
def _maxpool(self, bottom, stride):
return max_pooling2d(bottom, [stride,stride], [stride,stride])
def MNIST_model(features, labels, mode):
#input features
input_layer = tf.reshape(features["x"], [-1, 28, 28, 1])
if mode == tf.estimator.ModeKeys.EVAL:
#input_data = tf.constant(1.0) - input_layer
input_data = input_layer
else:
input_data = input_layer
#convolution 1
conv1 = conv2d(inputs=input_data,
filters=32,
kernel_size=[5, 5],
padding="same",
activation=relu,
use_bias=True)
#max pooling 1
pool1 = max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)
#convolution 2
conv2 = conv2d(inputs=pool1,
filters=64,
kernel_size=[5, 5],
padding="same",
activation=relu,
use_bias=True)
#max pooling 2
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2)
#Fully connected
pool2_flat = tf.reshape(pool2, [-1, 7 * 7 * 64])
dense_out = dense(inputs=pool2_flat, units=hidden_size, activation=relu)
dropout_out = dropout(inputs=dense_out,
rate=drop_rate,
training=mode == tf.estimator.ModeKeys.TRAIN)
#Generate a [28 * 28, 10] matrix as context
#initialize
w_a_1 = weight_variable(name="w_a_1", shape=[hidden_size, 28 * 28 * 10])
b_a_1 = bias_variable(name="b_a_1", shape=[28 * 28 * 10])
context = tf.add(tf.matmul(dropout_out, w_a_1), b_a_1)
context_matrix = tf.reshape(context, [-1, 28 * 28, 10])
#dot product layer
input_data_flat = tf.reshape(input_data, [-1, 28 * 28, 1])
input_data_tiled = tf.tile(input=input_data_flat, multiples=[1, 1, 10])
weighted_context = tf.multiply(input_data_tiled, context_matrix)
#Generate softmax result
logits = tf.reduce_sum(weighted_context, axis=[1])
#a dictionary of prediction operators
predictions = {
"logits": tf.multiply(logits, tf.constant(1.0), name="logit_out"),
#class prediction
"classes": tf.argmax(input=logits, axis=1),
#probability prediction
"probabilities": tf.nn.softmax(logits, name="softmax_tensor")
}
#prediction mode
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
#regularization
l1_regularizer = tf.contrib.layers.l1_regularizer(scale=reg_scale)
regularization_cost = tf.contrib.layers.apply_regularization(
l1_regularizer, [context_matrix])
#Calculate Loss (for both TRAIN and EVAL modes)
onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=10)
error_cost = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,
logits=logits)
#total lost
loss = regularization_cost + error_cost
#train mode
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
train_op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
#evaluation metrics (for EVAL mode)
eval_metric_ops = {
"accuracy":
tf.metrics.accuracy(labels=labels, predictions=predictions["classes"])
}
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
eval_metric_ops=eval_metric_ops)
def MNIST_model(features, labels, mode):
#input features
input_layer = tf.reshape(features["x"], [-1, 28, 28, 1])
if mode == tf.estimator.ModeKeys.EVAL:
input_data = tf.constant(1.0) - input_layer
else:
input_data = input_layer
#convolution 1
conv1 = conv2d(inputs=input_data,
filters=32,
kernel_size=[5, 5],
padding="same",
activation=relu,
use_bias=True)
#max pooling 1
pool1 = max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)
#convolution 2
conv2 = conv2d(inputs=pool1,
filters=64,
kernel_size=[5, 5],
padding="same",
activation=relu,
use_bias=True)
#max pooling 2
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2], strides=2)
#Fully connected
pool2_flat = tf.reshape(pool2, [-1, 7 * 7 * 64])
dense_out = dense(inputs=pool2_flat, units=hidden_size, activation=relu)
dropout_out = dropout(inputs=dense_out,
rate=drop_rate,
training=mode == tf.estimator.ModeKeys.TRAIN)
logits = dense(inputs=dropout_out, units=10)
#a dictionary of prediction operators
predictions = {
"logits": tf.multiply(logits, tf.constant(1.0), name="logit_out"),
#class prediction
"classes": tf.argmax(input=logits, axis=1),
#probability prediction
"probabilities": tf.nn.softmax(logits, name="softmax_tensor")
}
#prediction mode
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
#Calculate Loss (for both TRAIN and EVAL modes)
onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=10)
loss = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,
logits=logits)
#train mode
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer = tf.train.AdamOptimizer(learning_rate=1e-3)
train_op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
#evaluation metrics (for EVAL mode)
eval_metric_ops = {
"accuracy":
tf.metrics.accuracy(labels=labels, predictions=predictions["classes"])
}
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
eval_metric_ops=eval_metric_ops)
def _maxpool(self, bottom):
return max_pooling2d(bottom, [2, 2], [2, 2])
def _create_model(self):
self.input_image = tf.placeholder(tf.float32, shape=(None, None, None, self.img_input_channels), name='input_image_placeholder')
self.gt_image = tf.placeholder(tf.int32, shape=(None, None, None, self.num_classes), name='gt_image_placeholder')
self.gt_contours = tf.placeholder(tf.int32, shape=(None, None, None, self.num_classes), name='gt_contours_placeholder')
self.dropout_prob = tf.placeholder(dtype=tf.float32, shape=None, name='dropout_prob_placeholder')
self.lr = tf.placeholder(dtype=tf.float32, shape=None, name='learning_rate_placeholder')
scale_nc = self.hps.get('scale_nc')
with tf.variable_scope("encoder"):
with tf.variable_scope("block_1"):
conv1 = self._add_common_layers(conv2d(self.input_image, filters=32*scale_nc, kernel_size=3, padding='same'))
conv2 = self._add_common_layers(conv2d(conv1, filters=32*scale_nc, kernel_size=3, padding='same'))
with tf.variable_scope("block_2"):
mp2 = max_pooling2d(conv2, pool_size=2, strides=2, padding='same')
bn1 = self._add_common_layers(self._bottleneck(mp2, size=64*scale_nc))
bn2 = self._add_common_layers(self._bottleneck(bn1, size=64*scale_nc))
with tf.variable_scope("block_3"):
mp3 = max_pooling2d(bn2, pool_size=2, strides=2, padding='same')
bn3 = self._add_common_layers(self._bottleneck(mp3, size=128*scale_nc))
bn4 = self._add_common_layers(self._bottleneck(bn3, size=128*scale_nc))
with tf.variable_scope("block_4"):
mp4 = max_pooling2d(bn4, pool_size=2, strides=2, padding='same')
bn5 = self._add_common_layers(self._bottleneck(mp4, size=256*scale_nc))
bn6 = self._add_common_layers(self._bottleneck(bn5, size=256*scale_nc))
d1 = dropout(bn6, rate=self.dropout_prob)
with tf.variable_scope("block_5"):
mp5 = max_pooling2d(d1, pool_size=2, strides=2, padding='same')
bn7 = self._add_common_layers(self._bottleneck(mp5, size=256*scale_nc))
bn8 = self._add_common_layers(self._bottleneck(bn7, size=256*scale_nc))
d2 = dropout(bn8, rate=self.dropout_prob)
with tf.variable_scope("block_6"):
mp6 = max_pooling2d(d2, pool_size=2, strides=2, padding='same')
bn9 = self._add_common_layers(self._bottleneck(mp6, size=256*scale_nc))
bn10 = self._add_common_layers(self._bottleneck(bn9, size=256*scale_nc))
d3 = dropout(bn10, rate=self.dropout_prob)
self.img_descriptor = tf.reduce_mean(d3, axis=(1, 2))
with tf.variable_scope("decoder_seg"):
deconvs = []
deconvs.append(conv2d(conv2, filters=self.num_classes, kernel_size=3,
padding='same'))
deconvs.append(self._upsample(bn2, k=1))
deconvs.append(self._upsample(bn4, k=2))
deconvs.append(self._upsample(d1, k=3))
deconvs.append(self._upsample(d2, k=4))
deconvs.append(self._upsample(d3, k=5))
concat = tf.concat(deconvs, axis=3)
conv3 = conv2d(concat, filters=self.num_classes, kernel_size=3, padding='same')
ac1 = self._add_common_layers(conv3)
conv4 = conv2d(ac1, filters=self.num_classes, kernel_size=1, padding='same')
ac2 = self._add_common_layers(conv4)
self.preds_seg = softmax(ac2)
with tf.variable_scope("decoder_cont"):
deconvs = []
deconvs.append(conv2d(conv2, filters=self.num_classes, kernel_size=3,
padding='same'))
deconvs.append(self._upsample(bn2, k=1))
deconvs.append(self._upsample(bn4, k=2))
deconvs.append(self._upsample(d1, k=3))
deconvs.append(self._upsample(d2, k=4))
deconvs.append(self._upsample(d3, k=5))
concat = tf.concat(deconvs, axis=3)
conv3 = conv2d(concat, filters=self.num_classes, kernel_size=3, padding='same')
ac1 = self._add_common_layers(conv3)
conv4 = conv2d(ac1, filters=self.num_classes, kernel_size=1, padding='same')
ac2 = self._add_common_layers(conv4)
self.preds_cont = softmax(ac2)
cond1 = tf.greater_equal(self.preds_seg, self.threshold)
cond2 = tf.less(self.preds_cont, self.threshold)
conditions = tf.logical_and(cond1, cond2)
self.preds = tf.where(conditions, tf.ones_like(conditions), tf.zeros_like(conditions))
self._add_train_op()
self.summaries = tf.summary.merge_all()
