def create(self):
conv1, _ = cnn.conv(self.input,
11,
11,
96,
4,
4,
padding='VALID',
name='conv1')
pool1 = cnn.max_pool(conv1, 3, 3, 2, 2, padding='VALID', name='pool1')
norm1 = cnn.lrn(pool1, 2, 2e-05, 0.75, name='norm1')
conv2, _ = cnn.conv(norm1, 5, 5, 256, 1, 1, groups=2, name='conv2')
pool2 = cnn.max_pool(conv2, 3, 3, 2, 2, padding='VALID', name='pool2')
norm2 = cnn.lrn(pool2, 2, 2e-05, 0.75, name='norm2')
conv3, _ = cnn.conv(norm2, 3, 3, 384, 1, 1, name='conv3')
conv4, _ = cnn.conv(conv3, 3, 3, 384, 1, 1, groups=2, name='conv4')
conv5, _ = cnn.conv(conv4, 3, 3, 256, 1, 1, groups=2, name='conv5')
pool5 = cnn.max_pool(conv5, 3, 3, 2, 2, padding='VALID', name='pool5')
flattened = cnn.flatten(pool5)
fc6, _ = cnn.fc(flattened, 4096, name='fc6')
dropout6 = tf.cond(
self.IS_TRAIN,
lambda: cnn.dropout(fc6, self.KEEP_PROB, name='dropout6'),
lambda: fc6)
self.fc7, _ = cnn.fc(dropout6, 4096, name='fc7')
dropout7 = tf.cond(
self.IS_TRAIN,
lambda: cnn.dropout(self.fc7, self.KEEP_PROB, name='dropout7'),
lambda: self.fc7)
self.fc8, _ = cnn.fc(dropout7,
self.NUM_CLASSES,
use_relu=False,
name='fc8')
def __init__(self, inputs, true_labels, is_train=False, num_classes=None):
self.true_labels = true_labels
self.NUM_CLASSES = num_classes
with slim.arg_scope(resnet_v2.resnet_arg_scope()):
self.output, self.features = resnet_v2.resnet_v2_50(
inputs=inputs, num_classes=None, is_training=False)
self.classifier, _ = cnn.fc(input=self.forward_pass(),
num_outputs=self.NUM_CLASSES,
use_relu=False,
name='classifier')
def compute_loss_for_progress(self, add_l2_reg=False, weight_decay=0.001):
prog_pred = tf.sigmoid(
cnn.fc(self.extract_features('fc7'),
num_outputs=1,
use_relu=False,
name='fc_prog')[0])
loss = tf.losses.huber_loss(self.true_labels, tf.squeeze(prog_pred))
if add_l2_reg:
loss += weight_decay * tf.add_n(
[tf.nn.l2_loss(var) for var in tf.trainable_variables()])
return loss
def __init__(self, inputs, true_labels, is_train=False, num_classes=None):
self.true_labels = true_labels
self.NUM_CLASSES = num_classes
with slim.arg_scope(vgg.vgg_arg_scope()):
self.output, self.features = vgg.vgg_16(inputs=inputs,
num_classes=1000,
is_training=False)
self.classifier, _ = cnn.fc(input=self.extract_features('fc7'),
num_outputs=self.NUM_CLASSES,
use_relu=False,
name='classifier')
def create(self):
conv1, _ = cnn.conv(self.input,
11,
11,
96,
4,
4,
padding='VALID',
name='conv1')
pool1 = cnn.max_pool(conv1, 3, 3, 2, 2, padding='VALID', name='pool1')
norm1 = cnn.lrn(pool1, 5, 1e-04, 0.75, name='norm1')
conv2, _ = cnn.conv(norm1, 5, 5, 256, 1, 1, groups=2, name='conv2')
pool2 = cnn.max_pool(conv2, 3, 3, 2, 2, padding='VALID', name='pool2')
norm2 = cnn.lrn(pool2, 5, 1e-04, 0.75, name='norm2')
conv3, _ = cnn.conv(norm2, 3, 3, 384, 1, 1, name='conv3')
conv4, _ = cnn.conv(conv3, 3, 3, 384, 1, 1, groups=2, name='conv4')
conv5, _ = cnn.conv(conv4, 3, 3, 256, 1, 1, groups=2, name='conv5')
pool5 = cnn.max_pool(conv5, 3, 3, 2, 2, padding='VALID', name='pool5')
flattened = cnn.flatten(pool5)
fc6, _ = cnn.fc(flattened, 4096, name='fc6')
dropout6 = tf.cond(
self.IS_TRAIN,
lambda: cnn.dropout(fc6, self.KEEP_PROB, name='dropout6'),
lambda: fc6)
fc7, _ = cnn.fc(dropout6, 4096, name='fc7')
dropout7 = tf.cond(
self.IS_TRAIN,
lambda: cnn.dropout(fc7, self.KEEP_PROB, name='dropout7'),
lambda: fc7)
fc8_clipper, _ = cnn.fc(dropout7,
1,
use_relu=False,
name='fc8_clipper')
fc8_hook, _ = cnn.fc(dropout7, 1, use_relu=False, name='fc8_hook')
fc8_grasperElec, _ = cnn.fc(dropout7,
1,
use_relu=False,
name='fc8_grasperElec')
fc8_scissorsElec, _ = cnn.fc(dropout7,
1,
use_relu=False,
name='fc8_scissorsElec')
fc8_grasperFen, _ = cnn.fc(dropout7,
1,
use_relu=False,
name='fc8_grasperFen')
fc8_irrigator, _ = cnn.fc(dropout7,
1,
use_relu=False,
name='fc8_irrigator')
fc8_specBag, _ = cnn.fc(dropout7,
1,
use_relu=False,
name='fc8_specBag')
self.fc8_concat = tf.concat([
fc7, fc8_clipper, fc8_hook, fc8_grasperFen, fc8_grasperElec,
fc8_scissorsElec, fc8_specBag, fc8_irrigator
],
axis=1)