def train_a_teacher_network():
x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
x_image = tf.reshape(x, [-1,28,28,1])
net = ops.conv2d(x_image, 32, [5, 5], scope='conv1', stddev=0.1, bias=0.1)
net = ops.max_pool(net, [2, 2], scope='pool1')
net = ops.conv2d(net, 64, [5, 5], scope='conv2', stddev=0.1, bias=0.1)
net = ops.max_pool(net, [2, 2], scope='pool2')
net = ops.flatten(net, scope='pool2_flat')
net = ops.fc(net, 1024, scope='fc1', stddev=0.1, bias=0.1)
net = ops.fc(net, 10, activation=None, scope='fc2', stddev=0.1, bias=0.1)
y_conv = tf.nn.softmax(net)
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), reduction_indices=[1]))
model = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.scalar_summary('loss', cross_entropy)
tf.scalar_summary('acc', accuracy)
merged = tf.merge_all_summaries()
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
writer = tf.train.SummaryWriter('./logs', sess.graph)
sess.run(tf.initialize_all_variables())
print 'Teacher Network...'
for i in range(MAX_ITER):
batch = mnist.train.next_batch(BATCH_SIZE)
sess.run(model, feed_dict={x: batch[0], y_: batch[1]})
if i % 100 == 0:
summary_str, acc = sess.run([merged, accuracy], feed_dict={x: mnist.test.images, y_: mnist.test.labels})
writer.add_summary(summary_str, i)
print '[Iter: {}] Validation Accuracy : {:.4f}'.format(i,acc)
saver.save(sess, 'my-model', global_step=TEST_ITER)
def testCreateConvWithoutWD(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height, width, 3), seed=1)
ops.conv2d(images, 32, [3, 3], weight_decay=0)
self.assertEquals(
tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES), [])
def coarse_layers(inputs):
with tf.variable_scope('coarse_layers'):
out = ops.conv2d(inputs, 12, [7,7], stride=2, padding='VALID', scope='coarse_conv1')
out = ops.conv2d(out, 24, [3,3], stride=2, padding='VALID', scope='coarse_conv2')
return out
def train_a_teacher_network():
x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
x_image = tf.reshape(x, [-1,28,28,1])
net = ops.conv2d(x_image, 32, [5, 5], scope='conv1', stddev=0.1, bias=0.1)
net = ops.max_pool(net, [2, 2], scope='pool1')
net = ops.conv2d(net, 64, [5, 5], scope='conv2', stddev=0.1, bias=0.1)
net = ops.max_pool(net, [2, 2], scope='pool2')
net = ops.flatten(net, scope='pool2_flat')
net = ops.fc(net, 1024, scope='fc1', stddev=0.1, bias=0.1)
net = ops.fc(net, 10, activation=None, scope='fc2', stddev=0.1, bias=0.1)
y_conv = tf.nn.softmax(net)
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), axis=[1]))
model = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.summary.scalar('loss', cross_entropy)
tf.summary.scalar('acc', accuracy)
merged = tf.summary.merge_all()
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
writer = tf.summary.FileWriter('./logs', sess.graph)
sess.run(tf.global_variables_initializer())
print('Teacher Network...')
for i in range(MAX_ITER):
batch = mnist.train.next_batch(BATCH_SIZE)
sess.run(model, feed_dict={x: batch[0], y_: batch[1]})
# saver.save(sess, './my-model', global_step=TEST_ITER)
if i % 100 == 0:
summary_str, acc = sess.run([merged, accuracy], feed_dict={x: mnist.test.images, y_: mnist.test.labels})
writer.add_summary(summary_str, i)
print('[Iter: {}] Validation Accuracy : {:.4f}'.format(i,acc))
saver.save(sess, './my-model', global_step=TEST_ITER)
def train_a_student_network_wider():
new_width_conv = 128
new_w1, new_b1, new_w2, new_b2 = tf_net2wider(MODEL, WEIGHT, 'conv1',
'conv2', new_width_conv)
with tf.Graph().as_default():
with tf.Session(config=tf.ConfigProto(
allow_soft_placement=True)) as sess:
x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
x_image = tf.reshape(x, [-1, 28, 28, 1])
net = ops.conv2d(x_image,
new_width_conv, [5, 5],
scope='conv1',
initializer='constant',
weights=new_w1,
bias=new_b1,
restore=False)
net = ops.max_pool(net, [2, 2], scope='pool1')
net = ops.conv2d(net,
64, [5, 5],
scope='conv2',
initializer='constant',
weights=new_w2,
bias=new_b2,
restore=False)
net = ops.max_pool(net, [2, 2], scope='pool2')
net = ops.flatten(net, scope='pool2_flat')
net = ops.fc(net, 1024, scope='fc1')
net = ops.fc(net, 10, activation=None, scope='fc2')
y_conv = tf.nn.softmax(net)
cross_entropy = tf.reduce_mean(
-tf.reduce_sum(y_ * tf.log(y_conv), axis=[1]))
model = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1),
tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.summary.scalar('loss', cross_entropy)
tf.summary.scalar('acc', accuracy)
merged = tf.summary.merge_all()
saver = tf.train.Saver()
writer = tf.summary.FileWriter('./logs-wider', sess.graph)
sess.run(tf.global_variables_initializer())
variables_to_restore = tf.get_collection(
variables.VARIABLES_TO_RESTORE)
saver = tf.train.Saver(variables_to_restore)
saver.restore(sess, WEIGHT)
print('Net2Wider...')
for i in range(MAX_ITER):
batch = mnist.train.next_batch(BATCH_SIZE)
sess.run(model, feed_dict={x: batch[0], y_: batch[1]})
if i % 100 == 0:
summary_str, acc = sess.run([merged, accuracy],
feed_dict={
x: mnist.test.images,
y_: mnist.test.labels
})
writer.add_summary(summary_str, i)
print('[Iter: {}] Validation Accuracy : {:.4f}'.format(
i, acc))
def discrim16(inp):
with arg_scope([conv2d], batch_norm_params=batch_norm_params, stddev=0.02, activation=lrelu, weight_decay=1e-5):
o = conv2d(inp, num_filters_out=32, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=32, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=32, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=32, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=32, kernel_size=(3, 3), stride=1)
return fc(flatten(o), num_units_out=1, activation=tf.nn.sigmoid)
def testNonReuseVars(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height, width, 3), seed=1)
ops.conv2d(images, 32, [3, 3])
self.assertEquals(len(variables.get_variables()), 2)
ops.conv2d(images, 32, [3, 3])
self.assertEquals(len(variables.get_variables()), 4)
def testCreateConvCreatesWeightsAndBiasesVars(self):
height, width = 3, 3
images = tf.random_uniform((5, height, width, 3), seed=1)
with self.test_session():
self.assertFalse(variables.get_variables('conv1/weights'))
self.assertFalse(variables.get_variables('conv1/biases'))
ops.conv2d(images, 32, [3, 3], scope='conv1')
self.assertTrue(variables.get_variables('conv1/weights'))
self.assertTrue(variables.get_variables('conv1/biases'))
def testCreateConvWithWD(self):
height, width = 3, 3
with self.test_session() as sess:
images = tf.random_uniform((5, height, width, 3), seed=1)
ops.conv2d(images, 32, [3, 3], weight_decay=0.01)
wd = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)[0]
self.assertEquals(wd.op.name,
'Conv/weights/Regularizer/L2Regularizer/value')
sess.run(tf.global_variables_initializer())
self.assertTrue(sess.run(wd) <= 0.01)
def testReuseConvWithBatchNorm(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height, width, 32), seed=1)
with scopes.arg_scope([ops.conv2d],
batch_norm_params={'decay': 0.9}):
net = ops.conv2d(images, 32, [3, 3], scope='Conv')
net = ops.conv2d(net, 32, [3, 3], scope='Conv', reuse=True)
self.assertEquals(len(variables.get_variables()), 4)
self.assertEquals(len(variables.get_variables('Conv/BatchNorm')),
3)
self.assertEquals(len(variables.get_variables('Conv_1/BatchNorm')),
0)
def encoder(inp, z_dim):
n = 32
with arg_scope([conv2d], batch_norm_params=batch_norm_params, stddev=0.02, activation=lrelu, weight_decay=1e-5):
inp = inp-0.5
o = conv2d(inp, num_filters_out=n, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=1)
flat = flatten(o)
#flat = flatten(avg_pool(o, kernel_size=3))
z = fc(flat, num_units_out=z_dim, activation=tf.nn.tanh)/2+.5
return z
def generator_context(z):
n = 32
with arg_scope([conv2d, conv2d_transpose], batch_norm_params=batch_norm_params, stddev=0.02):
z = z*2-1
d = 8
z = fc(z, num_units_out=d*d*32, batch_norm_params=batch_norm_params)
c = z.get_shape()[1].value / (d*d)
z = tf.reshape(z, (-1, d, d, c))
o = conv2d_transpose(z, n, (3, 3), stride=(2, 2))
o = conv2d_transpose(o, n, (3, 3), stride=(2, 2))
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=4, kernel_size=(3, 3), stride=1)
attended = o
return attended
def generator(z):
n = 32
with arg_scope([conv2d, conv2d_transpose], batch_norm_params=batch_norm_params, stddev=0.02):
z = z*2-1
d = 8
z = fc(z, num_units_out=d*d*32, batch_norm_params=batch_norm_params)
c = z.get_shape()[1].value / (d*d)
z = tf.reshape(z, (-1, d, d, c))
o = conv2d_transpose(z, n, (3, 3), stride=(2, 2))
o = conv2d_transpose(o, n, (3, 3), stride=(2, 2))
o = conv2d(o, num_filters_out=n*2, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=1, kernel_size=(3, 3), stride=1, padding="VALID", batch_norm_params=None)
out = o[:, 1:29, 1:29, :]
return out
def coarse_layers(inputs):
with tf.variable_scope('coarse_layers'):
out = ops.conv2d(inputs,
12, [7, 7],
stride=2,
padding='VALID',
scope='coarse_conv1')
out = ops.conv2d(out,
24, [3, 3],
stride=2,
padding='VALID',
scope='coarse_conv2')
return out
def testReuseConvWithWD(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height, width, 3), seed=1)
ops.conv2d(images, 32, [3, 3], weight_decay=0.01, scope='conv1')
self.assertEquals(len(variables.get_variables()), 2)
self.assertEquals(
len(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)), 1)
ops.conv2d(images,
32, [3, 3],
weight_decay=0.01,
scope='conv1',
reuse=True)
self.assertEquals(len(variables.get_variables()), 2)
self.assertEquals(
len(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)), 1)
def testCreateHorizontalConv(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height, width, 3), seed=1)
output = ops.conv2d(images, 32, [1, 3])
self.assertEquals(output.op.name, 'Conv/Relu')
self.assertListEqual(output.get_shape().as_list(),
[5, height, width, 32])
def testCreateConvWithStride(self):
height, width = 6, 6
with self.test_session():
images = tf.random_uniform((5, height, width, 3), seed=1)
output = ops.conv2d(images, 32, [3, 3], stride=2)
self.assertEquals(output.op.name, 'Conv/Relu')
self.assertListEqual(output.get_shape().as_list(),
[5, height / 2, width / 2, 32])
def discriminator(inp):
n = 32
with arg_scope([conv2d], batch_norm_params=batch_norm_params, stddev=0.02, activation=lrelu, weight_decay=1e-5):
inp = inp-0.5
# todo make new set of ops?
o = conv2d(inp, num_filters_out=n, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
#o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=1)
flat = flatten(o)
#flat = flatten(avg_pool(o, kernel_size=3))
prob = fc(flat, num_units_out=1, activation=tf.nn.sigmoid)
#prob = tf.Print(prob, [prob])
return prob
def conv_model(inputs, is_training=True, scope=''):
# summaries or losses.
net = {}
with tf.op_scope([inputs], scope, 'mdm_conv'):
with scopes.arg_scope([ops.conv2d, ops.fc], is_training=is_training):
with scopes.arg_scope([ops.conv2d], activation=tf.nn.relu, padding='VALID'):
net['conv_1'] = ops.conv2d(inputs, 32, [3, 3], scope='conv_1')
net['pool_1'] = ops.max_pool(net['conv_1'], [2, 2])
net['conv_2'] = ops.conv2d(net['pool_1'], 32, [3, 3], scope='conv_2')
net['pool_2'] = ops.max_pool(net['conv_2'], [2, 2])
crop_size = net['pool_2'].get_shape().as_list()[1:3]
net['conv_2_cropped'] = utils.get_central_crop(net['conv_2'], box=crop_size)
net['concat'] = tf.concat(3, [net['conv_2_cropped'], net['pool_2']])
return net
def fine_layers(inputs):
with tf.variable_scope('fine_layers'):
out = ops.conv2d(inputs, 24, [3,3], stride=1, padding='VALID', scope='fine_conv1')
out = ops.conv2d(out, 24, [3,3], stride=1, padding='VALID', scope='fine_conv2')
out = tf.pad(out, [[0,0],[1,1],[1,1],[0,0]])
out = ops.max_pool(out, [2,2], stride=2, scope='fine_pool1')
out = ops.conv2d(out, 24, [3,3], stride=1, padding='VALID', scope='fine_conv3')
out = tf.pad(out, [[0,0],[1,1],[1,1],[0,0]])
out = ops.conv2d(out, 24, [3,3], stride=1, padding='VALID', scope='fine_conv4')
out = tf.pad(out, [[0,0],[1,1],[1,1],[0,0]])
out = ops.max_pool(out, [2,2], stride=2, scope='fine_pool2')
out = ops.conv2d(out, 24, [3,3], stride=1, padding='VALID', scope='fine_conv5')
return out
def conv_model(inputs, is_training=True, scope=''):
# summaries or losses.
net = {}
with tf.name_scope(scope, 'mdm_conv', [inputs]): # 给下面op_name 加前缀mdm_conv 用with 语句解决资源释放问题
with scopes.arg_scope([ops.conv2d, ops.fc], is_training=is_training):
with scopes.arg_scope([ops.conv2d], activation=tf.nn.relu, padding='VALID'):
net['conv_1'] = ops.conv2d(inputs, 32, [3, 3], scope='conv_1')
net['pool_1'] = ops.max_pool(net['conv_1'], [2, 2])
net['conv_2'] = ops.conv2d(net['pool_1'], 32, [3, 3], scope='conv_2')
net['pool_2'] = ops.max_pool(net['conv_2'], [2, 2])
# 两个卷积层 每层32个过滤器 3*3核
# 每层卷积层后有一个2*2 的最大池化层
crop_size = net['pool_2'].get_shape().as_list()[1:3]
net['conv_2_cropped'] = utils.get_central_crop(net['conv_2'], box=crop_size)
# 中央作物的激活与第二池化层的输出，
# 通过跳转链接concat连接起来，以保留更多相关本地信息，否则使用max池化层会丢失这些信息
net['concat'] = tf.concat([net['conv_2_cropped'], net['pool_2']], 3) # axis=3
return net
def testCreateFullyConv(self):
height, width = 6, 6
with self.test_session():
images = tf.random_uniform((5, height, width, 32), seed=1)
output = ops.conv2d(images,
64,
images.get_shape()[1:3],
padding='VALID')
self.assertEquals(output.op.name, 'Conv/Relu')
self.assertListEqual(output.get_shape().as_list(), [5, 1, 1, 64])
def fine_layers(inputs):
with tf.variable_scope('fine_layers'):
out = ops.conv2d(inputs, 24, [3,3], stride=1, padding='VALID', scope='fine_conv1')
out = ops.conv2d(out, 24, [3,3], stride=1, padding='VALID', scope='fine_conv2')
#out = tf.pad(out, [[0,0],[1,1],[1,1],[0,0]])
out = ops.max_pool(out, [2,2], stride=2, scope='fine_pool1')
out = ops.conv2d(out, 24, [3,3], stride=1, padding='VALID', scope='fine_conv3')
#out = tf.pad(out, [[0,0],[1,1],[1,1],[0,0]])
out = ops.conv2d(out, 24, [3,3], stride=1, padding='VALID', scope='fine_conv4')
#out = tf.pad(out, [[0,0],[1,1],[1,1],[0,0]])
out = ops.max_pool(out, [2,2], stride=2, scope='fine_pool2')
out = ops.conv2d(out, 24, [3,3], stride=1, padding='VALID', scope='fine_conv5')
return out
def top_layers(inputs):
with tf.variable_scope('top_layers'):
out = ops.conv2d(inputs, 96, [4,4], stride=2, scope='top_conv1')
_,fm_size,fm_size,_ = out.get_shape()
out = ops.max_pool(out, [fm_size,fm_size], stride=1, scope='top_gpool')
out = ops.flatten(out, scope='top_flatten')
out = ops.fc(out, 10, activation=None, bias=0.0, scope='top_logits')
return out
def top_layers(inputs):
with tf.variable_scope('top_layers'):
out = ops.conv2d(inputs, 96, [4,4], stride=2, padding='VALID', scope='top_conv1')
_,fm_size,fm_size,_ = out.get_shape()
out = ops.max_pool(out, [fm_size,fm_size], stride=1, scope='top_gpool')
out = ops.flatten(out, scope='top_flatten')
#out = ops.fc(out, 10, activation=None, bias=0.0, batch_norm_params=None, scope='top_logits')
bn_out = {'decay': 0.99, 'epsilon': 0.001, 'scale':True}
out = ops.fc(out, 10, activation=None, batch_norm_params=bn_out, scope='top_logits')
return out
def conv_model(inputs, is_training=True, scope=''):
# summaries or losses.
net = {}
with tf.name_scope(scope, 'Conv_lay', [inputs]):
with scopes.arg_scope([ops.conv2d, ops.fc], is_training=is_training):
with scopes.arg_scope([ops.conv2d],
activation=tf.nn.relu,
padding='VALID'):
net['conv_1'] = ops.conv2d(inputs, 32, [7, 7], scope='conv_1')
net['pool_1'] = ops.max_pool(net['conv_1'], [2, 2])
net['conv_2'] = ops.conv2d(net['pool_1'],
64, [3, 3],
scope='conv_2')
net['pool_2'] = ops.max_pool(net['conv_2'], [2, 2])
net['conv_3'] = ops.conv2d(net['pool_2'],
64, [3, 3],
scope='conv_3')
net['pool_3'] = ops.max_pool(net['conv_3'], [2, 2])
net['concat'] = net['pool_3']
return net
def top_layers(inputs):
with tf.variable_scope('top_layers'):
out = ops.conv2d(inputs, 96, [4, 4], stride=2, scope='top_conv1')
_, fm_size, fm_size, _ = out.get_shape()
out = ops.max_pool(out, [fm_size, fm_size],
stride=1,
scope='top_gpool')
out = ops.flatten(out, scope='top_flatten')
out = ops.fc(out, 10, activation=None, bias=0.0, scope='top_logits')
return out
def top_layers(inputs):
with tf.variable_scope('top_layers'):
out = ops.conv2d(inputs, 96, [4,4], stride=2, padding='VALID', scope='top_conv1')
_,fm_size,fm_size,_ = out.get_shape()
out = ops.max_pool(out, [fm_size,fm_size], stride=1, scope='top_gpool')
out = ops.flatten(out, scope='top_flatten')
#out = ops.fc(out, 10, activation=None, bias=0.0, batch_norm_params=None, scope='top_logits')
bn_out = {'decay': 0.99, 'epsilon': 0.001, 'scale':True}
out = ops.fc(out, 10, activation=None, batch_norm_params=bn_out, scope='top_logits')
return out
def discriminator(inp):
n = 32
with arg_scope([conv2d], batch_norm_params=batch_norm_params, stddev=0.02, activation=lrelu, weight_decay=1e-5):
inp = inp-0.5
with tf.device("/gpu:%d"%FLAGS.gpu_num):
o = conv2d(inp, num_filters_out=32, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=64, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=64, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=128, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=128, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=128, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=256, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=256, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=256, kernel_size=(3, 3), stride=1)
flat = flatten(o)
#flat = flatten(avg_pool(o, kernel_size=3))
prob = fc(flat, num_units_out=1, activation=tf.nn.sigmoid)
#prob = tf.Print(prob, [prob])
return prob
def conv_model(inputs, is_training=True, scope=''):
# summaries or losses.
net = {}
with tf.name_scope(scope, 'rdn_conv', [inputs]):
with scopes.arg_scope([ops.conv2d, ops.fc], is_training=is_training):
with scopes.arg_scope([ops.conv2d],
activation=tf.nn.relu,
padding='VALID'):
net['conv_1'] = ops.conv2d(inputs, 32, [3, 3], scope='conv_1')
net['pool_1'] = ops.max_pool(net['conv_1'], [2, 2])
net['conv_2'] = ops.conv2d(net['pool_1'],
32, [3, 3],
scope='conv_2')
net['pool_2'] = ops.max_pool(net['conv_2'], [2, 2])
crop_size = net['pool_2'].get_shape().as_list()[1:3]
net['conv_2_cropped'] = utils.get_central_crop(net['conv_2'],
box=crop_size)
net['concat'] = tf.concat(
[net['conv_2_cropped'], net['pool_2']], 3)
return net
def train_a_student_network_deeper():
new_w1, new_b1 = tf_net2deeper(MODEL, WEIGHT, 'conv1')
with tf.Graph().as_default():
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
x_image = tf.reshape(x, [-1,28,28,1])
net = ops.conv2d(x_image, 32, [5, 5], scope='conv1')
net = ops.conv2d(net, 32, [5, 5], scope='conv1_new', initializer='constant', weights=new_w1, bias=new_b1, restore=False)
net = ops.max_pool(net, [2, 2], scope='pool1')
net = ops.conv2d(net, 64, [5, 5], scope='conv2')
net = ops.max_pool(net, [2, 2], scope='pool2')
net = ops.flatten(net, scope='pool2_flat')
net = ops.fc(net, 1024, scope='fc1')
net = ops.fc(net, 10, activation=None, scope='fc2')
y_conv = tf.nn.softmax(net)
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), axis=[1]))
model = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.summary.scalar('loss', cross_entropy)
tf.summary.scalar('acc', accuracy)
merged = tf.summary.merge_all()
saver = tf.train.Saver()
writer = tf.summary.FileWriter('./logs-deeper', sess.graph)
sess.run(tf.global_variables_initializer())
variables_to_restore = tf.get_collection(variables.VARIABLES_TO_RESTORE)
saver = tf.train.Saver(variables_to_restore)
saver.restore(sess, WEIGHT)
print('Net2Deeper...')
for i in range(MAX_ITER):
batch = mnist.train.next_batch(BATCH_SIZE)
sess.run(model, feed_dict={x: batch[0], y_: batch[1]})
if i % 100 == 0:
summary_str, acc = sess.run([merged, accuracy], feed_dict={x: mnist.test.images, y_: mnist.test.labels})
writer.add_summary(summary_str, i)
print('[Iter: {}] Validation Accuracy : {:.4f}'.format(i,acc))
def build_a_new_graph():
with tf.Graph().as_default() as graph:
x = tf.placeholder(tf.float32, shape=[None, 784])
tf.add_to_collection('input', x)
y_ = tf.placeholder(tf.float32, shape=[None, 10])
tf.add_to_collection('label', y_)
x_image = tf.reshape(x, [-1,28,28,1])
net = ops.conv2d(x_image, 64, [5, 5], scope='conv1')
net = ops.max_pool(net, [2, 2], scope='pool1')
net = ops.conv2d(net, 64, [5, 5], scope='conv2')
net = ops.conv2d(net, 64, [5, 5], scope='conv2_new', stddev=0.1, bias=0.1)
net = ops.max_pool(net, [2, 2], scope='pool2')
net = ops.flatten(net, scope='pool2_flat')
net = ops.fc(net, 1024, scope='fc1')
net = ops.fc(net, 1024, scope='fc1_new')
net = ops.fc(net, 10, activation=None, scope='fc2')
y_conv = tf.nn.softmax(net)
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), reduction_indices=[1]))
model = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
tf.add_to_collection('objective', model)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.add_to_collection('accuracy', accuracy)
return graph
def train_a_student_network_deeper_rand_init():
with tf.Graph().as_default():
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
x_image = tf.reshape(x, [-1,28,28,1])
net = ops.conv2d(x_image, 32, [5, 5], scope='conv1', stddev=0.1, bias=0.1)
net = ops.conv2d(net, 32, [5, 5], scope='conv1_new', stddev=0.1, bias=0.1, restore=False)
net = ops.max_pool(net, [2, 2], scope='pool1')
net = ops.conv2d(net, 64, [5, 5], scope='conv2', stddev=0.1, bias=0.1)
net = ops.max_pool(net, [2, 2], scope='pool2')
net = ops.flatten(net, scope='pool2_flat')
net = ops.fc(net, 1024, scope='fc1', stddev=0.1, bias=0.1)
net = ops.fc(net, 10, activation=None, scope='fc2', stddev=0.1, bias=0.1)
y_conv = tf.nn.softmax(net)
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), reduction_indices=[1]))
model = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.scalar_summary('loss', cross_entropy)
tf.scalar_summary('acc', accuracy)
merged = tf.merge_all_summaries()
saver = tf.train.Saver()
writer = tf.train.SummaryWriter('./logs-deeper-rand', sess.graph)
sess.run(tf.initialize_all_variables())
variables_to_restore = tf.get_collection(variables.VARIABLES_TO_RESTORE)
saver = tf.train.Saver(variables_to_restore)
saver.restore(sess, WEIGHT)
print 'Net2Deeper Baseline (Rand init)...'
for i in range(MAX_ITER):
batch = mnist.train.next_batch(BATCH_SIZE)
sess.run(model, feed_dict={x: batch[0], y_: batch[1]})
if i % 100 == 0:
summary_str, acc = sess.run([merged, accuracy], feed_dict={x: mnist.test.images, y_: mnist.test.labels})
writer.add_summary(summary_str, i)
print '[Iter: {}] Validation Accuracy : {:.4f}'.format(i,acc)
def encoder(inp, z_dim):
#n = 32
with arg_scope([conv2d, fc], batch_norm_params=batch_norm_params, stddev=0.02, activation=lrelu, weight_decay=1e-5):
with tf.device("/gpu:%d"%FLAGS.gpu_num):
inp = inp-0.5
o = conv2d(inp, num_filters_out=32, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=32, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=64, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=64, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=128, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=128, kernel_size=(3, 3), stride=1)
flat = flatten(o)
z = fc(flat, num_units_out=z_dim, activation=None)
# normalized between -2 and 2 because of batchnorm
return tf.nn.sigmoid(z * 2)
def encoder(inp, z_dim):
#n = 32
with arg_scope([conv2d], batch_norm_params=batch_norm_params, stddev=0.02, activation=lrelu, weight_decay=1e-5):
with tf.device("/gpu:%d"%FLAGS.gpu_num):
inp = inp-0.5
n = 64*4
o = conv2d(inp, num_filters_out=n, kernel_size=(3, 3), stride=1)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=1)
n = 128*4
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=2)
o = conv2d(o, num_filters_out=n, kernel_size=(3, 3), stride=1)
flat = flatten(o)
#flat = flatten(avg_pool(o, kernel_size=3))
#z = fc(flat, num_units_out=z_dim, activation=tf.nn.tanh)/2+.5
z = fc(flat, num_units_out=z_dim, activation=tf.nn.sigmoid)
return z
def testCreateConvValid(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height, width, 3), seed=1)
output = ops.conv2d(images, 32, [3, 3], padding='VALID')
self.assertListEqual(output.get_shape().as_list(), [5, 1, 1, 32])
def testCreateConvWithScope(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height, width, 3), seed=1)
output = ops.conv2d(images, 32, [3, 3], scope='conv1')
self.assertEquals(output.op.name, 'conv1/Relu')
def testCreateConvWithoutActivation(self):
height, width = 3, 3
with self.test_session():
images = tf.random_uniform((5, height, width, 3), seed=1)
output = ops.conv2d(images, 32, [3, 3], activation=None)
self.assertEquals(output.op.name, 'Conv/BiasAdd')
def inception_v3(inputs,
dropout_keep_prob=0.8,
num_classes=1000,
is_training=True,
restore_logits=True,
scope=''):
"""Latest Inception from http://arxiv.org/abs/1512.00567.
"Rethinking the Inception Architecture for Computer Vision"
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens,
Zbigniew Wojna
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: Optional scope for name_scope.
Returns:
a list containing 'logits', 'aux_logits' Tensors.
"""
# end_points will collect relevant activations for external use, for example
# summaries or losses.
end_points = {}
with tf.name_scope(scope, 'inception_v3', [inputs]):
with scopes.arg_scope(
[ops.conv2d, ops.fc, ops.batch_norm, ops.dropout],
is_training=is_training):
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1,
padding='VALID'):
# 299 x 299 x 3
end_points['conv0'] = ops.conv2d(inputs,
32, [3, 3],
stride=2,
scope='conv0')
# 149 x 149 x 32
end_points['conv1'] = ops.conv2d(end_points['conv0'],
32, [3, 3],
scope='conv1')
# 147 x 147 x 32
end_points['conv2'] = ops.conv2d(end_points['conv1'],
64, [3, 3],
padding='SAME',
scope='conv2')
# 147 x 147 x 64
end_points['pool1'] = ops.max_pool(end_points['conv2'], [3, 3],
stride=2,
scope='pool1')
# 73 x 73 x 64
end_points['conv3'] = ops.conv2d(end_points['pool1'],
80, [1, 1],
scope='conv3')
# 73 x 73 x 80.
end_points['conv4'] = ops.conv2d(end_points['conv3'],
192, [3, 3],
scope='conv4')
# 71 x 71 x 192.
end_points['pool2'] = ops.max_pool(end_points['conv4'], [3, 3],
stride=2,
scope='pool2')
# 35 x 35 x 192.
net = end_points['pool2']
# Inception blocks
with scopes.arg_scope([ops.conv2d, ops.max_pool, ops.avg_pool],
stride=1,
padding='SAME'):
# mixed: 35 x 35 x 256.
with tf.variable_scope('mixed_35x35x256a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 32, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch5x5, branch3x3dbl,
branch_pool
])
end_points['mixed_35x35x256a'] = net
# mixed_1: 35 x 35 x 288.
with tf.variable_scope('mixed_35x35x288a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 64, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch5x5, branch3x3dbl,
branch_pool
])
end_points['mixed_35x35x288a'] = net
# mixed_2: 35 x 35 x 288.
with tf.variable_scope('mixed_35x35x288b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 64, [1, 1])
with tf.variable_scope('branch5x5'):
branch5x5 = ops.conv2d(net, 48, [1, 1])
branch5x5 = ops.conv2d(branch5x5, 64, [5, 5])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 64, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch5x5, branch3x3dbl,
branch_pool
])
end_points['mixed_35x35x288b'] = net
# mixed_3: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768a'):
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net,
384, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 64, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 96, [3, 3])
branch3x3dbl = ops.conv2d(branch3x3dbl,
96, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch_pool'):
branch_pool = ops.max_pool(net, [3, 3],
stride=2,
padding='VALID')
net = tf.concat(
axis=3, values=[branch3x3, branch3x3dbl, branch_pool])
end_points['mixed_17x17x768a'] = net
# mixed4: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 128, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 128, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 128, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 128, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch7x7, branch7x7dbl,
branch_pool
])
end_points['mixed_17x17x768b'] = net
# mixed_5: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768c'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 160, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 160, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 160, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch7x7, branch7x7dbl,
branch_pool
])
end_points['mixed_17x17x768c'] = net
# mixed_6: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768d'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 160, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 160, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 160, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 160, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch7x7, branch7x7dbl,
branch_pool
])
end_points['mixed_17x17x768d'] = net
# mixed_7: 17 x 17 x 768.
with tf.variable_scope('mixed_17x17x768e'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 192, [1, 1])
with tf.variable_scope('branch7x7'):
branch7x7 = ops.conv2d(net, 192, [1, 1])
branch7x7 = ops.conv2d(branch7x7, 192, [1, 7])
branch7x7 = ops.conv2d(branch7x7, 192, [7, 1])
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = ops.conv2d(net, 192, [1, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [7, 1])
branch7x7dbl = ops.conv2d(branch7x7dbl, 192, [1, 7])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch7x7, branch7x7dbl,
branch_pool
])
end_points['mixed_17x17x768e'] = net
# Auxiliary Head logits
aux_logits = tf.identity(end_points['mixed_17x17x768e'])
with tf.variable_scope('aux_logits'):
aux_logits = ops.avg_pool(aux_logits, [5, 5],
stride=3,
padding='VALID')
aux_logits = ops.conv2d(aux_logits,
128, [1, 1],
scope='proj')
# Shape of feature map before the final layer.
shape = aux_logits.get_shape()
aux_logits = ops.conv2d(aux_logits,
768,
shape[1:3],
stddev=0.01,
padding='VALID')
aux_logits = ops.flatten(aux_logits)
aux_logits = ops.fc(aux_logits,
num_classes,
activation=None,
stddev=0.001,
restore=restore_logits)
end_points['aux_logits'] = aux_logits
# mixed_8: 8 x 8 x 1280.
# Note that the scope below is not changed to not void previous
# checkpoints.
# (TODO) Fix the scope when appropriate.
with tf.variable_scope('mixed_17x17x1280a'):
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 192, [1, 1])
branch3x3 = ops.conv2d(branch3x3,
320, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch7x7x3'):
branch7x7x3 = ops.conv2d(net, 192, [1, 1])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [1, 7])
branch7x7x3 = ops.conv2d(branch7x7x3, 192, [7, 1])
branch7x7x3 = ops.conv2d(branch7x7x3,
192, [3, 3],
stride=2,
padding='VALID')
with tf.variable_scope('branch_pool'):
branch_pool = ops.max_pool(net, [3, 3],
stride=2,
padding='VALID')
net = tf.concat(
axis=3, values=[branch3x3, branch7x7x3, branch_pool])
end_points['mixed_17x17x1280a'] = net
# mixed_9: 8 x 8 x 2048.
with tf.variable_scope('mixed_8x8x2048a'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [1, 1])
branch3x3 = tf.concat(
axis=3,
values=[
ops.conv2d(branch3x3, 384, [1, 3]),
ops.conv2d(branch3x3, 384, [3, 1])
])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 448, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 384, [3, 3])
branch3x3dbl = tf.concat(
axis=3,
values=[
ops.conv2d(branch3x3dbl, 384, [1, 3]),
ops.conv2d(branch3x3dbl, 384, [3, 1])
])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch3x3, branch3x3dbl,
branch_pool
])
end_points['mixed_8x8x2048a'] = net
# mixed_10: 8 x 8 x 2048.
with tf.variable_scope('mixed_8x8x2048b'):
with tf.variable_scope('branch1x1'):
branch1x1 = ops.conv2d(net, 320, [1, 1])
with tf.variable_scope('branch3x3'):
branch3x3 = ops.conv2d(net, 384, [1, 1])
branch3x3 = tf.concat(
axis=3,
values=[
ops.conv2d(branch3x3, 384, [1, 3]),
ops.conv2d(branch3x3, 384, [3, 1])
])
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = ops.conv2d(net, 448, [1, 1])
branch3x3dbl = ops.conv2d(branch3x3dbl, 384, [3, 3])
branch3x3dbl = tf.concat(
axis=3,
values=[
ops.conv2d(branch3x3dbl, 384, [1, 3]),
ops.conv2d(branch3x3dbl, 384, [3, 1])
])
with tf.variable_scope('branch_pool'):
branch_pool = ops.avg_pool(net, [3, 3])
branch_pool = ops.conv2d(branch_pool, 192, [1, 1])
net = tf.concat(axis=3,
values=[
branch1x1, branch3x3, branch3x3dbl,
branch_pool
])
end_points['mixed_8x8x2048b'] = net
# Final pooling and prediction
with tf.variable_scope('logits'):
shape = net.get_shape()
net = ops.avg_pool(net,
shape[1:3],
padding='VALID',
scope='pool')
# 1 x 1 x 2048
net = ops.dropout(net, dropout_keep_prob, scope='dropout')
net = ops.flatten(net, scope='flatten')
# 2048
logits = ops.fc(net,
num_classes,
activation=None,
scope='logits',
restore=restore_logits)
# 1000
end_points['logits'] = logits
end_points['predictions'] = tf.nn.softmax(
logits, name='predictions')
return logits, end_points
