def aux_classifier(self, inputs, labels, input_channels, is_training, scope=None):
"""
Auxiliary Classifier used in Inception Module to help propagate
gradients backward.
"""
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
# pooling layer with 5x5 kernel and stride 3 (new size: 4x4xC)
network = tf.nn.avg_pool(inputs, 5, 3, 'VALID', name='pool')
# convolution with 1x1 kernel and stride 1 (new size: 4x4x128)
network = ops.convolution(network, input_channels, 128, 1, 128, batch_norm=False,
is_training=is_training, scope='auxconv')
# flatten (new size: 2048)
network = ops.flatten(network, scope='flatten')
# fully connected layer (new size: 1024)
network = ops.dense(network, 2048, 1024, dropout=True, dropout_rate=0.7,
is_training=is_training, scope='fc1')
# output layer (new size: 10) -- Original Paper Size: 1000 (for ImageNet)
network = ops.dense(network, 1024, 10, activation=None, is_training=is_training,
scope='fc2')
# loss of auxiliary classifier
loss = ops.loss(network, labels, scope='auxloss')
return loss
def build_model(self, inputs, labels, is_training):
# pad inputs to size 224x224x3 - NOTE: may change to bilinear upsampling
pad = int((self.image_size - self.height) / 2)
inputs = tf.pad(inputs, [[0, 0], [pad, pad], [pad, pad], [0, 0]])
# convolution with 11x11 kernel and stride 4 (new size: 55x55x96)
self.network = ops.convolution(inputs, self.channels, 96, 11, 96, stride=4,
padding='VALID', is_training=is_training, scope='conv1')
# pooling with 3x3 kernel and stride 2 (new size: 27x27x96)
self.network = ops.pooling(self.network, k_size=3, scope='pool1')
# convolution with 5x5 kernel and stride 1 (new size: 27x27x256)
self.network = ops.convolution(self.network, 96, 256, 5, 256,
is_training=is_training, scope='conv2')
# pooling with 3x3 kernel and stride 2 (new size: 13x13x256)
self.network = ops.pooling(self.network, k_size=3, scope='pool2')
# convolution with 3x3 kernel and stride 1 (new size: 13x13x384)
self.network = ops.convolution(self.network, 256, 384, 3, 384, batch_norm=False,
is_training=is_training, scope='conv3')
# convolution with 3x3 kernel and stride 1 (new size: 13x13x384)
self.network = ops.convolution(self.network, 384, 384, 3, 384, batch_norm=False,
is_training=is_training, scope='conv4')
# convolution with 3x3 kernel and stride 1 (new size: 13x13x256)
self.network = ops.convolution(self.network, 384, 256, 3, 256, batch_norm=False,
is_training=is_training, scope='conv5')
# pooling with 3x3 kernel and stride 2 (new size: 6x6x256)
self.network = ops.pooling(self.network, k_size=3, scope='pool3')
# flatten (new size: 9216)
self.network = ops.flatten(self.network, scope='flatten')
# fully connected layer (new size: 4096)
self.network = ops.dense(self.network, 9216, 4096, dropout=True, dropout_rate=0.2,
is_training=is_training, scope='fc1')
# fully connected layer (new size: 1024) -- Original Paper Size: 4096 (for ImageNet)
self.network = ops.dense(self.network, 4096, 1024, dropout=True, dropout_rate=0.2,
is_training=is_training, scope='fc2')
# output layer (new size: 10) -- Original Paper Size: 1000 (for ImageNet)
self.network = ops.dense(self.network, 1024, 10, activation=None,
is_training=is_training, scope='fc3')
self.loss = ops.loss(self.network, labels, scope='loss')
if is_training:
self.optimizer = ops.optimize(self.loss, self.learning_rate, scope='update')
def __init__(self,
lr=0.0001,
optimizer=tf.train.Optimizer,
fine_tuning=True,
dropout=False,
adaptive_ratio=1.0):
'''
----------Hyperparameters -------------
:param fine_tuning: If True, the parameters of CNN layers will also be fine-tuned.
Otherwise, only the parameters of FC layers will be trained.
:param dropout: If True, dropout is applied to all fully connected layers except for the last one.
Also, dropout_keep_prob should be fed. (default value is 1.0)
:param adaptive_ratio: If True, the learning rate of convolutional layer will be learning rate * adaptive_ratio
:return:
'''
self.desc = "Learning rate : {}, optimizer : {}, fine_tuning : {}, dropout : {}, adaptive ratio : {}"\
.format(lr, optimizer.__name__, fine_tuning, dropout, adaptive_ratio)
print(self.desc)
self.params = {
'lr': lr,
'optimizer': optimizer,
'fine_tuning': fine_tuning,
'dropout': dropout,
'adaptive_ratio': adaptive_ratio
}
self.xs = tf.placeholder(tf.float32, [None, 32, 32, 3])
self.ys = tf.placeholder(tf.int32, [None])
self.dropout_keep_prob = tf.placeholder_with_default(1.0, None)
pool5 = self.build_convnet(fine_tuning)
fc3 = self.build_fcnet(pool5, dropout)
self.probs = tf.nn.softmax(fc3, name='softmax')
self.loss = ops.loss(logits=self.probs, labels=self.ys, one_hot=False)
self.accuracy = ops.accuracy(logits=self.probs,
labels=self.ys,
one_hot=False)
if adaptive_ratio < 1.0:
self.train = ops.train(self.loss,
optimizer=optimizer,
conv_lr=lr * adaptive_ratio,
fc_lr=lr)
else:
self.train = optimizer(learning_rate=lr).minimize(self.loss)
def build_model(self, inputs, labels, is_training=False):
self.network = ops.convolution(inputs,
self.channels,
50,
5,
50,
is_training=is_training,
scope='conv1')
self.network = ops.pooling(self.network, scope='pool1')
self.network = ops.convolution(self.network,
50,
20,
5,
20,
is_training=is_training,
scope='conv2')
self.network = ops.pooling(self.network, scope='pool2')
self.network = ops.flatten(self.network, scope='flatten')
self.network = ops.dense(self.network,
self.network.get_shape().as_list()[1],
200,
scope='fc1')
self.network = ops.dense(self.network, 200, 50, scope='fc2')
self.network = ops.dense(self.network,
50,
10,
activation=None,
scope='fc3')
self.loss = ops.loss(self.network, labels, scope='loss')
self.accuracy = ops.accuracy(self.network, labels, scope='accuracy')
if is_training:
self.optimizer = ops.optimize(self.loss,
self.learning_rate,
scope='update')
def train(args, data, params):
train = data['train']
valid = data['valid']
learning_rate = args.learning_rate
with tf.Graph().as_default():
input_ph = tf.placeholder(tf.int32,
shape=[args.batch_size, params['gram_size']])
targ_ph = tf.placeholder(tf.int32, shape=[args.batch_size])
learning_rate_ph = tf.placeholder(tf.float32, shape=[])
if args.w2v:
with h5py.File(args.w2v, 'r') as datafile:
embeds = datafile['w2v'][:]
scores, normalize_op, vars = ops.model(input_ph, params, embeds)
else:
scores, normalize_op, vars = ops.model(input_ph, params)
loss = ops.loss(scores, targ_ph)
train_op, print_op = ops.train(loss, learning_rate_ph, args)
#sess = tf.Session(config=tf.ConfigProto(inter_op_parallelism_threads=NUM_THREADS,\
# intra_op_parallelism_threads=NUM_THREADS))
sess = tf.Session()
init = tf.initialize_all_variables(
) # initialize variables before they can be used
saver = tf.train.Saver()
sess.run(init)
if args.modelfile:
saver.restore(sess, args.modelfile)
print "Model restored from %s" % args.modelfile
valid_loss = 0.
for i in xrange(valid.nbatches):
valid_feed_dict = get_feed_dict(valid, i, input_ph, targ_ph,
learning_rate_ph)
batch_loss = sess.run([loss], feed_dict=valid_feed_dict)[0]
valid_loss += batch_loss
last_valid = valid_loss
print 'Initial valid loss: %.3f' % math.exp(
valid_loss / valid.nbatches)
for epoch in xrange(args.nepochs):
print "Training epoch %d with learning rate %.3f" % (epoch + 1,
learning_rate)
vals = sess.run(vars)
start_time = time.time()
train_loss = 0.
valid_loss = 0.
for i in xrange(train.nbatches):
train_feed_dict = get_feed_dict(train, i, input_ph, targ_ph, \
learning_rate_ph, learning_rate)
#grads = sess.run(print_op, feed_dict=train_feed_dict)
_, batch_loss = sess.run([train_op, loss],
feed_dict=train_feed_dict)
train_loss += batch_loss
for i in xrange(valid.nbatches):
valid_feed_dict = get_feed_dict(valid, i, input_ph, targ_ph,
learning_rate_ph)
batch_loss = sess.run([loss], feed_dict=valid_feed_dict)[0]
valid_loss += batch_loss
if args.normalize:
_ = sess.run(normalize_op)
duration = time.time() - start_time
print "\tloss = %.3f, valid ppl = %.3f, %.3f s" % \
(math.exp(train_loss/train.nbatches), \
math.exp(valid_loss/valid.nbatches), duration)
if last_valid < valid_loss:
learning_rate /= 2.
elif args.outfile:
saver.save(sess, args.outfile)
if epoch >= args.decay_after:
learning_rate /= 1.2
last_valid = valid_loss
return sess.run([normalize_op
])[0] # return final normalized embeddings
weights = tf.Variable(
tf.truncated_normal(shape=[num_hidden, num_classes],
stddev=1.0 / math.sqrt(float(num_classes))))
biases = tf.Variable(tf.zeros([num_classes]))
# Define TensorFlow Operations
# ===========================================================================
# Assigns values for each example across all classes corresponding to
# likelihood that the sequence belongs to that class (logits are unscaled
# values, not a probability distribution)
logit = ops.inference(sequence, weights, biases, num_hidden)
# Initialize tensorflow operations to train network:
# Loss function: cross entropy
loss = ops.loss(logit, target)
# Add a scalar summary for the snapshot loss.
tf.scalar_summary(loss.op.name, loss)
# Gradient calculations:
# Initialize gradient descent optimizer
optimizer = ops.optimizer(learning_rate)
# Step 1: Calculate gradient
gradients = ops.calc_gradient(optimizer, loss)
# Step 2: Calculate gradient norm for stopping criteria
gradient_norm = ops.gradient_norm(gradients)
# Step 3: Apply gradients and update model
def __init__(self,
img_shape,
model_path,
logdir,
sampledir,
epochs=200,
gen_lr=0.001,
dis_lr=0.001,
z_shape=100,
batch_size=64,
beta1=0.5,
SampleAfter=100,
SaveAfter=1000):
## Loading parameters so other methods can access them easily
self.height, self.width, self.channels = img_shape
self.epochs = epochs
self.gen_lr = gen_lr
self.dis_lr = dis_lr
self.z_shape = z_shape #single integer value
self.batch_size = batch_size
self.beta1 = beta1
self.SampleAfter = SampleAfter
self.SaveAfter = SaveAfter
self.model_path = model_path
self.logdir = logdir
self.sampledir = sampledir
# Initiating genrator and discriminator object
self.genrator = genrator(img_shape, z_shape)
self.discriminator = discriminator(img_shape)
# Loading Dataset
mnist = tf.keras.datasets.mnist
(x_train, _), (x_test, _) = mnist.load_data()
x_train = np.concatenate([x_train, x_test])
self.x_train = x_train / 127.5 - 1
## Input placeholders
self.in_x = tf.placeholder(tf.float32, [None, self.height, self.width])
self.in_z = tf.placeholder(tf.float32, [None, z_shape])
## genrate images
self.genrated = self.genrator.feed(self.in_z)
## Feeding both fake and real images into discriminator
DisFake = self.discriminator.feed(self.genrated)
DiscReal = self.discriminator.feed(self.in_x)
## Calculating loss , trying to predict genrated images as fake and real images as real
FakeLoss = loss(tf.zeros_like(DisFake), DisFake)
RealLoss = loss(tf.ones_like(DiscReal), DiscReal)
#Defining genrator and discriminator loss
self.DisLoss = tf.add(FakeLoss, RealLoss)
self.GenLoss = loss(tf.ones_like(DisFake), DisFake)
## Adding summary for tensorboard visualization
tf.summary.scalar("DisLos", self.DisLoss)
#tf.summary.scalar("GenLoss",self.GenLoss)
## Seprating descriminator and genrator trainable variables
TrainVar = tf.trainable_variables()
DisVar = [var for var in TrainVar if 'DIS' in var.name]
GenVar = [var for var in TrainVar if 'GEN' in var.name]
self.DisOpt = tf.train.AdamOptimizer(self.dis_lr, self.beta1).minimize(
self.DisLoss, var_list=DisVar)
self.GenOpt = tf.train.AdamOptimizer(self.gen_lr, self.beta1).minimize(
self.GenLoss, var_list=GenVar)
self.SummaryOp = tf.summary.merge_all()
self.saver = tf.train.Saver()
def build_model(self, inputs, labels, is_training):
pad = int((self.image_size - self.height) / 2)
inputs = tf.pad(inputs, [[0, 0], [pad, pad], [pad, pad], [0, 0]])
# convolution with 7x7 kernel and stride 2 (new size: 112x112x64)
self.network = ops.convolution(inputs,
self.channels,
64,
7,
64,
stride=2,
is_training=is_training,
scope='conv1')
# pooling with 3x3 kernel and stride 2 (new size: 56x56x64)
self.network = ops.pooling(self.network, k_size=3, scope='pool1')
# convolution with 1x1 kernel and stride 1 (new size: 56x56x192)
self.network = ops.convolution(self.network,
64,
192,
1,
192,
batch_norm=False,
is_training=is_training,
scope='conv2')
# convolution with 3x3 kernel and stride 1 (new size: 56x56x192)
self.network = ops.convolution(self.network,
192,
192,
3,
192,
is_training=is_training,
scope='conv3')
# pooling with 3x3 kernel and stride 2 (new size: 28x28x192)
self.network = ops.pooling(self.network, k_size=3, scope='pool2')
# inception module (3a)
self.network = self.inception_module(self.network,
[[64, 96, 16], [128, 32, 32]],
scope='incept1')
# inception module (3b)
self.network = self.inception_module(self.network,
[[128, 128, 32], [192, 96, 64]],
final_pool=True,
scope='incept' + str(i))
# inception module (4a)
self.network = self.inception_module(self.network,
[[192, 96, 16], [208, 48, 64]],
scope='incept' + str(i))
# auxiliary classifier
if is_training:
aux_loss1 = self.aux_classifier(self.network,
labels,
512,
is_training,
scope='auxclass1')
# inception module (4b)
self.network = self.inception_module(self.network,
[[160, 112, 24], [224, 64, 64]],
scope='incept' + str(i))
# inception module (4c)
self.network = self.inception_module(self.network,
[[128, 128, 24], [256, 64, 64]],
scope='incept' + str(i))
# inception module (4d)
self.network = self.inception_module(self.network,
[[112, 144, 32], [288, 64, 64]],
scope='incept' + str(i))
# auxiliary classifier
if is_training:
aux_loss2 = self.aux_classifier(self.network,
labels,
528,
is_training,
scope='auxclass2')
# inception module (4e)
self.network = self.inception_module(self.network,
[[256, 160, 32], [320, 128, 128]],
final_pool=True,
scope='incept' + str(i))
# inception module (5a)
self.network = self.inception_module(self.network,
[[256, 160, 32], [320, 128, 128]],
scope='incept' + str(i))
# inception module (5b)
self.network = self.inception_module(self.network,
[[384, 192, 48], [384, 128, 128]],
scope='incept' + str(i))
# pooling with 7x7 kernel and stride 1 (new size: 1x1x1024)
with tf.variable_scope('final_pool', reuse=tf.AUTO_REUSE):
self.network = tf.nn.avg_pool(self.network,
7,
1,
'SAME',
scope='pool')
# flatten (new size: 1024)
self.network = ops.flatten(self.network, scope='flatten')
# fully connected layer (new size: 1024)
self.network = ops.dense(self.network,
1024,
1024,
dropout=True,
dropout_rate=0.4,
is_training=is_training,
scope='fc1')
# output layer (new size: 10) -- Original Paper Size: 1000 (for ImageNet)
self.network = ops.dense(self.network,
1024,
10,
activation=None,
is_training=is_training,
scope='fc2')
loss = ops.loss(self.network, labels, scope='loss')
self.accuracy = ops.accuracy(self.network, labels, scope='accuracy')
if is_training: # if training use auxiliary classifiers as well
self.loss = loss + aux_loss1 + aux_loss2
self.optimizer = ops.optimize(self.loss,
self.learning_rate,
scope='update')
else:
self.loss = loss
def build_model(self, inputs, labels, is_training):
def res_block(inputs, in_channels, out_channels, is_training, idx):
net = ops.convolution(inputs,
in_channels[0],
out_channels[0],
1,
out_channels[0],
is_training=is_training,
scope='res%s_conv1' % idx)
net = ops.convolution(net,
in_channels[1],
out_channels[1],
3,
out_channels[1],
is_training=is_training,
scope='res%s_conv2' % idx)
net = ops.convolution(net,
in_channels[2],
out_channels[2],
1,
out_channels[2],
activation=None,
is_training=is_training,
scope='res%s_conv3' % idx)
return tf.nn.relu(inputs + net, scope='res%s_relu' % idx)
def res_conv_block(inputs, in_channel, out_channel, stride,
is_training, idx):
skip = ops.convolution(inputs,
in_channels[0],
out_channels[2],
1,
out_channels[2],
stride=stride,
activation=None,
is_training=is_training,
scope='res%s_skip' % idx)
net = ops.convolution(inputs,
in_channels[0],
out_channels[0],
1,
out_channels[0],
is_training=is_training,
scope='res%s_conv1' % idx)
net = ops.convolution(net,
in_channels[1],
out_channels[1],
3,
out_channels[1],
is_training=is_training,
scope='res%s_conv2' % idx)
net = ops.convolution(net,
in_channels[2],
out_channels[2],
1,
out_channels[2],
stride=stride,
activation=None,
is_training=is_training,
scope='res%s_conv3' % idx)
return tf.nn.relu(skip + net, scope='res%s_relu' % idx)
# pad inputs to size 224x224x3 - NOTE: may change to bilinear upsampling
pad = int((self.image_size - self.height) / 2)
inputs = tf.pad(inputs, [[0, 0], [pad, pad], [pad, pad], [0, 0]])
# convolution with 7x7 kernel and stride 2 (new size: 112x112x64)
self.network = ops.convolution(inputs,
self.channels,
64,
7,
64,
stride=2,
is_training=is_training,
scope='conv1')
# pooling with 3x3 kernel and stride 2 (new size: 56x56x64)
self.network = ops.pooling(self.network, k_size=3, scope='pool1')
# residual block 1
stride = 1
out_channels = [64, 64, 256]
self.network = res_conv_block(self.network, [64, 64, 64], out_channels,
stride, is_training, 1)
self.network = res_block(self.network, [256, 64, 64], out_channels,
is_training, 2)
self.network = res_block(self.network, [256, 64, 64], out_channels,
is_training, 3)
# residual block 2
stride = 2
out_channels = [128, 128, 512]
self.network = res_conv_block(self.network, [256, 128, 128],
out_channels, stride, is_training, 4)
self.network = res_block(self.network, [512, 128, 128], out_channels,
is_training, 5)
self.network = res_block(self.network, [512, 128, 128], out_channels,
is_training, 6)
self.network = res_block(self.network, [512, 128, 128], out_channels,
is_training, 7)
# residual block 3
stride = 2
out_channels = [256, 256, 1024]
self.network = res_conv_block(self.network, [512, 256, 256],
out_channels, stride, is_training, 8)
self.network = res_block(self.network, [1024, 256, 256], out_channels,
is_training, 9)
self.network = res_block(self.network, [1024, 256, 256], out_channels,
is_training, 10)
self.network = res_block(self.network, [1024, 256, 256], out_channels,
is_training, 11)
self.network = res_block(self.network, [1024, 256, 256], out_channels,
is_training, 12)
self.network = res_block(self.network, [1024, 256, 256], out_channels,
is_training, 13)
# residual block 4
stride = 2
out_channels = [512, 512, 2048]
self.network = res_conv_block(self.network, [1024, 512, 512],
out_channels, stride, is_training, 14)
self.network = res_block(self.network, [2048, 512, 512], out_channels,
is_training, 15)
self.network = res_block(self.network, [2048, 512, 512], out_channels,
is_training, 16)
# average pooling
self.network = tf.nn.avg_pool(self.network,
7,
1,
'SAME',
scope='avg_pool')
self.network = ops.flatten(self.network, scope='flatten')
# fully connected
self.network = ops.dense(self.network,
2048,
10,
activation=None,
is_training=is_training,
scope='fc')
self.loss = ops.loss(self.network, labels, scope='loss')
self.accuracy = ops.accuracy(self.network, labels, scope='accuracy')
if is_training:
self.optimizer = ops.optimize(self.loss,
self.learning_rate,
scope='update')
def build_model(self, inputs, labels, is_training):
# pad inputs to size 224x224x3 - NOTE: may change to bilinear upsampling
pad = int((self.image_size - self.height) / 2)
inputs = tf.pad(inputs, [[0, 0], [pad, pad], [pad, pad], [0, 0]])
# convolution with 3x3 kernel and stride 1 (new size: 224x224x64)
self.network = ops.convolution(inputs,
self.channels,
64,
3,
64,
is_training=is_training,
scope='conv1')
# convolution with 3x3 kernel and stride 1 (new size: 224x224x64)
self.network = ops.convolution(self.network,
64,
64,
3,
64,
is_training=is_training,
scope='conv2')
# pooling with 2x2 kernel and stride 2 (new size: 112x112x64)
self.network = ops.pooling(self.network, scope='pool1')
# convolution with 3x3 kernel and stride 1 (new size: 112x112x128)
self.network = ops.convolution(self.network,
64,
128,
3,
128,
is_training=is_training,
scope='conv3')
# convolution with 3x3 kernel and stride 1 (new size: 112x112x128)
self.network = ops.convolution(self.network,
128,
128,
3,
128,
is_training=is_training,
scope='conv4')
# pooling with 2x2 kernel and stride 2 (new size: 56x56x128)
self.network = ops.pooling(self.network, scope='pool2')
# convolution with 3x3 kernel and stride 1 (new size: 56x56x256)
self.network = ops.convolution(self.network,
128,
256,
3,
256,
is_training=is_training,
scope='conv5')
# 3 convolutions with 3x3 kernel and stride 1 (new size: 56x56x256)
for idx in range(6, 9):
self.network = ops.convolution(self.network,
256,
256,
3,
256,
is_training=is_training,
scope='conv' + str(idx))
# pooling with 2x2 kernel and stride 2 (new size: 28x28x256)
self.network = ops.pooling(self.network, scope='pool3')
# convolution with 3x3 kernel and stride 1 (new size: 28x28x512)
self.network = ops.convolution(self.network,
256,
512,
3,
512,
is_training=is_training,
scope='conv9')
# 3 convolutions with 3x3 kernel and stride 1 (new size: 28x28x512)
for idx in range(10, 13):
self.network = ops.convolution(self.network,
512,
512,
3,
512,
is_training=is_training,
scope='conv' + str(idx))
# pooling with 2x2 kernel and stride 2 (new size: 14x14x512)
self.network = ops.pooling(self.network, scope='pool4')
# 4 convolutions with 3x3 kernel and stride 1 (new size: 14x14x512)
for idx in range(13, 17):
self.network = ops.convolution(self.network,
512,
512,
3,
512,
is_training=is_training,
scope='conv' + str(idx))
# pooling with 2x2 kernel and stride 2 (new size: 7x7x512)
self.network = ops.pooling(self.network, scope='pool5')
# flatten (new size: 25088)
self.network = ops.flatten(self.network, scope='flatten')
# fully connected layer (new size: 4096)
self.network = ops.dense(self.network,
25088,
4096,
dropout=True,
dropout_rate=0.2,
is_training=is_training,
scope='fc1')
# fully connected layer (new size: 1024) -- Original Paper Size: 4096 (for ImageNet)
self.network = ops.dense(self.network,
4096,
1024,
dropout=True,
dropout_rate=0.2,
is_training=is_training,
scope='fc2')
# output layer (new size: 10) -- Original Paper Size: 1000 (for ImageNet)
self.network = ops.dense(self.network,
1024,
10,
activation=None,
is_training=is_training,
scope='fc3')
self.loss = ops.loss(self.network, labels, scope='loss')
self.accuracy = ops.accuracy(self.network, labels, scope='accuracy')
if is_training:
self.optimizer = ops.optimize(self.loss,
self.learning_rate,
scope='update')
Y = tf.placeholder(tf.float32, shape=[BATCH_SIZE, 6, 1])
img1, img2 = getBatch('2001',ID, BATCH_SIZE, WIDTH, HEIGHT)
output_data = np.genfromtxt('2001/results.csv', delimiter=',')
output = getGT(output_data, ID, BATCH_SIZE)[:,:,np.newaxis]
arg = {}
net = KLTNet( arg )
p, dp = net( X1, X2 )
warp = get_warp(p)
# print p.shape
# loss_op = tf.reduce_mean( tf.square( X2[:,EPS:(HEIGHT-EPS), EPS:(WIDTH-EPS),:] - spatial_transformer_network(X1, warp)[:,5:67, 5:99,:] ), axis=[1,2,3] )
loss_op = loss( p, Y, BATCH_SIZE, HEIGHT, WIDTH )
# transform_op = spatial_transformer_network(X1, warp)
# loss2 = tf.reduce_mean(dp)
init = tf.global_variables_initializer()
config = tf.ConfigProto(
device_count = {'GPU': 0}
)
with tf.Session( config=config ) as sess:
sess.run(init)
# print( sess.run(p,feed_dict={X1:img1, X2:img2} ) )
def run_training(datasets):
"""Train the autoencoder for a number of steps."""
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
placeholders = placeholder_inputs(FLAGS.batch_size, datasets.emb_size)
inputs_placeholder, labels_placeholder, mask_placeholder = placeholders
# Build a Graph that computes predictions from the inference model.
logits = ops.inference(inputs_placeholder,
datasets.emb_size,
FLAGS.hidden1,
FLAGS.hidden2,
FLAGS.dropout_rate)
# Add to the Graph the Ops for loss calculation.
loss = ops.loss(logits, labels_placeholder, mask_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = ops.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct, eval_total = ops.evaluation(logits,
labels_placeholder,
mask_placeholder)
# keep track of the epoch
count = tf.Variable(0)
increment = tf.count_up_to(count, FLAGS.num_epochs)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
if os.path.exists(CP_INFO) and not FLAGS.retrain:
restore_variables(sess)
# redo epoch that we last quit in the middle of
sess.run(count.assign_sub(1))
else:
init = tf.initialize_all_variables()
sess.run(init)
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.train.SummaryWriter(FLAGS.summary_dir,
graph_def=sess.graph_def)
def do_eval(data_set):
"""Runs one evaluation against the full epoch of data.
:param data_set: The data_set which we will use to retrieve batches
"""
# And run one epoch of eval.
steps_per_epoch = data_set.num_examples // FLAGS.batch_size
counts = np.zeros(2)
for _ in xrange(steps_per_epoch):
feed_dict = fill_feed_dict(data_set, placeholders)
run = sess.run([eval_correct, eval_total], feed_dict=feed_dict)
counts += run
correct, total = counts
print(' Num examples: %d Num correct: %d Precision @ 1: %0.04f' %
(int(total), int(correct), correct / total))
# And then after everything is built, start the training loop.
steps_per_epoch = datasets.train.num_examples // FLAGS.batch_size
start_time = time.time()
# TODO: make epoch a saved variable so that training picks up where it left off
for epoch in xrange(FLAGS.num_epochs):
for step in xrange(steps_per_epoch):
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
feed_dict = fill_feed_dict(datasets.train, placeholders)
# Run one step of the model. The return values are the activations
# from the `train_op` (which is discarded) and the `loss` Op. To
# inspect the values of your Ops or variables, you may include them
# in the list passed to sess.run() and the value tensors will be
# returned in the tuple from the call.
_, loss_value = sess.run([train_op, loss],
feed_dict=feed_dict)
if step == 0:
# Write the summaries and print an overview fairly often.
duration = time.time() - start_time
# Print status to stdout.
print('Epoch %d: loss = %.2f (%.3f sec)' % (epoch, loss_value, duration))
# Update the events file.
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, epoch)
# create the save directory if not already there
if not os.path.isdir(FLAGS.save_dir):
os.mkdir(FLAGS.save_dir)
save_dir = os.path.join(FLAGS.save_dir, 'model.ckpt')
saver.save(sess, save_dir, global_step=step)
if epoch % 10 == 0 or (epoch + 1) == FLAGS.num_epochs:
# Evaluate against the training set.
print('Training Data Eval:')
do_eval(datasets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
do_eval(datasets.validation)
# Evaluate against the test set.
print('Test Data Eval:')
do_eval(datasets.test)