def __init__(self, batch_size):
# NUM_CORES = 4
# session = tf.Session(config=tf.ConfigProto(inter_op_parallelism_threads=NUM_CORES, intra_op_parallelism_threads=NUM_CORES))
num_hidden_units = 2048
session = tf.InteractiveSession()
input_units = 28
output_units = 1
x = tf.placeholder("float", shape=[None, input_units], name='x')
true_y = tf.placeholder("float", shape=[None, output_units], name='y_')
# W_fc1 = weight_variable([input_units, num_hidden_units])
W_fc1 = weight_variable([input_units, num_hidden_units])
b_fc1 = bias_variable([num_hidden_units])
h_fc1 = tf.nn.relu(tf.matmul(x, W_fc1) + b_fc1)
# W_fc2 = weight_variable([num_hidden_units, num_hidden_units])
W_fc2 = weight_variable([num_hidden_units, num_hidden_units])
b_fc2 = bias_variable([num_hidden_units])
h_fc2 = tf.nn.relu(tf.matmul(h_fc1, W_fc2) + b_fc2)
# W_fc3 = weight_variable([num_hidden_units, num_hidden_units])
W_fc3 = weight_variable([num_hidden_units, num_hidden_units])
b_fc3 = bias_variable([num_hidden_units])
h_fc3 = tf.nn.relu(tf.matmul(h_fc2, W_fc3) + b_fc3)
# W_fc4 = weight_variable([num_hidden_units, output_units])
W_fc4 = weight_variable([num_hidden_units, output_units])
b_fc4 = bias_variable([output_units])
keep_prob = tf.Variable(0.5, name='keep_prob', trainable=False)
h_fc3_dropout = tf.nn.dropout(h_fc3, keep_prob)
guess_y = tf.matmul(h_fc3, W_fc4) + b_fc4
guess_y_dropout = tf.matmul(h_fc3_dropout, W_fc4) + b_fc4
variables = [W_fc1, b_fc1, W_fc2, b_fc2, W_fc3, b_fc3, W_fc4, b_fc4]
# loss = tf.nn.l2_loss(guess_y_dropout - true_y)
#loss = tf.reduce_mean(tf.square(tf.sign(guess_y_dropout) - tf.sign(true_y - 0.5)))
loss = tf.reduce_mean(tf.square(guess_y_dropout - true_y))
# optimizer = tf.train.AdamOptimizer(learning_rate=0.00001, beta1=0.99, beta2=0.999, epsilon=1e-06, use_locking=False, name='Adam')
#optimizer = tf.train.MomentumOptimizer(learning_rate=0.0000001, momentum=0.9, use_locking=False, name='momentum')
optimizer = tf.train.MomentumOptimizer(learning_rate=0.0000001, momentum=0.9995, use_locking=False, name='Momentum')
# optimizer = tf.train.RMSPropOptimizer(1e-4, decay=0.9, momentum=0.0, epsilon=1e-10, use_locking=False, name='RMSProp')
compute_gradients = optimizer.compute_gradients(loss, variables)
apply_gradients = optimizer.apply_gradients(compute_gradients)
minimize = optimizer.minimize(loss)
#correct_prediction = tf.equal(tf.clip_by_value(tf.round(guess_y), 0.0, 1.0), true_y)
#correct_prediction = tf.equal(tf.round(guess_y), true_y)
#correct_prediction = tf.equal((tf.sign(guess_y) + 1)/2., true_y)
error_rate = loss
#error_rate = 1 - tf.reduce_mean(tf.cast(correct_prediction, "float"))
# correct_prediction = tf.equal(tf.argmax(guess_y,1), tf.argmax(true_y,1))
ParameterServerModel.__init__(self, x, true_y, compute_gradients, apply_gradients, minimize, error_rate, session,batch_size)
def __init__(self, batch_size, gpu=True):
#NUM_CORES = 4
self.gpu = gpu
self.batch_size = batch_size
session = tf.InteractiveSession()
# session = tf.Session(config=tf.ConfigProto(log_device_placement=True))
#session = tf.Session(config=tf.ConfigProto(inter_op_parallelism_threads=NUM_CORES, intra_op_parallelism_threads=NUM_CORES))
# with session.graph.device(self.device_for_node):
input_units = 784
output_units = 10
hidden_units = 1024
x = tf.placeholder("float", shape=[None, input_units], name='x')
#x_image = tf.reshape(x, [-1,28,28,1], name='reshape')
true_y = tf.placeholder("float", shape=[None, output_units], name='y_')
W_fc0 = weight_variable([input_units, hidden_units], 'W_fc0')
b_fc0 = bias_variable([hidden_units], 'b_fc0')
h_fc0 = tf.nn.relu(tf.matmul(x, W_fc0) + b_fc0)
W_fc1 = weight_variable([hidden_units, hidden_units], 'W_fc1')
b_fc1 = bias_variable([hidden_units], 'b_fc1')
h_fc1 = tf.nn.relu(tf.matmul(h_fc0, W_fc1) + b_fc1)
W_fc2 = weight_variable([hidden_units, output_units], 'W_fc2')
b_fc2 = bias_variable([output_units], 'b_fc2')
keep_prob = tf.Variable(0.5, name='keep_prob', trainable=False)
h_fc1_dropout = tf.nn.dropout(h_fc1, keep_prob)
guess_y = tf.matmul(h_fc1, W_fc2) + b_fc2
guess_y_dropout = tf.matmul(h_fc1_dropout, W_fc2) + b_fc2
variables = [W_fc0, b_fc0, W_fc1, b_fc1, W_fc2, b_fc2]
loss = tf.nn.softmax_cross_entropy_with_logits(guess_y_dropout, true_y)
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
compute_gradients = optimizer.compute_gradients(loss, variables)
apply_gradients = optimizer.apply_gradients(compute_gradients)
minimize = optimizer.minimize(loss)
correct_prediction = tf.equal(tf.argmax(guess_y, 1),
tf.argmax(true_y, 1))
error_rate = 1 - tf.reduce_mean(tf.cast(correct_prediction, "float"))
ParameterServerModel.__init__(self, x, true_y, compute_gradients,
apply_gradients, minimize, error_rate,
session, batch_size)
def __init__(self):
session = tf.InteractiveSession()
x = tf.placeholder("float", shape=[None, 784], name='x')
x_image = tf.reshape(x, [-1,28,28,1], name='reshape')
y_ = tf.placeholder("float", shape=[None, 10], name='y_')
W_conv1 = weight_variable([5, 5, 1, 32], 'W_conv1')
b_conv1 = bias_variable([32], 'b_conv1')
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
W_conv2 = weight_variable([5, 5, 32, 64], 'W_conv2')
b_conv2 = bias_variable([64], 'b_conv2')
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
W_fc1 = weight_variable([7 * 7 * 64, 1024], 'W_fc1')
b_fc1 = bias_variable([1024], 'b_fc1')
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
keep_prob = tf.Variable(0.5, name='keep_prob', trainable=False)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = weight_variable([1024, 10], 'W_fc2')
b_fc2 = bias_variable([10], 'b_fc2')
# not using dropout for testing, only training
y_conv_dropout = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
y_conv = tf.matmul(h_fc1, W_fc2) + b_fc2
variables = [W_conv1, b_conv1, W_conv2, b_conv2, W_fc1, b_fc1, W_fc2, b_fc2]
loss = tf.nn.softmax_cross_entropy_with_logits(y_conv_dropout, y_)
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
compute_gradients = optimizer.compute_gradients(loss, variables)
apply_gradients = optimizer.apply_gradients(compute_gradients)
minimize = optimizer.minimize(loss)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
ParameterServerModel.__init__(self,
x,
y_,
compute_gradients,
apply_gradients,
minimize,
accuracy,
session)
def __init__(self, batch_size, gpu=True):
#NUM_CORES = 4
self.gpu = gpu
self.batch_size = batch_size
session = tf.InteractiveSession()
# session = tf.Session(config=tf.ConfigProto(log_device_placement=True))
#session = tf.Session(config=tf.ConfigProto(inter_op_parallelism_threads=NUM_CORES, intra_op_parallelism_threads=NUM_CORES))
# with session.graph.device(self.device_for_node):
input_units = 784
output_units = 10
hidden_units = 1024
x = tf.placeholder("float", shape=[None, input_units], name='x')
#x_image = tf.reshape(x, [-1,28,28,1], name='reshape')
true_y = tf.placeholder("float", shape=[None, output_units], name='y_')
W_fc0 = weight_variable([input_units, hidden_units], 'W_fc0')
b_fc0 = bias_variable([hidden_units], 'b_fc0')
h_fc0 = tf.nn.relu(tf.matmul(x, W_fc0) + b_fc0)
W_fc1 = weight_variable([hidden_units, hidden_units], 'W_fc1')
b_fc1 = bias_variable([hidden_units], 'b_fc1')
h_fc1 = tf.nn.relu(tf.matmul(h_fc0, W_fc1) + b_fc1)
W_fc2 = weight_variable([hidden_units, output_units], 'W_fc2')
b_fc2 = bias_variable([output_units], 'b_fc2')
keep_prob = tf.Variable(0.5, name='keep_prob', trainable=False)
h_fc1_dropout = tf.nn.dropout(h_fc1, keep_prob)
guess_y = tf.matmul(h_fc1, W_fc2) + b_fc2
guess_y_dropout = tf.matmul(h_fc1_dropout, W_fc2) + b_fc2
variables = [W_fc0, b_fc0, W_fc1, b_fc1, W_fc2, b_fc2]
loss = tf.nn.softmax_cross_entropy_with_logits(guess_y_dropout, true_y)
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
compute_gradients = optimizer.compute_gradients(loss, variables)
apply_gradients = optimizer.apply_gradients(compute_gradients)
minimize = optimizer.minimize(loss)
correct_prediction = tf.equal(tf.argmax(guess_y,1), tf.argmax(true_y,1))
error_rate = 1 - tf.reduce_mean(tf.cast(correct_prediction, "float"))
ParameterServerModel.__init__(self, x, true_y, compute_gradients, apply_gradients, minimize, error_rate, session, batch_size)
def __init__(self):
session = tf.InteractiveSession()
x = tf.placeholder("float", shape=[None, 784], name='x')
x_image = tf.reshape(x, [-1, 28, 28, 1], name='reshape')
y_ = tf.placeholder("float", shape=[None, 10], name='y_')
W_conv1 = weight_variable([5, 5, 1, 32], 'W_conv1')
b_conv1 = bias_variable([32], 'b_conv1')
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
W_conv2 = weight_variable([5, 5, 32, 64], 'W_conv2')
b_conv2 = bias_variable([64], 'b_conv2')
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
W_fc1 = weight_variable([7 * 7 * 64, 1024], 'W_fc1')
b_fc1 = bias_variable([1024], 'b_fc1')
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
keep_prob = tf.Variable(0.5, name='keep_prob', trainable=False)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = weight_variable([1024, 10], 'W_fc2')
b_fc2 = bias_variable([10], 'b_fc2')
# not using dropout for testing, only training
y_conv_dropout = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
y_conv = tf.matmul(h_fc1, W_fc2) + b_fc2
variables = [
W_conv1, b_conv1, W_conv2, b_conv2, W_fc1, b_fc1, W_fc2, b_fc2
]
loss = tf.nn.softmax_cross_entropy_with_logits(y_conv_dropout, y_)
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
compute_gradients = optimizer.compute_gradients(loss, variables)
apply_gradients = optimizer.apply_gradients(compute_gradients)
minimize = optimizer.minimize(loss)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
ParameterServerModel.__init__(self, x, y_, compute_gradients,
apply_gradients, minimize, accuracy,
session)
def __init__(self, batch_size, gpu=False):
session = tf.Session()
self.gpu = gpu
#self.batch_size = batch_size
input_units = 784
output_units = 10
hidden_units = 100 #700
x = tf.placeholder('float32', shape=[None, input_units], name='x')
true_y = tf.placeholder('float32',
shape=[None, output_units],
name='y')
W_fc0 = weight_variable([input_units, hidden_units], 'W_fc0')
b_fc0 = bias_variable([hidden_units], 'b_fc0')
h_fc0 = tf.nn.relu(tf.matmul(x, W_fc0) + b_fc0)
W_fc1 = weight_variable([hidden_units, output_units], 'W_fc1')
b_fc1 = bias_variable([output_units], 'b_fc1')
guess_y = tf.matmul(h_fc0, W_fc1) + b_fc1
variables = [W_fc0, b_fc0, W_fc1, b_fc1]
loss = tf.nn.softmax_cross_entropy_with_logits(guess_y, true_y)
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
compute_gradients = optimizer.compute_gradients(loss, variables)
apply_gradients = optimizer.apply_gradients(compute_gradients)
minimize = optimizer.minimize(loss)
correct_prediction = tf.equal(tf.argmax(guess_y, 1),
tf.argmax(true_y, 1))
error_rate = 1 - tf.reduce_mean(tf.cast(correct_prediction, 'float32'))
ParameterServerModel.__init__(self, x, true_y, compute_gradients,
apply_gradients, minimize, error_rate,
session, batch_size)
def __init__(self, input_size, output_size, conv_filter_sizes, dense_layer_sizes, image_width,
image_height, dropout = 0.5, learning_rate = 1e-4, channels = 1, l2norm = 1e-5, gpu=False,
recall = 'input_x', dropname = 'drop_pb', predict = 'predictions', batch_size=64, dump=sys.stdout):
session = tf.InteractiveSession()
training_accuracies = []
validation_accuracies = []
training_steps = []
features_weights = []
test_accuracy = 0
x = tf.placeholder('float', shape = [None, image_width, image_height, channels], name=recall)
y_ = tf.placeholder('float', shape = [None, output_size])
keep_prob = tf.placeholder('float', name=dropname)
with tf.device('/gpu:0' if gpu else '/cpu:0'):
# Build the network structure
previous_inputs_size = input_size
inputs_next = x
for conv_layer in conv_filter_sizes:
cur_filter = weight_variable(conv_layer)
cur_conv_bias = bias_variable([conv_layer[-1]])
out_conv = tf.nn.conv2d(inputs_next, cur_filter, strides=[1,1,1,1], padding='SAME')
inputs_next = tf.nn.relu(out_conv + cur_conv_bias)
new_width = out_conv.get_shape()[1]
new_height = out_conv.get_shape()[2]
new_channels = out_conv.get_shape()[3]
out_pool_flat = tf.reshape(out_conv, [-1, int(new_channels*new_height*new_width)])
# build some densely connected layers
inputs_next = out_pool_flat
previous_inputs_size = int(new_height*new_width*new_channels)
for layer_size in dense_layer_sizes[1:-1]:
cur_weights = weight_variable([previous_inputs_size, layer_size])
cur_bias = bias_variable([layer_size])
cur_neurons = tf.nn.relu(tf.matmul(inputs_next, cur_weights) + cur_bias)
cur_dropout = tf.nn.dropout(cur_neurons, keep_prob = keep_prob)
inputs_next = cur_dropout
previous_inputs_size = layer_size
final_inputs = inputs_next
prev_add_size = 0
# This is the last layer
out_weights = weight_variable([previous_inputs_size + prev_add_size, output_size])
out_bias = bias_variable([output_size])
out_final = tf.nn.softmax(tf.matmul(final_inputs, out_weights) + out_bias, name=predict)
# define cost function
cost_entropy = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(out_final,1e-10,1.0))) + tf.reduce_sum(tf.clip_by_value(l2norm*cur_weights,1e-10,1.0))
# define optimiser
opt = tf.train.AdamOptimizer(learning_rate)
compute_gradients = opt.compute_gradients(cost_entropy, tf.trainable_variables())
apply_gradients = opt.apply_gradients(compute_gradients)
minimize = opt.minimize(cost_entropy)
# define accuracy
correct_prediction = tf.equal(tf.argmax(out_final,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
ParameterServerModel.__init__(self, x, y_, keep_prob, dropout, compute_gradients, apply_gradients, minimize, accuracy, session, batch_size, out_final, gpu, dump)
def __init__(self, batch_size):
# NUM_CORES = 4
# session = tf.Session(config=tf.ConfigProto(inter_op_parallelism_threads=NUM_CORES, intra_op_parallelism_threads=NUM_CORES))
num_hidden_units = 2048
session = tf.InteractiveSession()
input_units = 28
output_units = 1
x = tf.placeholder("float", shape=[None, input_units], name='x')
true_y = tf.placeholder("float", shape=[None, output_units], name='y_')
# W_fc1 = weight_variable([input_units, num_hidden_units])
W_fc1 = weight_variable([input_units, num_hidden_units])
b_fc1 = bias_variable([num_hidden_units])
h_fc1 = tf.nn.relu(tf.matmul(x, W_fc1) + b_fc1)
# W_fc2 = weight_variable([num_hidden_units, num_hidden_units])
W_fc2 = weight_variable([num_hidden_units, num_hidden_units])
b_fc2 = bias_variable([num_hidden_units])
h_fc2 = tf.nn.relu(tf.matmul(h_fc1, W_fc2) + b_fc2)
# W_fc3 = weight_variable([num_hidden_units, num_hidden_units])
W_fc3 = weight_variable([num_hidden_units, num_hidden_units])
b_fc3 = bias_variable([num_hidden_units])
h_fc3 = tf.nn.relu(tf.matmul(h_fc2, W_fc3) + b_fc3)
# W_fc4 = weight_variable([num_hidden_units, output_units])
W_fc4 = weight_variable([num_hidden_units, output_units])
b_fc4 = bias_variable([output_units])
keep_prob = tf.Variable(0.5, name='keep_prob', trainable=False)
h_fc3_dropout = tf.nn.dropout(h_fc3, keep_prob)
guess_y = tf.matmul(h_fc3, W_fc4) + b_fc4
guess_y_dropout = tf.matmul(h_fc3_dropout, W_fc4) + b_fc4
variables = [W_fc1, b_fc1, W_fc2, b_fc2, W_fc3, b_fc3, W_fc4, b_fc4]
# loss = tf.nn.l2_loss(guess_y_dropout - true_y)
#loss = tf.reduce_mean(tf.square(tf.sign(guess_y_dropout) - tf.sign(true_y - 0.5)))
loss = tf.reduce_mean(tf.square(guess_y_dropout - true_y))
# optimizer = tf.train.AdamOptimizer(learning_rate=0.00001, beta1=0.99, beta2=0.999, epsilon=1e-06, use_locking=False, name='Adam')
#optimizer = tf.train.MomentumOptimizer(learning_rate=0.0000001, momentum=0.9, use_locking=False, name='momentum')
optimizer = tf.train.MomentumOptimizer(learning_rate=0.0000001,
momentum=0.9995,
use_locking=False,
name='Momentum')
# optimizer = tf.train.RMSPropOptimizer(1e-4, decay=0.9, momentum=0.0, epsilon=1e-10, use_locking=False, name='RMSProp')
compute_gradients = optimizer.compute_gradients(loss, variables)
apply_gradients = optimizer.apply_gradients(compute_gradients)
minimize = optimizer.minimize(loss)
#correct_prediction = tf.equal(tf.clip_by_value(tf.round(guess_y), 0.0, 1.0), true_y)
#correct_prediction = tf.equal(tf.round(guess_y), true_y)
#correct_prediction = tf.equal((tf.sign(guess_y) + 1)/2., true_y)
error_rate = loss
#error_rate = 1 - tf.reduce_mean(tf.cast(correct_prediction, "float"))
# correct_prediction = tf.equal(tf.argmax(guess_y,1), tf.argmax(true_y,1))
ParameterServerModel.__init__(self, x, true_y, compute_gradients,
apply_gradients, minimize, error_rate,
session, batch_size)