def get_model(point_cloud, is_training,Transform=False, bn_decay=None):
forloss = []
idx=None
if point_cloud.get_shape()[-1].value>2:
l0_xyz = point_cloud[:,:,0:2]
l0_points = point_cloud [:,:,2:]
else:
l0_xyz = point_cloud
l0_points = None
if Transform:
l0_xyz, tnet,l0_points= transform_moudule(l0_xyz,l0_points, is_training,idx, [64,128,1024],[[1, 1], [1, 1],[1, 1]], [512,256],num_channle=32, nsample=32, scope='Tnet1', bn_decay=bn_decay, bn=True, K=2)
#forloss=tnet
l0_points, plane_feature, _ = plane_module(l0_xyz, l0_points, idx, centralize=True, num_channle=32, npoint=512,
have_sample=False,
nsample=32, is_training=is_training, scope='plane', pool='avg',
use_xyz=True, bn=True, bn_decay=bn_decay, weight_decay=0.0)
new_points,_ = conv_module(l0_points,plane_feature,mlp=[64,128,128,1024], is_training=is_training, scope='conv', bn=True, bn_decay=bn_decay)
# Fully connected layers
net = util.fully_connected(new_points, 512, bn=True, is_training=is_training, scope='fc1', bn_decay=bn_decay)
net = util.dropout(net, keep_prob=0.5, is_training=is_training, scope='dp1')
net = util.fully_connected(net, 256, bn=True, is_training=is_training, scope='fc2', bn_decay=bn_decay)
net = util.dropout(net, keep_prob=0.5, is_training=is_training, scope='dp2')
net = util.fully_connected(net, 10, activation_fn=None, scope='fc3')
return net, forloss
def __init__(self, optimizer, activation):
super().__init__(optimizer, activation)
############################################################################
# Define the graph #
############################################################################
# It turns out that this network from ex03 is already capable of memorizing
# the entire training or validation set, so we need to tweak generalization,
# not capacity
# In order to speed up convergence, we added batch normalization.
# Our best effort was Adam optimizer with bs=32, lr=0.001 (only possible
# because of norm)
x = tf.placeholder(tf.float32, shape=(None, 32, 32, 1), name='input')
y_ = tf.placeholder(dtype=tf.int32, shape=(None,), name='labels')
self.x = x
self.y_ = y_
kernel_shape1 = (5, 5, 1, 8)
activation1 = conv_layer(x, kernel_shape1, activation=activation)
normalize1 = batch_norm_layer(activation1)
pool1 = weighted_pool_layer(
normalize1, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1)
)
kernel_shape2 = (3, 3, 8, 10)
activation2 = conv_layer(pool1, kernel_shape2, activation=activation)
normalize2 = batch_norm_layer(activation2)
pool2 = weighted_pool_layer(
normalize2, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1)
)
pool2_reshaped = tf.reshape(pool2, (-1, 8*8*10), name='reshaped1')
fc1 = fully_connected(pool2_reshaped, 512, with_activation=True,
activation=activation)
fc2_logit = fully_connected(fc1, 10, activation=activation)
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=fc2_logit,
labels=y_)
mean_cross_entropy = tf.reduce_mean(cross_entropy)
self.mean_cross_entropy = mean_cross_entropy
train_step = optimizer.minimize(mean_cross_entropy)
self.train_step = train_step
self.prediction = tf.cast(tf.argmax(fc2_logit, 1), tf.int32)
# check if neuron firing strongest coincides with max value position in real
# labels
correct_prediction = tf.equal(self.prediction, y_)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
self.accuracy = accuracy
def __init__(self, optimizer, activation):
super().__init__(optimizer, activation)
############################################################################
# Define the graph #
############################################################################
# It turns out that this network from ex03 is already capable of memorizing
# the entire training or validation set, so we need to tweak generalization,
# not capacity
x = tf.placeholder(tf.float32, shape=(None, 32, 32, 1), name='input')
y_ = tf.placeholder(dtype=tf.int32, shape=(None, ), name='labels')
self.x = x
self.y_ = y_
kernel_shape1 = (5, 5, 1, 16)
activation1 = conv_layer(x, kernel_shape1, activation=activation)
pool1 = max_pool_layer(activation1,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1))
kernel_shape2 = (3, 3, 16, 32)
activation2 = conv_layer(pool1, kernel_shape2, activation=activation)
pool2 = max_pool_layer(activation2,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1))
pool2_reshaped = tf.reshape(pool2, (-1, 2048), name='reshaped1')
fc1 = fully_connected(pool2_reshaped,
512,
with_activation=True,
activation=activation)
fc2_logit = fully_connected(fc1, 10, activation=activation)
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=fc2_logit, labels=y_)
mean_cross_entropy = tf.reduce_mean(cross_entropy)
self.mean_cross_entropy = mean_cross_entropy
train_step = optimizer.minimize(mean_cross_entropy)
self.train_step = train_step
# check if neuron firing strongest coincides with max value position in real
# labels
correct_prediction = tf.equal(
tf.argmax(fc2_logit, 1, output_type=tf.int32), y_)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
self.accuracy = accuracy
def transform_moudule(xyz,points, is_training,idx, num_out_conv, size_conv, num_out_fc,num_channle,nsample, scope, bn_decay, bn=True,K=3):
with tf.variable_scope(scope) as sc:
new_points, net ,idx = plane_module(xyz, points, idx,centralize=True, num_channle=num_channle, npoint=1024,
nsample=nsample,
is_training=is_training, scope='Tnet-plane', pool='avg',
use_xyz=True, bn=bn, bn_decay=bn_decay,weight_decay=0.0)
batch_size = tf.shape(xyz)[0]
for i in range(len(num_out_conv)):
net = util.conv2d(net, num_out_conv[i], size_conv[i], scope='tconv' + str(i + 1), stride=[1, 1], bn=True,
bn_decay=bn_decay, is_training=is_training, padding='VALID')
net = tf.squeeze(net, 2)
net = tf.reduce_max(net, axis=1)
for i in range(len(num_out_fc)):
net = util.fully_connected(net, num_out_fc[i], scope='tfc' + str(i + 1), bn_decay=bn_decay, bn=True,
is_training=is_training)
with tf.variable_scope(scope) as sc:
weights = tf.get_variable('weights', [256, K * K],
initializer=tf.constant_initializer(0.0),
dtype=tf.float32)
biases = tf.get_variable('biases', [K * K],
initializer=tf.constant_initializer(0.0),
dtype=tf.float32)
biases += tf.constant(np.eye(K).flatten(), dtype=tf.float32)
transform = tf.matmul(net, weights)
transform = tf.nn.bias_add(transform, biases)
transform = tf.reshape(transform, [batch_size, K, K])
xyz = tf.matmul(xyz, transform)
if new_points is not None:
new_points = tf.matmul(new_points, transform)
return xyz,transform,new_points
def action_value_function(self, input, reuse=True, scope='Action_Value'):
#should be created with invarianve to batch size
#input should be a placeholder
#outputs are actions shape [batch_size, num_actions]
with tf.variable_scope(scope, reuse=reuse):
out = tf.cast(input, tf.float32)
conv_params = self.params['conv_layers']
for i in range(len(conv_params)):
clp = conv_params[i]
out = ut.conv2d(out, clp[0], clp[1], clp[2],
'conv2d%d' % (i + 1))
out = tf.nn.relu(out)
out = ut.fully_connected(out, self.params['FC_layer'],
'fully_con1')
out = tf.nn.relu(out)
out = ut.fully_connected(out, self.params['n_output'],
'fully_con2', False)
return out
def farword(self, inputs_image):
inputs = tf.reshape(inputs_image, [-1, 58, 58, 1])
with tf.variable_scope('conv1'):
W1 = tf.Variable(tf.random_normal([3, 3, 1, 16]), name='w1')
net = conv2d(inputs, W1)
net = pooling(net)
with tf.variable_scope('conv2'):
W2 = tf.Variable(tf.random_normal([3, 3, 16, 16]), name='w2')
net = conv2d(net, W2)
net = pooling(net)
with tf.variable_scope('fc'):
logits = fully_connected(net)
return logits
def __init__(self, learning_rate, **kwargs):
'''Respected kwargs are
Parameters
----------
steps : int
Decay learning rate according to `exponential_decay` every `steps` steps
decay : decay factor for `exponential_decay`
learning_rate : base learning rate for `exponential_decay`
'''
######################################
# Create the dynamic learning rate #
######################################
step_counter = tf.Variable(0,
trainable=False,
dtype=tf.int32,
name='step_counter')
steps = kwargs.get('steps', 100)
decay = kwargs.get('decay', 0.8)
learning_rate = tf.train.exponential_decay(learning_rate, step_counter,
steps, decay)
############################################################################################
# Define the network #
############################################################################################
self.observations = tf.placeholder(tf.float32,
shape=[1, 4],
name='observations')
hidden_layer = fully_connected(self.observations,
8,
with_activation=True,
activation=tf.nn.relu)
probability = fully_connected(hidden_layer,
1,
with_activation=True,
activation=tf.nn.sigmoid)
complementary = tf.subtract(1.0, probability)
output = tf.concat([probability, complementary],
1,
name='action_probabilities')
log_likelihoods = tf.log(output)
self.action = tf.multinomial(log_likelihoods, num_samples=1)[0][0]
log_likelihood = log_likelihoods[:, tf.to_int32(self.action)]
optimizer = tf.train.AdamOptimizer(learning_rate)
grads_and_vars = optimizer.compute_gradients(log_likelihood)
self.gradients = [grad * -1 for (grad, _) in grads_and_vars]
# Gradients must be fed for training
self.grad_placeholders = []
for i, gradient in enumerate(self.gradients):
self.grad_placeholders.append(
tf.placeholder(tf.float32,
gradient.shape,
name=f'gradient_{i}'))
self.grad_dummies = [
np.zeros(grad.shape) for grad in self.grad_placeholders
]
tvars = tf.trainable_variables()
self.training_step = optimizer.apply_gradients(
zip(self.grad_placeholders, tvars), global_step=step_counter)
def __init__(self, optimizer, activation):
super().__init__(optimizer, activation)
x = tf.placeholder(tf.float32, shape=(None, 32, 32, 1), name='input')
y_ = tf.placeholder(dtype=tf.int32, shape=(None, ), name='labels')
self.x = x
self.y_ = y_
# logits = (LinearWrap(image)
# .Conv2D('conv1', 24, 5, padding='VALID')
# .MaxPooling('pool1', 2, padding='SAME')
# .Conv2D('conv2', 32, 3, padding='VALID')
# .Conv2D('conv3', 32, 3, padding='VALID')
# .MaxPooling('pool2', 2, padding='SAME')
# .Conv2D('conv4', 64, 3, padding='VALID')
# .Dropout('drop', 0.5)
# .FullyConnected('fc0', 512,
# b_init=tf.constant_initializer(0.1), nl=tf.nn.relu)
# .FullyConnected('linear', out_dim=10, nl=tf.identity)())
# tf.nn.softmax(logits, name='output')
l1 = conv_layer(x, (5, 5, 1, 24),
activation=None,
padding='VALID',
use_bias=False)
l2 = tf.nn.relu(batch_norm_layer(l1))
l3 = tf.nn.max_pool(l2, (1, 2, 2, 1), (1, 2, 2, 1), padding='SAME')
l4 = conv_layer(l3, (3, 3, 24, 32),
padding='VALID',
activation=None,
use_bias=False)
l5 = tf.nn.relu(batch_norm_layer(l4))
l6 = conv_layer(l5, (3, 3, 32, 32),
padding='VALID',
activation=None,
use_bias=False)
l7 = tf.nn.relu(batch_norm_layer(l6))
l8 = tf.nn.max_pool(l7, (1, 2, 2, 1), (1, 2, 2, 1), padding='SAME')
l8_ = conv_layer(l8, (3, 3, 32, 64),
padding='VALID',
activation=None,
use_bias=False)
l9 = tf.nn.relu(batch_norm_layer(l8_))
l10 = tf.nn.dropout(l9, 0.5)
l11 = tf.reshape(l10, (-1, 3 * 3 * 64), name='reshaped1')
l12 = fully_connected(l11,
512,
with_activation=True,
activation=tf.nn.relu)
l13 = fully_connected(l12, 10, with_activation=False, use_bias=False)
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=l13, labels=y_)
mean_cross_entropy = tf.reduce_mean(cross_entropy)
self.mean_cross_entropy = mean_cross_entropy
train_step = optimizer.minimize(mean_cross_entropy)
self.train_step = train_step
self.prediction = tf.cast(tf.argmax(l13, 1), tf.int32)
# check if neuron firing strongest coincides with max value position in real
# labels
correct_prediction = tf.equal(self.prediction, y_)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
self.accuracy = accuracy
def __init__(self, **kwargs):
'''The following arguments are accepted:
Parameters
----------
vocab_size : int
Size of the vocabulary for creating embeddings
embedding_matrix : int
Dimensionality of the embedding space
memory_size : int
LSTM memory size
keep_prob : float
Inverse of dropout percentage for embedding and LSTM
subsequence_length : int
Length of the subsequences (all embeddings are padded to this
length)
optimizer : OptimizerSpec
'''
############################################################################################
# Get all hyperparameters #
############################################################################################
vocab_size = kwargs['vocab_size']
embedding_size = kwargs['embedding_size']
memory_size = kwargs['memory_size']
keep_prob = kwargs['keep_prob']
subsequence_length = kwargs['subsequence_length']
optimizer_spec = kwargs['optimizer']
optimizer = optimizer_spec.create()
self.learning_rate = optimizer_spec.learning_rate
self.step_counter = optimizer_spec.step_counter
############################################################################################
# Net inputs #
############################################################################################
self.batch_size = placeholder(tf.int32, shape=[], name='batch_size')
self.is_training = placeholder(tf.bool, shape=[], name='is_training')
self.word_ids = placeholder(tf.int32,
shape=(None, subsequence_length),
name='word_ids')
self.labels = placeholder(tf.int32, shape=(None, ), name='labels')
self.hidden_state = placeholder(tf.float32,
shape=(None, memory_size),
name='hidden_state')
self.cell_state = placeholder(tf.float32,
shape=(None, memory_size),
name='cell_state')
lengths = sequence_lengths(self.word_ids)
############################################################################################
# Embedding #
############################################################################################
self.embedding_matrix, _bias = get_weights_and_bias(
(vocab_size, embedding_size))
embeddings = cond(
self.is_training, lambda: nn.dropout(nn.embedding_lookup(
self.embedding_matrix, self.word_ids),
keep_prob=keep_prob),
lambda: nn.embedding_lookup(self.embedding_matrix, self.word_ids))
############################################################################################
# LSTM layer #
############################################################################################
cell = BasicLSTMCell(memory_size, activation=tf.nn.tanh)
# during inference, use entire ensemble
keep_prob = cond(self.is_training, lambda: constant(keep_prob),
lambda: constant(1.0))
cell = DropoutWrapper(cell, output_keep_prob=keep_prob)
# what's the difference to just creating a zero-filled tensor tuple?
self.zero_state = cell.zero_state(self.batch_size, tf.float32)
state = LSTMStateTuple(h=self.cell_state, c=self.hidden_state)
# A dynamic rnn creates the graph on the fly, so it can deal with embeddings of different
# lengths. We do not need to unstack the embedding tensor to get rows, instead we compute
# the actual sequence lengths and pass that
# We are not sure how any of this works. Do we need to mask the cost function so the cell
# outputs for _NOT_A_WORD_ inputs are ignored? Is the final cell state really relevant if it
# was last updated with _NOT_A_WORD_ input? Does static_rnn absolve us of any of those
# issues?
outputs, self.state = nn.dynamic_rnn(cell,
embeddings,
sequence_length=lengths,
initial_state=state)
# Recreate tensor from list
outputs = reshape(concat(outputs, 1),
[-1, subsequence_length * memory_size])
self.outputs = reduce_mean(outputs)
############################################################################################
# Fully connected layer, loss, and training #
############################################################################################
ff1 = fully_connected(outputs, 2, with_activation=False, use_bias=True)
loss = reduce_mean(
nn.sparse_softmax_cross_entropy_with_logits(labels=self.labels,
logits=ff1))
self.train_step = optimizer.minimize(loss,
global_step=self.step_counter)
self.predictions = nn.softmax(ff1)
correct_prediction = equal(cast(argmax(self.predictions, 1), tf.int32),
self.labels)
self.accuracy = reduce_mean(cast(correct_prediction, tf.float32))
############################################################################################
# Create summaraies #
############################################################################################
with tf.variable_scope('summary'):
self.summary_loss = tf.summary.scalar('loss', loss)
self.summary_accuracy = tf.summary.scalar('accuracy',
self.accuracy)
