def __init__(self, training, development, synthetic, config):
self.Training = training
self.Development = development
self.config = config
self.num_objs = 20 # 20 blocks + Background
with tf.variable_scope('language'):
self.text_embeddings = Embedding(
self.Training.vocab_size,
name='txt',
one_hot=False,
embedding_size=self.config.txt_dim)
with tf.variable_scope('ops'):
self.op_embeddings = Embedding(
self.config.num_ops,
name='ops',
one_hot=False,
embedding_size=self.config.pixel_dim)
self.locs = []
self.vision_bn_phase = tf.placeholder(tf.bool, name='train_phase')
self.dropout = tf.placeholder(tf.float32, name='dropout')
self.interactive_mode = tf.placeholder(tf.bool, name='interactive')
# Create language pipeline
with tf.variable_scope('language') as scope:
l_phs, self.vision_loss, (pred_x, pred_y, pred_z,
pred_t) = self._model()
scope.reuse_variables()
self.locs.extend([pred_x, pred_y, pred_z, pred_t])
self.placeholders = l_phs
with tf.variable_scope('lang_op'):
self.language_op = self._language_loss()
self.opts = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.hist_summary = tf.summary.merge_all()
self.sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True))
# Tensorboard
self.train_reg_summary = tf.summary.scalar('train_reg', self.reg_loss)
self.train_rot_summary = tf.summary.scalar('train_rot',
self.rotation_loss)
self.val_cost_summary = tf.summary.scalar('val_cost', self.reg_loss)
self.val_rot_summary = tf.summary.scalar('val_rot', self.rotation_loss)
self.writer = tf.summary.FileWriter(self.config.summary_path,
graph=tf.get_default_graph())
self.saver = tf.train.Saver()
if config.load_model is not None:
print "Restoring ", config.load_model
self.saver.restore(self.sess, config.load_model)
else:
print "Initializing"
self.sess.run(tf.global_variables_initializer())
def __init__(self, training, development, synthetic, config):
self.Training = training
self.Development = development
self.config = config
self.num_objs = 20
with tf.variable_scope('language'):
self.text_embeddings = Embedding(
self.Training.vocab_size,
name='txt',
one_hot=False,
embedding_size=self.config.txt_dim)
self.dropout = tf.placeholder(tf.float32, name='dropout')
self.interactive_mode = tf.placeholder(tf.bool, name='interactive')
# Create language pipeline
with tf.variable_scope('language') as scope:
self.placeholders, self.source_loss, self.prediction = self._model(
)
scope.reuse_variables()
with tf.variable_scope('lang_op'):
self.language_op = self._language_loss()
self.opts = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.hist_summary = tf.summary.merge_all()
self.sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True))
# Tensorboard
self.train_summary = tf.summary.scalar('train', self.source_loss)
self.val_summary = tf.summary.scalar('val', self.source_loss)
self.writer = tf.summary.FileWriter(self.config.summary_path,
graph=tf.get_default_graph())
self.saver = tf.train.Saver()
if config.load_model is not None:
print "Restoring ", config.load_model
self.saver.restore(self.sess, config.load_model)
else:
print "Initializing"
self.sess.run(tf.global_variables_initializer())
training, training_labels = Sparse.matrix(train)
development, development_labels = Sparse.matrix(dev)
testing, testing_labels = Sparse.matrix(test)
## TODO:
## MutiCellLSTM
batch_size = 128
hiddendim = 256
embeddingdim = 100
onehot = False
graph = tf.Graph()
dropout=0.5
# Define embeddings matrix
embeddings = Embedding(vocabsize, one_hot=onehot, embedding_size=embeddingdim)
# Input -> LSTM -> Outstate
inputs = tf.placeholder(tf.int32, [batch_size, maxlength])
labels = tf.placeholder(tf.float32, [batch_size, labelspace])
lengths = tf.placeholder(tf.int32, [batch_size])
# RNN
lstm = tf.nn.rnn_cell.LSTMCell(hiddendim,
initializer=tf.contrib.layers.xavier_initializer(seed=20160501))
lstm = tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=dropout)
# Prediction
output_layer = Layer.W(2*hiddendim, labelspace, 'Output')
output_bias = Layer.b(labelspace, 'OutputBias')
test, _, _ = Sparse.read("JSONReader/data/2016-NAACL/SRD/Test.mat")
## Create sparse arrays
training, training_labels = Sparse.matrix(train)
development, development_labels = Sparse.matrix(dev)
testing, testing_labels = Sparse.matrix(test)
batch_size = 128
hiddendim = 100
embeddingdim = 100
graph = tf.Graph()
onehot = True
inputdim = maxlength * vocabsize if onehot else maxlength * embeddingdim
# Define embeddings matrix
embeddings = Embedding(vocabsize, one_hot=onehot, embedding_size=embeddingdim)
# Input data.
dataset = tf.placeholder(tf.int32, shape=[batch_size, maxlength], name='Train')
labels = tf.placeholder(tf.float32,
shape=[batch_size, labelspace],
name='Label')
# Model
hidden_layer = Layer.W(inputdim, hiddendim, 'Hidden')
hidden_bias = Layer.b(hiddendim, 'HiddenBias')
# Prediction
output_layer = Layer.W(hiddendim, labelspace, 'Output')
output_bias = Layer.b(labelspace, 'OutputBias')
embedded = tf.reshape(embeddings.lookup(dataset), [batch_size, inputdim])
forward = tf.nn.relu(tf.matmul(embedded, hidden_layer) + hidden_bias)
dropout = tf.nn.dropout(forward, 0.5)
## Create sparse arrays
t, t_l = Sparse.matrix(train)
training[prediction] = t
training_labels[prediction] = t_l
d, d_l = Sparse.matrix(dev)
development[prediction] = d
development_labels[prediction] = d_l
t, t_l = Sparse.matrix(test)
testing[prediction] = t
testing_labels[prediction] = t_l
# Define embeddings matrix
embeddings = {}
embeddings[0] = Embedding(vocabsize,
one_hot=onehot,
embedding_size=embeddingdim)
embeddings[1] = Embedding(vocabsize,
one_hot=onehot,
embedding_size=embeddingdim)
embeddings[2] = Embedding(vocabsize,
one_hot=onehot,
embedding_size=embeddingdim)
# Input -> LSTM -> Outstate
inputs = tf.placeholder(tf.int32, [batch_size, maxlength])
labels = {}
labels[0] = tf.placeholder(tf.float32, [batch_size, 20])
labels[1] = tf.placeholder(tf.float32, [batch_size, 20])
labels[2] = tf.placeholder(tf.float32, [batch_size, 9])
lengths = tf.placeholder(tf.int32, [batch_size])
test, _, _ = Sparse.read("JSONReader/data/2016-NAACL/SRD/Test.mat")
## Create sparse arrays
training, training_labels = Sparse.matrix(train)
development, development_labels = Sparse.matrix(dev)
testing, testing_labels = Sparse.matrix(test)
batch_size = 128
hiddendim = 100
embeddingdim = 100
graph = tf.Graph()
onehot = True
inputdim = maxlength*vocabsize if onehot else maxlength*embeddingdim
# Define embeddings matrix
embeddings = Embedding(vocabsize, one_hot=onehot, embedding_size=embeddingdim)
# Input data.
dataset = tf.placeholder(tf.int32, shape=[batch_size, maxlength], name='Train')
labels = tf.placeholder(tf.float32, shape=[batch_size, labelspace], name='Label')
# Model
hidden_layer = Layer.W(inputdim, hiddendim, 'Hidden')
hidden_bias = Layer.b(hiddendim, 'HiddenBias')
# Prediction
output_layer = Layer.W(hiddendim, labelspace, 'Output')
output_bias = Layer.b(labelspace, 'OutputBias')
embedded = tf.reshape(embeddings.lookup(dataset), [batch_size,inputdim])
forward = tf.nn.relu(tf.matmul(embedded, hidden_layer) + hidden_bias)
dropout = tf.nn.dropout(forward, 0.5)
logits = tf.matmul(dropout, output_layer) + output_bias
'out':
Layer.W(final_size * final_size * filters + 2 * hiddendim,
rep_dim * rep_dim)
}
B = {
'cb1': Layer.b(filters, init='Normal'),
'cb2': Layer.b(filters, init='Normal'),
'cb3': Layer.b(filters, init='Normal'),
'cb4': Layer.b(filters, init='Normal'),
'cb5': Layer.b(filters, init='Normal'),
'out': Layer.b(rep_dim * rep_dim)
}
# Define embeddings matrix
embeddings = Embedding(vocabsize, one_hot=False, embedding_size=hiddendim)
# RNN
dropout = 0.75
lstm = tf.nn.rnn_cell.LSTMCell(
hiddendim, initializer=tf.contrib.layers.xavier_initializer(seed=20160501))
lstm = tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=dropout)
# Encode from 18x18 to 12x12
l1 = conv2d('l1', cur_world, W['cl1'], B['cb1'],
padding='VALID') # -> 32->30 18->16
l2 = conv2d('l2', l1, W['cl2'], B['cb2'], padding='VALID') # -> 30-28 16->14
l3 = conv2d('l3', l2, W['cl3'], B['cb3'], padding='VALID') # -> 28->26 14->12
l4 = conv2d('l4', l3, W['cl4'], B['cb4'], padding='VALID') # -> 26->24 12->10
l5 = conv2d('l5', l4, W['cl5'], B['cb5'], padding='VALID') # -> 24->22 10->8
