def buid_sentence_expression():
sentence_tree = td.InputTransform(lambda sentence_json: WNJsonDecoder(sentence_json))
tree_rnn = td.ForwardDeclaration(td.PyObjectType())
leaf_case = td.GetItem('word_vec', name='leaf_in') >> td.Vector(embedding_size)
index_case = td.Record({'children': td.Map(tree_rnn()) >> td.Mean(), 'word_vec': td.Vector(embedding_size)}, name='index_in') >> td.Concat(name='concat_root_child') >> td.FC(embedding_size, name='FC_root_child')
expr_sentence = td.OneOf(td.GetItem('leaf'), {True: leaf_case, False: index_case}, name='recur_in')
tree_rnn.resolve_to(expr_sentence)
return sentence_tree >> expr_sentence
def tree_sum_blk(loss_blk):
# traverse the tree to sum up the loss
tree_sum_fwd = td.ForwardDeclaration(td.PyObjectType(), td.TensorType([]))
tree_sum = td.Composition()
with tree_sum.scope():
myloss = loss_blk().reads(tree_sum.input)
children = td.GetItem('children').reads(tree_sum.input)
mapped = td.Map(tree_sum_fwd()).reads(children)
summed = td.Reduce(td.Function(tf.add)).reads(mapped)
summed = td.Function(tf.add).reads(summed, myloss)
tree_sum.output.reads(summed)
tree_sum_fwd.resolve_to(tree_sum)
return tree_sum
def __init__(self, weight_matrix, word_idx, ModelConfig):
self.ModelConfig = ModelConfig
self.word_embedding = td.Embedding(*weight_matrix.shape,
initializer=weight_matrix,
name='word_embedding')
self.word_idx = word_idx
self.keep_prob_ph = tf.placeholder_with_default(1.0, [])
self.tree_lstm = td.ScopedLayer(tf.contrib.rnn.DropoutWrapper(
BinaryTreeLSTMCell(self.ModelConfig.lstm_num_units,
keep_prob=self.keep_prob_ph),
input_keep_prob=self.keep_prob_ph,
output_keep_prob=self.keep_prob_ph),
name_or_scope='tree_lstm')
self.output_layer = td.FC(self.ModelConfig.num_classes,
activation=None,
name='output_layer')
self.embed_subtree = td.ForwardDeclaration(name='embed_subtree')
self.model = self.embed_tree(is_root=True)
self.embed_subtree.resolve_to(self.embed_tree(is_root=False))
self.compiler = td.Compiler.create(self.model)
print('input type: %s' % self.model.input_type)
print('output type: %s' % self.model.output_type)
self.metrics = {
k: tf.reduce_mean(v)
for k, v in self.compiler.metric_tensors.items()
}
self.loss = tf.reduce_sum(self.compiler.metric_tensors['all_loss'])
opt = tf.train.AdagradOptimizer(ModelConfig.learning_rate)
grads_and_vars = opt.compute_gradients(self.loss)
found = 0
for i, (grad, var) in enumerate(grads_and_vars):
if var == self.word_embedding.weights:
found += 1
grad = tf.scalar_mul(ModelConfig.embedding_learning_rate, grad)
grads_and_vars[i] = (grad, var)
assert found == 1 # internal consistency check
self.train_op = opt.apply_gradients(grads_and_vars)
self.saver = tf.train.Saver()
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
def buid_sentence_expression():
sentence_tree = td.InputTransform(
lambda sentence_json: WordNode(sentence_json))
tree_rnn = td.ForwardDeclaration(td.PyObjectType())
leaf_case = td.GetItem(
'word_id', name='leaf_in') >> td.Scalar(dtype=tf.int32) >> embedding
index_case = td.Record({'left': tree_rnn(), 'right': tree_rnn()}) \
>> td.Concat(name='concat_root_child') \
>> fc
expr_sentence = td.OneOf(td.GetItem('leaf'), {
True: leaf_case,
False: index_case
},
name='recur_in')
tree_rnn.resolve_to(expr_sentence)
return sentence_tree >> expr_sentence
def __init__(self,
weights,
vocab,
tree_lstm_num_units,
tree_binarizer=None):
if not tree_binarizer:
tree_binarizer = TreeBinarizer(vocab, dict())
self.tree_binarizer = tree_binarizer
self.tree_lstm_keep_prob_ph = tf.placeholder_with_default(1.0, [])
self.tree_lstm_cell = td.ScopedLayer(
tf.contrib.rnn.DropoutWrapper(
BinaryTreeLSTMCell(tree_lstm_num_units,
self.tree_lstm_keep_prob_ph),
self.tree_lstm_keep_prob_ph, self.tree_lstm_keep_prob_ph),
name_or_scope=self._tree_lstm_cell_default_scope_name)
self.word_embedding = td.Embedding(
*weights.shape,
initializer=weights,
name=self._word_embedding_default_scope_name)
self.embed_subtree = td.ForwardDeclaration(name='embed_subtree')
self.vocab = vocab
def dynamic_pooling_blk():
"""Input: root node dic
Output: pooled, TensorType([hyper.conv_dim, ])
"""
leaf_case = feature_detector_blk()
pool_fwd = td.ForwardDeclaration(td.PyObjectType(),
td.TensorType([
hyper.conv_dim,
]))
pool = td.Composition()
with pool.scope():
cur_fea = feature_detector_blk().reads(pool.input)
children = td.GetItem('children').reads(pool.input)
mapped = td.Map(pool_fwd()).reads(children)
summed = td.Reduce(td.Function(tf.maximum)).reads(mapped)
summed = td.Function(tf.maximum).reads(summed, cur_fea)
pool.output.reads(summed)
pool = td.OneOf(lambda x: x['clen'] == 0, {True: leaf_case, False: pool})
pool_fwd.resolve_to(pool)
return pool
def __init__(self, image_feat_grid, text_seq_batch, seq_length_batch,
T_decoder, num_vocab_txt, embed_dim_txt, num_vocab_nmn,
embed_dim_nmn, lstm_dim, num_layers, assembler,
encoder_dropout, decoder_dropout, decoder_sampling,
num_choices, use_qpn, qpn_dropout, reduce_visfeat_dim=False, new_visfeat_dim=256,
use_gt_layout=None, gt_layout_batch=None,
scope='neural_module_network', reuse=None):
with tf.variable_scope(scope, reuse=reuse):
# Part 0: Visual feature from CNN
self.reduce_visfeat_dim = reduce_visfeat_dim
if reduce_visfeat_dim:
# use an extrac linear 1x1 conv layer (without ReLU)
# to reduce the feature dimension
with tf.variable_scope('reduce_visfeat_dim'):
image_feat_grid = conv('conv_reduce_visfeat_dim',
image_feat_grid, kernel_size=1, stride=1,
output_dim=new_visfeat_dim)
print('visual feature dimension reduced to %d' % new_visfeat_dim)
self.image_feat_grid = image_feat_grid
# Part 1: Seq2seq RNN to generate module layout tokensa
with tf.variable_scope('layout_generation'):
att_seq2seq = AttentionSeq2Seq(text_seq_batch,
seq_length_batch, T_decoder, num_vocab_txt,
embed_dim_txt, num_vocab_nmn, embed_dim_nmn, lstm_dim,
num_layers, assembler, encoder_dropout, decoder_dropout,
decoder_sampling, use_gt_layout, gt_layout_batch)
self.att_seq2seq = att_seq2seq
predicted_tokens = att_seq2seq.predicted_tokens
token_probs = att_seq2seq.token_probs
word_vecs = att_seq2seq.word_vecs
neg_entropy = att_seq2seq.neg_entropy
self.atts = att_seq2seq.atts
self.predicted_tokens = predicted_tokens
self.token_probs = token_probs
self.word_vecs = word_vecs
self.neg_entropy = neg_entropy
# log probability of each generated sequence
self.log_seq_prob = tf.reduce_sum(tf.log(token_probs), axis=0)
# Part 2: Neural Module Network
with tf.variable_scope('layout_execution'):
modules = Modules(image_feat_grid, word_vecs, None, num_choices)
self.modules = modules
# Recursion of modules
att_shape = image_feat_grid.get_shape().as_list()[1:-1] + [1]
# Forward declaration of module recursion
att_expr_decl = td.ForwardDeclaration(td.PyObjectType(), td.TensorType(att_shape))
# _Scene
case_scene = td.Record([('time_idx', td.Scalar(dtype='int32')),
('batch_idx', td.Scalar(dtype='int32'))])
case_scene = case_scene >> td.Function(modules.SceneModule)
# _Find
case_find = td.Record([('time_idx', td.Scalar(dtype='int32')),
('batch_idx', td.Scalar(dtype='int32'))])
case_find = case_find >> td.Function(modules.FindModule)
# _Filter
case_filter = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar(dtype='int32')),
('batch_idx', td.Scalar(dtype='int32'))])
case_filter = case_filter >> td.Function(modules.FilterModule)
# _FindSameProperty
case_find_same_property = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar(dtype='int32')),
('batch_idx', td.Scalar(dtype='int32'))])
case_find_same_property = case_find_same_property >> \
td.Function(modules.FindSamePropertyModule)
# _Transform
case_transform = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_transform = case_transform >> td.Function(modules.TransformModule)
# _And
case_and = td.Record([('input_0', att_expr_decl()),
('input_1', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_and = case_and >> td.Function(modules.AndModule)
# _Or
case_or = td.Record([('input_0', att_expr_decl()),
('input_1', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_or = case_or >> td.Function(modules.OrModule)
# _Exist
case_exist = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_exist = case_exist >> td.Function(modules.ExistModule)
# _Count
case_count = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_count = case_count >> td.Function(modules.CountModule)
# _EqualNum
case_equal_num = td.Record([('input_0', att_expr_decl()),
('input_1', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_equal_num = case_equal_num >> td.Function(modules.EqualNumModule)
# _MoreNum
case_more_num = td.Record([('input_0', att_expr_decl()),
('input_1', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_more_num = case_more_num >> td.Function(modules.MoreNumModule)
# _LessNum
case_less_num = td.Record([('input_0', att_expr_decl()),
('input_1', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_less_num = case_less_num >> td.Function(modules.LessNumModule)
# _SameProperty
case_same_property = td.Record([('input_0', att_expr_decl()),
('input_1', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_same_property = case_same_property >> \
td.Function(modules.SamePropertyModule)
# _Describe
case_describe = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_describe = case_describe >> \
td.Function(modules.DescribeModule)
recursion_cases = td.OneOf(td.GetItem('module'), {
'_Scene': case_scene,
'_Find': case_find,
'_Filter': case_filter,
'_FindSameProperty': case_find_same_property,
'_Transform': case_transform,
'_And': case_and,
'_Or': case_or})
att_expr_decl.resolve_to(recursion_cases)
# For invalid expressions, define a dummy answer
# so that all answers have the same form
dummy_scores = td.Void() >> td.FromTensor(np.zeros(num_choices, np.float32))
output_scores = td.OneOf(td.GetItem('module'), {
'_Exist': case_exist,
'_Count': case_count,
'_EqualNum': case_equal_num,
'_MoreNum': case_more_num,
'_LessNum': case_less_num,
'_SameProperty': case_same_property,
'_Describe': case_describe,
INVALID_EXPR: dummy_scores})
# compile and get the output scores
self.compiler = td.Compiler.create(output_scores)
self.scores_nmn = self.compiler.output_tensors[0]
# Add a question prior network if specified
self.use_qpn = use_qpn
self.qpn_dropout = qpn_dropout
if use_qpn:
self.scores_qpn = question_prior_net(att_seq2seq.encoder_states,
num_choices, qpn_dropout)
self.scores = self.scores_nmn + self.scores_qpn
else:
self.scores = self.scores_nmn
# Regularization: Entropy + L2
self.entropy_reg = tf.reduce_mean(neg_entropy)
module_weights = [v for v in tf.trainable_variables()
if (scope in v.op.name and
v.op.name.endswith('weights'))]
self.l2_reg = tf.add_n([tf.nn.l2_loss(v) for v in module_weights])
keep_prob_ph = tf.placeholder_with_default(1.0, [])
lstm_num_units = 256 # Tai et al. used 150
tree_lstm = td.ScopedLayer(binaryTreeLSTMCell(lstm_num_units,
keep_prob=keep_prob_ph),
name_or_scope='tree_lstm')
NUM_CLASSES = 6 # number of distinct labels
output_layer = td.FC(NUM_CLASSES, activation=None, name='output_layer')
word_embedding = td.Embedding(*weight_matrix.shape,
initializer=weight_matrix,
name='word_embedding',
trainable=False)
# declare recursive model
embed_subtree = td.ForwardDeclaration(name='embed_subtree')
def makeContextMat(input1):
input1 = int(input1)
if input1 == 0:
# if input1 < 2:
return [1 for i in range(10)]
else:
return [0 for i in range(10)]
def makeDepthMat(input2):
input1 = int(input2)
return [1 if i < input1 else 0 for i in range(20)]
def run(write_to, batch_size_setting):
startTime = time.time()
data_dir = "../senti/"
"""
def download_and_unzip(url_base, zip_name, *file_names):
zip_path = os.path.join(data_dir, zip_name)
url = url_base + zip_name
print('downloading %s to %s' % (url, zip_path))
urllib.request.urlretrieve(url, zip_path)
out_paths = []
with zipfile.ZipFile(zip_path, 'r') as f:
for file_name in file_names:
print('extracting %s' % file_name)
out_paths.append(f.extract(file_name, path=data_dir))
return out_paths
def download(url_base, zip_name):
zip_path = os.path.join(data_dir, zip_name)
url = url_base + zip_name
print('downloading %s to %s' % (url, zip_path))
urllib.request.urlretrieve(url, zip_path)
full_glove_path, = download_and_unzip(
'http://nlp.stanford.edu/data/', 'glove.840B.300d.zip',
'glove.840B.300d.txt')
train_path, dev_path, test_path = download_and_unzip(
'http://nlp.stanford.edu/sentiment/', 'trainDevTestTrees_PTB.zip',
'trees/train.txt', 'trees/dev.txt', 'trees/test.txt')
filtered_glove_path = os.path.join(data_dir, 'filtered_glove.txt')
def filter_glove():
vocab = set()
# Download the full set of unlabeled sentences separated by '|'.
sentence_path, = download_and_unzip(
'http://nlp.stanford.edu/~socherr/', 'stanfordSentimentTreebank.zip',
'stanfordSentimentTreebank/SOStr.txt')
with codecs.open(sentence_path, encoding='utf-8') as f:
for line in f:
# Drop the trailing newline and strip backslashes. Split into words.
vocab.update(line.strip().replace('\\', '').split('|'))
nread = 0
nwrote = 0
with codecs.open(full_glove_path, encoding='utf-8') as f:
with codecs.open(filtered_glove_path, 'w', encoding='utf-8') as out:
for line in f:
nread += 1
line = line.strip()
if not line: continue
if line.split(u' ', 1)[0] in vocab:
out.write(line + '\n')
nwrote += 1
print('read %s lines, wrote %s' % (nread, nwrote))
#filter_glove()
"""
dev_glove_path = os.path.join('./', 'small_glove.txt')
def load_embeddings(embedding_path):
"""Loads embedings, returns weight matrix and dict from words to indices."""
print('loading word embeddings from %s' % embedding_path)
weight_vectors = []
word_idx = {}
with codecs.open(embedding_path, encoding='utf-8') as f:
for line in f:
word, vec = line.split(u' ', 1)
word_idx[word] = len(weight_vectors)
weight_vectors.append(np.array(vec.split(), dtype=np.float32))
# Annoying implementation detail; '(' and ')' are replaced by '-LRB-' and
# '-RRB-' respectively in the parse-trees.
#word_idx[u'-LRB-'] = word_idx.pop(u'(')
#word_idx[u'-RRB-'] = word_idx.pop(u')')
# Random embedding vector for unknown words.
weight_vectors.append(
np.random.uniform(-0.05, 0.05,
weight_vectors[0].shape).astype(np.float32))
return np.stack(weight_vectors), word_idx
weight_matrix, word_idx = load_embeddings(dev_glove_path)
def load_trees(filename):
with codecs.open(filename, encoding='utf-8') as f:
# Drop the trailing newline and strip \s.
trees = [line.strip().replace('\\', '') for line in f]
print('loaded %s trees from %s' % (len(trees), filename))
return trees
#train_path = './senti/trees/train.txt'
#train_path = os.path.join(data_dir, 'trees/dev.txt')
train_path = './dev.txt'
#dev_path = './senti/trees/dev.txt'
#test_path = './senti/trees/test.txt'
train_trees = load_trees(train_path)
trainSIZE = len(train_trees)
#dev_trees = load_trees(dev_path)
#test_trees = load_trees(test_path)
class BinaryTreeLSTMCell(tf.contrib.rnn.BasicLSTMCell):
"""LSTM with two state inputs.
This is the model described in section 3.2 of 'Improved Semantic
Representations From Tree-Structured Long Short-Term Memory
Networks' <http://arxiv.org/pdf/1503.00075.pdf>, with recurrent
dropout as described in 'Recurrent Dropout without Memory Loss'
<http://arxiv.org/pdf/1603.05118.pdf>.
"""
def __init__(self, num_units, keep_prob=1.0):
"""Initialize the cell.
Args:
num_units: int, The number of units in the LSTM cell.
keep_prob: Keep probability for recurrent dropout.
"""
super(BinaryTreeLSTMCell, self).__init__(num_units)
self._keep_prob = keep_prob
def __call__(self, inputs, state, scope=None):
with tf.variable_scope(scope or type(self).__name__):
lhs, rhs = state
c0, h0 = lhs
c1, h1 = rhs
concat = tf.contrib.layers.linear(
tf.concat([inputs, h0, h1], 1), 5 * self._num_units)
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f0, f1, o = tf.split(value=concat,
num_or_size_splits=5,
axis=1)
j = self._activation(j)
if not isinstance(self._keep_prob,
float) or self._keep_prob < 1:
j = tf.nn.dropout(j, self._keep_prob)
new_c = (c0 * tf.sigmoid(f0 + self._forget_bias) +
c1 * tf.sigmoid(f1 + self._forget_bias) +
tf.sigmoid(i) * j)
new_h = self._activation(new_c) * tf.sigmoid(o)
new_state = tf.contrib.rnn.LSTMStateTuple(new_c, new_h)
return new_h, new_state
keep_prob_ph = tf.placeholder_with_default(1.0, [])
lstm_num_units = 150 # Tai et al. used 150, but our regularization strategy is more effective
tree_lstm = td.ScopedLayer(tf.contrib.rnn.DropoutWrapper(
BinaryTreeLSTMCell(lstm_num_units, keep_prob=keep_prob_ph),
input_keep_prob=keep_prob_ph,
output_keep_prob=keep_prob_ph),
name_or_scope='tree_lstm')
NUM_CLASSES = 5 # number of distinct sentiment labels
output_layer = td.FC(NUM_CLASSES, activation=None, name='output_layer')
word_embedding = td.Embedding(*weight_matrix.shape,
initializer=weight_matrix,
name='word_embedding',
trainable=False)
embed_subtree = td.ForwardDeclaration(name='embed_subtree')
def logits_and_state():
"""Creates a block that goes from tokens to (logits, state) tuples."""
unknown_idx = len(word_idx)
lookup_word = lambda word: word_idx.get(word, unknown_idx)
word2vec = (td.GetItem(0) >> td.InputTransform(lookup_word) >>
td.Scalar('int32') >> word_embedding)
pair2vec = (embed_subtree(), embed_subtree())
# Trees are binary, so the tree layer takes two states as its input_state.
zero_state = td.Zeros((tree_lstm.state_size, ) * 2)
# Input is a word vector.
zero_inp = td.Zeros(word_embedding.output_type.shape[0])
word_case = td.AllOf(word2vec, zero_state)
pair_case = td.AllOf(zero_inp, pair2vec)
tree2vec = td.OneOf(len, [(1, word_case), (2, pair_case)])
return tree2vec >> tree_lstm >> (output_layer, td.Identity())
def tf_node_loss(logits, labels):
return tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
labels=labels)
def tf_fine_grained_hits(logits, labels):
predictions = tf.cast(tf.argmax(logits, 1), tf.int32)
return tf.cast(tf.equal(predictions, labels), tf.float64)
def tf_binary_hits(logits, labels):
softmax = tf.nn.softmax(logits)
binary_predictions = (softmax[:, 3] + softmax[:, 4]) > (softmax[:, 0] +
softmax[:, 1])
binary_labels = labels > 2
return tf.cast(tf.equal(binary_predictions, binary_labels), tf.float64)
def add_metrics(is_root, is_neutral):
"""A block that adds metrics for loss and hits; output is the LSTM state."""
c = td.Composition(name='predict(is_root=%s, is_neutral=%s)' %
(is_root, is_neutral))
with c.scope():
# destructure the input; (labels, (logits, state))
labels = c.input[0]
logits = td.GetItem(0).reads(c.input[1])
state = td.GetItem(1).reads(c.input[1])
# calculate loss
loss = td.Function(tf_node_loss)
td.Metric('all_loss').reads(loss.reads(logits, labels))
if is_root: td.Metric('root_loss').reads(loss)
# calculate fine-grained hits
hits = td.Function(tf_fine_grained_hits)
td.Metric('all_hits').reads(hits.reads(logits, labels))
if is_root: td.Metric('root_hits').reads(hits)
# calculate binary hits, if the label is not neutral
if not is_neutral:
binary_hits = td.Function(tf_binary_hits).reads(logits, labels)
td.Metric('all_binary_hits').reads(binary_hits)
if is_root: td.Metric('root_binary_hits').reads(binary_hits)
# output the state, which will be read by our by parent's LSTM cell
c.output.reads(state)
return c
def tokenize(s):
label, phrase = s[1:-1].split(None, 1)
return label, sexpr.sexpr_tokenize(phrase)
def embed_tree(logits_and_state, is_root):
"""Creates a block that embeds trees; output is tree LSTM state."""
return td.InputTransform(tokenize) >> td.OneOf(
key_fn=lambda pair: pair[0] == '2', # label 2 means neutral
case_blocks=(add_metrics(is_root, is_neutral=False),
add_metrics(is_root, is_neutral=True)),
pre_block=(td.Scalar('int32'), logits_and_state))
model = embed_tree(logits_and_state(), is_root=True)
embed_subtree.resolve_to(embed_tree(logits_and_state(), is_root=False))
compiler = td.Compiler.create(model)
print('input type: %s' % model.input_type)
print('output type: %s' % model.output_type)
metrics = {
k: tf.reduce_mean(v)
for k, v in compiler.metric_tensors.items()
}
LEARNING_RATE = 0.05
KEEP_PROB = 1.0
BATCH_SIZE = batch_size_setting #20
EPOCHS = 6
EMBEDDING_LEARNING_RATE_FACTOR = 0
train_feed_dict = {keep_prob_ph: KEEP_PROB}
loss = tf.reduce_sum(compiler.metric_tensors['all_loss'])
opt = tf.train.AdagradOptimizer(LEARNING_RATE)
grads_and_vars = opt.compute_gradients(loss)
found = 0
for i, (grad, var) in enumerate(grads_and_vars):
if var == word_embedding.weights:
found += 1
grad = tf.scalar_mul(EMBEDDING_LEARNING_RATE_FACTOR, grad)
grads_and_vars[i] = (grad, var)
#assert found == 1 # internal consistency check
train = opt.apply_gradients(grads_and_vars)
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
def train_step(batch):
train_feed_dict[compiler.loom_input_tensor] = batch
_, batch_loss = sess.run([train, loss], train_feed_dict)
return batch_loss
def train_epoch(train_set):
return sum(
train_step(batch)
for batch in td.group_by_batches(train_set, BATCH_SIZE))
train_set = compiler.build_loom_inputs(train_trees)
"""
dev_feed_dict = compiler.build_feed_dict(dev_trees)
def dev_eval(epoch, train_loss):
dev_metrics = sess.run(metrics, dev_feed_dict)
dev_loss = dev_metrics['all_loss']
dev_accuracy = ['%s: %.2f' % (k, v * 100) for k, v in
sorted(dev_metrics.items()) if k.endswith('hits')]
print('epoch:%4d, train_loss: %.3e, dev_loss_avg: %.3e, dev_accuracy:\n [%s]'
% (epoch, train_loss, dev_loss, ' '.join(dev_accuracy)))
return dev_metrics['root_hits']
"""
best_accuracy = 0.0
save_path = os.path.join(data_dir, 'sentiment_model')
loopTime = time.time()
#print('prepare time %s ' % (loopTime - startTime))
loss_save = []
time_save = []
epoch_start_time = loopTime
for epoch, shuffled in enumerate(td.epochs(train_set, EPOCHS), 1):
train_loss = train_epoch(shuffled)
av_loss = train_loss / trainSIZE
epoch_end_time = time.time()
epoch_time = epoch_end_time - epoch_start_time
time_save.append(epoch_time)
epoch_start_time = epoch_end_time
print('train loss is %s at time %s' % (av_loss, epoch_time))
loss_save.append(av_loss)
#accuracy = dev_eval(epoch, train_loss)
#if accuracy > best_accuracy:
# best_accuracy = accuracy
# checkpoint_path = saver.save(sess, save_path, global_step=epoch)
# print('model saved in file: %s' % checkpoint_path)
loopEndTime = time.time()
#print('loop time %s ' % (loopEndTime - loopTime))
prepareTime = loopTime - startTime
loopTime = loopEndTime - loopTime
timePerEpoch = loopTime / EPOCHS
# use median time instead
time_save.sort()
median_time = time_save[int(EPOCHS / 2)]
with open(write_to, "w") as f:
f.write("unit: " + "1 epoch\n")
for loss in loss_save:
f.write(str(loss) + "\n")
f.write("run time: " + str(prepareTime) + " " + str(median_time) +
"\n")
def __init__(self,
config,
kb,
text_seq_batch,
seq_length_batch,
num_vocab_txt,
num_vocab_nmn,
EOS_idx,
num_choices,
decoder_sampling,
use_gt_layout=None,
gt_layout_batch=None,
scope='neural_module_network',
reuse=None):
with tf.variable_scope(scope, reuse=reuse):
# Part 1: Seq2seq RNN to generate module layout tokens
embedding_mat = tf.get_variable(
'embedding_mat', [num_vocab_txt, config.embed_dim_txt],
initializer=tf.contrib.layers.xavier_initializer())
with tf.variable_scope('layout_generation'):
att_seq2seq = netgen_att.AttentionSeq2Seq(
config, text_seq_batch, seq_length_batch, num_vocab_txt,
num_vocab_nmn, EOS_idx, decoder_sampling, embedding_mat,
use_gt_layout, gt_layout_batch)
self.att_seq2seq = att_seq2seq
predicted_tokens = att_seq2seq.predicted_tokens
token_probs = att_seq2seq.token_probs
word_vecs = att_seq2seq.word_vecs
neg_entropy = att_seq2seq.neg_entropy
self.atts = att_seq2seq.atts
self.predicted_tokens = predicted_tokens
self.token_probs = token_probs
self.word_vecs = word_vecs
self.neg_entropy = neg_entropy
# log probability of each generated sequence
self.log_seq_prob = tf.reduce_sum(tf.log(token_probs), axis=0)
# Part 2: Neural Module Network
with tf.variable_scope('layout_execution'):
modules = Modules(config, kb, word_vecs, num_choices,
embedding_mat)
self.modules = modules
# Recursion of modules
att_shape = [len(kb)]
# Forward declaration of module recursion
att_expr_decl = td.ForwardDeclaration(td.PyObjectType(),
td.TensorType(att_shape))
# _key_find
case_key_find = td.Record([
('time_idx', td.Scalar(dtype='int32')),
('batch_idx', td.Scalar(dtype='int32'))
])
case_key_find = case_key_find >> td.ScopedLayer(
modules.KeyFindModule, name_or_scope='KeyFindModule')
# _key_filter
case_key_filter = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))
])
case_key_filter = case_key_filter >> td.ScopedLayer(
modules.KeyFilterModule, name_or_scope='KeyFilterModule')
recursion_cases = td.OneOf(td.GetItem('module'), {
'_key_find': case_key_find,
'_key_filter': case_key_filter
})
att_expr_decl.resolve_to(recursion_cases)
# _val_desc: output scores for choice (for valid expressions)
predicted_scores = td.Record([('input_0', recursion_cases),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))
])
predicted_scores = predicted_scores >> td.ScopedLayer(
modules.ValDescribeModule,
name_or_scope='ValDescribeModule')
# For invalid expressions, define a dummy answer
# so that all answers have the same form
INVALID = assembler.INVALID_EXPR
dummy_scores = td.Void() >> td.FromTensor(
np.zeros(num_choices, np.float32))
output_scores = td.OneOf(td.GetItem('module'), {
'_val_desc': predicted_scores,
INVALID: dummy_scores
})
# compile and get the output scores
self.compiler = td.Compiler.create(output_scores)
self.scores = self.compiler.output_tensors[0]
# Regularization: Entropy + L2
self.entropy_reg = tf.reduce_mean(neg_entropy)
module_weights = [
v for v in tf.trainable_variables()
if (scope in v.op.name and v.op.name.endswith('weights'))
]
self.l2_reg = tf.add_n([tf.nn.l2_loss(v) for v in module_weights])
def __init__(self,
image_data_batch,
image_mean,
text_seq_batch,
seq_length_batch,
T_decoder,
num_vocab_txt,
embed_dim_txt,
num_vocab_nmn,
embed_dim_nmn,
lstm_dim,
num_layers,
assembler,
encoder_dropout,
decoder_dropout,
decoder_sampling,
num_choices,
use_qpn,
qpn_dropout,
reduce_visfeat_dim=False,
new_visfeat_dim=128,
use_gt_layout=None,
gt_layout_batch=None,
map_dim=1024,
scope='neural_module_network',
reuse=None):
with tf.variable_scope(scope, reuse=reuse):
# Part 0: Visual feature from CNN
with tf.variable_scope('image_feature_cnn'):
image_data_batch = image_data_batch / 255.0 - image_mean
image_feat_grid = nlvr_convnet(image_data_batch)
self.image_feat_grid = image_feat_grid
# Part 1: Seq2seq RNN to generate module layout tokensa
with tf.variable_scope('layout_generation'):
att_seq2seq = AttentionSeq2Seq(
text_seq_batch, seq_length_batch, T_decoder, num_vocab_txt,
embed_dim_txt, num_vocab_nmn, embed_dim_nmn, lstm_dim,
num_layers, assembler, encoder_dropout, decoder_dropout,
decoder_sampling, use_gt_layout, gt_layout_batch)
self.att_seq2seq = att_seq2seq
predicted_tokens = att_seq2seq.predicted_tokens
token_probs = att_seq2seq.token_probs
word_vecs = att_seq2seq.word_vecs
neg_entropy = att_seq2seq.neg_entropy
self.atts = att_seq2seq.atts
self.predicted_tokens = predicted_tokens
self.token_probs = token_probs
self.word_vecs = word_vecs
self.neg_entropy = neg_entropy
# log probability of each generated sequence
self.log_seq_prob = tf.reduce_sum(tf.log(token_probs), axis=0)
# Part 2: Neural Module Network
with tf.variable_scope('layout_execution'):
modules = Modules(image_feat_grid, word_vecs, None,
num_choices, map_dim)
self.modules = modules
# Recursion of modules
att_shape = image_feat_grid.get_shape().as_list()[1:-1] + [1]
# Forward declaration of module recursion
att_expr_decl = td.ForwardDeclaration(td.PyObjectType(),
td.TensorType(att_shape))
# _Find
case_find = td.Record([('time_idx', td.Scalar(dtype='int32')),
('batch_idx', td.Scalar(dtype='int32'))
])
case_find = case_find >> td.Function(modules.FindModule)
# _Transform
case_transform = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_transform = case_transform >> td.Function(
modules.TransformModule)
# _And
case_and = td.Record([('input_0', att_expr_decl()),
('input_1', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_and = case_and >> td.Function(modules.AndModule)
# _Describe
case_describe = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_describe = case_describe >> \
td.Function(modules.DescribeModule)
recursion_cases = td.OneOf(
td.GetItem('module'), {
'_Find': case_find,
'_Transform': case_transform,
'_And': case_and
})
att_expr_decl.resolve_to(recursion_cases)
# For invalid expressions, define a dummy answer
# so that all answers have the same form
dummy_scores = td.Void() >> td.FromTensor(
np.zeros(num_choices, np.float32))
output_scores = td.OneOf(td.GetItem('module'), {
'_Describe': case_describe,
INVALID_EXPR: dummy_scores
})
# compile and get the output scores
self.compiler = td.Compiler.create(output_scores)
self.scores_nmn = self.compiler.output_tensors[0]
# Add a question prior network if specified
self.use_qpn = use_qpn
self.qpn_dropout = qpn_dropout
if use_qpn:
self.scores_qpn = question_prior_net(
att_seq2seq.encoder_states, num_choices, qpn_dropout)
self.scores = self.scores_nmn + self.scores_qpn
#self.scores = self.scores_nmn
else:
self.scores = self.scores_nmn
# Regularization: Entropy + L2
self.entropy_reg = tf.reduce_mean(neg_entropy)
#tf.check_numerics(self.entropy_reg, 'entropy NaN/Inf ')
#print(self.entropy_reg.eval())
module_weights = [
v for v in tf.trainable_variables()
if (scope in v.op.name and v.op.name.endswith('weights'))
]
self.l2_reg = tf.add_n([tf.nn.l2_loss(v) for v in module_weights])
def __init__(self,
image_batch,
text_seq_batch,
seq_length_batch,
T_decoder,
num_vocab_txt,
embed_dim_txt,
num_vocab_nmn,
embed_dim_nmn,
lstm_dim,
num_layers,
EOS_idx,
encoder_dropout,
decoder_dropout,
decoder_sampling,
num_choices,
use_gt_layout=None,
gt_layout_batch=None,
scope='neural_module_network',
reuse=None):
with tf.variable_scope(scope, reuse=reuse):
# Part 0: Visual feature from CNN
with tf.variable_scope('image_feature_cnn'):
image_feat_grid = shapes_convnet(image_batch)
self.image_feat_grid = image_feat_grid
# Part 1: Seq2seq RNN to generate module layout tokens
with tf.variable_scope('layout_generation'):
att_seq2seq = nmn3_netgen_att.AttentionSeq2Seq(
text_seq_batch, seq_length_batch, T_decoder, num_vocab_txt,
embed_dim_txt, num_vocab_nmn, embed_dim_nmn, lstm_dim,
num_layers, EOS_idx, encoder_dropout, decoder_dropout,
decoder_sampling, use_gt_layout, gt_layout_batch)
self.att_seq2seq = att_seq2seq
predicted_tokens = att_seq2seq.predicted_tokens
token_probs = att_seq2seq.token_probs
word_vecs = att_seq2seq.word_vecs
neg_entropy = att_seq2seq.neg_entropy
self.atts = att_seq2seq.atts
self.predicted_tokens = predicted_tokens
self.token_probs = token_probs
self.word_vecs = word_vecs
self.neg_entropy = neg_entropy
# log probability of each generated sequence
self.log_seq_prob = tf.reduce_sum(tf.log(token_probs), axis=0)
# Part 2: Neural Module Network
with tf.variable_scope('layout_execution'):
modules = Modules(image_feat_grid, word_vecs, num_choices)
self.modules = modules
# Recursion of modules
att_shape = image_feat_grid.get_shape().as_list()[1:-1] + [1]
# Forward declaration of module recursion
att_expr_decl = td.ForwardDeclaration(td.PyObjectType(),
td.TensorType(att_shape))
# _Find
case_find = td.Record([('time_idx', td.Scalar(dtype='int32')),
('batch_idx', td.Scalar(dtype='int32'))
])
case_find = case_find >> \
td.ScopedLayer(modules.FindModule, name_or_scope='FindModule')
# _Transform
case_transform = td.Record([('input_0', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_transform = case_transform >> \
td.ScopedLayer(modules.TransformModule, name_or_scope='TransformModule')
# _And
case_and = td.Record([('input_0', att_expr_decl()),
('input_1', att_expr_decl()),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))])
case_and = case_and >> \
td.ScopedLayer(modules.AndModule, name_or_scope='AndModule')
recursion_cases = td.OneOf(
td.GetItem('module'), {
'_Find': case_find,
'_Transform': case_transform,
'_And': case_and
})
att_expr_decl.resolve_to(recursion_cases)
# _Answer: output scores for choice (for valid expressions)
predicted_scores = td.Record([('input_0', recursion_cases),
('time_idx', td.Scalar('int32')),
('batch_idx', td.Scalar('int32'))
])
predicted_scores = predicted_scores >> \
td.ScopedLayer(modules.AnswerModule, name_or_scope='AnswerModule')
# For invalid expressions, define a dummy answer
# so that all answers have the same form
INVALID = nmn3_assembler.INVALID_EXPR
dummy_scores = td.Void() >> td.FromTensor(
np.zeros(num_choices, np.float32))
output_scores = td.OneOf(td.GetItem('module'), {
'_Answer': predicted_scores,
INVALID: dummy_scores
})
# compile and get the output scores
self.compiler = td.Compiler.create(output_scores)
self.scores = self.compiler.output_tensors[0]
# Regularization: Entropy + L2
self.entropy_reg = tf.reduce_mean(neg_entropy)
module_weights = [
v for v in tf.trainable_variables()
if (scope in v.op.name and v.op.name.endswith('weights'))
]
self.l2_reg = tf.add_n([tf.nn.l2_loss(v) for v in module_weights])
