def composed_embed_blk():
leaf_case = direct_embed_blk()
nonleaf_case = td.Composition(name='composed_embed_nonleaf')
with nonleaf_case.scope():
children = td.GetItem('children').reads(nonleaf_case.input)
clen = td.Scalar().reads(td.GetItem('clen').reads(nonleaf_case.input))
cclens = td.Map(td.GetItem('clen') >> td.Scalar()).reads(children)
fchildren = td.Map(direct_embed_blk()).reads(children)
initial_state = td.Composition()
with initial_state.scope():
initial_state.output.reads(
td.FromTensor(tf.zeros(hyper.word_dim)),
td.FromTensor(tf.zeros([])),
)
summed = td.Zip().reads(fchildren, cclens, td.Broadcast().reads(clen))
summed = td.Fold(continous_weighted_add_blk(),
initial_state).reads(summed)[0]
added = td.Function(tf.add, name='add_bias').reads(
summed, td.FromTensor(param.get('B')))
normed = clip_by_norm_blk().reads(added)
act_fn = tf.nn.relu if hyper.use_relu else tf.nn.tanh
relu = td.Function(act_fn).reads(normed)
nonleaf_case.output.reads(relu)
return td.OneOf(lambda node: node['clen'] == 0, {
True: leaf_case,
False: nonleaf_case
})
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))
def logits_and_state(self):
"""Creates a block that goes from tokens to (logits, state) tuples."""
unknown_idx = len(self.vocab)
def lookup_word(word):
return self.vocab.get(word, unknown_idx)
#(GetItem(key) >> block).eval(inp) => block.eval(inp[key])
# InputTransform(funk): A Python function, lifted to a block.
# Scalar - input to scalar
word2vec = (td.GetItem(0) >> td.InputTransform(lookup_word) >>
td.Scalar('int32') >> self.word_embedding)
#
pair2vec = (self.embed_subtree(), self.embed_subtree())
# Trees are binary, so the tree layer takes two states as its
# input_state.
zero_state = td.Zeros((self.tree_lstm_cell.state_size, ) * 2)
# Input is a word vector.
zero_inp = td.Zeros(self.word_embedding.output_type.shape[0])
# AllOf(a, b, c).eval(inp) => (a.eval(inp), b.eval(inp), c.eval(inp))
word_case = td.AllOf(word2vec, zero_state)
pair_case = td.AllOf(zero_inp, pair2vec)
# OneOf(func, [(key, block),(key,block)])) where funk(input) => key and
# OneOf returns one of blocks
tree2vec = td.OneOf(len, [(1, word_case), (2, pair_case)])
return tree2vec >> self.tree_lstm_cell
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 l2loss_blk():
# rewrite using metric
leaf_case = td.Composition()
with leaf_case.scope():
leaf_case.output.reads(td.FromTensor(tf.constant(1.)))
nonleaf_case = td.Composition()
with nonleaf_case.scope():
direct = direct_embed_blk().reads(nonleaf_case.input)
com = composed_embed_blk().reads(nonleaf_case.input)
loss = td.Function(batch_nn_l2loss).reads(direct, com)
nonleaf_case.output.reads(loss)
return td.OneOf(lambda node: node['clen'] != 0, {
False: leaf_case,
True: nonleaf_case
})
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 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 is the default return value
word2vec = (
td.GetItem(0) >> td.GetItem(0) >> td.InputTransform(lookup_word) >>
td.Scalar('int32') >> word_embedding
) # <td.Pipe>: None -> TensorType((200,), 'float32')
context2vec1 = td.GetItem(1) >> td.InputTransform(
makeContextMat) >> td.Vector(10)
context2vec2 = td.GetItem(1) >> td.InputTransform(
makeContextMat) >> td.Vector(10)
ent1posit1 = td.GetItem(2) >> td.InputTransform(
makeEntPositMat) >> td.Vector(10)
ent1posit2 = td.GetItem(2) >> td.InputTransform(
makeEntPositMat) >> td.Vector(10)
ent2posit1 = td.GetItem(3) >> td.InputTransform(
makeEntPositMat) >> td.Vector(10)
ent2posit2 = td.GetItem(3) >> td.InputTransform(
makeEntPositMat) >> td.Vector(10)
pairs2vec = td.GetItem(0) >> (embed_subtree(), embed_subtree())
# our binary Tree can have two child nodes, therefore, we assume the zero state have two child nodes.
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_embedding.output_type.shape[0] == 200
word_case = td.AllOf(word2vec, zero_state, context2vec1, ent1posit1,
ent2posit1)
children_case = td.AllOf(zero_inp, pairs2vec, context2vec2, ent1posit2,
ent2posit2)
# if leaf case, go to word case...
tree2vec = td.OneOf(lambda x: 1
if len(x[0]) == 1 else 2, [(1, word_case),
(2, children_case)])
# tree2vec = td.OneOf(lambda pair: len(pair[0]), [(1, word_case), (2, children_case)])
# logits and lstm states
return tree2vec >> tree_lstm >> (output_layer, td.Identity())
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 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])
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])
