def build_decoder_block_for_analysis(z_size, token_emb_size, decoder_cell, input_size): c = td.Composition() c.set_input_type( td.TupleType(td.TensorType((z_size, )), td.SequenceType(td.TensorType((input_size, ))))) with c.scope(): hidden_state = td.GetItem(0).reads(c.input) rnn_input = td.GetItem(1).reads(c.input) # decoder_output = build_program_decoder_for_analysis( # token_emb_size, default_gru_cell(z_size) # ) decoder_output = decoder_cell decoder_output.reads(rnn_input, hidden_state) decoder_rnn_output = td.GetItem(1).reads(decoder_output) un_normalised_token_probs = td.GetItem(0).reads(decoder_output) # get the first output (meant to only compute one interation) c.output.reads( td.GetItem(0).reads(un_normalised_token_probs), td.GetItem(0).reads(decoder_rnn_output)) return td.Record((td.Vector(z_size), td.Map(td.Vector(input_size)))) >> c
def feature_detector_blk(max_depth=2): """Input: node dict Output: TensorType([hyper.conv_dim, ]) Single patch of the conv. Depth is max_depth """ blk = td.Composition() with blk.scope(): nodes_in_patch = collect_node_for_conv_patch_blk( max_depth=max_depth).reads(blk.input) # map from python object to tensors mapped = td.Map( td.Record((coding_blk(), td.Scalar(), td.Scalar(), td.Scalar(), td.Scalar()))).reads(nodes_in_patch) # mapped = [(feature, idx, depth, max_depth), (...)] # compute weighted feature for each elem weighted = td.Map(weighted_feature_blk()).reads(mapped) # weighted = [fea, fea, fea, ...] # add together added = td.Reduce(td.Function(tf.add)).reads(weighted) # added = TensorType([hyper.conv_dim, ]) # add bias biased = td.Function(tf.add).reads(added, td.FromTensor(param.get('Bconv'))) # biased = TensorType([hyper.conv_dim, ]) # tanh tanh = td.Function(tf.nn.tanh).reads(biased) # tanh = TensorType([hyper.conv_dim, ]) blk.output.reads(tanh) return blk
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 create_compiler(file_queue): expr_left_sentence, expr_right_sentence = buid_sentence_expression(), buid_sentence_expression() expr_label = td.InputTransform(lambda label: int(label)) >> td.OneHot(2, dtype=tf.float32) one_record = td.InputTransform(lambda record: json.loads(record.decode('utf-8'))) >> \ td.Record((expr_left_sentence, expr_right_sentence, expr_label), name='instance') #td.AllOf(td.slice(start=0, stop=2) >> td.Concat() >> td.FC(2), td.Slice(start=2, stop=3)) batch = make_batch(file_queue) compiler = td.Compiler().create(one_record, input_tensor=batch) return compiler
def create_compiler(): expr_left_sentence, expr_right_sentence = buid_sentence_expression( ), buid_sentence_expression() expr_label = td.InputTransform(lambda label: int(label)) >> td.OneHot( 2, dtype=tf.float32) id = td.Scalar(dtype=tf.int32) one_record = td.InputTransform(lambda record: json.loads(record)) >> \ td.Record((expr_left_sentence, expr_right_sentence, expr_label, id), name='instance') compiler = td.Compiler().create(one_record) return compiler
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, 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 _compile(self): with self.sess.as_default(): import tensorflow_fold as td output_size = len(self.labels) self.keep_prob = tf.placeholder_with_default(tf.constant(1.0),shape=None) char_emb = td.Embedding(num_buckets=self.char_buckets, num_units_out=self.embedding_size) #initializer=tf.truncated_normal_initializer(stddev=0.15)) char_cell = td.ScopedLayer(tf.contrib.rnn.LSTMCell(num_units=self.rnn_dim), 'char_cell') char_lstm = (td.InputTransform(lambda s: [ord(c) for c in s]) >> td.Map(td.Scalar('int32') >> char_emb) >> td.RNN(char_cell) >> td.GetItem(1) >> td.GetItem(1)) rnn_fwdcell = td.ScopedLayer(tf.contrib.rnn.LSTMCell(num_units=self.rnn_dim), 'lstm_fwd') fwdlayer = td.RNN(rnn_fwdcell) >> td.GetItem(0) rnn_bwdcell = td.ScopedLayer(tf.contrib.rnn.LSTMCell(num_units=self.rnn_dim), 'lstm_bwd') bwdlayer = (td.Slice(step=-1) >> td.RNN(rnn_bwdcell) >> td.GetItem(0) >> td.Slice(step=-1)) pos_emb = td.Embedding(num_buckets=300, num_units_out=32, initializer=tf.truncated_normal_initializer(stddev=0.1)) pos_x = (td.InputTransform(lambda x: x + 150) >> td.Scalar(dtype='int32') >> pos_emb) pos_y = (td.InputTransform(lambda x: x + 150) >> td.Scalar(dtype='int32') >> pos_emb) input_layer = td.Map(td.Record((char_lstm,pos_x,pos_y)) >> td.Concat()) maxlayer = (td.AllOf(fwdlayer, bwdlayer) >> td.ZipWith(td.Concat()) >> td.Max()) output_layer = (input_layer >> maxlayer >> td.FC(output_size, input_keep_prob=self.keep_prob, activation=None)) self.compiler = td.Compiler.create((output_layer, td.Vector(output_size,dtype=tf.int32))) self.y_out, self.y_true = self.compiler.output_tensors self.y_loss = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits( logits=self.y_out,labels=self.y_true)) self.y_prob = tf.nn.softmax(self.y_out) self.y_true_idx = tf.argmax(self.y_true,axis=1) self.y_pred_idx = tf.argmax(self.y_prob,axis=1) self.y_pred = tf.one_hot(self.y_pred_idx,depth=output_size,dtype=tf.int32) epoch_step = tf.Variable(0, trainable=False) self.epoch_step_op = tf.assign(epoch_step, epoch_step+1) lrate_decay = tf.train.exponential_decay(self.lrate, epoch_step, 1, self.decay) if self.optimizer == 'adam': self.opt = tf.train.AdamOptimizer(learning_rate=lrate_decay) elif self.optimizer == 'adagrad': self.opt = tf.train.AdagradOptimizer(learning_rate=lrate_decay, initial_accumulator_value=1e-08) elif self.optimizer == 'rmsprop' or self.optimizer == 'default': self.opt = tf.train.RMSPropOptimizer(learning_rate=lrate_decay, epsilon=1e-08) else: raise Exception(('The optimizer {} is not in list of available ' + 'optimizers: default, adam, adagrad, rmsprop.') .format(self.optimizer)) # apply learning multiplier on on embedding learning rate embeds = [pos_emb.weights, char_emb.weights] grads_and_vars = self.opt.compute_gradients(self.y_loss) found = 0 for i, (grad, var) in enumerate(grads_and_vars): if var in embeds: found += 1 grad = tf.scalar_mul(self.embedding_factor, grad) grads_and_vars[i] = (grad, var) assert found == len(embeds) # internal consistency check self.train_step = self.opt.apply_gradients(grads_and_vars) self.sess.run(tf.global_variables_initializer()) self.saver = tf.train.Saver(max_to_keep=100)
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])
forward_dir = (td.RNN(fw_cell) >> td.GetItem(0)).reads(fw_seq) back_dir = (td.RNN(bw_cell) >> td.GetItem(0)).reads(bw_seq) back_to_leftright = td.Slice(step=-1).reads(back_dir) output_transform = td.FC(1, activation=None) bidir_common = (td.ZipWith( td.Concat() >> output_transform >> td.Metric('logits'))).reads( forward_dir, back_to_leftright) bidir_conv_lstm.output.reads(bidir_common) return bidir_conv_lstm CONV_data = td.Record((td.Map( td.Vector(vsize) >> td.Function(lambda x: tf.reshape(x, [-1, vsize, 1]))), td.Map(td.Scalar()))) CONV_model = (CONV_data >> bidirectional_dynamic_CONV( multi_convLSTM_cell([vsize, vsize, vsize], [100, 100, 100]), multi_convLSTM_cell([vsize, vsize, vsize], [100, 100, 100])) >> td.Void()) FC_data = td.Record((td.Map(td.Vector(vsize)), td.Map(td.Scalar()))) FC_model = (FC_data >> bidirectional_dynamic_FC(multi_FC_cell( [1000] * 5), multi_FC_cell([1000] * 5), 1000) >> td.Void()) store = data(FLAGS.data_dir + FLAGS.data_type, FLAGS.truncate) if FLAGS.model == "lstm": model = FC_model elif FLAGS.model == "convlstm": model = CONV_model
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])