def step(prev_state_c, prev_state_h, tokens, seq_length,
stop_indicator):
input = tf.gather(tokens, tf.shape(tokens)[0] - 1)
# Look for new finish dialogue
is_stop_token = tf.equal(input, stop_token)
is_stop_dialogue_token = tf.equal(input, stop_dialogue_token)
is_stop = tf.logical_or(is_stop_token, is_stop_dialogue_token)
stop_indicator = tf.logical_or(
stop_indicator, is_stop) # flag to false new finished dialogue
# increment seq_length when the dialogue is not over
seq_length = tf.where(stop_indicator, seq_length,
tf.add(seq_length, 1))
# compute the next words. TODO: factorize with qgen.. but how?!
with tf.variable_scope(self.scope_name, reuse=True):
word_emb = utils.get_embedding(
input,
n_words=tokenizer.no_words,
n_dim=config['word_embedding_size'],
scope="word_embedding",
reuse=True)
inp_emb = tf.concat([word_emb, self.image_emb], axis=1)
with tf.variable_scope("word_decoder"):
lstm_cell = tf.contrib.rnn.LayerNormBasicLSTMCell(
config['num_lstm_units'],
layer_norm=False,
dropout_keep_prob=1.0,
reuse=True)
state = tf.contrib.rnn.LSTMStateTuple(c=prev_state_c,
h=prev_state_h)
out, state = lstm_cell(inp_emb, state)
# store/update the state when the dialogue is not finished (after sampling the <?> token)
cond = tf.greater_equal(
seq_length, tf.subtract(tf.reduce_max(seq_length), 1))
state_c = tf.where(cond, state.c, prev_state_c)
state_h = tf.where(cond, state.h, prev_state_h)
with tf.variable_scope('decoder_output'):
output = utils.fully_connected(state_h,
tokenizer.no_words,
reuse=True)
sampled_tokens = tf.cond(
self.greedy, lambda: tf.argmax(output, 1),
lambda: tf.reshape(tf.multinomial(output, 1), [-1]))
sampled_tokens = tf.cast(sampled_tokens, tf.int32)
tokens = tf.concat(
[tokens, tf.expand_dims(sampled_tokens, 0)],
axis=0) # check axis!
return state_c, state_h, tokens, seq_length, stop_indicator
def compute_attention(feature_maps, context, no_mlp_units, reuse=False):
with tf.variable_scope("attention"):
if len(feature_maps.get_shape()) == 3:
h = tf.shape(feature_maps)[
1] # when the shape is dynamic (attention over lstm)
w = 1
c = int(feature_maps.get_shape()[2])
else:
h = int(feature_maps.get_shape()[1])
w = int(feature_maps.get_shape()[2])
c = int(feature_maps.get_shape()[3])
s = int(context.get_shape()[1])
feature_maps = tf.reshape(feature_maps, shape=[-1, h * w, c])
context = tf.expand_dims(context, axis=1)
context = tf.tile(context, [1, h * w, 1])
embedding = tf.concat([feature_maps, context], axis=2)
embedding = tf.reshape(embedding, shape=[-1, s + c])
# compute the evidence from the embedding
with tf.variable_scope("mlp"):
e = utils.fully_connected(embedding,
no_mlp_units,
scope='hidden_layer',
activation="relu",
reuse=reuse)
e = utils.fully_connected(e, 1, scope='out', reuse=reuse)
e = tf.reshape(e, shape=[-1, h * w, 1])
# compute the softmax over the evidence
alpha = tf.nn.softmax(e, dim=1)
# apply soft attention
soft_attention = feature_maps * alpha
soft_attention = tf.reduce_sum(soft_attention, axis=1)
return soft_attention
def create_cbn_input(self, feature_maps):
no_features = int(feature_maps.get_shape()[3])
batch_size = tf.shape(feature_maps)[0]
if self.use_betas:
h_betas = utils.fully_connected(self.lstm_state,
self.cbn_embedding_size,
weight_initializer=RandomUniform(
-1e-4, 1e-4),
scope="hidden_betas",
activation='relu')
delta_betas = utils.fully_connected(
h_betas,
no_features,
scope="delta_beta",
weight_initializer=RandomUniform(-1e-4, 1e-4),
use_bias=False)
else:
delta_betas = tf.tile(tf.constant(0.0, shape=[1, no_features]),
tf.stack([batch_size, 1]))
if self.use_gammas:
h_gammas = utils.fully_connected(self.lstm_state,
self.cbn_embedding_size,
weight_initializer=RandomUniform(
-1e-4, 1e-4),
scope="hidden_gammas",
activation='relu')
delta_gammas = utils.fully_connected(
h_gammas,
no_features,
scope="delta_gamma",
weight_initializer=RandomUniform(-1e-4, 1e-4))
else:
delta_gammas = tf.tile(tf.constant(0.0, shape=[1, no_features]),
tf.stack([batch_size, 1]))
return delta_betas, delta_gammas
def __init__(self, config, num_words, policy_gradient, device='', reuse=False):
AbstractNetwork.__init__(self, "qgen", device=device)
# Create the scope for this graph
with tf.variable_scope(self.scope_name, reuse=reuse):
mini_batch_size = None
# Picture
self.images = tf.placeholder(tf.float32, [mini_batch_size] + config['image']["dim"], name='images')
# Question
self.dialogues = tf.placeholder(tf.int32, [mini_batch_size, None], name='dialogues')
self.answer_mask = tf.placeholder(tf.float32, [mini_batch_size, None], name='answer_mask') # 1 if keep and (1 q/a 1) for (START q/a STOP)
self.padding_mask = tf.placeholder(tf.float32, [mini_batch_size, None], name='padding_mask')
self.seq_length = tf.placeholder(tf.int32, [mini_batch_size], name='seq_length')
# Rewards
self.cum_rewards = tf.placeholder(tf.float32, shape=[mini_batch_size, None], name='cum_reward')
# DECODER Hidden state (for beam search)
zero_state = tf.zeros([1, config['num_lstm_units']]) # default LSTM state is a zero-vector
zero_state = tf.tile(zero_state, [tf.shape(self.images)[0], 1]) # trick to do a dynamic size 0 tensors
self.decoder_zero_state_c = tf.placeholder_with_default(zero_state, [mini_batch_size, config['num_lstm_units']], name="state_c")
self.decoder_zero_state_h = tf.placeholder_with_default(zero_state, [mini_batch_size, config['num_lstm_units']], name="state_h")
decoder_initial_state = tf.contrib.rnn.LSTMStateTuple(c=self.decoder_zero_state_c, h=self.decoder_zero_state_h)
# Misc
self.is_training = tf.placeholder(tf.bool, name='is_training')
self.greedy = tf.placeholder_with_default(False, shape=(), name="greedy") # use for graph
self.samples = None
# remove last token
input_dialogues = self.dialogues[:, :-1]
input_seq_length = self.seq_length - 1
# remove first token(=start token)
rewards = self.cum_rewards[:, 1:]
target_words = self.dialogues[:, 1:]
# to understand the padding:
# input
# <start> is it a blue <?> <yes> is it a car <?> <no> <stop_dialogue>
# target
# is it a blue <?> - is it a car <?> - <stop_dialogue> -
# image processing
if len(config["image"]["dim"]) == 1:
self.image_out = self.images
else:
self.image_out = get_attention(self.images, None, "none") #TODO: improve by using the previous lstm state?
# Reduce the embedding size of the image
with tf.variable_scope('picture_embedding'):
self.picture_emb = utils.fully_connected(self.image_out,
config['picture_embedding_size'])
picture_emb = tf.expand_dims(self.picture_emb, 1)
picture_emb = tf.tile(picture_emb, [1, tf.shape(input_dialogues)[1], 1])
# Compute the question embedding
input_words = utils.get_embedding(
input_dialogues,
n_words=num_words,
n_dim=config['word_embedding_size'],
scope="word_embedding")
# concat word embedding and picture embedding
decoder_input = tf.concat([input_words, picture_emb], axis=2, name="concat_full_embedding")
# encode one word+picture
decoder_lstm_cell = tf.contrib.rnn.LayerNormBasicLSTMCell(
config['num_lstm_units'],
layer_norm=False,
dropout_keep_prob=1.0,
reuse=reuse)
self.decoder_output, self.decoder_state = tf.nn.dynamic_rnn(
cell=decoder_lstm_cell,
inputs=decoder_input,
dtype=tf.float32,
initial_state=decoder_initial_state,
sequence_length=input_seq_length,
scope="word_decoder") # TODO: use multi-layer RNN
max_sequence = tf.reduce_max(self.seq_length)
# compute the softmax for evaluation
with tf.variable_scope('decoder_output'):
flat_decoder_output = tf.reshape(self.decoder_output, [-1, decoder_lstm_cell.output_size])
flat_mlp_output = utils.fully_connected(flat_decoder_output, num_words)
# retrieve the batch/dialogue format
mlp_output = tf.reshape(flat_mlp_output, [tf.shape(self.seq_length)[0], max_sequence - 1, num_words]) # Ignore th STOP token
self.softmax_output = tf.nn.softmax(mlp_output, name="softmax")
self.argmax_output = tf.argmax(mlp_output, axis=2)
self.cross_entropy_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=mlp_output, labels=target_words)
# compute the maximum likelihood loss
with tf.variable_scope('ml_loss'):
ml_loss = tf.identity(self.cross_entropy_loss)
ml_loss *= self.answer_mask[:, 1:] # remove answers (ignore the <stop> token)
ml_loss *= self.padding_mask[:, 1:] # remove padding (ignore the <start> token)
# Count number of unmask elements
count = tf.reduce_sum(self.padding_mask) - tf.reduce_sum(1 - self.answer_mask[:, :-1]) - 1 # no_unpad - no_qa - START token
ml_loss = tf.reduce_sum(ml_loss, axis=1) # reduce over dialogue dimension
ml_loss = tf.reduce_sum(ml_loss, axis=0) # reduce over minibatch dimension
self.ml_loss = ml_loss / count # Normalize
self.loss = self.ml_loss
# Compute policy gradient
if policy_gradient:
with tf.variable_scope('rl_baseline'):
decoder_out = tf.stop_gradient(self.decoder_output) # take the LSTM output (and stop the gradient!)
flat_decoder_output = tf.reshape(decoder_out, [-1, decoder_lstm_cell.output_size]) #
flat_h1 = utils.fully_connected(flat_decoder_output, n_out=100, activation='relu', scope='baseline_hidden')
flat_baseline = utils.fully_connected(flat_h1, 1, activation='relu', scope='baseline_out')
self.baseline = tf.reshape(flat_baseline, [tf.shape(self.seq_length)[0], max_sequence-1])
self.baseline *= self.answer_mask[:, 1:]
self.baseline *= self.padding_mask[:, 1:]
with tf.variable_scope('policy_gradient_loss'):
# Compute log_prob
self.log_of_policy = tf.identity(self.cross_entropy_loss)
self.log_of_policy *= self.answer_mask[:, 1:] # remove answers (<=> predicted answer has maximum reward) (ignore the START token in the mask)
# No need to use padding mask as the discounted_reward is already zero once the episode terminated
# Policy gradient loss
rewards *= self.answer_mask[:, 1:]
self.score_function = tf.multiply(self.log_of_policy, rewards - self.baseline) # score function
self.baseline_loss = tf.reduce_sum(tf.square(rewards - self.baseline))
self.policy_gradient_loss = tf.reduce_sum(self.score_function, axis=1) # sum over the dialogue trajectory
self.policy_gradient_loss = tf.reduce_mean(self.policy_gradient_loss, axis=0) # reduce over minibatch dimension
self.loss = self.policy_gradient_loss
def compute_glimpse(feature_maps,
context,
no_glimpse,
glimpse_embedding_size,
keep_dropout,
reuse=False):
with tf.variable_scope("glimpse"):
h = int(feature_maps.get_shape()[1])
w = int(feature_maps.get_shape()[2])
c = int(feature_maps.get_shape()[3])
# reshape state to perform batch operation
context = tf.nn.dropout(context, keep_dropout)
projected_context = utils.fully_connected(context,
glimpse_embedding_size,
scope='hidden_layer',
activation="tanh",
use_bias=False,
reuse=reuse)
projected_context = tf.expand_dims(projected_context, axis=1)
projected_context = tf.tile(projected_context, [1, h * w, 1])
projected_context = tf.reshape(projected_context,
[-1, glimpse_embedding_size])
feature_maps = tf.reshape(feature_maps, shape=[-1, h * w, c])
glimpses = []
with tf.variable_scope("glimpse"):
g_feature_maps = tf.reshape(
feature_maps,
shape=[-1, c]) # linearise the feature map as as single batch
g_feature_maps = tf.nn.dropout(g_feature_maps, keep_dropout)
g_feature_maps = utils.fully_connected(g_feature_maps,
glimpse_embedding_size,
scope='image_projection',
activation="tanh",
use_bias=False,
reuse=reuse)
hadamard = g_feature_maps * projected_context
hadamard = tf.nn.dropout(hadamard, keep_dropout)
e = utils.fully_connected(hadamard,
no_glimpse,
scope='hadamard_projection',
reuse=reuse)
e = tf.reshape(e, shape=[-1, h * w, no_glimpse])
for i in range(no_glimpse):
ev = e[:, :, i]
alpha = tf.nn.softmax(ev)
# apply soft attention
soft_glimpses = feature_maps * tf.expand_dims(alpha, -1)
soft_glimpses = tf.reduce_sum(soft_glimpses, axis=1)
glimpses.append(soft_glimpses)
full_glimpse = tf.concat(glimpses, axis=1)
return full_glimpse
def compute_all_attention(question_states,
caption,
history_states,
image_feature,
no_mlp_units,
reuse=False,
config=None):
print("image_feature = {}", image_feature)
print("Question = {}", question_states)
print("caption = {}", caption)
print("history_sta = {}", history_states)
##### 1 ####
# recupere les données
#### 2 ####
#step_dictionnaire
##### loop step_dictionnaire ########
# feature_input = mlp(x,g1,g2) [1.. feature_shape]
# soft_feature = softmax(feature_input) [1.. feature_input_shape]
# x = soft_feature * x [1.. feature_input_shape]
with tf.variable_scope("coattention"):
if image_feature != None:
if len(image_feature.get_shape()) == 3:
h = tf.shape(image_feature)[
1] # when the shape is dynamic (attention over lstm)
w = 1
c = int(image_feature.get_shape()[2])
else:
h = int(image_feature.get_shape()[1])
w = int(image_feature.get_shape()[2])
c = int(image_feature.get_shape()[3])
s = int(question_states.get_shape()[2])
image_feature = tf.reshape(image_feature, shape=[
-1, (h * w), c
]) # input image_feature ?,7,7,2048 => ?,49,2048
print("******************** B Image_feature = {} ".format(
image_feature))
image_feature = tf.reduce_sum(image_feature, axis=1)
print("******************** A Image_feature = {} ".format(
image_feature))
set_img(image_feature)
question_shape = question_states.get_shape()
question_states = tf.reshape(
question_states,
shape=[-1, int(question_shape[1]) * int(question_shape[2])])
set_question(question_states)
if history_states != None:
if caption != None:
caption = tf.expand_dims(caption, axis=1)
if caption != None:
print("caption = {} ,history_states = {} ".format(
caption, history_states))
history_states = tf.reshape(history_states, [-1, 6, 1024])
hist = tf.concat([caption, history_states], axis=1)
print("hist B= {} ".format(hist))
hist = tf.reshape(history_states, [-1, 7 * 1024])
else:
# hist = tf.reshape(history_states,[-1,6*1024])
hist = history_states
# print("hist = {} ".format(hist))
# exit()
set_history(hist)
# return question_states,hist,image_feature
dict_step = {0: "img", 1: "question", 3: "hist"}
step_attention = {
0: [0, 1, None],
1: [3, 0, 1],
2: [1, 0, 3],
3: [0, 3, 1]
}
# step_attention = {0:[0,1,None]}
for key, value in step_attention.items():
input_data, g1, g2 = get_input_g1_g2(value[0], value[1], value[2])
dimension_two = int(input_data.get_shape()[1])
print("---- input_shape = {} ".format(input_data.get_shape()))
print("---- g1_shape = {} ".format(g1.get_shape()))
# print("g2_shape = {} ".format(g2.get_shape()))
hidden_mlp, weight = utils.fully_connected(
input_data,
no_mlp_units,
scope='hidden_layer_256_{}'.format(key),
activation="tanh",
reuse=reuse,
co_attention=True,
g1=g1,
g2=g2,
key_input=key)
hidden_mlp = utils.fully_connected(
hidden_mlp,
1,
scope='hidden_layer_1_{}'.format(key),
reuse=reuse,
co_attention=False)
alpha = tf.nn.softmax(hidden_mlp, axis=1)
input_data = input_data * alpha
if value[0] == 0: set_img(input_data)
elif value[0] == 1: set_question(input_data)
elif value[0] == 2: set_history(input_data)
# print("-- {} -- hidden = {} ,ALPHA= {} ,INPUT_DATA = {}".format(key,hidden_mlp,alpha,input_data))
# print("Data_ouput = ",get_img(),get_question(),get_history())
# if key == 2:
# exit()
# img_shape = get_img().get_shape()
# question_shape = get_question().get_shape()
# history_shape = get_history().get_shape()
# img = tf.reshape(get_img(), shape=[-1,int(img_shape[1]) * int(img_shape[2]) ])
# question = tf.reshape(get_question(), shape=[-1, int(question_shape[1]) * int(question_shape[2]) ])
# history = tf.reshape(get_history(), shape=[-1,int(history_shape[1]) * int(history_shape[2]) ])
question_states = get_question()
history = get_history()
image_feature = get_img()
# print(" history = {} , image_feature = {} , question = {}".format(history,image_feature,question_states))
# exit()
return question_states, history, image_feature
def compute_attention(feature_maps, context, no_mlp_units, reuse=False):
with tf.variable_scope("attention"):
if len(feature_maps.get_shape()) == 3:
h = tf.shape(feature_maps)[
1] # when the shape is dynamic (attention over lstm)
w = 1
c = int(feature_maps.get_shape()[2])
else:
h = int(feature_maps.get_shape()[1])
w = int(feature_maps.get_shape()[2])
c = int(feature_maps.get_shape()[3])
s = int(context.get_shape()[1])
feature_maps = tf.reshape(
feature_maps, shape=[-1, h * w, c]
) #Tensor("oracle/attention/Reshape:0", shape=(?, 49, 2048), dtype=float32)
context = tf.expand_dims(
context, axis=1
) # Tensor("oracle/attention/ExpandDims:0", shape=(?, 1, 6144), dtype=float32)
context = tf.tile(
context,
[1, h * w, 1]) # tf.tile([a,b,c],dimension=[2]) => [a,b,c,a,b,c]
# Tensor("oracle/attention/Tile:0", shape=(?, 49, 6144), dtype=float32)
embedding = tf.concat([feature_maps, context], axis=2)
embedding = tf.reshape(
embedding, shape=[-1, s + c]
) #Tensor("oracle/attention/Reshape_1:0", shape=(?, 8192), dtype=float32)
# compute the evidence from the embedding
with tf.variable_scope("mlp"):
e = utils.fully_connected(
embedding,
no_mlp_units,
scope='hidden_layer',
activation="relu",
reuse=reuse
) #Tensor("oracle/attention/mlp/Relu:0", shape=(?, 256), dtype=float32)
# print(" Before E = {}".format(e))
e = utils.fully_connected(
e, 1, scope='out', reuse=reuse
) # Tensor("oracle/attention/mlp/out/add:0", shape=(?, 1), dtype=float32)
# print(" After E = {}".format(e))
e = tf.reshape(
e, shape=[-1, h * w, 1]
) #Tensor("oracle/attention/Reshape_2:0", shape=(?, 49, 1), dtype=float32)
# compute the softmax over the evidence
alpha = tf.nn.softmax(
e, dim=1
) # Tensor("oracle/attention/transpose_1:0", shape=(?, 49, 1), dtype=float32)
# apply soft attention
soft_attention = feature_maps * alpha # Tensor("oracle/attention/mul:0", shape=(?, 49, 2048), dtype=float32)
soft_attention = tf.reduce_sum(
soft_attention, axis=1
) # Tensor("oracle/attention/Sum:0", shape=(?, 2048), dtype=float32)
return soft_attention
def __init__(self, config, num_words, device='', reuse=False):
AbstractNetwork.__init__(self, "guesser", device=device)
mini_batch_size = None
with tf.variable_scope(self.scope_name, reuse=reuse):
# Dialogues
self.dialogues = tf.placeholder(tf.int32, [mini_batch_size, None],
name='dialogues')
self.seq_length = tf.placeholder(tf.int32, [mini_batch_size],
name='seq_length')
# Objects
self.obj_mask = tf.placeholder(tf.float32, [mini_batch_size, None],
name='obj_mask')
self.obj_cats = tf.placeholder(tf.int32, [mini_batch_size, None],
name='obj_cats')
self.obj_spats = tf.placeholder(
tf.float32, [mini_batch_size, None, config['spat_dim']],
name='obj_spats')
# Targets
self.targets = tf.placeholder(tf.int32, [mini_batch_size],
name="targets_index")
self.object_cats_emb = utils.get_embedding(
self.obj_cats,
config['no_categories'] + 1,
config['cat_emb_dim'],
scope='cat_embedding')
self.objects_input = tf.concat(
[self.object_cats_emb, self.obj_spats], axis=2)
self.flat_objects_inp = tf.reshape(
self.objects_input,
[-1, config['cat_emb_dim'] + config['spat_dim']])
with tf.variable_scope('obj_mlp'):
h1 = utils.fully_connected(self.flat_objects_inp,
n_out=config['obj_mlp_units'],
activation='relu',
scope='l1')
h2 = utils.fully_connected(h1,
n_out=config['dialog_emb_dim'],
activation='relu',
scope='l2')
obj_embs = tf.reshape(
h2,
[-1, tf.shape(self.obj_cats)[1], config['dialog_emb_dim']])
# Compute the word embedding
input_words = utils.get_embedding(self.dialogues,
n_words=num_words,
n_dim=config['word_emb_dim'],
scope="input_word_embedding")
last_states, _ = rnn.variable_length_LSTM(
input_words,
num_hidden=config['num_lstm_units'],
seq_length=self.seq_length)
last_states = tf.reshape(last_states,
[-1, config['num_lstm_units'], 1])
scores = tf.matmul(obj_embs, last_states)
scores = tf.reshape(scores, [-1, tf.shape(self.obj_cats)[1]])
def masked_softmax(scores, mask):
# subtract max for stability
scores = scores - tf.tile(
tf.reduce_max(scores, axis=(1, ), keep_dims=True),
[1, tf.shape(scores)[1]])
# compute padded softmax
exp_scores = tf.exp(scores)
exp_scores *= mask
exp_sum_scores = tf.reduce_sum(exp_scores,
axis=1,
keep_dims=True)
return exp_scores / tf.tile(exp_sum_scores,
[1, tf.shape(exp_scores)[1]])
self.softmax = masked_softmax(scores, self.obj_mask)
self.selected_object = tf.argmax(self.softmax, axis=1)
self.loss = tf.reduce_mean(
utils.cross_entropy(self.softmax, self.targets))
self.error = tf.reduce_mean(utils.error(self.softmax,
self.targets))
def __init__(self, config, no_words, no_answers, reuse=False, device=''):
ResnetModel.__init__(self, "vqa", device=device)
with tf.variable_scope(self.scope_name, reuse=reuse) as scope:
self.batch_size = None
#####################
# QUESTION
#####################
self._question = tf.placeholder(tf.int32, [self.batch_size, None],
name='question')
self._seq_length = tf.placeholder(tf.int32, [self.batch_size],
name='seq_length')
self._answer_count = tf.placeholder(tf.float32,
[self.batch_size, no_answers],
name='answer_count')
self._is_training = tf.placeholder(tf.bool, name="is_training")
dropout_keep = float(config.get("dropout_keep_prob", 1.0))
dropout_keep = tf.cond(self._is_training,
lambda: tf.constant(dropout_keep),
lambda: tf.constant(1.0))
word_emb = utils.get_embedding(self._question,
n_words=no_words,
n_dim=int(
config["word_embedding_dim"]),
scope="word_embedding",
reuse=reuse)
if config['glove']:
self._glove = tf.placeholder(tf.float32, [None, None, 300],
name="glove")
word_emb = tf.concat([word_emb, self._glove], axis=2)
self.question_lstm, self.all_lstm_states = rnn.variable_length_LSTM(
word_emb,
num_hidden=int(config["no_hidden_LSTM"]),
dropout_keep_prob=dropout_keep,
seq_length=self._seq_length,
depth=int(config["no_LSTM_cell"]),
scope="question_lstm",
reuse=reuse)
#####################
# IMAGES
#####################
self._image = tf.placeholder(tf.float32, [self.batch_size] +
config['image']["dim"],
name='image')
self.image_out = get_image_features(image=self._image,
question=self.question_lstm,
is_training=self._is_training,
scope_name=scope.name,
config=config['image'],
dropout_keep=dropout_keep)
#####################
# COMBINE
#####################
activation_name = config["activation"]
with tf.variable_scope('final_mlp'):
self.question_embedding = utils.fully_connected(
self.question_lstm,
config["no_question_mlp"],
activation=activation_name,
scope='question_mlp')
self.image_embedding = utils.fully_connected(
self.image_out,
config["no_image_mlp"],
activation=activation_name,
scope='image_mlp')
full_embedding = self.image_embedding * self.question_embedding
full_embedding = tf.nn.dropout(full_embedding, dropout_keep)
out = utils.fully_connected(full_embedding,
config["no_hidden_final_mlp"],
scope='layer1',
activation=activation_name)
out = tf.nn.dropout(out, dropout_keep)
out = utils.fully_connected(out,
no_answers,
activation='linear',
scope='layer2')
# improve soft loss
answer_count = tf.minimum(self._answer_count, 3)
normalizing_sum = tf.maximum(
1.0, tf.reduce_sum(answer_count, 1, keep_dims=True))
self.answer_prob = answer_count / normalizing_sum
self.soft_cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
logits=out, labels=self.answer_prob, name='soft_cross_entropy')
self.soft_loss = self.soft_cross_entropy
self.target_answer = tf.argmax(self._answer_count, axis=1)
# unmorm_log_prob = tf.log(self._answer_count)
# self.target_answer = tf.multinomial(unmorm_log_prob, num_samples=1)
# self.target_answer = tf.reshape(self.target_answer, shape=[-1])
self.hard_cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=out,
labels=self.target_answer,
name='hard_cross_entropy')
self.hard_loss = self.hard_cross_entropy
if config['loss'] == 'soft':
self.loss = self.soft_loss
else:
self.loss = self.hard_loss
self.loss = tf.reduce_mean(self.loss)
self.softmax = tf.nn.softmax(out, name='answer_prob')
self.prediction = tf.argmax(
out, axis=1,
name='predicted_answer') # no need to compute the softmax
with tf.variable_scope('accuracy'):
ind = tf.range(tf.shape(
self.prediction)[0]) * no_answers + tf.cast(
self.prediction, tf.int32)
pred_count = tf.gather(tf.reshape(self._answer_count, [-1]),
ind)
self.extended_accuracy = tf.minimum(pred_count / 3.0,
1.0,
name="extended_accuracy")
self.accuracy = tf.reduce_mean(self.extended_accuracy)
tf.summary.scalar('soft_loss', self.soft_loss)
tf.summary.scalar('hard_loss', self.hard_loss)
tf.summary.scalar('accuracy', self.accuracy)
print('Model... build!')
def __init__(self, config, num_words, device='', reuse=False):
ResnetModel.__init__(self, "oracle", device=device)
with tf.variable_scope(self.scope_name, reuse=reuse) as scope:
embeddings = []
self.batch_size = None
# QUESTION
self._is_training = tf.placeholder(tf.bool, name="is_training")
self._question = tf.placeholder(tf.int32, [self.batch_size, None],
name='question')
self._seq_length = tf.placeholder(tf.int32, [self.batch_size],
name='seq_length')
word_emb = utils.get_embedding(
self._question,
n_words=num_words,
n_dim=int(config['model']['question']["embedding_dim"]),
scope="word_embedding")
lstm_states, _ = rnn.variable_length_LSTM(
word_emb,
num_hidden=int(config['model']['question']["no_LSTM_hiddens"]),
seq_length=self._seq_length)
embeddings.append(lstm_states)
# CATEGORY
if config['inputs']['category']:
self._category = tf.placeholder(tf.int32, [self.batch_size],
name='category')
cat_emb = utils.get_embedding(
self._category,
int(config['model']['category']["n_categories"]) +
1, # we add the unkwon category
int(config['model']['category']["embedding_dim"]),
scope="cat_embedding")
embeddings.append(cat_emb)
print("Input: Category")
# SPATIAL
if config['inputs']['spatial']:
self._spatial = tf.placeholder(tf.float32,
[self.batch_size, 8],
name='spatial')
embeddings.append(self._spatial)
print("Input: Spatial")
# IMAGE
if config['inputs']['image']:
self._image = tf.placeholder(tf.float32, [self.batch_size] +
config['model']['image']["dim"],
name='image')
self.image_out = get_image_features(
image=self._image,
question=lstm_states,
is_training=self._is_training,
scope_name=scope.name,
config=config['model']['image'])
embeddings.append(self.image_out)
print("Input: Image")
# CROP
if config['inputs']['crop']:
self._crop = tf.placeholder(tf.float32, [self.batch_size] +
config['model']['crop']["dim"],
name='crop')
self.crop_out = get_image_features(
image=self._crop,
question=lstm_states,
is_training=self._is_training,
scope_name=scope.name,
config=config["model"]['crop'])
embeddings.append(self.crop_out)
print("Input: Crop")
# Compute the final embedding
emb = tf.concat(embeddings, axis=1)
# OUTPUT
num_classes = 3
self._answer = tf.placeholder(tf.float32,
[self.batch_size, num_classes],
name='answer')
with tf.variable_scope('mlp'):
num_hiddens = config['model']['MLP']['num_hiddens']
l1 = utils.fully_connected(emb,
num_hiddens,
activation='relu',
scope='l1')
self.pred = utils.fully_connected(l1,
num_classes,
activation='softmax',
scope='softmax')
self.best_pred = tf.argmax(self.pred, axis=1)
self.loss = tf.reduce_mean(
utils.cross_entropy(self.pred, self._answer))
self.error = tf.reduce_mean(utils.error(self.pred, self._answer))
print('Model... Oracle build!')
def __init__(self, config, num_words_question ,num_words_description=None, device='', reuse=False):
ResnetModel.__init__(self, "oracle", device=device)
with open("data/dict_word_embedding_{}_{}.pickle".format("fasttext",config["model"]["question"]["embedding_type"]),"rb") as f:
dict_all_embedding = pickle.load(f)
with tf.variable_scope(self.scope_name, reuse=reuse) as scope:
embeddings = []
co_attention = [None,None,None,None]
self.batch_size = None
max_seq_length = 12
# QUESTION
if config['inputs']['question']:
self._is_training = tf.placeholder(tf.bool, name="is_training")
# self._question_word = tf.placeholder(tf.int32, [self.batch_size], name='question_word') #
self._question = tf.placeholder(tf.int32, [self.batch_size, 14], name='question')
self.seq_length_question = tf.placeholder(tf.int32, [self.batch_size], name='seq_length_question')
if config["model"]["glove"] == True or config["model"]["fasttext"] == True:
print("****** WITH EMBEDDING ******")
word_emb = utils.get_embedding(self._question,
n_words=num_words_question,
n_dim=int(config["model"]["word_embedding_dim"]),
scope="word_embedding",
dict_all_embedding=dict_all_embedding)
else:
print("****** NOT EMBEDDING ******")
word_emb = utils.get_embedding(self._question,
n_words=num_words_question,
n_dim=int(config["model"]["word_embedding_dim"]),
scope="word_embedding",
dict_all_embedding=[])
print(".... word_emb 1 = {} ".format(word_emb))
self.out_question = None
if config['model']['question']['lstm']:
self.lstm_states_question, self.lstm_all_state_ques = rnn.variable_length_LSTM(word_emb,
num_hidden=int(config['model']['question']["no_LSTM_hiddens"]),
seq_length=self.seq_length_question)
self.out_question = self.lstm_all_state_ques
# print("out_queston = {} ".format(self.lstm_states_question))
# exit()
# self.out_question = tf.reshape(self.out_question,[-1, self.out_question.get_shape()[1] * self.out_question.get_shape()[2] ])
else:
self.out_question = word_emb
if config["model"]["attention"]["co-attention"]:
co_attention[0] = self.out_question # Tensor("oracle/lstm/lstmcell0/concat:0", shape=(?, 14, 1024), dtype=float32)
embeddings.append(self.lstm_states_question)
# print("question_lstm = {} ".format(self.out_question ))
# exit()
else:
embeddings.append(self.lstm_states_question)
# QUESTION-Pos
if config['model']['question'] ['pos']:
print("----------------------------------------")
print("**** Oracle_network | input = question-pos ")
self._question_pos = tf.placeholder(tf.int32, [self.batch_size, None], name='question_pos')
self.seq_length_pos = tf.placeholder(tf.int32, [self.batch_size], name='seq_length_ques_pos')
word_emb = utils.get_embedding(self._question_pos,
n_words=num_words_question,
n_dim=100,
scope="word_embedding_pos")
if config["model"]["glove"] == True or config["model"]["fasttext"] == True:
self._glove = tf.placeholder(tf.float32, [None, None,int(config["model"]["word_embedding_dim"])], name="embedding_vector_ques_pos")
word_emb = tf.concat([word_emb, self._glove], axis=2)
else:
print("None ****************")
lstm_states, _ = rnn.variable_length_LSTM(word_emb,
num_hidden=int(config['model']['question']["no_LSTM_hiddens"]),
seq_length=self.seq_length_pos,scope="lstm2")
# embeddings.append(lstm_states)
# DESCRIPTION
if config['inputs']['description']:
print("**** Oracle_network | input = Description ")
self._description = tf.placeholder(tf.int32, [self.batch_size, None], name='description')
self.seq_length_description = tf.placeholder(tf.int32, [self.batch_size], name='seq_length_description')
word_emb = utils.get_embedding(self._description,
n_words=num_words_question,
n_dim=100,
reuse=True,
scope="word_embedding")
# print("word_emb = {} ".format(word_emb))
if config['model']['question']['lstm']:
self.lstm_states_des, self.lstm_all_state_des = rnn.variable_length_LSTM(word_emb,
num_hidden=int(config['model']['question']["no_LSTM_hiddens"]),
seq_length=self.seq_length_description,scope="lstm3")
self.out_question = self.lstm_states_des
# print("self.out_question_emb = {} ".format(self.out_question))
# self.out_question = tf.reshape(self.out_question,[-1, self.out_question.get_shape()[1] * self.out_question.get_shape()[2] ])
else:
self.out_question = word_emb
# print("self.out_question = {} ".format(self.out_question))
if config["model"]["attention"]["co-attention"]:
# co_attention[1] = self.out_question # embeddings.append(self.lstm_all_state_ques)
embeddings.append(self.lstm_states_des)
else:
embeddings.append(self.lstm_states_des)
if config['inputs']['history_question']:
placeholders_lstmQuestion = []
placeholders_lstmLength = []
for i in range(6):
self._embWord = tf.placeholder(tf.int32, [self.batch_size, 14], name="ques_hist_H{}".format(i))
self.seq_length_question_history = tf.placeholder(tf.int32, [self.batch_size], name='seq_length_question_history_H{}'.format(i))
self.word_emb = utils.get_embedding(self._embWord,
n_words=num_words_question,
n_dim=100,
reuse=True,
scope="word_embedding")
placeholders_lstmQuestion.append(self.word_emb)
placeholders_lstmLength.append(self.seq_length_question_history)
self.lstm_states, self.lstm_all_state_ques_hist = rnn.variable_length_LSTM(placeholders_lstmQuestion,
num_hidden=int(config['model']['question']["no_LSTM_hiddens"]),
seq_length=placeholders_lstmLength,scope="lstm4",dim_4=True)
if config["model"]["attention"]["co-attention"]:
co_attention[2] = self.lstm_states
else:
embeddings.append(self.lstm_states)
# Description-Pos
if config['model']['description'] ['pos']:
print("----------------------------------------")
print("**** Oracle_network | inpurt = question-pos ")
self._question_pos = tf.placeholder(tf.int32, [self.batch_size, None], name='des_pos')
self.seq_length_pos = tf.placeholder(tf.int32, [self.batch_size], name='seq_length_des_pos')
word_emb = utils.get_embedding(self._question_pos,
n_words=num_words_question,
n_dim=300,
scope="word_embedding_pos")
if config["model"]["glove"] == True or config["model"]["fasttext"] == True:
self._glove = tf.placeholder(tf.float32, [None, None, int(config["model"]["word_embedding_dim"])], name="embedding_vector_des_pos")
word_emb = tf.concat([word_emb, self._glove], axis=2)
else:
print("None ****************")
lstm_states, _ = rnn.variable_length_LSTM(word_emb,
num_hidden=int(config['model']['question']["no_LSTM_hiddens"]),
seq_length=self.seq_length_pos,scope="lstm5")
# embeddings.append(lstm_states)
# CATEGORY
if config['inputs']['category']:
print("**** Oracle_network | input = category ")
if config["model"]["category"]["use_embedding"]:
self._category = tf.placeholder(tf.float32, [self.batch_size,int(config["model"]["word_embedding_dim"])], name='category')
cat_emb = self._category
# cat_emb = utils.get_embedding(self._category,
# int(config['model']['category']["n_categories"]) + 1,
# n_dim=int(config["model"]["word_embedding_dim"]),
# scope="cat_embedding",
# dict_all_embedding=dict_all_embedding
# )
else:
self._category = tf.placeholder(tf.int32, [self.batch_size], name='category')
cat_emb = utils.get_embedding(self._category,
int(config['model']['category']["n_categories"]) + 1, # we add the unkwon category
int(config["model"]["word_embedding_dim"]),
scope="cat_embedding",
)
# cat_emb = tf.expand_dims(cat_emb,1)
embeddings.append(cat_emb)
print("Input: Category")
# ALLCATEGORY
if config['inputs']['allcategory']:
print("**** Oracle_network | input = allcategory ")
self._allcategory = tf.placeholder(tf.float32, [self.batch_size,90], name='allcategory')
# self.seq_length_allcategory = tf.placeholder(tf.int32, [self.batch_size], name='seq_length_allcategory')
# word_emb = utils.get_embedding(self._allcategory,
# n_words=int(config['model']['category']["n_categories"]) + 1,
# n_dim=int(config['model']['description']["embedding_dim"]),
# scope="word_embedding_allcategory")
#print(" SeqDescription = ",self.seq_length_description)
# lstm_states, _ = rnn.variable_length_LSTM(word_emb,
# num_hidden=int(config['model']['question']["no_LSTM_hiddens"]),
# seq_length=self.seq_length_allcategory,scope="lstm3")
print(" Oracle_network | embdedding all_cate=",word_emb)
# embeddings.append(self._allcategory)
print("Input: allcategory")
# SPATIAL
if config['inputs']['spatial']:
print("**** Oracle_network | input = spatial ")
self._spatial = tf.placeholder(tf.float32, [self.batch_size, 8], name='spatial')
embeddings.append(self._spatial)
print("Input: Spatial")
# IMAGE
if config['inputs']['image']:
print("**** Oracle_network | input = image ")
self._image_id = tf.placeholder(tf.float32, [self.batch_size], name='image_id')
self._image = tf.placeholder(tf.float32, [self.batch_size] + config['model']['image']["dim"], name='image')
# self.image_out = tf.reshape(self._image,shpe=[224*224*3])
# print("question = {} ".format(self.lstm_states_question))
# exit()
self.image_out = get_image_features(
image=self._image, question=self.lstm_states_question,
is_training=self._is_training,
scope_name=scope.name,
scope_feature="Image/",
config=config['model']['image']
)
# embeddings.append(self.image_out)
print("Input: Image")
co_attention[3] = self.image_out
print(" -- image_int ={}".format(self._image))
# exit()
image_feature = tf.reshape(self.image_out, shape=[-1, (7 * 7) * 2048]) # input image_feature ?,7,7,2048 => ?,49,2048
embeddings.append(image_feature)
# print("... Image Features = {}".format(self.image_out))
# CROP
if config['inputs']['crop']:
print("**** Oracle_network | input = crop ")
self._image_id = tf.placeholder(tf.float32, [self.batch_size], name='image_id')
# self._crop_id = tf.placeholder(tf.float32, [self.batch_size], name='crop_id')
self._crop = tf.placeholder(tf.float32, [self.batch_size] + config['model']['crop']["dim"], name='crop')
if config["model"]["attention"]["co-attention"]:
self.crop_out = get_image_features(
image=self._crop, question=self.lstm_states_question,
is_training=self._is_training,
scope_name=scope.name,
scope_feature="Crop/",
config=config["model"]['crop'])
co_attention[3] = self.crop_out
else:
self.crop_out = get_image_features(
image=self._crop, question=self.lstm_states_question,
is_training=self._is_training,
scope_name=scope.name,
scope_feature="Crop/",
co_attention=False,
config=config["model"]['crop'])
embeddings.append(self.crop_out)
if config["model"]["crop"]["segment_crop"]["use"]:
all_segment_crop = []
# for i in range(10):
self._segment_crop = tf.placeholder(tf.float32, [self.batch_size] + config['model']['crop']["dim"], name='crop_segment'.format(0))
self.crop_out = get_image_features(
image=self._segment_crop, question=self.lstm_states_question,
is_training=self._is_training,
scope_name="test",
scope_feature="Segment/",
config=config["model"]['crop'])
print("self.crop_out = {} ".format(self.crop_out))
# all_segment_crop.add(self.crop_out)
# print("-- crop = {},image_features = {} ".format(self.crop_out, image_feature))
# exit()
if config["model"]["attention"]["co-attention"]:
question_feature,history_feature , image_feature = compute_all_attention(question_states=co_attention[0],
caption=co_attention[1],
history_states=co_attention[2],
image_feature=co_attention[3],
no_mlp_units=config['model']['attention']['no_attention_mlp'],
config = config
)
embeddings.append(history_feature)
embeddings.append(question_feature)
embeddings.append(image_feature)
# embeddings.append(question_feature)
print("*** All Embedding = ",embeddings)
self.emb = tf.concat(embeddings, axis=1)
print("*** self.emb = ",self.emb)
# Compute the final embedding
# print("---------- Embeddings=",embeddings)
# self.emb = tf.concat(embeddings, axis=1)
# OUTPUT
num_classes = 3
self._answer = tf.placeholder(tf.float32, [self.batch_size, num_classes], name='answer')
with tf.variable_scope('mlp'):
num_hiddens = config['model']['MLP']['num_hiddens']
# emb = tf.print(emb, [emb], "input: ")
l1 = utils.fully_connected(self.emb, num_hiddens, activation='relu', scope='l1')
self.pred = utils.fully_connected(l1, num_classes, activation='softmax', scope='softmax')
self.best_pred = tf.argmax(self.pred, axis=1)
# self.best_pred = tf.reduce_mean(self.best_pred)
print("--- predict = {} ,answer = {} ".format(self.pred,self._answer))
# exit()
# self.loss = None
self.loss = tf.reduce_mean(utils.cross_entropy(self.pred, self._answer))
self.error = tf.reduce_mean(utils.error(self.pred, self._answer))
print("loss = {} ,error = {} ".format(self.loss,self.error))
print('Model... Oracle build!')
def __init__(self, config, num_words, policy_gradient, device='', reuse=False):
# AbstractNetwork.__init__(self, "qgen_guesser", device=device)
ResnetModel.__init__(self, "qgen_guesser", device=device)
# Create the scope for this graph
with tf.variable_scope(self.scope_name, reuse=reuse):
# We set batch size to be none as the batch size for the validation set and train set are different
# mini_batch_size = None
# mini_batch_size = config['batch_size']
mini_batch_size = None
self.guesser_loss_weight = tf.constant(config["guesser_loss_weight"], dtype = tf.float32, name = "guesser_loss_weight")
self.qgen_loss_weight = tf.constant(config["qgen_loss_weight"], dtype = tf.float32, name = "qgen_loss_weight")
self.loss = 0
# *********************************************************
# Placeholders specific for guesser and its processing
# *********************************************************
# Objects
self.obj_mask = tf.placeholder(tf.float32, [mini_batch_size, None], name='obj_mask')
self.obj_cats = tf.placeholder(tf.int32, [mini_batch_size, None], name='obj_cats')
self.obj_spats = tf.placeholder(tf.float32, [mini_batch_size, None, config['spat_dim']], name='obj_spats')
# Targets
self.targets = tf.placeholder(tf.int32, [mini_batch_size], name="targets_index")
self.object_cats_emb = utils.get_embedding(
self.obj_cats,
config['no_categories'] + 1,
config['cat_emb_dim'],
scope='cat_embedding')
self.objects_input = tf.concat([self.object_cats_emb, self.obj_spats], axis=2)
self.flat_objects_inp = tf.reshape(self.objects_input, [-1, config['cat_emb_dim'] + config['spat_dim']])
with tf.variable_scope('obj_mlp'):
h1 = utils.fully_connected(
self.flat_objects_inp,
n_out=config['obj_mlp_units'],
activation='relu',
scope='l1')
h2 = utils.fully_connected(
h1,
n_out=config['no_hidden_final_mlp'],
activation='relu',
scope='l2')
# print
# print
# print h2
# TODO: Object Embeddings do not have image features right now
obj_embs = tf.reshape(h2, [-1, tf.shape(self.obj_cats)[1], config['no_hidden_final_mlp']])
# *********************************************************
# Placeholders for Qgen and common placeholder for guesser and its processing
# *********************************************************
# Image
self.images = tf.placeholder(tf.float32, [mini_batch_size] + config['image']["dim"], name='images')
# Question
self.dialogues = tf.placeholder(tf.int32, [mini_batch_size, None], name='dialogues')
self.answer_mask = tf.placeholder(tf.float32, [mini_batch_size, None], name='answer_mask') # 1 if keep and (1 q/a 1) for (START q/a STOP)
self.padding_mask = tf.placeholder(tf.float32, [mini_batch_size, None], name='padding_mask')
self.seq_length = tf.placeholder(tf.int32, [mini_batch_size], name='seq_length')
# Rewards
self.cum_rewards = tf.placeholder(tf.float32, shape=[mini_batch_size, None], name='cum_reward')
# DECODER Hidden state (for beam search)
zero_state = tf.zeros([1, config['num_lstm_units']]) # default LSTM state is a zero-vector
zero_state = tf.tile(zero_state, [tf.shape(self.images)[0], 1]) # trick to do a dynamic size 0 tensors
self.decoder_zero_state_c = tf.placeholder_with_default(zero_state, [mini_batch_size, config['num_lstm_units']], name="state_c")
self.decoder_zero_state_h = tf.placeholder_with_default(zero_state, [mini_batch_size, config['num_lstm_units']], name="state_h")
decoder_initial_state = tf.contrib.rnn.LSTMStateTuple(c=self.decoder_zero_state_c, h=self.decoder_zero_state_h)
# *******
# Misc
# *******
self.is_training = tf.placeholder(tf.bool, name='is_training')
self.greedy = tf.placeholder_with_default(False, shape=(), name="greedy") # use for graph
self.samples = None
# For each length of the answer, we are finding the next token
# remove last token
input_dialogues = self.dialogues[:, :-1]
input_seq_length = self.seq_length - 1
# remove first token(=start token)
rewards = self.cum_rewards[:, 1:]
target_words = self.dialogues[:, 1:]
# to understand the padding:
# input
# <start> is it a blue <?> <yes> is it a car <?> <no> <stop_dialogue>
# target
# is it a blue <?> - is it a car <?> - <stop_dialogue> -
# TODO:
# 1. Include Film in guesser (Check if you are using film or cbn????)
# Add finetune in the training (see the training and config file of clever)
# Check the use of finetune (should we input pretrained model), normalize etc
# See in the config file, if the attention has to be put inside image block
# As of now, in the first part where we get image embedding, the we only flatten the image. Use RCNN or other method to get the image features.
# Include attention on image given the dialog embedding in the guesser part
# Include dropout om lstm (option for inside and outside) and image
# Include attention on words given the image features
# 2. Use tf.gather and use all the lstm states where there was yes or no (-) in the target and the stop dialog
# 3. Make the code run
# Check how does the is_training flag works
# image processing
with tf.variable_scope('image_feature') as img_scope:
if len(config["image"]["dim"]) == 1:
self.image_out = self.images
else:
# TODO: Create a different config for this attention
# Putting images
tf.summary.image("image", self.images)
self.image_out = get_image_features(
image=self.images, question = None,
is_training=self.is_training,
scope_name=img_scope.name,
config=config['image'],
att = False
)
image_pooling_size = [int((self.image_out).get_shape()[1]), int((self.image_out).get_shape()[2])]
image_feature_depth = int((self.image_out).get_shape()[3])
self.image_out = tf.layers.max_pooling2d(self.image_out,
image_pooling_size,
1,
padding='valid',
data_format='channels_last',
name='max_pooling_image_out')
self.image_out = tf.reshape(self.image_out, [-1, image_feature_depth])
# self.filmed_picture_out = tf.layers.average_pooling2d( self.filmed_picture_out,
# final_pooling_size,
# 1,
# padding='valid',
# data_format='channels_last',
# name='average_pooling_filmed_picture_out')
# self.image_out = get_attention(self.images, None, config["image"]["attention"]) #TODO: improve by using the previous lstm state?
# self.image_out = tf.contrib.layers.flatten(self.image_out)
print self.image_out
print
print
# Reduce the embedding size of the image
with tf.variable_scope('image_embedding'):
self.image_emb = utils.fully_connected(self.image_out,
config['image_embedding_size'])
image_emb = tf.expand_dims(self.image_emb, 1)
image_emb = tf.tile(image_emb, [1, tf.shape(input_dialogues)[1], 1])
# Compute the question embedding
input_words = utils.get_embedding(
input_dialogues,
n_words=num_words,
n_dim=config['word_embedding_size'],
scope="word_embedding")
# concat word embedding and image embedding
# TODO: Check the size (see if input_seq_length is increased or not)
decoder_input = tf.concat([input_words, image_emb], axis=2, name="concat_full_embedding")
# encode one word+image
decoder_lstm_cell = tf.contrib.rnn.LayerNormBasicLSTMCell(
config['num_lstm_units'],
layer_norm=False,
dropout_keep_prob=1.0,
reuse=reuse)
# TODO: Since we have concatinated image, check if the input_seq_length should be increased by one
# Decoding the states to generate questions
self.decoder_output, self.decoder_state = tf.nn.dynamic_rnn(
cell=decoder_lstm_cell,
inputs=decoder_input,
dtype=tf.float32,
initial_state=decoder_initial_state,
sequence_length=input_seq_length,
scope="word_decoder") # TODO: use multi-layer RNN
max_sequence = tf.reduce_max(self.seq_length)
# For the Guesser
# Adding extra layers of LSTM
# TODO: There are several default parameters in the fuction. Try using them
# TODO: as of now, not using it.
# TODO, as of now only using the hidden state, you may include the other state too
last_states = self.decoder_state.h
# last_states, _ = rnn.variable_length_LSTM_extension(
# self.decoder_output,
# self.decoder_state,
# num_hidden = config['num_lstm_units'],
# seq_length = input_seq_length
# )
last_states = tf.reshape(last_states, [-1, config['num_lstm_units']])
# TODO: Can be moved to utils
def masked_softmax(scores, mask):
# subtract max for stability
scores = scores - tf.tile(tf.reduce_max(scores, axis=(1,), keepdims=True), [1, tf.shape(scores)[1]])
# compute padded softmax
exp_scores = tf.exp(scores)
exp_scores *= mask
exp_sum_scores = tf.reduce_sum(exp_scores, axis=1, keepdims=True)
return exp_scores / tf.tile(exp_sum_scores, [1, tf.shape(exp_scores)[1]])
# compute the softmax for evaluation (on all the words on dialogue)
with tf.variable_scope('decoder_output'):
flat_decoder_output = tf.reshape(self.decoder_output, [-1, decoder_lstm_cell.output_size])
flat_mlp_output = utils.fully_connected(flat_decoder_output, num_words)
# retrieve the batch/dialogue format
mlp_output = tf.reshape(flat_mlp_output, [tf.shape(self.seq_length)[0], max_sequence - 1, num_words]) # Ignore th STOP token
self.softmax_output = tf.nn.softmax(mlp_output, name="softmax")
self.argmax_output = tf.argmax(mlp_output, axis=2)
self.cross_entropy_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=mlp_output, labels=target_words)
# compute the maximum likelihood loss for the dialogues (for valid words)
with tf.variable_scope('ml_loss'):
ml_loss = tf.identity(self.cross_entropy_loss)
ml_loss *= self.answer_mask[:, 1:] # remove answers (ignore the <stop> token)
ml_loss *= self.padding_mask[:, 1:] # remove padding (ignore the <start> token)
# Count number of unmask elements
count = tf.reduce_sum(self.padding_mask) - tf.reduce_sum(1 - self.answer_mask[:, :-1]) - 1 # no_unpad - no_qa - START token
ml_loss = tf.reduce_sum(ml_loss, axis=1) # reduce over dialogue dimension
ml_loss = tf.reduce_sum(ml_loss, axis=0) # reduce over minibatch dimension
self.ml_loss = ml_loss / count # Normalize
self.qgen_loss = self.ml_loss
self.loss += self.qgen_loss_weight * self.qgen_loss
tf.summary.scalar("qgen_loss", self.qgen_loss)
# TODO NOTE: IMP in config file, under the image section, set cbn to be true
with tf.variable_scope('guesser_input') as scope:
# Getting the CBN image features
self.CBN_picture_out = get_image_features(
image=self.images, question = last_states,
is_training=self.is_training,
scope_name=scope.name,
config=config['image']
)
# FILMING the Features
# self.filmed_picture_out = film_layer(ft=self.CBN_picture_out, context=last_states)
# TODO: Make n a hyperparameter and add it to network parameters
self.filmed_picture_out = self.CBN_picture_out
n = 1
for i in range(n):
with tf.variable_scope('film_layer_' + str(i)):
self.filmed_picture_out = FiLMResblock(features=self.filmed_picture_out, context=last_states, is_training=self.is_training).get()
# self.filmed_picture_out_3 = FiLMResblock(features=self.filmed_picture_out_2, context=last_states, is_training=self.is_training).get()
# self.filmed_picture_out = FiLMResblock(features=self.filmed_picture_out_3, context=last_states, is_training=self.is_training).get()
# self.filmed_picture_out = FiLMResblock(features=self.filmed_picture_out, context=last_states, is_training=self.is_training).get()
# self.filmed_picture_out = FiLMResblock(features=self.filmed_picture_out, context=last_states, is_training=self.is_training).get()
# self.filmed_picture_out = FiLMResblock(features=self.CBN_picture_out, context=last_states, is_training=self.is_training).get()
# self.filmed_picture_out = film_layer(features=self.CBN_picture_out, context=last_states)
# TODO: Doing a convolution over the feature maps (before the classifier)
# Do a max pooling over the feature maps
final_pooling_size = [int((self.filmed_picture_out).get_shape()[1]), int((self.filmed_picture_out).get_shape()[2])]
final_feature_depth = int((self.filmed_picture_out).get_shape()[3])
if str(config["pooling"]).lower() == 'max':
self.filmed_picture_out = tf.layers.max_pooling2d( self.filmed_picture_out,
final_pooling_size,
1,
padding='valid',
data_format='channels_last',
name='max_pooling_filmed_picture_out')
elif str(config["pooling"]).lower() == 'avg':
self.filmed_picture_out = tf.layers.average_pooling2d( self.filmed_picture_out,
final_pooling_size,
1,
padding='valid',
data_format='channels_last',
name='average_pooling_filmed_picture_out')
else:
print "No Pooling defined"
sys.exit()
self.filmed_picture_out = tf.reshape(self.filmed_picture_out, [-1, final_feature_depth])
# Combining filmed image and dialog features into one
#####################
activation_name = config["activation"]
self.question_embedding = utils.fully_connected(last_states, config["no_question_mlp"], activation=activation_name, scope='question_mlp')
self.picture_embedding = utils.fully_connected(self.filmed_picture_out, config["no_picture_mlp"], activation=activation_name, scope='picture_mlp')
self.full_embedding = self.picture_embedding * self.question_embedding
# self.full_embedding = tf.nn.dropout(full_embedding, dropout_keep)
# self.guesser_out_0 = utils.fully_connected(self.full_embedding, config["no_hidden_prefinal_mlp"], scope='hidden_prefinal', activation=activation_name)
self.guesser_out_0 = self.full_embedding
# out = tf.nn.dropout(out, dropout_keep)
# Since we are not having
# out = utils.fully_connected(out, no_answers, activation='linear', scope='layer_softmax')
self.guesser_out = utils.fully_connected(self.guesser_out_0, config["no_hidden_final_mlp"], scope='hidden_final', activation=activation_name)
self.guesser_out = tf.reshape(self.guesser_out, [-1, config["no_hidden_final_mlp"], 1])
# TODO DONE: Add all these losses to tensorboard
with tf.variable_scope('guesser_output'):
# TODO: In paper they do dot product, but in code they do matmul !!
scores = tf.matmul(obj_embs, self.guesser_out)
scores = tf.reshape(scores, [-1, tf.shape(self.obj_cats)[1]])
self.softmax = masked_softmax(scores, self.obj_mask)
self.selected_object = tf.argmax(self.softmax, axis=1)
self.guesser_error = tf.reduce_mean(utils.error(self.softmax, self.targets))
self.guesser_loss = tf.reduce_mean(utils.cross_entropy(self.softmax, self.targets))
self.loss += self.guesser_loss_weight * self.guesser_loss
tf.summary.scalar("guesser loss", self.guesser_loss)
# Compute policy gradient
if policy_gradient:
with tf.variable_scope('rl_baseline'):
decoder_out = tf.stop_gradient(self.decoder_output) # take the LSTM output (and stop the gradient!)
flat_decoder_output = tf.reshape(decoder_out, [-1, decoder_lstm_cell.output_size]) #
flat_h1 = utils.fully_connected(flat_decoder_output, n_out=100, activation='relu', scope='baseline_hidden')
flat_baseline = utils.fully_connected(flat_h1, 1, activation='relu', scope='baseline_out')
self.baseline = tf.reshape(flat_baseline, [tf.shape(self.seq_length)[0], max_sequence-1])
self.baseline *= self.answer_mask[:, 1:]
self.baseline *= self.padding_mask[:, 1:]
with tf.variable_scope('policy_gradient_loss'):
# Compute log_prob
self.log_of_policy = tf.identity(self.cross_entropy_loss)
self.log_of_policy *= self.answer_mask[:, 1:] # remove answers (<=> predicted answer has maximum reward) (ignore the START token in the mask)
# No need to use padding mask as the discounted_reward is already zero once the episode terminated
# Policy gradient loss
rewards *= self.answer_mask[:, 1:]
self.score_function = tf.multiply(self.log_of_policy, rewards - self.baseline) # score function
self.baseline_loss = tf.reduce_sum(tf.square(rewards - self.baseline))
self.policy_gradient_loss = tf.reduce_sum(self.score_function, axis=1) # sum over the dialogue trajectory
self.policy_gradient_loss = tf.reduce_mean(self.policy_gradient_loss, axis=0) # reduce over minibatch dimension
self.loss = self.policy_gradient_loss
tf.summary.scalar("total network loss", self.loss)
self.summary = tf.summary.merge_all()
print('Model... build!')