def __init__(self,
img_size,
glimpse_size,
n_what,
glimpse_encoder,
scale_offset=0.,
masked_glimpse=False,
debug=False):
super(AIREncoder, self).__init__(img_size, glimpse_size, inverse=False)
self.n_what = n_what
self._masked_glimpse = masked_glimpse
with self._enter_variable_scope():
self._glimpse_encoder = glimpse_encoder
self._what_distrib = GaussianFromParamVec(
n_what,
scale_offset=scale_offset,
validate_args=debug,
allow_nan_stats=not debug)
if self._masked_glimpse:
self._mask_mlp = MLP(
128,
n_out=np.prod(glimpse_size),
transfer=tf.nn.sigmoid,
output_initializers={'b': tf.constant_initializer(1.)})
def _build(self, inpt):
n = np.prod(self._output_size)
mlp = MLP(self._n_hidden, n_out=n)
reshape = snt.BatchReshape(self._output_size)
seq = snt.Sequential([mlp, reshape])
return seq(inpt) * tf.get_variable('output_scale', initializer=self._output_scale)
def _make_priors(self, time_step, prior_conditioning):
"""Instantiates prior distributions for discovery.
"""
is_first_timestep = tf.to_float(tf.equal(time_step, 0))
if self._disc_prior_type == 'geom':
num_steps_prior = tfd.Geometric(probs=1. - self._init_disc_step_success_prob)
elif self._disc_prior_type == 'cat':
init = [0.] * (self._n_steps + 1)
step_logits = tf.Variable(init, trainable=True, dtype=tf.float32, name='step_prior_bias')
# increase probability of zero steps when t>0
init = [10.] + [0] * self._n_steps
timstep_bias = tf.Variable(init, trainable=True, dtype=tf.float32, name='step_prior_timestep_bias')
step_logits += (1. - is_first_timestep) * timstep_bias
if prior_conditioning is not None:
step_logits = tf.expand_dims(step_logits, 0) + MLP(10, n_out=self._n_steps + 1)(prior_conditioning)
step_logits = tf.nn.elu(step_logits)
num_steps_prior = tfd.Categorical(logits=step_logits)
else:
raise ValueError('Invalid prior type: {}'.format(self._disc_prior_type))
return self._what_prior, self._where_prior, num_steps_prior
def _build(self, inpt):
flatten = snt.BatchFlatten()
mlp = MLP(self._n_hidden, n_out=8)
seq = snt.Sequential([flatten, mlp])
params = seq(inpt)
return params[..., :4], params[..., 4:] + self._scale_bias
def _build(self, inpt):
flatten = snt.BatchFlatten()
mlp = MLP(self._n_hidden, n_out=8)
seq = snt.Sequential([flatten, mlp])
params = seq(inpt)
scale_offset = tf.get_variable('scale_offset', initializer=self._scale_offset)
return params[..., :4], params[..., 4:] + scale_offset
def _build(self, img, what, where, presence_prob):
batch_size = int(img.get_shape()[0])
parts = [
tf.reshape(tf.transpose(i, (1, 0, 2)), (batch_size, -1))
for i in (what, where, presence_prob)
]
img_flat = tf.reshape(img, (batch_size, -1))
baseline_inpts = [img_flat] + parts
baseline_inpts = tf.concat(baseline_inpts, -1)
mlp = MLP(self._n_hidden, n_out=1)
baseline = mlp(baseline_inpts)
return baseline
def __init__(self, n_steps, n_latent_code=0, relation_embedding=False):
"""Initialises the module.
:param n_steps: Integer, number of inference steps to perform at this time-step.
:param n_latent_code: Integer, dimensionality of summary of latent variables.
:param relation_embedding: Boolean; computes DeepSet-like embedding of latent variables if True.
"""
super(AbstractTimstepModule, self).__init__()
self._n_steps = n_steps
self._n_latent_code = n_latent_code
self._relation_embedding = relation_embedding
with self._enter_variable_scope():
if n_latent_code > 0:
self._latent_encoder = MLP([n_latent_code] * 2)
def _build(self, previous_presence, previois_logit, *features):
init = {'b': tf.constant_initializer(self._steps_bias)}
mlp = MLP(self._n_hidden, n_out=1, output_initializers=init)
features = ops.maybe_concat(features)
logit = mlp(features)
logit = previous_presence * logit + (previous_presence - 1.) * 88.
if previois_logit is not None:
if self._max_rel_logit_change != np.inf:
min_logit = (1. - self._max_rel_logit_change) * previois_logit
max_logit = (1. + self._max_rel_logit_change) * previois_logit
logit = tf.clip_by_value(logit, min_logit, max_logit)
elif self._max_logit_change != np.inf:
logit = previois_logit + self._max_logit_change * tf.nn.tanh(logit)
return self._bernoulli(logit)
def _build(self, inpt):
mlp = MLP(self._n_hidden, n_out=1)
logit = mlp(inpt) + self._steps_bias
return tf.nn.sigmoid(logit)
def _build(self, inpt):
n = np.prod(self._output_size)
mlp = MLP(self._n_hidden, n_out=n)
reshape = snt.BatchReshape(self._output_size)
seq = snt.Sequential([mlp, reshape])
return seq(inpt)
def _build(self, inpt):
flat = snt.BatchFlatten()
mlp = MLP(self._n_hidden)
seq = snt.Sequential([flat, mlp])
return seq(inpt)
def _embed(self, inpt):
flatten = snt.BatchFlatten()
mlp = MLP(self._n_hidden, n_out=self._n_param)
seq = snt.Sequential([flatten, mlp])
return seq(inpt)
def build(self):
with tf.variable_scope("Embeddings"):
self.embeddings = tf.get_variable(
"emb", [self.config.n_embed, self.config.d_embed],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
embeddings_root = tf.get_variable(
"emb_root", [1, 1, 2 * self.config.dim_sem],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
embeddings_root_s = tf.get_variable(
"emb_root_s", [1, 1, 2 * self.config.dim_sem],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
with tf.variable_scope("Model"):
w_comb = tf.get_variable(
"w_comb", [4 * self.config.dim_sem, 2 * self.config.dim_sem],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
b_comb = tf.get_variable("bias_comb", [2 * self.config.dim_sem],
dtype=tf.float32,
initializer=tf.constant_initializer())
w_comb_s = tf.get_variable(
"w_comb_s", [4 * self.config.dim_sem, 2 * self.config.dim_sem],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
b_comb_s = tf.get_variable("bias_comb_s",
[2 * self.config.dim_sem],
dtype=tf.float32,
initializer=tf.constant_initializer())
w_softmax = tf.get_variable(
"w_softmax", [2 * self.config.dim_sem, self.config.dim_output],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
b_softmax = tf.get_variable(
"bias_softmax", [self.config.dim_output],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
with tf.variable_scope("Structure/doc"):
tf.get_variable("w_parser_p",
[2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("w_parser_c",
[2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("w_parser_s",
[2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("bias_parser_p", [2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("bias_parser_c", [2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("w_parser_root", [2 * self.config.dim_str, 1],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
with tf.variable_scope("Structure/sent"):
tf.get_variable("w_parser_p",
[2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("w_parser_c",
[2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("bias_parser_p", [2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("bias_parser_c", [2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("w_parser_s",
[2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
tf.get_variable("w_parser_root", [2 * self.config.dim_str, 1],
dtype=tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
sent_l = self.t_variables['sent_l']
doc_l = self.t_variables['doc_l']
max_sent_l = self.t_variables['max_sent_l']
max_doc_l = self.t_variables['max_doc_l']
batch_l = self.t_variables['batch_l']
tokens_input = tf.nn.embedding_lookup(
self.embeddings,
self.t_variables['token_idxs'][:, :max_doc_l, :max_sent_l])
tokens_input = tf.nn.dropout(tokens_input,
self.t_variables['keep_prob'])
mask_tokens = self.t_variables[
'mask_tokens'][:, :max_doc_l, :max_sent_l]
mask_sents = self.t_variables['mask_sents'][:, :max_doc_l]
[_, _, _, rnn_size] = tokens_input.get_shape().as_list()
tokens_input_do = tf.reshape(
tokens_input, [batch_l * max_doc_l, max_sent_l, rnn_size])
sent_l = tf.reshape(sent_l, [batch_l * max_doc_l])
mask_tokens = tf.reshape(mask_tokens, [batch_l * max_doc_l, -1])
tokens_output, _ = dynamicBiRNN(tokens_input_do,
sent_l,
n_hidden=self.config.dim_hidden,
cell_type=self.config.rnn_cell,
cell_name='Model/sent')
tokens_sem = tf.concat([
tokens_output[0][:, :, :self.config.dim_sem],
tokens_output[1][:, :, :self.config.dim_sem]
], 2)
tokens_str = tf.concat([
tokens_output[0][:, :, self.config.dim_sem:],
tokens_output[1][:, :, self.config.dim_sem:]
], 2)
temp1 = tf.zeros([batch_l * max_doc_l, max_sent_l, 1], tf.float32)
temp2 = tf.zeros([batch_l * max_doc_l, 1, max_sent_l], tf.float32)
mask1 = tf.ones([batch_l * max_doc_l, max_sent_l, max_sent_l - 1],
tf.float32)
mask2 = tf.ones([batch_l * max_doc_l, max_sent_l - 1, max_sent_l],
tf.float32)
mask1 = tf.concat([temp1, mask1], 2)
mask2 = tf.concat([temp2, mask2], 1)
str_scores_s_ = get_structure('sent', tokens_str, max_sent_l, mask1,
mask2) # batch_l, sent_l+1, sent_l
str_scores_s = tf.matrix_transpose(str_scores_s_) # soft parent
tokens_sem_root = tf.concat([
tf.tile(embeddings_root_s, [batch_l * max_doc_l, 1, 1]), tokens_sem
], 1)
tokens_output_ = tf.matmul(str_scores_s, tokens_sem_root)
tokens_output = LReLu(
tf.tensordot(tf.concat([tokens_sem, tokens_output_], 2), w_comb_s,
[[2], [0]]) + b_comb_s)
if (self.config.sent_attention == 'sum'):
tokens_output = tokens_output * tf.expand_dims(mask_tokens, 2)
tokens_output = tf.reduce_sum(tokens_output, 1)
elif (self.config.sent_attention == 'mean'):
tokens_output = tokens_output * tf.expand_dims(mask_tokens, 2)
tokens_output = tf.reduce_sum(tokens_output, 1) / tf.expand_dims(
tf.cast(sent_l, tf.float32), 1)
elif (self.config.sent_attention == 'max'):
tokens_output = tokens_output + tf.expand_dims(
(mask_tokens - 1) * 999, 2)
tokens_output = tf.reduce_max(tokens_output, 1)
sents_input = tf.reshape(tokens_output,
[batch_l, max_doc_l, 2 * self.config.dim_sem])
sents_output, _ = dynamicBiRNN(sents_input,
doc_l,
n_hidden=self.config.dim_hidden,
cell_type=self.config.rnn_cell,
cell_name='Model/doc')
sents_sem = tf.concat([
sents_output[0][:, :, :self.config.dim_sem],
sents_output[1][:, :, :self.config.dim_sem]
], 2)
sents_str = tf.concat([
sents_output[0][:, :, self.config.dim_sem:],
sents_output[1][:, :, self.config.dim_sem:]
], 2)
str_scores_ = get_structure(
'doc', sents_str, max_doc_l, self.t_variables['mask_parser_1'],
self.t_variables['mask_parser_2']) #batch_l, sent_l+1, sent_l
str_scores = tf.matrix_transpose(str_scores_) # soft parent
sents_sem_root = tf.concat(
[tf.tile(embeddings_root, [batch_l, 1, 1]), sents_sem], 1)
sents_output_ = tf.matmul(str_scores, sents_sem_root)
sents_output = LReLu(
tf.tensordot(tf.concat([sents_sem, sents_output_], 2), w_comb,
[[2], [0]]) + b_comb)
if (self.config.doc_attention == 'sum'):
sents_output = sents_output * tf.expand_dims(mask_sents, 2)
sents_output = tf.reduce_sum(sents_output, 1)
elif (self.config.doc_attention == 'mean'):
sents_output = sents_output * tf.expand_dims(mask_sents, 2)
sents_output = tf.reduce_sum(sents_output, 1) / tf.expand_dims(
tf.cast(doc_l, tf.float32), 1)
elif (self.config.doc_attention == 'max'):
sents_output = sents_output + tf.expand_dims(
(mask_sents - 1) * 999, 2)
sents_output = tf.reduce_max(sents_output, 1)
final_output = MLP(sents_output, 'output',
self.t_variables['keep_prob'])
self.final_output = tf.matmul(final_output, w_softmax) + b_softmax
from neural import MLP
df = pd.read_csv('data.csv')
df = df.iloc[np.random.permutation(len(df))]
y = df.iloc[0:100, 4].values
y = np.where(y == "Iris-setosa", 1, 0).reshape(-1, 1)
X = df.iloc[0:100, [0, 2]].values
inputSize = X.shape[
1] # количество входных сигналов равно количеству признаков задачи
hiddenSizes = 10 # задаем число нейронов скрытого (А) слоя
outputSize = 1 if len(y.shape) else y.shape[
1] # количество выходных сигналов равно количеству классов задачи
iterations = 50
learning_rate = 0.1
net = MLP(inputSize, outputSize, learning_rate, hiddenSizes)
# обучаем сеть (фактически сеть это вектор весов weights)
for i in range(iterations):
net.train(X, y)
if i % 10 == 0:
print("На итерации: " + str(i) + ' || ' + "Средняя ошибка: " +
str(np.mean(np.square(y - net.predict(X)))))
# считаем ошибку на обучающей выборке
pr = net.predict(X)
print(sum(abs(y - (pr > 0.5))))
class PropagationCore(BaseSQAIRCore):
"""Recurrent propagation core.
It is run iteratively to propagate several objects.
"""
_output_names = 'what what_sample what_loc what_scale where where_sample where_loc where_scale presence_prob' \
' presence presence_logit temporal_state'.split()
_init_presence_value = 0. # at the beginning we assume no objects
_what_scale_bias = -3.
def __init__(self,
img_size,
crop_size,
n_what,
transition,
input_encoder,
glimpse_encoder,
transform_estimator,
steps_predictor,
temporal_cell,
where_update_scale=1.0,
debug=False):
"""Initialises the model.
If argument is not covered here, see BaseSQAIRCore for documentation.
:param temporal_cell: RNNCore for the temporal rnn.
:param where_update_scale: Float, rescales the update of the `where` variables.
"""
super(PropagationCore, self).__init__(img_size,
crop_size,
n_what,
transition,
input_encoder,
glimpse_encoder,
transform_estimator,
steps_predictor,
debug=debug)
self._temporal_cell = temporal_cell
with self._enter_variable_scope():
self._where_update_scale = tf.get_variable(
'where_update_scale',
shape=[],
dtype=tf.float32,
initializer=tf.constant_initializer(where_update_scale),
trainable=False)
self._where_distrib = AffineDiagNormal(
validate_args=self._debug, allow_nan_stats=not self._debug)
@property
def output_size(self):
return [
self._n_what, # what code
self._n_what, # what sample
self._n_what, # what loc
self._n_what, # what scale
self._n_transform_param, # where code
self._n_transform_param, # where sample
self._n_transform_param, # where loc
self._n_transform_param, # where scale
1, # presence prob
1, # presence
1, # presence_logit,
self._temporal_cell.state_size,
]
def _build(self, (z_tm1, temporal_hidden_state), state):
"""Input is unused; it's only to force a maximum number of steps"""
# same object, previous timestep
what_tm1, where_tm1, presence_tm1, presence_logit_tm1 = z_tm1
temporal_state = nest.flatten(temporal_hidden_state)[-1]
# different object, current timestep
img_flat, what_km1, where_km1, presence_km1, hidden_state = state
img = tf.reshape(img_flat, (-1, ) + tuple(self._img_size))
with tf.variable_scope('rnn_inpt'):
where_bias = MLP(128, n_out=4)(temporal_state) * .1
what_distrib = self._glimpse_encoder(img,
where_tm1 + where_bias,
mask_inpt=temporal_state)[0]
rnn_inpt = what_distrib.loc
rnn_inpt = [
rnn_inpt, # img
what_km1,
where_km1,
presence_km1, # explaining away
what_tm1,
where_tm1,
presence_tm1,
temporal_state # previous state
]
rnn_inpt = tf.concat(rnn_inpt, -1)
hidden_output, hidden_state = self._cell(rnn_inpt, hidden_state)
with tf.variable_scope('where'):
where, where_sample, where_loc, where_scale = self._compute_where(
where_tm1, hidden_output, temporal_state)
with tf.variable_scope('what'):
what, what_sample, what_loc, what_scale, temporal_hidden_state\
= self._compute_what(img, what_tm1, where, hidden_output, temporal_hidden_state, temporal_state)
with tf.variable_scope('presence'):
presence, presence_prob, presence_logit \
= self._compute_presence(presence_tm1, presence_logit_tm1, hidden_output, temporal_state, what)
output = [
what, what_sample, what_loc, what_scale, where, where_sample,
where_loc, where_scale, presence_prob, presence, presence_logit,
temporal_hidden_state
]
new_state = [img_flat, what, where, presence, hidden_state]
return output, new_state
class PropagationCore(BaseSQAIRCore):
"""Recurrent propagation core.
It is run iteratively to propagate several objects.
"""
_output_names = 'what what_sample what_loc what_scale where where_sample where_loc where_scale presence_prob' \
' presence presence_logit temporal_state'.split()
_init_presence_value = 0. # at the beginning we assume no objects
_what_scale_bias = -3.
def __init__(self, img_size, crop_size, n_what,
transition, input_encoder, glimpse_encoder, transform_estimator, steps_predictor, temporal_cell,
where_update_scale=1.0, debug=False):
"""Initialises the model.
If argument is not covered here, see BaseSQAIRCore for documentation.
:param temporal_cell: RNNCore for the temporal rnn.
:param where_update_scale: Float, rescales the update of the `where` variables.
"""
#inhereting from BaseSQAIRCore
super(PropagationCore, self).__init__(img_size, crop_size, n_what,
transition, input_encoder, glimpse_encoder, transform_estimator, steps_predictor, debug = debug)
#adding temporal RNN which we get as input parameter
self._temporal_cell = temporal_cell
with tf._enter_variable_scope():
#getting st deviation for where latent variable distribution
self._where_update_scale = tf.get_variable('where_update_scale', shape[], dtype = tf.float32,
initializer = tf.constant_initializer(where_update_scale), trainable = False)
#specifying distribution for where latent variables
self._where_distrib = AffineDiagNormal()
@property
def output_size(self):
return [
self._n_what, # what code
self._n_what, # what sample
self._n_what, # what loc
self._n_what, # what scale
self._n_transform_param, # where code
self._n_transform_param, # where sample
self._n_transform_param, # where loc
self._n_transform_param, # where scale
1, # presence prob
1, # presence
1, # presence_logit,
self._temporal_cell.state_size,
]
def _build(self, (z_tm1, temporal_hidden_state), state):
"""Input is unused; it's only to force a maximum number of steps"""
# same object, previous timestep
#splitting z_tm1 into what where and presense along with the presense logits
#getting the object from the previous time step
what_tm1, where_tm1, presence_tm1, presence_logit_tm1 = z_tm1
#getting numpy list of flattened temporal_hidden_state
temporal_state = nest.flatten(temporal_hidden_state)[-1]
#initialize another object itself,and latent variables for another object at the current time step
img_flat, what_km1, where_km1, presence_km1, hidden state = state
#transforming vector of image pixels into matrix of pixels
img = tf.reshape(img_flat, (-1,) + tuple(self._img_size))
with tf.variable_scope('rnn_inpt'):
#achieving the bias value that we add to where latent variable from the previous time step in order to get new
#proposal where latent variable for the current step
where_bias = MLP(128, n_out=4)(temporal_state) * .1
#extracting and encoding proposal glimpse(includes spatial transformer inside)
what_distrib = self._glimpse_encoder(img, where_tm1 + where_bias, mask_inpt=temporal_state)[0]
#taking the mean of what dustribution for the object which corresponds to our object
rnn_inpt = what_distrib.loc
#constructing the input to relational RNN
rnn_inpt = [
rnn_inpt, # img
what_km1, where_km1, presence_km1, # explaining away
what_tm1, where_tm1, presence_tm1, temporal_state # previous state
]
#making tensor from array
rnn_inpt = tf.concat(rnn_inpt, -1)
#getting the output from relational RNN
hidden_output, hidden_state = self._cell(rnn_inpt, hidden_state)
#sample latent variable 'where' for the current time step using 'where' from the previous time step and weights from
#relational rnn
with tf.variable_scope('where'):
where, where_sample, where_loc, where_scale = self._compute_where(where_tm1, hidden_output, temporal_state)
#sample latent variable 'what' for the current time step using 'what' from the previous time step and weights from
#relational and temporal rnns alonf with encoded glimpse of the object
with tf.variable_scope('what'):
what, what_sample, what_loc, what_scale, temporal_hidden_state\
= self._compute_what(img, what_tm1, where, hidden_output, temporal_hidden_state, temporal_state)
#compute the presense of the oobject for the current time step using
#presence from the previous time step, 'what' and 'where' latent variable from the current timestep
#and weights from relational and temporal RNNs
with tf.variable_scope('presence'):
presence, presence_prob, presence_logit \
= self._compute_presence(presence_tm1, presence_logit_tm1, hidden_output, temporal_state, what)
output = [what, what_sample, what_loc, what_scale, where, where_sample, where_loc, where_scale,
presence_prob, presence, presence_logit, temporal_hidden_state]
new_state = [img_flat, what, where, presence, hidden_state]
return output, new_state
def build(self):
with tf.variable_scope("Embeddings"):
self.embeddings = tf.get_variable("emb", [self.config.n_embed, self.config.d_embed], dtype=tf.float64,
initializer=self.xavier_init)
embeddings_root = tf.get_variable("emb_root", [1, 1, 2 * self.config.dim_sem], dtype=tf.float64,
initializer=self.xavier_init)
embeddings_root_s = tf.get_variable("emb_root_s", [1, 1,2* self.config.dim_sem], dtype=tf.float64,
initializer=self.xavier_init)
with tf.variable_scope("Model"):
w_comb = tf.get_variable("w_comb", [4 * self.config.dim_sem, 2 * self.config.dim_sem], dtype=tf.float64,
initializer=self.xavier_init)
w_comb_both = tf.get_variable("w_comb_both", [6 * self.config.dim_sem, 2 * self.config.dim_sem], dtype=tf.float64,
initializer=self.xavier_init)
b_comb = tf.get_variable("bias_comb", [2 * self.config.dim_sem], dtype=tf.float64, initializer=tf.constant_initializer())
w_comb_s = tf.get_variable("w_comb_s", [4 * self.config.dim_sem, 2 * self.config.dim_sem], dtype=tf.float64,
initializer=self.xavier_init)
b_comb_s = tf.get_variable("bias_comb_s", [2 * self.config.dim_sem], dtype=tf.float64, initializer=tf.constant_initializer())
w_softmax = tf.get_variable("w_softmax", [2 * self.config.dim_sem, self.config.dim_output], dtype=tf.float64,
initializer=self.xavier_init)
b_softmax = tf.get_variable("bias_softmax", [self.config.dim_output], dtype=tf.float64,
initializer=self.xavier_init)
w_sem_doc = tf.get_variable("w_sem_doc", [2 * self.config.dim_sem, 2 * self.config.dim_sem], dtype=tf.float64,
initializer=self.xavier_init)
w_str_doc = tf.get_variable("w_str_doc", [2 * self.config.dim_sem, 2 * self.config.dim_str], dtype=tf.float64,
initializer=self.xavier_init)
with tf.variable_scope("Structure/doc"):
tf.get_variable("w_parser_p", [2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("w_parser_c", [2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("w_parser_s", [2 * self.config.dim_str, 2 * self.config.dim_str], dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("bias_parser_p", [2 * self.config.dim_str], dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("bias_parser_c", [2 * self.config.dim_str], dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("w_parser_root", [2 * self.config.dim_str, 1], dtype=tf.float64,
initializer=self.xavier_init)
with tf.variable_scope("Structure/sent"):
tf.get_variable("w_parser_p", [2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("w_parser_c", [2 * self.config.dim_str, 2 * self.config.dim_str],
dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("bias_parser_p", [2 * self.config.dim_str], dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("bias_parser_c", [2 * self.config.dim_str], dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("w_parser_s", [2 * self.config.dim_str, 2 * self.config.dim_str], dtype=tf.float64,
initializer=self.xavier_init)
tf.get_variable("w_parser_root", [2 * self.config.dim_str, 1], dtype=tf.float64,
initializer=self.xavier_init)
sent_l = self.t_variables['sent_l']
doc_l = self.t_variables['doc_l']
max_sent_l = self.t_variables['max_sent_l']
max_doc_l = self.t_variables['max_doc_l']
batch_l = self.t_variables['batch_l']
tokens_input = tf.nn.embedding_lookup(self.embeddings, self.t_variables['token_idxs'][:, :max_doc_l, :max_sent_l])
tokens_input = tf.nn.dropout(tokens_input, self.t_variables['keep_prob']) # [batch_size, doc_l, sent_l, d_embed]
mask_tokens = self.t_variables['mask_tokens'][:, :max_doc_l, :max_sent_l]
mask_sents = self.t_variables['mask_sents'][:, :max_doc_l] # [batch_size, doc_l]
tokens_input_do = tf.reshape(tokens_input, [batch_l * max_doc_l, max_sent_l, self.config.d_embed])
sent_l = tf.reshape(sent_l, [batch_l * max_doc_l])
mask_tokens = tf.reshape(mask_tokens, [batch_l * max_doc_l, -1])
tokens_output, _ = dynamicBiRNN(tokens_input_do, sent_l, n_hidden=self.config.dim_hidden, xavier_init=self.xavier_init,
cell_type=self.config.rnn_cell, cell_name='Model/sent')
tokens_sem = tf.concat([tokens_output[0][:,:,:self.config.dim_sem], tokens_output[1][:,:,:self.config.dim_sem]], 2)
tokens_str = tf.concat([tokens_output[0][:,:,self.config.dim_sem:], tokens_output[1][:,:,self.config.dim_sem:]], 2)
if self.config.skip_sent_attention:
tokens_output = LReLu(tf.tensordot(tf.concat([tokens_sem, tokens_input_do], 2), w_comb_s, [[2], [0]]) + b_comb_s)
else:
temp1 = tf.zeros([batch_l * max_doc_l, max_sent_l,1], tf.float64)
temp2 = tf.zeros([batch_l * max_doc_l,1,max_sent_l], tf.float64)
mask1 = tf.ones([batch_l * max_doc_l, max_sent_l, max_sent_l-1], tf.float64)
mask2 = tf.ones([batch_l * max_doc_l, max_sent_l-1, max_sent_l], tf.float64)
mask1 = tf.concat([temp1,mask1],2)
mask2 = tf.concat([temp2,mask2],1)
if self.config.skip_mask_bug_fix:
str_scores_s_, _, LL_tokens = get_structure('sent', tokens_str, mask1, mask2, None, None, None) # batch_l, sent_l+1, sent_l
else:
# create mask for setting all padded cells to 0
mask_ll_tokens = tf.expand_dims(mask_tokens, 2)
mask_ll_tokens_trans = tf.transpose(mask_ll_tokens, perm=[0, 2, 1])
mask_ll_tokens = mask_ll_tokens
mask_tokens_mult = mask_ll_tokens * mask_ll_tokens_trans
# create mask for setting the padded diagonals to 1
mask_diags = tf.matrix_diag_part(mask_tokens_mult)
mask_diags_invert = tf.cast(tf.logical_not(tf.cast(mask_diags, tf.bool)), tf.float64)
zero_matrix = tf.zeros([batch_l * max_doc_l, max_sent_l, max_sent_l], tf.float64)
mask_tokens_add = tf.matrix_set_diag(zero_matrix, mask_diags_invert)
str_scores_s_, _, LL_tokens = get_structure('sent', tokens_str, mask1, mask2, mask_tokens_mult,
mask_tokens_add, tf.expand_dims(mask_tokens,
2)) # batch_l, sent_l+1, sent_l
str_scores_s = tf.matrix_transpose(str_scores_s_) # soft parent
tokens_sem_root = tf.concat([tf.tile(embeddings_root_s, [batch_l * max_doc_l, 1, 1]), tokens_sem], 1)
tokens_output_ = tf.matmul(str_scores_s, tokens_sem_root)
tokens_output = LReLu(tf.tensordot(tf.concat([tokens_sem, tokens_output_], 2), w_comb_s, [[2], [0]]) + b_comb_s)
if (self.config.sent_attention == 'sum'):
tokens_output = tokens_output * tf.expand_dims(mask_tokens,2)
tokens_output = tf.reduce_sum(tokens_output, 1)
elif (self.config.sent_attention == 'mean'):
tokens_output = tokens_output * tf.expand_dims(mask_tokens,2)
tokens_output = tf.reduce_sum(tokens_output, 1)/tf.expand_dims(tf.cast(sent_l,tf.float64),1)
elif (self.config.sent_attention == 'max'):
tokens_output = tokens_output + tf.expand_dims((mask_tokens-1)*999,2)
tokens_output = tf.reduce_max(tokens_output, 1)
# batch_l * max_doc_l, 200
if self.config.skip_doc_bilstm:
if self.config.use_positional_encoding:
tokens_output = tf.reshape(tokens_output, [batch_l, max_doc_l, 2 * self.config.dim_sem])
tokens_output = self.add_timing_signal(tokens_output, max_doc_l, num_timescales=self.config.dim_sem)
tokens_output = tf.reshape(tokens_output, [batch_l * max_doc_l, 2 * self.config.dim_sem])
sents_sem = tf.matmul(tokens_output, w_sem_doc)
sents_sem = tf.reshape(sents_sem, [batch_l, max_doc_l, 2 * self.config.dim_sem])
sents_str = tf.matmul(tokens_output, w_str_doc)
sents_str = tf.reshape(sents_str, [batch_l, max_doc_l, 2 * self.config.dim_str])
else:
sents_input = tf.reshape(tokens_output, [batch_l, max_doc_l, 2 * self.config.dim_sem])
sents_output, _ = dynamicBiRNN(sents_input, doc_l, n_hidden=self.config.dim_hidden, xavier_init=self.xavier_init,
cell_type=self.config.rnn_cell, cell_name='Model/doc')
sents_sem = tf.concat([sents_output[0][:,:,:self.config.dim_sem], sents_output[1][:,:,:self.config.dim_sem]], 2) # [batch_l, doc+l, dim_sem*2]
sents_str = tf.concat([sents_output[0][:,:,self.config.dim_sem:], sents_output[1][:,:,self.config.dim_sem:]], 2) # [batch_l, doc+l, dim_str*2]
if self.config.skip_doc_attention:
if self.config.skip_doc_bilstm:
sents_input = tf.reshape(tokens_output, [batch_l, max_doc_l, 2 * self.config.dim_sem])
sents_output = LReLu(tf.tensordot(tf.concat([sents_sem, sents_input], 2), w_comb, [[2], [0]]) + b_comb)
else:
sents_output = LReLu(tf.tensordot(tf.concat([sents_sem, sents_input], 2), w_comb, [[2], [0]]) + b_comb)
else:
if self.config.skip_mask_bug_fix:
str_scores_, str_scores_no_root, LL_sents = get_structure('doc', sents_str, self.t_variables['mask_parser_1'],
self.t_variables['mask_parser_2'], None, None, None) # [batch_size, doc_l+1, doc_l]
else:
# create mask for setting all padded cells to 0
mask_ll_sents = tf.expand_dims(mask_sents, 2)
mask_ll_sents_trans = tf.transpose(mask_ll_sents, perm=[0, 2, 1])
mask_ll_sents = mask_ll_sents
mask_sents_mult = mask_ll_sents * mask_ll_sents_trans
# create mask for setting the padded diagonals to 1
mask_sents_diags = tf.matrix_diag_part(mask_sents_mult)
mask_sents_diags_invert = tf.cast(tf.logical_not(tf.cast(mask_sents_diags, tf.bool)), tf.float64)
zero_matrix_sents = tf.zeros([batch_l, max_doc_l, max_doc_l], tf.float64)
mask_sents_add = tf.matrix_set_diag(zero_matrix_sents, mask_sents_diags_invert)
str_scores_, str_scores_no_root, LL_sents = get_structure('doc', sents_str, self.t_variables['mask_parser_1'],
self.t_variables['mask_parser_2'], mask_sents_mult,
mask_sents_add, tf.expand_dims(mask_sents,
2)) # [batch_size, doc_l+1, doc_l]
str_scores = tf.matrix_transpose(str_scores_)
self.str_scores = str_scores # shape is [batch_size, doc_l, doc_l+1]
sents_children = tf.matmul(str_scores_no_root, sents_sem)
if self.config.tree_percolation == "child":
sents_output = LReLu(tf.tensordot(tf.concat([sents_sem, sents_children], 2), w_comb, [[2], [0]]) + b_comb)
else:
sents_sem_root = tf.concat([tf.tile(embeddings_root, [batch_l, 1, 1]), sents_sem], 1)
sents_parents = tf.matmul(str_scores, sents_sem_root)
if self.config.tree_percolation == "parent":
sents_output = LReLu(tf.tensordot(tf.concat([sents_sem, sents_parents], 2), w_comb, [[2], [0]]) + b_comb)
elif self.config.tree_percolation == "both":
sents_output = LReLu(tf.tensordot(tf.concat([sents_sem, sents_parents, sents_children], 2), w_comb_both, [[2], [0]]) + b_comb)
# percolation is only supported for "child" option
if self.config.tree_percolation_levels > 0:
count = 0
while count < self.config.tree_percolation_levels:
sents_children_2 = tf.matmul(str_scores_no_root, sents_output)
sents_output = LReLu(tf.tensordot(tf.concat([sents_output, sents_children_2], 2), w_comb, [[2], [0]]) + b_comb)
count += 1
if (self.config.doc_attention == 'sum'):
sents_output = sents_output * tf.expand_dims(mask_sents, 2) # mask is [batch_size, doc_l, 1]
sents_output = tf.reduce_sum(sents_output, 1) # [batch_size, dim_sem*2]
elif (self.config.doc_attention == 'mean'):
sents_output = sents_output * tf.expand_dims(mask_sents, 2)
sents_output = tf.reduce_sum(sents_output, 1)/tf.expand_dims(tf.cast(doc_l,tf.float64),1)
elif (self.config.doc_attention == 'max'):
sents_output = sents_output + tf.expand_dims((mask_sents-1)*999,2)
sents_output = tf.reduce_max(sents_output, 1)
elif (self.config.doc_attention == 'weighted_sum'):
sents_weighted = sents_output * tf.expand_dims(str_scores[:,:,0], 2)
sents_output = sents_weighted * tf.expand_dims(mask_sents, 2) # apply mask
sents_output = tf.reduce_sum(sents_output, 1)
final_output = MLP(sents_output, 'output', self.t_variables['keep_prob'], self.config.seed, self.xavier_init)
self.final_output = tf.matmul(final_output, w_softmax) + b_softmax
