def inference(inputs, keep_prob, is_training=True):
# initialization
# h1_out = affine_transform(inputs, num_hidden_1, name="hidden_1")
h1_out = utils.batch_norm_affine_transform(inputs,
num_hidden_1,
name="hidden_1",
decay=decay,
is_training=is_training)
h1_out = tf.nn.relu(h1_out)
h1_out = tf.nn.dropout(h1_out, keep_prob=keep_prob)
# h2_out = utils.batch_norm_affine_transform(h1_out, num_hidden_2, name="hidden_2")
h2_out = utils.batch_norm_affine_transform(h1_out,
num_hidden_2,
name="hidden_2",
decay=decay,
is_training=is_training)
h2_out = tf.nn.relu(h2_out)
h2_out = tf.nn.dropout(h2_out, keep_prob=keep_prob)
logits = affine_transform(h2_out, bdnn_outputsize, name="output")
logits = tf.sigmoid(logits)
logits = tf.reshape(logits, [-1, int(bdnn_outputsize)])
return logits
def action_network(self, outputs):
is_training = self.is_training
with tf.variable_scope("action_network"):
h1_out = tf.nn.relu(
utils.batch_norm_affine_transform(outputs,
action_hidden_1,
decay=decay,
name='action_hidden_1',
seed=SEED,
is_training=is_training))
h1_out = tf.nn.dropout(h1_out,
keep_prob=self.keep_probability,
seed=SEED)
h2_out = tf.nn.relu(
utils.batch_norm_affine_transform(h1_out,
action_hidden_2,
decay=decay,
name='action_hidden_2',
seed=SEED,
is_training=is_training))
h2_out = tf.nn.dropout(h2_out,
keep_prob=self.keep_probability,
seed=SEED)
return h2_out
def get_next_input(self, cell_output, reuse=None):
raw_inputs = self.inputs
is_training = self.is_training
with tf.variable_scope("baseline", reuse=reuse):
baseline = tf.sigmoid(affine_transform(((cell_output)), 1, name='baseline'))
self.baselines.append(baseline)
with tf.variable_scope("selection_network", reuse=reuse):
mean_bp = smooth_softmax(
utils.batch_norm_affine_transform(cell_output, int(bdnn_winlen), decay=decay, name='selection',
is_training=is_training))
# mean_bp = softmax(
# utils.batch_norm_affine_transform(cell_output, int(bdnn_winlen), decay=decay, name='selection',
# is_training=is_training), beta)
self.mean_bps.append(mean_bp)
# rand_seq = tf.random_uniform(mean_bp.get_shape().as_list(), minval=0, maxval=1, seed=SEED)
if is_training:
sampled_bp = tf.multinomial(mean_bp, num_samples=1, seed=SEED)
sampled_bp = utils.onehot_tensor(sampled_bp, bdnn_winlen)
else:
sampled_bp = mean_bp
sampled_bp = tf.stop_gradient(sampled_bp)
self.sampled_bps.append(sampled_bp)
return self.get_glimpse(raw_inputs, mean_bp, reuse=True)
def inference(self, inputs, keep_prob, is_training=True):
# initialization
# h1_out = affine_transform(inputs, num_hidden_1, name="hidden_1")
h1_out = utils.batch_norm_affine_transform(inputs, num_hidden_1, name="hidden_1", decay=decay,
is_training=is_training)
h1_out = tf.nn.relu(h1_out)
h1_out = tf.nn.dropout(h1_out, keep_prob=keep_prob)
# h2_out = utils.batch_norm_affine_transform(h1_out, num_hidden_2, name="hidden_2")
h2_out = utils.batch_norm_affine_transform(h1_out, num_hidden_2, name="hidden_2", decay=decay,
is_training=is_training)
h2_out = tf.nn.relu(h2_out)
h2_out = tf.nn.dropout(h2_out, keep_prob=keep_prob)
logits = affine_transform(h2_out, 2, name="output")
return logits
def get_glimpse(self, inputs, bp, reuse=None):
is_training = self.is_training
with tf.variable_scope("glimpse_net", reuse=reuse):
glimpse_input = sw_sensor(inputs, bp)
act_glimpse_hidden = tf.nn.relu(
utils.batch_norm_affine_transform(glimpse_input, glimpse_hidden, decay=decay,
name='glimpse_hidden', seed=SEED, is_training=is_training))
act_bp_hidden = tf.nn.relu(utils.batch_norm_affine_transform(bp, bp_hidden, decay=decay, name='bp_hidden',
seed=SEED, is_training=is_training))
glimpse_feature = tf.nn.relu(utils.batch_norm_affine_transform(act_glimpse_hidden, glimpse_out, decay=decay,
name='glimpse_out',
seed=SEED, is_training=is_training) +
utils.batch_norm_affine_transform(act_bp_hidden, bp_out, decay=decay,
seed=SEED, name='bp_out', is_training=is_training))
return glimpse_feature