def construct(M, N):
# placeholders = {
# 'embedding': tf.placeholder(tf.float32, shape=(None, FLAGS.dim)),
# 'train_mask': tf.placeholder(tf.int32, shape=(N,)),
# 'labels': tf.placeholder(tf.float32, shape=(None, M))
#
# }
placeholders = {
# 'support': [tf.sparse_placeholder(tf.float32) for _ in range(num_supports)],
'support_': None,
'output_dim': 39,
}
if FLAGS.embed == 0 or FLAGS.embed == 3:
placeholders['embedding'] = tf.placeholder(tf.float32,
shape=(None, 100))
placeholders['train_mask'] = tf.placeholder(tf.int32, shape=(N, ))
placeholders['labels'] = tf.placeholder(tf.float32, shape=(None, 39))
optimizer = tf.train.AdamOptimizer(learning_rate=0.01)
preds = fully_connected(placeholders['embedding'],
M,
activation_fn=lambda x: x)
loss = masked_softmax_cross_entropy(preds, placeholders['labels'],
placeholders['train_mask'])
opt_op = optimizer.minimize(loss)
return placeholders, opt_op, loss, preds
def _loss(self):
# Weight decay loss
for var in self.layers[0].vars.values():
self.loss += FLAGS.weight_decay * tf.nn.l2_loss(var)
# Cross entropy error
self.loss += masked_softmax_cross_entropy(
self.outputs, self.placeholders['labels'],
self.placeholders['labels_mask'])
def set_loss(self):
"""
损失函数
"""
# Weight decay loss
with tf.name_scope("loss"):
self.loss = 0
for var in self.vars:
self.loss += self.params.weight_decay * tf.nn.l2_loss(var)
# Cross entropy error
self.loss += metrics.masked_softmax_cross_entropy(self.outputs, self.labels, self.labels_mask)
def _loss(self):
# Weight decay loss
for var in self.layers[0].vars.values():
self.loss += FLAGS.weight_decay * tf.nn.l2_loss(var)
# Cross entropy error
if self.categorical:
self.loss += metrics.masked_softmax_cross_entropy(self.outputs, self.placeholders['labels'],
self.placeholders['labels_mask'])
# L2
else:
diff = self.labels - self.outputs
self.loss += tf.reduce_sum(tf.sqrt(tf.reduce_sum(diff * diff, axis=1)))
def _loss(self):
"""Construct the loss function."""
# Weight decay loss
if FLAGS.weight_decay > 0.0:
for var in self.layers[0].vars.values():
self.loss += FLAGS.weight_decay * tf.nn.l2_loss(var)
# Cross entropy error
if self.multilabel:
self.loss += metrics.masked_sigmoid_cross_entropy(
self.outputs, self.placeholders['labels'],
self.placeholders['labels_mask'])
else:
self.loss += metrics.masked_softmax_cross_entropy(
self.outputs, self.placeholders['labels'],
self.placeholders['labels_mask'])
