def train_interactiongraph(subgraph):
linear1 = fully_connected_layer.apply(subgraph["X_1"], [1], subgraph,
relu_in=False, relu_mid=False, relu_out=False,
dropout_in=None, dropout_mid=None, dropout_out=None,
bias_in=True, bias_mid=True, bias_out=True, batch_norm=False,
train=False, l2_reg=l2_reg, scope="LinearComponent")
linear1 = tf.reshape(linear1, shape=[-1]) # shaped [None, ]
interactive1 = fm_layer.apply(subgraph["X_1"], factor_dim, l2_weight=0.01, scope="InteractiveComponent")
interactive1 = tf.reshape(tf.math.reduce_sum(interactive1, axis=1), shape=[-1])
linear2 = fully_connected_layer.apply(subgraph["X_2"], [1], subgraph,
relu_in=False, relu_mid=False, relu_out=False,
dropout_in=None, dropout_mid=None, dropout_out=None,
bias_in=True, bias_mid=True, bias_out=True, batch_norm=False,
train=False, l2_reg=l2_reg, scope="LinearComponent")
linear2 = tf.reshape(linear2, shape=[-1]) # shaped [None, ]
interactive2 = fm_layer.apply(subgraph["X_2"], factor_dim, l2_weight=0.01, scope="InteractiveComponent")
interactive2 = tf.reshape(tf.math.reduce_sum(interactive2, axis=1), shape=[-1])
dy_tilde = (linear1 + interactive1) - (linear2 + interactive2)
loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=subgraph["dy"], logits=dy_tilde, name="loss")
subgraph.register_global_loss(tf.reduce_mean(loss))
tf.summary.scalar('loss', tf.reduce_mean(loss))
summary = tf.summary.merge_all()
subgraph.register_global_summary(summary)
subgraph.register_global_output(subgraph["dy"])
subgraph.register_global_output(dy_tilde)
pass
def apply(in_tensor,
dims,
subgraph,
item_bias=None,
extra=None,
l2_reg=None,
labels=None,
dropout=None,
train=None,
scope=None):
"""
MLP softmax layer
final layer, registers loss for training and prediction for serving
"""
with tf.variable_scope(scope,
default_name='MLPSoftmax',
reuse=tf.AUTO_REUSE):
if train:
logits = fully_connected_layer.apply(in_tensor=in_tensor,
dims=dims,
subgraph=subgraph,
bias_in=True,
bias_mid=True,
bias_out=False,
dropout_mid=dropout,
l2_reg=l2_reg,
scope='mlp_reg')
else:
logits = fully_connected_layer.apply(in_tensor=in_tensor,
dims=dims,
subgraph=subgraph,
bias_in=True,
bias_mid=True,
bias_out=False,
l2_reg=l2_reg,
scope='mlp_reg')
if item_bias is not None:
logits += item_bias
if train:
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=labels, logits=logits)
subgraph.register_global_loss(tf.reduce_mean(loss))
else:
subgraph.register_global_output(logits)
def train_fushion_graph(subgraph):
logits_1 = fully_connected_layer.apply(subgraph['X1'], [1],
subgraph,
relu_in=False,
relu_mid=False,
relu_out=False,
dropout_in=None,
dropout_mid=None,
dropout_out=None,
bias_in=True,
bias_mid=True,
bias_out=True,
batch_norm=False,
train=False,
l2_reg=l2_reg,
scope='Weights1dTensor')
logits_2 = fully_connected_layer.apply(subgraph['X2'], [1],
subgraph,
relu_in=False,
relu_mid=False,
relu_out=False,
dropout_in=None,
dropout_mid=None,
dropout_out=None,
bias_in=True,
bias_mid=True,
bias_out=True,
batch_norm=False,
train=False,
l2_reg=l2_reg,
scope='Weights1dTensor')
dy_tilde = tf.squeeze(logits_1 - logits_2)
loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=subgraph['dy'],
logits=dy_tilde,
name='loss')
subgraph.register_global_loss(tf.reduce_mean(loss))
subgraph.register_global_output(subgraph['dy'])
subgraph.register_global_output(dy_tilde)
tf.summary.scalar('loss', tf.reduce_mean(loss))
summary = tf.summary.merge_all()
subgraph.register_global_summary(summary)
def serve_interactiongraph(subgraph):
linear = fully_connected_layer.apply(subgraph["X"], [1], subgraph,
relu_in=False, relu_mid=False, relu_out=False,
dropout_in=None, dropout_mid=None, dropout_out=None,
bias_in=True, bias_mid=True, bias_out=True, batch_norm=False,
train=False, l2_reg=l2_reg, scope="LinearComponent")
linear = tf.reshape(linear, shape=[-1]) # shaped [None, ]
interactive = fm_layer.apply(subgraph["X"], factor_dim, l2_weight=0.01, scope="InteractiveComponent") # shaped [None, factor_dim]
interactive = tf.reshape(tf.math.reduce_sum(interactive, axis=1), shape=[-1])
score = linear + interactive
subgraph.register_global_output(score)
pass
def serve_fusion_graph(subgraph):
logit = fully_connected_layer.apply(subgraph['X'], [1],
subgraph,
relu_in=False,
relu_mid=False,
relu_out=False,
dropout_in=None,
dropout_mid=None,
dropout_out=None,
bias_in=True,
bias_mid=True,
bias_out=True,
batch_norm=False,
train=False,
l2_reg=l2_reg,
scope='Weights1dTensor')
subgraph.register_global_output(logit)