W1 = tf.Variable(w2v, trainable=True)
X_embedding = tf.nn.embedding_lookup(W1, X)
X_embedding = X_embedding[..., tf.newaxis]
if args.tf_embedding_type == 'multi-channel':
W1 = tf.Variable(w2v, trainable=True)
W2 = tf.Variable(w2v, trainable=False)
X_1 = tf.nn.embedding_lookup(W1, X)
X_2 = tf.nn.embedding_lookup(W2, X)
X_1 = X_1[..., tf.newaxis]
X_2 = X_2[..., tf.newaxis]
X_embedding = tf.concat([X_1, X_2], axis=-1)
tf.logging.info("input dimension:{}".format(X_embedding.get_shape()))
if args.tf_model_type == 'capsule-A':
poses, activations = capsule_model_A(X_embedding, args.num_classes)
if args.tf_model_type == 'capsule-B':
poses, activations = capsule_model_B(X_embedding, args.num_classes)
if args.tf_model_type == 'CNN':
poses, activations = baseline_model_cnn(X_embedding, args.num_classes)
if args.tf_model_type == 'KIMCNN':
poses, activations = baseline_model_kimcnn(X_embedding, args.max_sent,
args.num_classes)
if args.tf_loss_type == 'spread_loss':
loss = spread_loss(y, activations, margin)
if args.tf_loss_type == 'margin_loss':
loss = margin_loss(y, activations)
if args.tf_loss_type == 'cross_entropy':
loss = cross_entropy(y, activations)
W1 = tf.Variable(w2v, trainable = True)
X_embedding = tf.nn.embedding_lookup(W1, X)
X_embedding = X_embedding[...,tf.newaxis]
if args.embedding_type == 'multi-channel':
W1 = tf.Variable(w2v, trainable = True)
W2 = tf.Variable(w2v, trainable = False)
X_1 = tf.nn.embedding_lookup(W1, X)
X_2 = tf.nn.embedding_lookup(W2, X)
X_1 = X_1[...,tf.newaxis]
X_2 = X_2[...,tf.newaxis]
X_embedding = tf.concat([X_1,X_2],axis=-1)
tf.logging.info("input dimension:{}".format(X_embedding.get_shape()))
if args.model_type == 'capsule-A':
poses, activations = capsule_model_A(X_embedding, args.num_classes)
if args.model_type == 'capsule-B':
poses, activations = capsule_model_B(X_embedding, args.num_classes)
if args.model_type == 'CNN':
poses, activations = baseline_model_cnn(X_embedding, args.num_classes)
if args.model_type == 'KIMCNN':
poses, activations = baseline_model_kimcnn(X_embedding, args.max_sent, args.num_classes)
if args.loss_type == 'spread_loss':
loss = spread_loss(y, activations, margin)
if args.loss_type == 'margin_loss':
loss = margin_loss(y, activations)
if args.loss_type == 'cross_entropy':
loss = cross_entropy(y, activations)
y_pred = tf.argmax(activations, axis=1, name="y_proba")