import numpy as np import tensorflow as tf class OptimizerCycle(object): def __init__(self, preds, labels, pos_weight, norm, d_real, d_fake, GD_real, GD_fake, preds_z2g, labels_z2g, preds_cycle, labels_cycle, gradient, gradient_z, settings): preds_sub = preds labels_sub = labels self.real = d_real self.settings = settings # Discrimminator Loss self.dc_loss_real = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(self.real), logits=self.real, name='dclreal')) # self.dc_loss_real = - tf.reduce_mean(self.real) self.dc_loss_fake = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.zeros_like(d_fake), logits=d_fake, name='dcfake')) # self.dc_loss_fake = tf.reduce_mean(d_fake) # GP_loss = tf.reduce_mean(tf.square(tf.sqrt(tf.reduce_mean(tf.square(gradient), axis = [0, 1])) - 1)) # GP_loss_z = tf.reduce_mean(tf.square(tf.sqrt(tf.reduce_mean(tf.square(gradient_z), axis = [0, 1])) - 1)) # self.dc_loss = self.dc_loss_fake + self.dc_loss_real + 10.0 * GP_loss self.GD_loss_real = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(GD_real), logits=GD_real, name='GD_real')) # self.GD_loss_real = - tf.reduce_mean(GD_real) self.GD_loss_fake = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.zeros_like(GD_fake), logits=GD_fake, name='GD_fake')) # self.GD_loss_fake = tf.reduce_mean(GD_fake) self.dc_loss = self.dc_loss_fake + self.dc_loss_real self.GD_loss = self.GD_loss_fake + self.GD_loss_real # Generator loss generator_loss = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(d_fake), logits=d_fake, name='gl')) # generator_loss = -self.dc_loss_fake generator_loss_z2g = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(GD_fake), logits=GD_fake, name='G_z2g')) # generator_loss_z2g = -self.GD_loss_fake # pos_weight，允许人们通过向上或向下加权相对于负误差的正误差的成本来权衡召回率和精确度 self.cost = norm * tf.reduce_mean(tf.nn.weighted_cross_entropy_with_logits(logits=preds_sub, targets=labels_sub, pos_weight=pos_weight)) cost_cycle = norm * tf.reduce_mean(tf.square(preds_cycle - labels_cycle)) cost_z2g = norm * tf.reduce_mean(tf.square(preds_z2g - labels_z2g)) self.cost = self.cost + cost_cycle self.generator_loss = generator_loss + self.cost self.generator_loss_z2g = generator_loss_z2g all_variables = tf.trainable_variables() dc_var = [var for var in all_variables if 'dc_' in var.name] en_var = [var for var in all_variables if 'e_' in var.name] GG_var = [var for var in all_variables if 'GG' in var.name] GD_var = [var for var in all_variables if 'GD' in var.name] with tf.variable_scope(tf.get_variable_scope()): self.discriminator_optimizer = tf.train.AdamOptimizer(learning_rate=self.settings.discriminator_learning_rate, beta1=0.9, name='adam1').minimize(self.dc_loss, var_list=dc_var) # minimize(dc_loss_real, var_list=dc_var) self.generator_optimizer = tf.train.AdamOptimizer(learning_rate=self.settings.discriminator_learning_rate, beta1=0.9, name='adam2').minimize(self.generator_loss, var_list=en_var) self.discriminator_optimizer_z2g = tf.train.AdamOptimizer(learning_rate=self.settings.discriminator_learning_rate, beta1=0.9, name='adam1').minimize(self.GD_loss, var_list=GD_var) self.generator_optimizer_z2g = tf.train.AdamOptimizer(learning_rate=self.settings.discriminator_learning_rate, beta1=0.9, name='adam2').minimize(self.generator_loss_z2g, var_list=GG_var) # 值得注意的是，这个地方，除了对抗优化之外， # 还单纯用cost损失又优化了一遍， # 待会儿看训练的时候注意看是在哪部分进行的这部分优化操作 self.optimizer = tf.train.AdamOptimizer(learning_rate=self.settings.learning_rate) # Adam Optimizer self.opt_op = self.optimizer.minimize(self.cost) # self.grads_vars = self.optimizer.compute_gradients(self.cost) # self.optimizer_z2g = tf.train.AdamOptimizer(learning_rate=FLAGS.learning_rate) # Adam Optimizer # self.opt_op_z2g = self.optimizer.minimize(cost_z2g) # self.grads_vars_z2g = self.optimizer.compute_gradients(cost_z2g) class Optimizer(object): def __init__(self, model, model_z2g, D_Graph, discriminator, placeholders, pos_weight, norm, d_real, num_nodes, GD_real, settings): self.opt = self.construct_optimizer(model, model_z2g, D_Graph, discriminator, placeholders, pos_weight, norm, d_real, num_nodes, GD_real, settings) def construct_optimizer(self, model, model_z2g, D_Graph, discriminator, placeholders, pos_weight, norm, d_real, num_nodes, GD_real, settings): z2g = model_z2g.construct() hidden = z2g[1] z2g = z2g[0] preds_z2g = model.construct(hidden=hidden, reuse=True)[0] g2z = model.construct() embeddings = g2z[0] reconstructions = g2z[1] d_fake = discriminator.construct(embeddings, reuse=True) GD_fake = D_Graph.construct(z2g, reuse=True) epsilon = tf.random_uniform(shape=[1], minval=0.0, maxval=1.0) interpolated_input = epsilon * placeholders['real_distribution'] + (1 - epsilon) * embeddings gradient = tf.gradients(discriminator.construct(interpolated_input, reuse=True), [interpolated_input])[0] epsilon = tf.random_uniform(shape=[1], minval=0.0, maxval=1.0) interpolated_input = epsilon * placeholders['features_dense'] + (1 - epsilon) * z2g gradient_z = tf.gradients(D_Graph.construct(interpolated_input, reuse=True), [interpolated_input])[0] opt = OptimizerCycle(preds=reconstructions, labels=tf.reshape(tf.sparse_tensor_to_dense(placeholders['adj_orig'], validate_indices=False), [-1]), pos_weight=pos_weight, norm=norm, d_real=d_real, d_fake=d_fake, GD_real=GD_real, GD_fake=GD_fake, preds_z2g=preds_z2g, labels_z2g=placeholders['real_distribution'], preds_cycle=model_z2g.construct(embeddings, reuse=True)[0], labels_cycle=placeholders['features_dense'], gradient=gradient, gradient_z=gradient_z, settings=settings) return opt def construct_feed_dict(adj_normalized, adj, features, placeholders): # construct feed dictionary # .update()用法就是将()内的字段增加到dict当中 feed_dict = dict() # 创建一个空字典 feed_dict.update({placeholders['features']: features}) feed_dict.update({placeholders['adj']: adj_normalized}) feed_dict.update({placeholders['adj_orig']: adj}) return feed_dict def update(model, opt, sess, adj_norm, adj_label, features, placeholders, adj, distribution, adj_dense, settings): # Construct feed dictionary feed_dict = construct_feed_dict(adj_norm, adj_label, features, placeholders) feed_dict.update({placeholders['dropout']: settings.dropout}) feed_dict.update({placeholders['features_dense']: adj_dense}) feed_dict.update({placeholders['dropout']: 0}) z_real_dist = np.random.randn(adj.shape[0], settings.hidden2) z_real_dist = distribution.sample(adj.shape[0]) feed_dict.update({placeholders['real_distribution']: z_real_dist}) for j in range(5): _, reconstruct_loss = sess.run([opt.opt_op, opt.cost], feed_dict=feed_dict) g_loss, _ = sess.run([opt.generator_loss, opt.generator_optimizer], feed_dict=feed_dict) d_loss, _ = sess.run([opt.dc_loss, opt.discriminator_optimizer], feed_dict=feed_dict) GD_loss, _ = sess.run([opt.GD_loss, opt.discriminator_optimizer_z2g], feed_dict=feed_dict) GG_loss, _ = sess.run([opt.generator_loss_z2g, opt.generator_optimizer_z2g], feed_dict=feed_dict) # GD_loss = sess.run(opt.GD_loss, feed_dict=feed_dict) # GG_loss = sess.run(opt.generator_loss_z2g, feed_dict=feed_dict) # g_loss = sess.run(opt.generator_loss, feed_dict=feed_dict) # d_loss = sess.run(opt.dc_loss, feed_dict=feed_dict) emb = sess.run(model.z_mean, feed_dict=feed_dict) avg_cost = [reconstruct_loss, d_loss, g_loss, GD_loss, GG_loss] return emb, avg_cost