import tensorflow as tf from tensorflow.contrib.layers import l2_regularizer import numpy as np import time import os os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" #### id of gpu to use os.environ['CUDA_VISIBLE_DEVICES'] = "0" os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' #### training data #### shape=(n_sample, n_code=854) REAL = np.load('') #### demographic for training data #### shape=(n_sample, 6) #### if sample_x is male, then LABEL[x,0]=1, else LABEL[x,1]=1 #### if sample_x's is within 0-17, then LABEL[x,2]=1 #### elif sample_x's is within 18-44, then LABEL[x,3]=1 #### elif sample_x's is within 45-64, then LABEL[x,4]=1 #### elif sample_x's is within 64-, then LABEL[x,5]=1 LABEL = np.load('') #### training parameters NUM_GPUS = 1 BATCHSIZE_PER_GPU = 2000 TOTAL_BATCHSIZE = BATCHSIZE_PER_GPU * NUM_GPUS STEPS_PER_EPOCH = int(np.load('ICD9/train.npy').shape[0] / 2000) g_structure = [128, 128] d_structure = [854, 256, 128] z_dim = 128 def _variable_on_cpu(name, shape, initializer=None): with tf.device('/cpu:0'): var = tf.get_variable(name, shape, initializer=initializer) return var def batchnorm(inputs, name, labels=None, n_labels=None): mean, var = tf.nn.moments(inputs, [0], keep_dims=True) shape = mean.shape[1].value offset_m = _variable_on_cpu(shape=[n_labels,shape], name='offset'+name, initializer=tf.zeros_initializer) scale_m = _variable_on_cpu(shape=[n_labels,shape], name='scale'+name, initializer=tf.ones_initializer) offset = tf.nn.embedding_lookup(offset_m, labels) scale = tf.nn.embedding_lookup(scale_m, labels) result = tf.nn.batch_normalization(inputs, mean, var, offset, scale, 1e-8) return result def layernorm(inputs, name, labels=None, n_labels=None): mean, var = tf.nn.moments(inputs, [1], keep_dims=True) shape = inputs.shape[1].value offset_m = _variable_on_cpu(shape=[n_labels,shape], name='offset'+name, initializer=tf.zeros_initializer) scale_m = _variable_on_cpu(shape=[n_labels,shape], name='scale'+name, initializer=tf.ones_initializer) offset = tf.nn.embedding_lookup(offset_m, labels) scale = tf.nn.embedding_lookup(scale_m, labels) result = tf.nn.batch_normalization(inputs, mean, var, offset, scale, 1e-8) return result def input_fn(): features_placeholder = tf.placeholder(shape=REAL.shape, dtype=tf.float32) labels_placeholder = tf.placeholder(shape=LABEL.shape, dtype=tf.float32) dataset = tf.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder)) dataset = dataset.repeat(10000) dataset = dataset.batch(batch_size=BATCHSIZE_PER_GPU) dataset = dataset.prefetch(1) iterator = dataset.make_initializable_iterator() # next_element = iterator.get_next() # init_op = iterator.initializer return iterator, features_placeholder, labels_placeholder def generator(z, label): x = z tmp_dim = z_dim with tf.variable_scope('G', reuse=tf.AUTO_REUSE, regularizer=l2_regularizer(0.00001)): for i, dim in enumerate(g_structure[:-1]): kernel = _variable_on_cpu('W_' + str(i), shape=[tmp_dim, dim]) h1 = batchnorm(tf.matmul(x, kernel), name='cbn' + str(i), labels=label, n_labels=8) h2 = tf.nn.relu(h1) x = x + h2 tmp_dim = dim i = len(g_structure) - 1 kernel = _variable_on_cpu('W_' + str(i), shape=[tmp_dim, g_structure[-1]]) h1 = batchnorm(tf.matmul(x, kernel), name='cbn' + str(i), labels=label, n_labels=8) h2 = tf.nn.tanh(h1) x = x + h2 kernel = _variable_on_cpu('W_' + str(i+1), shape=[128, 854]) bias = _variable_on_cpu('b_' + str(i+1), shape=[854]) x = tf.nn.sigmoid(tf.add(tf.matmul(x, kernel), bias)) return x def discriminator(x, label): with tf.variable_scope('D', reuse=tf.AUTO_REUSE, regularizer=l2_regularizer(0.00001)): for i, dim in enumerate(d_structure[1:]): kernel = _variable_on_cpu('W_' + str(i), shape=[d_structure[i], dim]) bias = _variable_on_cpu('b_' + str(i), shape=[dim]) x = tf.nn.relu(tf.add(tf.matmul(x, kernel), bias)) x = layernorm(x, name='cln' + str(i), labels=label, n_labels=8) i = len(d_structure) kernel = _variable_on_cpu('W_' + str(i), shape=[d_structure[-1], 1]) bias = _variable_on_cpu('b_' + str(i), shape=[1]) y = tf.add(tf.matmul(x, kernel), bias) return y def compute_dloss(real, fake, label): epsilon = tf.random_uniform( shape=[BATCHSIZE_PER_GPU, 1], minval=0., maxval=1.) x_hat = real + epsilon * (fake - real) y_hat_fake = discriminator(fake, label) y_hat_real = discriminator(real, label) y_hat = discriminator(x_hat, label) grad = tf.gradients(y_hat, [x_hat])[0] slopes = tf.sqrt(tf.reduce_sum(tf.square(grad), 1)) gradient_penalty = tf.reduce_mean((slopes - 1.) ** 2) all_regs = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES) w_distance = -tf.reduce_mean(y_hat_real) + tf.reduce_mean(y_hat_fake) loss = w_distance + 10 * gradient_penalty + sum(all_regs) tf.add_to_collection('dlosses', loss) return w_distance, loss def compute_gloss(fake, label): y_hat_fake = discriminator(fake, label) all_regs = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES) loss = -tf.reduce_mean(y_hat_fake) + sum(all_regs) tf.add_to_collection('glosses', loss) return loss, loss def tower_loss(scope, stage, real, label): label = tf.cast(label, tf.int32) label = label[:, 1] * 4 + tf.squeeze( tf.matmul(label[:, 2:], tf.constant([[0], [1], [2], [3]], dtype=tf.int32))) z = tf.random_normal(shape=[BATCHSIZE_PER_GPU, z_dim]) fake = generator(z, label) if stage == 'D': w, loss = compute_dloss(real, fake, label) losses = tf.get_collection('dlosses', scope) else: w, loss = compute_gloss(fake, label) losses = tf.get_collection('glosses', scope) total_loss = tf.add_n(losses, name='total_loss') # loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg') # loss_averages_op = loss_averages.apply(losses + [total_loss]) # # with tf.control_dependencies([loss_averages_op]): # total_loss = tf.identity(total_loss) return total_loss, w def average_gradients(tower_grads): average_grads = [] for grad_and_vars in zip(*tower_grads): grads = [] for g, _ in grad_and_vars: expanded_g = tf.expand_dims(g, 0) grads.append(expanded_g) grad = tf.concat(axis=0, values=grads) grad = tf.reduce_mean(grad, 0) v = grad_and_vars[0][1] grad_and_var = (grad, v) average_grads.append(grad_and_var) return average_grads def graph(stage, opt): # global_step = tf.get_variable(stage+'_step', [], initializer=tf.constant_initializer(0), trainable=False) tower_grads = [] per_gpu_w = [] iterator, features_placeholder, labels_placeholder = input_fn() with tf.variable_scope(tf.get_variable_scope()): for i in range(NUM_GPUS): with tf.device('/gpu:%d' % i): with tf.name_scope('%s_%d' % ('TOWER', i)) as scope: (real, label) = iterator.get_next() loss, w = tower_loss(scope, stage, real, label) tf.get_variable_scope().reuse_variables() vars_ = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=stage) grads = opt.compute_gradients(loss, vars_) tower_grads.append(grads) per_gpu_w.append(w) grads = average_gradients(tower_grads) apply_gradient_op = opt.apply_gradients(grads) mean_w = tf.reduce_mean(per_gpu_w) train_op = apply_gradient_op return train_op, mean_w, iterator, features_placeholder, labels_placeholder def train(max_epochs, train_dir): with tf.device('/cpu:0'): opt_d = tf.train.AdamOptimizer(1e-4) opt_g = tf.train.AdamOptimizer(1e-4) train_d, w_distance, iterator_d, features_placeholder_d, labels_placeholder_d = graph('D', opt_d) train_g, _, iterator_g, features_placeholder_g, labels_placeholder_g = graph('G', opt_g) saver = tf.train.Saver() init = tf.global_variables_initializer() with tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False)) as sess: sess.run(init) sess.run(iterator_d.initializer, feed_dict={features_placeholder_d: REAL, labels_placeholder_d: LABEL}) sess.run(iterator_g.initializer, feed_dict={features_placeholder_g: REAL, labels_placeholder_g: LABEL}) for epoch in range(1, max_epochs + 1): start_time = time.time() w_sum = 0 for i in range(STEPS_PER_EPOCH): for _ in range(2): _, w = sess.run([train_d, w_distance]) w_sum += w sess.run(train_g) duration = time.time() - start_time assert not np.isnan(w_sum), 'Model diverged with loss = NaN' format_str = 'epoch: %d, w_distance = %f (%.1f)' print(format_str % (epoch, -w_sum/(STEPS_PER_EPOCH*2), duration)) if epoch % 500 == 0: # checkpoint_path = os.path.join(train_dir, 'multi') saver.save(sess, train_dir, write_meta_graph=False, global_step=epoch) # saver.save(sess, train_dir, global_step=epoch) def generate(model_dir, synthetic_dir, demo): tf.reset_default_graph() z = tf.random_normal(shape=[BATCHSIZE_PER_GPU, z_dim]) y = tf.placeholder(shape=[BATCHSIZE_PER_GPU, 6], dtype=tf.int32) label = y[:, 1] * 4 + tf.squeeze(tf.matmul(y[:, 2:], tf.constant([[0], [1], [2], [3]], dtype=tf.int32))) fake = generator(z, label) saver = tf.train.Saver() with tf.Session() as sess: saver.restore(sess, model_dir) for m in range(2): for n in range(2, 6): idx1 = (demo[:, m] == 1) idx2 = (demo[:, n] == 1) idx = [idx1[j] and idx2[j] for j in range(len(idx1))] num = np.sum(idx) nbatch = int(np.ceil(num / BATCHSIZE_PER_GPU)) label_input = np.zeros((nbatch*BATCHSIZE_PER_GPU, 6)) label_input[:, n] = 1 label_input[:, m] = 1 output = [] for i in range(nbatch): f = sess.run(fake,feed_dict={y: label_input[i*BATCHSIZE_PER_GPU:(i+1)*BATCHSIZE_PER_GPU]}) output.extend(np.round(f)) output = np.array(output)[:num] np.save(synthetic_dir + str(m) + str(n), output) if __name__ == '__main__': #### args_1: number of training epochs #### args_2: dir to save the trained model train(500, '') #### args_1: dir of trained model #### args_2: dir to save synthetic data #### args_3, label of data-to-be-generated generate('', '', demo=LABEL)