基于注意力机制的疾病诊断预测模型

import tensorflow as tf class AttentionModel(object): def __init__(self, n_disease, n_drug, n_visit, n_disease_category, n_drug_category, n_embed=500, n_rnn=(200, 200), dropout_rate=0.2, batch_size=100): self.n_disease = n_disease self.n_drug = n_drug self.n_visit = n_visit self.n_disease_category = n_disease_category self.n_drug_category = n_drug_category self.n_embed = n_embed self.n_rnn = n_rnn self.dropout_rate = dropout_rate self.batch_size = batch_size def gru_unit(self, hidden_size, dropout_rate): gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size) gru_cell = tf.nn.rnn_cell.DropoutWrapper(cell=gru_cell, input_keep_prob=1.0, output_keep_prob=1 - dropout_rate) return gru_cell def gru_unit_0(self, hidden_size, dropout_rate): gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size, name="0") gru_cell = tf.nn.rnn_cell.DropoutWrapper(cell=gru_cell, input_keep_prob=1.0, output_keep_prob=1 - dropout_rate) return gru_cell # def build_model(self, x_disease, x_drug, disease_label, drug_label): # x_code = tf.concat([x_disease, x_drug], axis=2) # x_code = tf.reshape(x_code, [-1, self.n_disease + self.n_drug]) # w_embed = tf.get_variable('w_embed', # shape=[self.n_disease + self.n_drug, self.n_embed]) # b_embed = tf.get_variable('b_mebed', shape=[self.n_embed]) # hidden = tf.nn.relu(tf.add(tf.matmul(x_code, w_embed), b_embed)) # hidden = tf.reshape(hidden, [-1, self.n_visit, self.n_embed]) # stacked_gru = tf.nn.rnn_cell.MultiRNNCell( # [self.gru_unit(self.n_rnn[0], self.dropout_rate) for i in range(len(self.n_rnn))], state_is_tuple=True) # outputs, states = tf.nn.dynamic_rnn(cell=stacked_gru, inputs=hidden, dtype=tf.float32) # w_alpha = tf.get_variable('w_alpha', # shape=[self.n_rnn[-1], 1]) # b_alpha = tf.get_variable('b_alpha', shape=[1]) # w_code = tf.get_variable('w_code', # shape=[self.n_rnn[-1], self.n_disease_category + self.n_drug_category]) # b_code = tf.get_variable('b_code', shape=[self.n_disease_category + self.n_drug_category]) # code_loss = tf.Tensor # code_y_last = tf.Tensor # code_label = tf.concat([disease_label, drug_label], axis=2) # for i in range(1, self.n_visit): # outputs_reshape = tf.reshape(outputs, [-1, self.n_rnn[-1]]) # alpha = tf.reshape( # tf.reduce_mean(tf.nn.softmax(tf.reshape(tf.add(tf.matmul(outputs_reshape, w_alpha), b_alpha), # [-1, self.n_visit])[:, 0:i]), axis=1), [-1, 1]) # h_ = alpha * outputs[:, i, :] # code_loss = tf.reduce_sum( # tf.nn.softmax_cross_entropy_with_logits_v2(logits=tf.add(tf.matmul(h_, w_code), b_code), # labels=code_label[:, i, :])) # code_y_last = tf.nn.softmax(tf.add(tf.matmul(h_, w_code), b_code)) def build_model(self, x_disease, x_drug, disease_label, drug_label): x_code = tf.concat([x_disease, x_drug], axis=2) x_code = tf.reshape(x_code, [-1, self.n_disease + self.n_drug]) w_embed = tf.get_variable('w_embed', shape=[self.n_disease + self.n_drug, self.n_embed]) b_embed = tf.get_variable('b_mebed', shape=[self.n_embed]) hidden = tf.nn.relu(tf.add(tf.matmul(x_code, w_embed), b_embed)) hidden = tf.reshape(hidden, [-1, self.n_visit, self.n_embed]) stacked_gru = tf.nn.rnn_cell.MultiRNNCell( [self.gru_unit(self.n_rnn[0], self.dropout_rate) for i in range(len(self.n_rnn))], state_is_tuple=True) outputs, states = tf.nn.dynamic_rnn(cell=stacked_gru, inputs=hidden, dtype=tf.float32) stacked_gru_0 = tf.nn.rnn_cell.MultiRNNCell( [self.gru_unit_0(self.n_rnn[0], self.dropout_rate) for i in range(len(self.n_rnn))], state_is_tuple=True) outputs_0, states_0 = tf.nn.dynamic_rnn(cell=stacked_gru_0, inputs=hidden, dtype=tf.float32) w_alpha = tf.get_variable('w_alpha', shape=[self.n_rnn[-1], 1]) b_alpha = tf.get_variable('b_alpha', shape=[1]) w_code = tf.get_variable('w_code', shape=[self.n_rnn[-1], self.n_disease_category + self.n_drug_category]) b_code = tf.get_variable('b_code', shape=[self.n_disease_category + self.n_drug_category]) code_loss = tf.Tensor code_y_last = tf.Tensor code_label = tf.concat([disease_label, drug_label], axis=2) for i in range(1, self.n_visit): outputs_reshape = tf.reshape(outputs, [-1, self.n_rnn[-1]]) alpha = tf.reshape( tf.reduce_mean(tf.nn.softmax(tf.reshape(tf.add(tf.matmul(outputs_reshape, w_alpha), b_alpha), [-1, self.n_visit])[:, 0:i]), axis=1), [-1, 1]) h_ = alpha * outputs_0[:, i, :] code_loss = tf.reduce_sum( tf.nn.softmax_cross_entropy_with_logits_v2(logits=tf.add(tf.matmul(h_, w_code), b_code), labels=code_label[:, i, :])) code_y_last = tf.nn.softmax(tf.add(tf.matmul(h_, w_code), b_code)) # x_code = tf.concat([x_disease, x_drug], axis=2) # x_code = tf.reshape(x_code, [-1, self.n_disease + self.n_drug]) # w_embed = tf.get_variable('w_embed', # shape=[self.n_disease + self.n_drug, self.n_embed]) # b_embed = tf.get_variable('b_mebed', shape=[self.n_embed]) # hidden = tf.nn.relu(tf.add(tf.matmul(x_code, w_embed), b_embed)) # hidden = tf.reshape(hidden, [-1, self.n_visit, self.n_embed]) # stacked_gru = tf.nn.rnn_cell.MultiRNNCell( # [self.gru_unit(self.n_rnn[0], self.dropout_rate) for i in range(len(self.n_rnn))], state_is_tuple=True) # outputs, states = tf.nn.dynamic_rnn(cell=stacked_gru, inputs=hidden, dtype=tf.float32) # w_alpha = tf.get_variable('w_alpha', # shape=[self.n_rnn[-1], 1]) # b_alpha = tf.get_variable('b_alpha', shape=[1]) # w_code = tf.get_variable('w_code', # shape=[self.n_rnn[-1], self.n_disease_category + self.n_drug_category]) # b_code = tf.get_variable('b_code', shape=[self.n_disease_category + self.n_drug_category]) # code_loss = tf.Tensor # code_y_last = tf.Tensor # code_label = tf.concat([disease_label, drug_label], axis=2) # for i in range(1, self.n_visit): # outputs_reshape = tf.reshape(outputs, [-1, self.n_rnn[-1]]) # alpha = tf.reshape(tf.nn.softmax(tf.reshape(tf.add(tf.matmul(outputs_reshape, w_alpha), b_alpha), # [-1, self.n_visit])[:, 0:i]), [-1, i, 1]) # c = tf.reduce_sum(alpha * outputs[:, 0:i, :], axis=1) # h_ = tf.concat([c, outputs[:, i, :]], axis=1) # # h_ = c * outputs[:, i, :] # code_loss = tf.reduce_sum( # tf.nn.softmax_cross_entropy_with_logits_v2(logits=tf.add(tf.matmul(h_, w_code), b_code), # labels=code_label[:, i, :])) # code_y_last = tf.nn.softmax(tf.add(tf.matmul(h_, w_code), b_code)) return code_loss, code_y_last def print2file(self, content, out_file): outfd = open(out_file, 'a') outfd.write(content + '\n') outfd.close()