def train(): # Training procedure # ====================================================== # 设定最小显存使用量 config = tf.ConfigProto() config.gpu_options.allow_growth = True with tf.Session(config=config) as sess: config = CNNConfig('CHAR-RANDOM') cnn = TextCNN(config) cnn.prepare_data() cnn.setCNN() print('Setting Tensorboard and Saver...') # 设置Saver和checkpoint来保存模型 # =================================================== checkpoint_dir = os.path.join(os.path.abspath("checkpoints"), "textcnn") checkpoint_prefix = os.path.join(checkpoint_dir, cnn.train_mode) if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables()) # ===================================================== # 配置Tensorboard,重新训练时,请将tensorboard文件夹删除,不然图会覆盖 # ==================================================================== train_tensorboard_dir = 'tensorboard/textcnn/train/' + config.train_mode valid_tensorboard_dir = 'tensorboard/textcnn/valid/' + config.train_mode if not os.path.exists(train_tensorboard_dir): os.makedirs(train_tensorboard_dir) if not os.path.exists(valid_tensorboard_dir): os.makedirs(valid_tensorboard_dir) # 训练结果记录 log_file = open(valid_tensorboard_dir + '/log.txt', mode='w') merged_summary = tf.summary.merge([ tf.summary.scalar('Trainloss', cnn.loss), tf.summary.scalar('Trainaccuracy', cnn.accuracy) ]) merged_summary_t = tf.summary.merge([ tf.summary.scalar('Testloss', cnn.loss), tf.summary.scalar('Testaccuracy', cnn.accuracy) ]) train_summary_writer = tf.summary.FileWriter(train_tensorboard_dir, sess.graph) # ========================================================================= global_step = tf.Variable(0, trainable=False) # 保证Batch normalization的执行 update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) with tf.control_dependencies( update_ops): # 保证train_op在update_ops执行之后再执行。 train_op = tf.train.AdamOptimizer(config.learning_rate).minimize( cnn.loss, global_step) # 训练步骤 def train_step(batch_x, batch_y, keep_prob=config.dropout_keep_prob): feed_dict = { cnn.input_x: batch_x, cnn.labels: batch_y, cnn.dropout_keep_prob: keep_prob, cnn.training: True } sess.run(train_op, feed_dict=feed_dict) step, loss, accuracy, summery = sess.run( [global_step, cnn.loss, cnn.accuracy, merged_summary], feed_dict={ cnn.input_x: batch_x, cnn.labels: batch_y, cnn.dropout_keep_prob: 1.0, cnn.training: False }) t = datetime.datetime.now().strftime('%m-%d %H:%M') print('TRAIN %s: epoch: %d, step: %d, loss: %f, accuracy: %f' % (t, epoch, step, loss, accuracy)) # 把结果写入Tensorboard中 train_summary_writer.add_summary(summery, step) def test_step(batch_x, batch_y): step, loss, accuracy, summery = sess.run( [global_step, cnn.loss, cnn.accuracy, merged_summary_t], feed_dict={ cnn.input_x: batch_x, cnn.labels: batch_y, cnn.dropout_keep_prob: 1.0, cnn.training: False }) t = datetime.datetime.now().strftime('%m-%d %H:%M') print('TEST %s: epoch: %d, step: %d, loss: %f, accuracy: %f' % (t, epoch, step, loss, accuracy)) # 把结果写入Tensorboard中 train_summary_writer.add_summary(summery, step) return accuracy print('Start training TextCNN, training mode=' + cnn.train_mode) sess.run(tf.global_variables_initializer()) last = 0 # Training loop for epoch in range(1000000): batch_x, batch_y = train_dataset.next_batch(128) train_step(batch_x, batch_y, config.dropout_keep_prob) if epoch % 10 == 0: batch_x, batch_y = test_dataset.next_batch(128) accuracy = test_step(batch_x, batch_y) if accuracy > last: path = saver.save(sess, checkpoint_prefix, global_step=global_step) print("Saved model checkpoint to {}\n".format(path)) last = accuracy train_summary_writer.close() log_file.close()
def train(): # Training procedure # ====================================================== # 设定最小显存使用量 config = tf.ConfigProto() config.gpu_options.allow_growth = True with tf.Session(config=config) as sess: config = CNNConfig('MULTI') cnn = TextCNN(config) cnn.prepare_data() cnn.setCNN() print('Setting Tensorboard and Saver...') # 设置Saver和checkpoint来保存模型 # =================================================== checkpoint_dir = os.path.join(os.path.abspath("checkpoints"), "textcnn") checkpoint_prefix = os.path.join(checkpoint_dir, cnn.train_mode) if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) saver = tf.train.Saver(tf.global_variables()) # ===================================================== # 配置Tensorboard,重新训练时,请将tensorboard文件夹删除,不然图会覆盖 # ==================================================================== train_tensorboard_dir = 'tensorboard/textcnn/train/' + config.train_mode valid_tensorboard_dir = 'tensorboard/textcnn/valid/' + config.train_mode if not os.path.exists(train_tensorboard_dir): os.makedirs(train_tensorboard_dir) if not os.path.exists(valid_tensorboard_dir): os.makedirs(valid_tensorboard_dir) # 训练结果记录 log_file = open(valid_tensorboard_dir + '/log.txt', mode='w') merged_summary = tf.summary.merge([ tf.summary.scalar('loss', cnn.loss), tf.summary.scalar('accuracy', cnn.accuracy) ]) train_summary_writer = tf.summary.FileWriter(train_tensorboard_dir, sess.graph) # ========================================================================= global_step = tf.Variable(0, trainable=False) # 保证Batch normalization的执行 update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) with tf.control_dependencies( update_ops): # 保证train_op在update_ops执行之后再执行。 train_op = tf.train.AdamOptimizer(config.learning_rate).minimize( cnn.loss, global_step) # 训练步骤 def train_step(batch_x, batch_y, keep_prob=config.dropout_keep_prob): feed_dict = { cnn.input_x: batch_x, cnn.labels: batch_y, cnn.dropout_keep_prob: keep_prob, cnn.training: True } sess.run(train_op, feed_dict=feed_dict) step, loss, accuracy, summery = sess.run( [global_step, cnn.loss, cnn.accuracy, merged_summary], feed_dict={ cnn.input_x: batch_x, cnn.labels: batch_y, cnn.dropout_keep_prob: 1.0, cnn.training: False }) t = datetime.datetime.now().strftime('%m-%d %H:%M') print('%s: epoch: %d, step: %d, loss: %f, accuracy: %f' % (t, epoch, step, loss, accuracy)) # 把结果写入Tensorboard中 train_summary_writer.add_summary(summery, step) # 验证步骤 def valid_step(next_valid_element): # 把valid_loss和valid_accuracy归0 valid_loss = 0.0 valid_accuracy = 0.0 valid_precision = 0.0 valid_recall = 0.0 valid_f1_score = 0.0 i = 0 while True: try: lines = sess.run(next_valid_element) batch_x, batch_y = cnn.convert_input(lines) feed_dict = { cnn.input_x: batch_x, cnn.labels: batch_y, cnn.dropout_keep_prob: 1.0, cnn.training: False } loss, accuracy, prediction, y_true = sess.run( [cnn.loss, cnn.accuracy, cnn.prediction, cnn.labels], feed_dict) precision = sk.metrics.precision_score(y_true=y_true, y_pred=prediction, average='weighted') recall = sk.metrics.recall_score(y_true=y_true, y_pred=prediction, average='weighted') f1_score = sk.metrics.f1_score(y_true=y_true, y_pred=prediction, average='weighted') valid_loss += loss valid_accuracy += accuracy valid_precision += precision valid_recall += recall valid_f1_score += f1_score i += 1 except tf.errors.OutOfRangeError: # 遍历完验证集,然后对loss和accuracy求平均值 valid_loss /= i valid_accuracy /= i valid_precision /= i valid_recall /= i valid_f1_score /= i t = datetime.datetime.now().strftime('%m-%d %H:%M') log = '%s: epoch %d, validation loss: %0.6f, accuracy: %0.6f' % ( t, epoch, valid_loss, valid_accuracy) log = log + '\n' + ( 'precision: %0.6f, recall: %0.6f, f1_score: %0.6f' % (valid_precision, valid_recall, valid_f1_score)) print(log) log_file.write(log + '\n') time.sleep(3) return print('Start training TextCNN, training mode=' + cnn.train_mode) sess.run(tf.global_variables_initializer()) # Training loop for epoch in range(config.epoch_num): train_init_op, valid_init_op, next_train_element, next_valid_element = cnn.shuffle_datset( ) sess.run(train_init_op) while True: try: lines = sess.run(next_train_element) batch_x, batch_y = cnn.convert_input(lines) train_step(batch_x, batch_y, config.dropout_keep_prob) except tf.errors.OutOfRangeError: # 初始化验证集迭代器 sess.run(valid_init_op) # 计算验证集准确率 valid_step(next_valid_element) break train_summary_writer.close() log_file.close() # 训练完成后保存参数 path = saver.save(sess, checkpoint_prefix, global_step=global_step) print("Saved model checkpoint to {}\n".format(path))