我想测量正向传播中操作的时间，因此我提到了Tensorflow's tutorial。这在训练阶段效果很好，但是在我的预测阶段存在错误。
预测阶段，我构建测试集的管道
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(
handle, predict_dataset.output_types, predict_dataset.output_shapes)
predict_init_op = predict_dataset.make_one_shot_iterator()
next_element = iterator.get_next()
在会话中运行：
merged = tf.summary.merge_all()
predict_result = []
with tf.Session() as sess:
global_step = 0
saver.restore(sess, './output/msra_24_crf_result_dir/model.ckpt-26999')
# sess.run(tf.global_variables_initializer())
predict_writer = tf.summary.FileWriter(FLAGS.output_dir+'/predict' , sess.graph)
predict_init_handle = sess.run(predict_init_op.string_handle())
while True:
try:
tf.logging.info("======================= the %d step starts ==================="%global_step)
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, label, result= sess.run([merged, label_ids, pred_ids], feed_dict={handle:predict_init_handle},
options=run_options, run_metadata=run_metadata)
train_writer.add_run_metadata(run_metadata, 'step%03d' % step)
predict_writer.add_summary(summary, global_step=global_step)
predict_result.append(result)
print("the label is", label)
print("the result is", result)
# result_to_pair(predict_examples, result)
global_step += 1
except tf.errors.OutOfRangeError:
break
predict_writer.close()
我在运行阶段也做同样的事情，但是没有任何问题。训练阶段的数据管道：
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(
handle, train_dataset.output_types, train_dataset.output_shapes)
train_init_op = train_dataset.make_one_shot_iterator()
dev_init_op = dev_dataset.make_initializable_iterator()
next_element = iterator.get_next()
会话运行：
merged = tf.summary.merge_all()
tf.logging.info("==================Entering Session Running=========================")
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_writer = tf.summary.FileWriter(FLAGS.output_dir , sess.graph)
dev_writer = tf.summary.FileWriter(FLAGS.output_dir + '/eval')
train_iterator_handle = sess.run(train_init_op.string_handle())
dev_iterator_handle = sess.run(dev_init_op.string_handle())
for step in range(num_train_steps):
if step % 100 == 0:
tf.logging.info("===============evaluate at %d step=============="%step)
sess.run(running_vars_initializer)
sess.run(dev_init_op.initializer)
# while True:
while True:
try:
# print(sess.run([label_ids, pred_ids], feed_dict={handle: dev_iterator_handle}))
summary,_,_,_ = sess.run([merged, prec_op, recall_op, f1_op], feed_dict={handle: dev_iterator_handle})
except tf.errors.OutOfRangeError:
break
dev_writer.add_summary(summary, step)
_precision, _recall, _f1 = sess.run([prec_scalar, recall_scalar, f1_scalar])
print("At step {}, the precision is {:.2f}%,the recall is {:.2f}%,the f1 is {:.2f}%".format(step, _precision*100, _recall*100, _f1*100))
else:
if step % 1000 == 999:
tf.logging.info("===============save model at %d step==============" % step)
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, _ , _total_loss= sess.run([merged, train_op, total_loss],
feed_dict={handle: train_iterator_handle},
options=run_options,
run_metadata=run_metadata)
train_writer.add_run_metadata(run_metadata, 'step%03d' % step)
train_writer.add_summary(summary, step)
tf.logging.info("========== the total loss is %.5f ===============" %(_total_loss))
print('Adding run metadata for', step)
save_path = saver.save(sess, os.path.join(FLAGS.output_dir, "model.ckpt"), global_step=step)
print("Model saved in path: %s" % save_path)
else:
# print(sess.run([pred_ids, label_ids], feed_dict={handle: train_iterator_handle}))
summary, _ = sess.run([merged, train_op], feed_dict={handle: train_iterator_handle})
train_writer.add_summary(summary, step)
train_writer.close()
dev_writer.close()
我想知道训练阶段的总计算时间。剂量是指向前和向后传播的整个计算时间，或者只是向前传播的整个计算时间。我正在尝试修复此问题，因此将不胜感激。