Question

我看着the interactive demo in tensorflow，似乎他们有单独的测试和训练图，可以切换。例如，他们对测试和训练有不同的准确度：

可以在它们之间切换：

他们甚至也有单独的百分位（hist）图表：

什么是提供这种有用功能的TensorFlow代码？

我尝试过测试和火车错误，但我没有得到这样的情节。这就是我试过的：

with open(path+errors_pretty, 'w+') as f_err_msgs:
    with tf.Session() as sess:
        merged = tf.merge_all_summaries()
        writer = tf.train.SummaryWriter(tensorboard_data_dump, sess.graph)

        sess.run( tf.initialize_all_variables() )
        for i in xrange(steps):
            ## Create fake data for y = W.x + b where W = 2, b = 0
            #(batch_xs, batch_ys) = get_batch_feed(X_train, Y_train, M, phase_train)
            feed_dict_batch = get_batch_feed(X_train, Y_train, M, phase_train)
            ## Train
            if i%report_error_freq == 0:
                (summary_str_train,train_error) = sess.run(fetches=[merged, l2_loss], feed_dict=feed_dict_train)
                (summary_str_test,test_error) = sess.run(fetches=[merged, l2_loss], feed_dict=feed_dict_test)
                writer.add_summary(summary_str_train, i)

                loss_msg = "Model *%s%s*, step %d/%d, training error %g, test error %g \n"%(model,nb_hidden_layers,i,steps,train_error,test_error)
                mdl_info_msg = "Opt: %s, BN %s, After %d/%d iteration, Init: %s \n" % (optimization_alg,bn,i,steps,init_type)
                print_messages(loss_msg, mdl_info_msg)
                # store results
                results['train_errors'].append(train_error)
                results['test_errors'].append(test_error)
                # write errors to pretty print
                f_err_msgs.write(loss_msg)
                f_err_msgs.write(mdl_info_msg)
                # save mdl
                #save_path = saver.save(sess, path+'/tmp_mdls/model.ckpt',global_step=i)
                if mdl_save:
                    save_path = saver.save(sess, path+mdl_dir+'/model.ckpt',global_step=i)
            sess.run(fetches=[merged,train_step], feed_dict=feed_dict_batch)
            #sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})

Answer 1

Tensor board利用日志文件夹结构，组织前端。要获得两个单独的准确度，在两个文件夹中有两个单独的摘要编写器，如下所示

     with open(path+errors_pretty, 'w+') as f_err_msgs:
with tf.Session() as sess:
    merged = tf.merge_all_summaries()
    ***train_writer = tf.train.SummaryWriter(tensorboard_train_data_dump, sess.graph)***
     ***test_writer = tf.train.SummaryWriter(tensboar_test_data_sump, sess.graph)***

    sess.run( tf.initialize_all_variables() )
    for i in xrange(steps):
        ## Create fake data for y = W.x + b where W = 2, b = 0
        #(batch_xs, batch_ys) = get_batch_feed(X_train, Y_train, M, phase_train)
        feed_dict_batch = get_batch_feed(X_train, Y_train, M, phase_train)
        ## Train
        if i%report_error_freq == 0:
            (summary_str_train,train_error) = sess.run(fetches=[merged, l2_loss], feed_dict=feed_dict_train)
            (summary_str_test,test_error) = sess.run(fetches=[merged, l2_loss], feed_dict=feed_dict_test)
            ***train_writer.add_summary(summary_str_train, i)
            test_writer.add_summary(summary_str_test, i)***

            loss_msg = "Model *%s%s*, step %d/%d, training error %g, test error %g \n"%(model,nb_hidden_layers,i,steps,train_error,test_error)
            mdl_info_msg = "Opt: %s, BN %s, After %d/%d iteration, Init: %s \n" % (optimization_alg,bn,i,steps,init_type)
            print_messages(loss_msg, mdl_info_msg)
            # store results
            results['train_errors'].append(train_error)
            results['test_errors'].append(test_error)
            # write errors to pretty print
            f_err_msgs.write(loss_msg)
            f_err_msgs.write(mdl_info_msg)
            # save mdl
            #save_path = saver.save(sess, path+'/tmp_mdls/model.ckpt',global_step=i)
            if mdl_save:
                save_path = saver.save(sess, path+mdl_dir+'/model.ckpt',global_step=i)
        sess.run(fetches=[merged,train_step], feed_dict=feed_dict_batch)
        #sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})

虽然这有点乱，但这是一个可以让你2＆＃34;运行＆＃34;的解决方法。我已经自己完成了这一点，如下图所示。紫色线对应于不同时期的训练精度，并且棕色线对应于不同时期的验证准确度。希望这可以帮助你：）

如何在交互式演示中为TensorBoard制作单独的测试和训练曲线？

1 个答案: