sess中的第二个feed\u dict。run(),培训时评估

臭名昭著的:

*InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'Placeholder_1' with dtype float and shape [100,8]
         [[Node: Placeholder_1 = Placeholder[dtype=DT_FLOAT, shape=[100,8], _device="/job:localhost/replica:0/task:0/device:GPU:0"]()]]*

我已经训练CNN有一段时间了。我已经为培训和评估方法编写了单独的python脚本,并希望将它们结合起来。我希望网络进行N次迭代的训练,然后运行评估,检查模型是否优于以前的任何模型,如果是,请保存,然后恢复训练。

由于将它们结合在一起,网络将进行训练,但一旦进入循环的“验证”部分,它将在sess失败。在我必须输入验证数据的地方运行。

sess中有一些调试打印(“hi”)。运行时,错误出现在第3行之后,但在第4个“hi”之前。这是验证数据的feed\u dict发生的地方。

我试着只使用一个图,也用它自己的v\u image\u out、v\u label\u out等构建一个单独的图,但都没有用。我已经为这个错误挣扎了好几天,改变了很多事情。非常感谢您的帮助。该代码已在此github中进行了修改,以符合我的目的: https://github.com/yeephycho/tensorflow_input_image_by_tfrecord/blob/master/src/flower_train_cnn.py

*编辑:它调用错误的精确行是:

     File "train_cnn_4_pools.py", line 295, in train
        v_accuracy_out, v_logits_batch_out, v_summary = sess.run(
[accuracy, logits_batch, merged_summary_op],
 feed_dict={image_batch_placeholder: v_image_out, label_tensor_placeholder: `v_label_out})`

原因:

  File "train_cnn_4_pools.py", line 317, in <module>
    train()
  File "train_cnn_4_pools.py", line 226, in train
    label_batch_placeholder = tf.placeholder(tf.float32, shape=[BATCH_SIZE, NUM_CLASSES])

这非常奇怪,因为我在验证循环中调用label\u tensor\u占位符时失败了。不是因为label\U batch\U占位符而发生错误!为什么它会尝试调用错误的占位符?!

def train():
image_batch_out, label_batch_out, filename_batch = input(if_eval = False)
v_image_batch_out, v_label_batch_out, v_filename_batch = input(if_eval = True)

image_batch_placeholder = tf.placeholder(tf.float32, shape=[BATCH_SIZE, None, None, 3])
image_batch = tf.reshape(image_batch_out, (BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, 3))
v_image_batch = tf.reshape(v_image_batch_out, (BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, 3))

label_batch_placeholder = tf.placeholder(tf.float32, shape=[BATCH_SIZE, NUM_CLASSES])
label_tensor_placeholder = tf.placeholder(tf.int64, shape=[BATCH_SIZE])
label_offset = -tf.ones([BATCH_SIZE], dtype=tf.int64, name="label_batch_offset")
label_batch_one_hot = tf.one_hot(tf.add(label_batch_out, label_offset), depth=NUM_CLASSES, on_value=1.0, off_value=0.0)
label_batch = tf.add(label_batch_out, label_offset)
v_label_batch = tf.add(v_label_batch_out, label_offset)
with tf.variable_scope("inference") as scope:

    logits_out = network(image_batch)
    scope.reuse_variables()
    v_logits_out = network(v_image_batch)

logits_batch = tf.to_int64(tf.arg_max(v_logits_out, dimension = 1))
#loss = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(labels=label_batch_one_hot, logits=logits_out))
prediction_op = tf.nn.softmax(logits_out)
v_prediction_op = tf.nn.softmax(v_logits_out)

correct_prediction = tf.equal(logits_batch, label_tensor_placeholder)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

loss = tf.losses.mean_squared_error(labels=label_batch_placeholder, predictions=prediction_op)
global_step = tf.Variable(0, name='global_step', trainable=False)
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)

variable_summaries(prediction_op, name="Predictions")
surity_summary(prediction_op, name="Certainty")
#surity = surity_calc(prediction_op)
tf.summary.scalar("loss", loss)
merged_summary_op = tf.summary.merge_all()

with tf.Session() as sess:
    # Visualize the graph through tensorboard.
    #file_writer = tf.summary.FileWriter("C:/logs", sess.graph)
    # op to write logs to Tensorboard
    summary_writer = tf.summary.FileWriter(logs_path, graph=tf.get_default_graph())
    accuracy_accu = 0
    best_accu = 0
    sess.run(tf.global_variables_initializer())
    saver = tf.train.Saver()
    saver.save(sess, chk_path)
    saver.restore(sess, chk_path)
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord, sess = sess)
    for epoch in range(1, 1000):
        for i in range(5):
            image_out, label_out, label_batch_one_hot_out, filename_out = sess.run([image_batch, label_batch_out, label_batch_one_hot, filename_batch])

            _, infer_out, loss_out, summary, predict_out, global_step_out = sess.run([train_step, logits_out, loss, merged_summary_op, prediction_op, global_step], feed_dict={image_batch_placeholder: image_out, label_batch_placeholder: label_batch_one_hot_out})


            #print(image_out.shape)
            print("label_out: ")
            #print(filename_out)
            print(label_out)
            #print(label_batch_one_hot_out)
            #print("infer_out: ")
            #print(infer_out)
            print("prediction: ")
            print(predict_out)
            print("loss: " + str(loss_out))
            print("local step: " + str(i))
            print("global step: " + str(global_step_out - 1))
            print("epoch: " + str(epoch))
            if(i%10 == 0):
                summary_writer.add_summary(summary, global_step_out)
        print("hi")
        for p in range(int(v_TRAINING_SET_SIZE/BATCH_SIZE)):
            print("hi")
            v_image_out, v_label_out, v_filename_out = sess.run([v_image_batch, v_label_batch, v_filename_batch])
            print("hi")
            v_accuracy_out, v_logits_batch_out, v_summary = sess.run([accuracy, logits_batch, merged_summary_op], feed_dict={image_batch_placeholder: v_image_out, label_tensor_placeholder: label_out})
            print("hi")
            accuracy_accu += accuracy_out

            print(p)
            print(v_image_out.shape)
            print("label_out: ")
            print(v_filename_out)
            print(v_label_out)
            print(v_logits_batch_out)
            print("accuracy: ", v_accuracy_out)
            summary_writer.add_summary(v_summary, p)

        print("Accuracy: ")
        print((accuracy_accu/TRAINING_SET_SIZE)*100)
        if(accuracy_accu > best_accu):
            saver.save(sess, chk_path)
            best_accu = accuracy_accu
    coord.request_stop()
    coord.join(threads)
    sess.close()

列车()