Question

我有一个名为cnn-bn-relu的函数，该函数执行一系列卷积-批处理-归一化-Relu并返回输出。通过上述功能获得训练有素的模型，我试图不仅使用测试图像而且还使用其增强副本来测试模型

第1步。将原始图像输入到函数cnn-bn-relu

test_reconstructed_image_1_output=cnn-bn-relu(test_image_1,weights_layer_1,weights_layer_2,weights_layer_3,weights_layer_4)

第2步。接下来，将其扩展副本传递给函数cnn-bn-relu

test_reconstructed_aug_image_1_output=cnn-bn-relu(test_aug_image_1,weights_layer_1,weights_layer_2,weights_layer_3,weights_layer_4)

经过上述两个步骤，我恢复了训练后的模型权重，并尝试获取步骤1和步骤2的输出。但是，出现以下错误

NotFoundError（请参阅上面的回溯）：从检查点还原失败了这很可能是由于变量名或其他图形键检查点缺少的内容。请确保您没有根据检查点更改了期望的图形。原始错误：

我发现计算图像的增强副本与原始图像的卷积会导致上述错误，但是我不知道如何解决。

下面是我使用的代码。

Step 1:
Reading the images from tensorflow record using dataset api
test_dataset = (tf.data.TFRecordDataset(test_tfrecords_filename_label)
      .map(read_and_decode,num_parallel_calls=multiprocessing.cpu_count())
      .prefetch(1))
iterator = test_dataset.make_one_shot_iterator()
test_image_1,aug_test_image_1= iterator.get_next()

Step 2:
# Calculating convolution for original image
test_reconstructed_image_1_output=cnn-bn-relu(test_image_1,weights_layer_1,weights_layer_2,weights_layer_3,weights_layer_4)
# Calculating convolution for augmented copy of original image
test_reconstructed_aug_image_1_output=cnn-bn-relu(test_patches_image_1,weights_layer_1,weights_layer_2,weights_layer_3,weights_layer_4)

Step 3:restoring the model and running the tensorflow session
saver1 = tf.train.Saver()
sess1 = tf.Session()
saver1.restore(sess1,  "model.ckpt")
for i in range(0,3):
   output_1,output_2 = sess1.run( test_reconstructed_image_1_output,test_reconstructed_aug_image_1)

上面的代码会产生上述错误。

If i comment the line
test_reconstructed_aug_image_1_output=cnn-bn- relu(test_patches_image_1,weights_layer_1,weights_layer_2,weights_layer_3,weights_layer_4)

and replace 
for i in range(0,3):
       output_1,output_2 = sess1.run( 
test_reconstructed_image_1_output,test_reconstructed_aug_image_1)

with
for i in range(0,3):
       output_1= sess1.run( test_reconstructed_image_1_output)

我没有收到错误消息。

谁能指导我如何修改代码以同时获得output_1和output_2

从检查点还原失败-TensorFlow

0 个答案: