Tensor Tensor("predictions/Softmax:0", shape=(?, 1000), dtype=float32) 不是這個圖的元素

Question

我正在嘗試遵循有關如何使用預訓練 VGG 模型進行圖像分類的簡單教程。 我擁有的代碼：

from keras.applications.vgg16 import VGG16
from keras.preprocessing.image import load_img, img_to_array
from keras.applications.vgg16 import preprocess_input, decode_predictions

import numpy as np

class KerasModel(object):
    def __init__(self):
        self.model = VGG16()
    def evaluate(self):
        image = load_img('mug.jpg', target_size=(224,224))
        image = img_to_array(image)
        image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
        image = preprocess_input(image)
        yhat = self.model.predict(image)
        label = decode_predictions(yhat)
        label = label[0][0]
        return ('%s (%.2f%%)' % (label[1]), label[2]*100)

這給出了錯誤：Tensor Tensor("predictions/Softmax:0", shape=(?, 1000), dtype=float32) is not an element of this graph。

在搜索了這個錯誤之后，我得到了這個代碼：

from keras.applications.vgg16 import VGG16
from keras.preprocessing.image import load_img, img_to_array
from keras.applications.vgg16 import preprocess_input, decode_predictions

import numpy as np

import tensorflow as tf
graph = tf.get_default_graph()


class KerasModel(object):
    def __init__(self):
        self.model = VGG16()
    def evaluate(self):
        image = load_img('mug.jpg', target_size=(224,224))
        image = img_to_array(image)
        image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
        image = preprocess_input(image)
        with graph.as_default():
            yhat = self.model.predict(image)
        label = decode_predictions(yhat)
        label = label[0][0]
        return ('%s (%.2f%%)' % (label[1]), label[2]*100)

但這仍然會導致相同的錯誤。 有人可以幫我嗎？ 我不明白我做錯了什么，因為教程似乎對每個人都有效。

型號概要：

 _________________________________________________________________
xvision | Layer (type)                 Output Shape              Param #   
xvision | =================================================================
xvision | input_1 (InputLayer)         (None, 224, 224, 3)       0         
xvision | _________________________________________________________________
xvision | block1_conv1 (Conv2D)        (None, 224, 224, 64)      1792      
xvision | _________________________________________________________________
xvision | block1_conv2 (Conv2D)        (None, 224, 224, 64)      36928     
xvision | _________________________________________________________________
xvision | block1_pool (MaxPooling2D)   (None, 112, 112, 64)      0         
xvision | _________________________________________________________________
xvision | block2_conv1 (Conv2D)        (None, 112, 112, 128)     73856     
xvision | _________________________________________________________________
xvision | block2_conv2 (Conv2D)        (None, 112, 112, 128)     147584    
xvision | _________________________________________________________________
xvision | block2_pool (MaxPooling2D)   (None, 56, 56, 128)       0         
xvision | _________________________________________________________________
xvision | block3_conv1 (Conv2D)        (None, 56, 56, 256)       295168    
xvision | _________________________________________________________________
xvision | block3_conv2 (Conv2D)        (None, 56, 56, 256)       590080    
xvision | _________________________________________________________________
xvision | block3_conv3 (Conv2D)        (None, 56, 56, 256)       590080    
xvision | _________________________________________________________________
xvision | block3_pool (MaxPooling2D)   (None, 28, 28, 256)       0         
xvision | _________________________________________________________________
xvision | block4_conv1 (Conv2D)        (None, 28, 28, 512)       1180160   
xvision | _________________________________________________________________
xvision | block4_conv2 (Conv2D)        (None, 28, 28, 512)       2359808   
xvision | _________________________________________________________________
xvision | block4_conv3 (Conv2D)        (None, 28, 28, 512)       2359808   
xvision | _________________________________________________________________
xvision | block4_pool (MaxPooling2D)   (None, 14, 14, 512)       0         
xvision | _________________________________________________________________
xvision | block5_conv1 (Conv2D)        (None, 14, 14, 512)       2359808   
xvision | _________________________________________________________________
xvision | block5_conv2 (Conv2D)        (None, 14, 14, 512)       2359808   
xvision | _________________________________________________________________
xvision | block5_conv3 (Conv2D)        (None, 14, 14, 512)       2359808   
xvision | _________________________________________________________________
xvision | block5_pool (MaxPooling2D)   (None, 7, 7, 512)         0         
xvision | _________________________________________________________________
xvision | flatten (Flatten)            (None, 25088)             0         
xvision | _________________________________________________________________
xvision | fc1 (Dense)                  (None, 4096)              102764544 
xvision | _________________________________________________________________
xvision | fc2 (Dense)                  (None, 4096)              16781312  
xvision | _________________________________________________________________
xvision | predictions (Dense)          (None, 1000)              4097000   
xvision | =================================================================
xvision | Total params: 138,357,544
xvision | Trainable params: 138,357,544
xvision | Non-trainable params: 0
xvision | _________________________________________________________________
xvision | None

Answer 1

似乎 Keras 不是線程安全的，因此您需要在每個線程中初始化模型。 修復調用：_make_predict_function()

它確實對我有用。 這是一個干凈的例子：

from keras.models import load_model

def load_model():
  model = load_model('./my_model.h5')
  model._make_predict_function() 
  print('model loaded') # just to keep track in your server
return model

希望這可以幫助。

Answer 2

由於您的代碼很好，在干凈的環境中運行應該可以解決它。

• 清除~/.keras/ keras 緩存

• 使用正確的包在新環境中運行（使用 anaconda 可以輕松完成）

• 確保您處於新會話中， keras.backend.clear_session()應刪除所有現有的 tf 圖。