Tensor(..., shape=(), dtype=int64) 必须与 Tensor(..., shape=(), dtype=resource) Keras 来自同一个图

Question

我正在尝试使用 Keras 来运行 Conv2D 网络以读取一组包含来自200 亿 Jester的手势图像的文件夹在更改太多代码之前正确运行。 但是，我一直遇到

ValueError: Tensor("training/Adamax/Const:0", shape=(), dtype=int64) must be from the same graph as Tensor("Adamax/iterations:0", shape=(), dtype=resource).

并且不够了解以解决它。 我已经尝试过有关重置图表的其他答案

import keras 
keras.backend.clear_session()

或者

tf.reset_default_graph()

但两者都不起作用，或者两者都不起作用。

我的图像文件结构类似于：../images/train/[Gesture]/[Sample] ../images/train/[Gesture]/[Sample]/Image001.png

这比我以前使用的更深，但 flow_from_directory 正确输出图像和 class 计数，用于训练和验证集

Found 3456570 images belonging to 27 classes.
Found 532578 images belonging to 27 classes.

康达清单：

...
cudatoolkit               10.0.130                      0  
cudnn                     7.6.4                cuda10.0_0 
...
keras                     2.3.1                         0  
keras-applications        1.0.8                      py_0  
keras-base                2.3.1                    py37_0  
keras-gpu                 2.3.1                         0  
keras-preprocessing       1.1.0                      py_1  
...
tensorboard               1.14.0           py37hf484d3e_0  
tensorflow                1.14.0          gpu_py37h4491b45_0  
tensorflow-base           1.14.0          gpu_py37h8d69cac_0  
tensorflow-estimator      1.14.0                     py_0  
tensorflow-gpu            1.14.0               h0d30ee6_0  

代码：

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import os
import glob
import shutil
import pickle
import cv2
import numpy as np
import matplotlib.pyplot as plt
import random
from IPython.display import display
from PIL import Image

from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten, BatchNormalization, Activation
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.layers.convolutional import Conv3D, MaxPooling3D
from keras.constraints import maxnorm
from keras.utils import np_utils
from keras.preprocessing.image import ImageDataGenerator

import tensorflow as tf

os.environ["CUDA_VISIBLE_DEVICES"]="1"
tf.reset_default_graph()

# read in the training and validation labels
trainPairs = np.genfromtxt('/home/me/Videos/sign_language/jester-v1-train.csv', delimiter=';', skip_header=0, dtype=[('class', 'S12'),('sign','S50')])
trainLabels = [v for k,v in trainPairs]

validPairs = np.genfromtxt('/home/me/Videos/sign_language/jester-v1-validation.csv', delimiter=';', skip_header=0, dtype=[('class', 'S12'),('sign','S50')])
validLabels = [v for k,v in validPairs]

def copyDirectory(src, dest):
    try:
        shutil.copytree(src, dest)
    # Directories are the same
    except shutil.Error as e:
        print('Directory not copied. Error: %s' % e)
    # Any error saying that the directory doesn't exist
    except OSError as e:
        print('Directory not copied. Error: %s' % e)

source = '/media/me/other/20bn-jester-v1/'
dest = '/media/me/other/jester/validation/'

# counter = 0
# for k,v in validPairs:
#     counter = counter + 1
#     source_folder = source + k.decode("utf-8")
#     dest_folder = dest + v.decode("utf-8") + "/" + k.decode("utf-8")

#     if counter%100 == 0:        
#         print(k)
#         print(v)
#         print(counter)
#         print(source_folder)
#         print(dest_folder)

#     if os.path.isdir(source_folder):
#         if os.path.isdir(dest + v.decode("utf-8")):
#             copyDirectory(source_folder, dest_folder)        

#     if counter%1000 == 0:
#         print(counter)

datagen = ImageDataGenerator()

train_it = datagen.flow_from_directory('/media/me/other/jester/train/', class_mode='categorical', batch_size=64)
valid_it = datagen.flow_from_directory('/media/me/other/jester/validation/', class_mode='categorical', batch_size=64)
# test_it = datagen.flow_from_directory('/media/me/other/jester/test/', class_mode='binary', batch_size=64)

seed = 21
epochs = 5
optimizer = 'Adamax'

with tf.device("/cpu:0"):
    model = Sequential()

model = Sequential()


#model.add(Conv2D(32,(3,3), input_shape=(X_train.shape[1:]), padding='same'))
#TODO is this the right shape??
model.add(Conv2D(32,(3,3), input_shape=(256, 256, 3), padding='same'))
model.add(Activation('relu'))
model.add(Conv2D(32, (3,3), input_shape=(3,32,32), activation='relu', padding='same'))
model.add(Dropout(0.2))
model.add(BatchNormalization())
model.add(Conv2D(64, (3,3), padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2,2)))

model.add(Conv2D(64, (3, 3), padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(BatchNormalization())
model.add(Conv2D(128, (3, 3), padding='same'))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Flatten())
model.add(Dropout(0.2))
model.add(Dense(256, kernel_constraint=maxnorm(3)))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Dense(128, kernel_constraint=maxnorm(3)))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(BatchNormalization())

#TODO make this a variable
model.add(Dense(27))
model.add(Activation('softmax'))

model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])

### I think everything up to here is ok???

global graph
graph = tf.get_default_graph()

for layer in model.layers:
    print(layer.output_shape)

print(model.summary())

np.random.seed(seed)

image_batch_train, label_batch_train = next(iter(train_it))
print("Image batch shape: ", image_batch_train.shape)
print("Label batch shape: ", label_batch_train.shape)
dataset_labels = sorted(train_it.class_indices.items(), key=lambda pair:pair[1])
dataset_labels = np.array([key.title() for key, value in dataset_labels])

print(dataset_labels)

from keras import backend as K
K.clear_session()

import keras 
keras.backend.clear_session()

tf.reset_default_graph()
model.fit_generator(train_it, steps_per_epoch=16, validation_data=valid_it, validation_steps=8)

#scores = model.evaluate(test_it, steps=24, verbose=0)
#print("Accuracy: %.2f%%" % (scores[1]*100))

编辑 1：添加日志

Traceback (most recent call last):

  File "<ipython-input-1-09b1bdd2e389>", line 152, in <module>
    model.fit_generator(train_it, steps_per_epoch=16, validation_data=valid_it, validation_steps=8)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/keras/engine/training.py", line 1732, in fit_generator
    initial_epoch=initial_epoch)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/keras/engine/training_generator.py", line 42, in fit_generator
    model._make_train_function()

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/keras/engine/training.py", line 316, in _make_train_function
    loss=self.total_loss)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/keras/optimizers.py", line 599, in get_updates
    self.updates = [K.update_add(self.iterations, 1)]

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py", line 1268, in update_add
    return tf_state_ops.assign_add(x, increment)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/tensorflow/python/ops/state_ops.py", line 195, in assign_add
    return ref.assign_add(value)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/tensorflow/python/ops/resource_variable_ops.py", line 1108, in assign_add
    name=name)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/tensorflow/python/ops/gen_resource_variable_ops.py", line 68, in assign_add_variable_op
    "AssignAddVariableOp", resource=resource, value=value, name=name)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/tensorflow/python/framework/op_def_library.py", line 366, in _apply_op_helper
    g = ops._get_graph_from_inputs(_Flatten(keywords.values()))

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/tensorflow/python/framework/ops.py", line 6135, in _get_graph_from_inputs
    _assert_same_graph(original_graph_element, graph_element)

  File "/home/me/Programs/anaconda3/envs/hand-gesture/lib/python3.7/site-packages/tensorflow/python/framework/ops.py", line 6071, in _assert_same_graph
    (item, original_item))

ValueError: Tensor("training/Adamax/Const:0", shape=(), dtype=int64) must be from the same graph as Tensor("Adamax/iterations:0", shape=(), dtype=resource).

Answer 1

问题是您在训练 model 之前重置了默认图：

tf.reset_default_graph()  # <-- remove this line
model.fit_generator(train_it, steps_per_epoch=16, validation_data=valid_it, validation_steps=8)

问题如下。 您首先在开始时重置默认图，除非您的实际脚本在此之前有更多代码，否则不会有任何区别，因为此时默认图是空的。 然后你制作你的 model，它会在新的默认图上创建操作。 新的默认图表是您稍后获得的图表：

graph = tf.get_default_graph()

问题是稍后，在清除 Keras session 两次（这也没有任何影响）之后，您再次重置默认图形。 当您调用fit时，训练过程开始，并创建了 Keras model 的一些新图形对象。 由于您的默认图表已更改（因为您重置了它），这些新对象是在与对象 rest 不同的图表中创建的，这会导致错误。 如果您在训练期间使用以前的默认图作为默认值，我认为您仍然可以重置图并使其正常工作：

tf.reset_default_graph()
with graph.as_default():  # Use former default graph as default
    model.fit_generator(train_it, steps_per_epoch=16, validation_data=valid_it, validation_steps=8)

我不完全确定 Keras 默认 session 是否会一直工作（因为它可能是为新的默认图表创建的），但我认为它应该......无论如何，如果你出于任何原因想要拥有你的Keras model 在自己的图中隔离，而不是重置图，您可以按如下方式执行：

with tf.Graph().as_default() as graph:  # Make a new graph and use it as default
    # Make dataset
    # Make model
    # Train