keras模型是课程的一部分时不起作用

Question

所以我有课

class Trainer:
    def __init__(self,episodes):
        self.factorModel()

    def factorModel(self):
        self.model = Sequential()
        self.model.add(Conv2D(50, (3, 3), activation='relu', input_shape=(3,200,200),dim_ordering="th",strides=4))
        self.model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2) ))
        self.model.add(Conv2D(64, (5, 5), activation='relu') )
        self.model.add(MaxPooling2D(pool_size=(2, 2) ))
        self.model.add(Dense(1000, activation='relu'))
        self.model.add(Flatten())
        self.model.add(Dense(4, activation='softmax'))
        self.model.compile(loss='categorical_crossentropy',optimizer=Adam(lr=0.01), metrics=['accuracy'])


    def do(self,state):
        self.model.predict(np.array(state))[0]

当我尝试调用do时遇到类似ValueError: Tensor Tensor("dense_2/Softmax:0", shape=(?, 4), dtype=float32) is not an element of this graph.的错误，当我尝试运行do时，如果使用相同的模型和相同的配置但我不运行do，则会出现问题。作为线程运行，一切正常

完整的错误消息

  File "/usr/lib/python2.7/threading.py", line 801, in __bootstrap_inner
    self.run()
  File "/usr/lib/python2.7/threading.py", line 754, in run
    self.__target(*self.__args, **self.__kwargs)
  File "path", line 141, in do
     self.model.predict_classes(state)[0]
  File "path/.local/lib/python2.7/site-packages/keras/engine/sequential.py", line 268, in predict_classes
    proba = self.predict(x, batch_size=batch_size, verbose=verbose)
  File "path/.local/lib/python2.7/site-packages/keras/engine/training.py", line 1456, in predict
    self._make_predict_function()
  File "path/.local/lib/python2.7/site-packages/keras/engine/training.py", line 378, in _make_predict_function
    **kwargs)
  File "path/.local/lib/python2.7/site-packages/keras/backend/tensorflow_backend.py", line 3009, in function
    **kwargs)
  File "path/.local/lib/python2.7/site-packages/tensorflow/python/keras/backend.py", line 3479, in function
    return GraphExecutionFunction(inputs, outputs, updates=updates, **kwargs)
  File "path/.local/lib/python2.7/site-packages/tensorflow/python/keras/backend.py", line 3142, in __init__
    with ops.control_dependencies([self.outputs[0]]):
  File "path/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 5426, in control_dependencies
    return get_default_graph().control_dependencies(control_inputs)
  File "path/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 4867, in control_dependencies
    c = self.as_graph_element(c)
  File "path/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 3796, in as_graph_element
    return self._as_graph_element_locked(obj, allow_tensor, allow_operation)
  File "path/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 3875, in _as_graph_element_locked
    raise ValueError("Tensor %s is not an element of this graph." % obj)
ValueError: Tensor Tensor("dense_2/Softmax:0", shape=(?, 4), dtype=float32) is not an element of this graph.

我尝试了通过此问题link来解决问题，因此我尝试在self.model._make_predict_function()之后调用self.factorModel()，但结果却出现此错误 InvalidArgumentError: Tensor conv2d_1_input:0, specified in either feed_devices or fetch_devices was not found in the Graph

好吧，我发现了这个问题link，所以可能无法在线程中进行预测

所以我根据代码的建议进行了一些更改，所以现在看起来像这样：

class Trainer:
    def __init__(self,episodes):
        self.factorModel()
        self.graph = tf.get_default_graph() 


    def factorModel(self):
        self.model = Sequential()
        self.model.add(Conv2D(50, (3, 3), activation='relu', input_shape=(3,200,200),dim_ordering="th",strides=4))
        self.model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2) ))
        self.model.add(Conv2D(64, (5, 5), activation='relu') )
        self.model.add(MaxPooling2D(pool_size=(2, 2) ))
        self.model.add(Dense(1000, activation='relu'))
        self.model.add(Flatten())
        self.model.add(Dense(4, activation='softmax'))
        self.model.compile(loss='categorical_crossentropy',optimizer=Adam(lr=0.01), metrics=['accuracy'])


    def do(self,state):
        with self.graph.as_default():
            self.model.predict(np.array(state))[0]

结果是我得到以下错误

Exception in thread Thread-1:
Traceback (most recent call last):
  File "/usr/lib/python2.7/threading.py", line 801, in __bootstrap_inner
    self.run()
  File "/usr/lib/python2.7/threading.py", line 754, in run
    self.__target(*self.__args, **self.__kwargs)
  File "path/Desktop/marioQProject/new_class_trainer.py", line 151, in do
    self.model.predict_classes(state)[0]
  File "path/.local/lib/python2.7/site-packages/keras/engine/sequential.py", line 268, in predict_classes
    proba = self.predict(x, batch_size=batch_size, verbose=verbose)
  File "path/.local/lib/python2.7/site-packages/keras/engine/training.py", line 1462, in predict
    callbacks=callbacks)
  File "path/.local/lib/python2.7/site-packages/keras/engine/training_arrays.py", line 324, in predict_loop
    batch_outs = f(ins_batch)
  File "patha/.local/lib/python2.7/site-packages/tensorflow/python/keras/backend.py", line 3292, in __call__
    run_metadata=self.run_metadata)
  File "path/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1458, in __call__
    run_metadata_ptr)
FailedPreconditionError: Error while reading resource variable conv2d_1/bias from Container: localhost. This could mean that the variable was uninitialized. Not found: Resource localhost/conv2d_1/bias/N10tensorflow3VarE does not exist.
         [[{{node conv2d_1/Reshape/ReadVariableOp}}]]

Answer 1

Tensorflow对多线程并不是很友好，但是有一种解决方法。

这样做

class Trainer:
    def __init__(self):
        self.factorModel()
        self.graph = tf.get_default_graph()  # [1]

    def do(self, state):
        with self.graph.as_default():  # [2]
            return self.model.predict(np.array(state))[0]

    def factorModel(self):
        self.model = Sequential()
        self.model.add(Conv2D(50, (3, 3), activation='relu', input_shape=(10, 10, 3), strides=4))
        self.model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

t = Trainer()
def fn():
    t.do(np.zeros((1, 10, 10, 3)))

if __name__ == '__main__':
    thread_one = threading.Thread(target=fn)
    thread_two = threading.Thread(target=fn)
    thread_one.start()
    thread_two.start()

顺便说一句，如果您不需要特别订购channel first，那么我建议您改用TF标准channel last。如果您使用opencv直接获取图像，或者使用numpy将Pillow图像转换为ndarray，则默认情况下将获得channel last。

编辑

您是否尝试过确保模型在发送到线程之前能够正常工作，例如

class Trainer:
    def __init__(self, episodes, model, graph):
        self.graph = graph
        self.model = model


model = Sequential()
model.add(Conv2D(...))
.
.
.
# make sure it runs here
model.predict(np.zeros((1, 3, 200, 200)))
# if you don't need to train then try not compile first
graph = tf.get_default_graph()
trainer = Trainer(episodes, model, graph)

也可以调用模型而不是顺序模型，例如

from keras import models, layers
inp = layers.Input((200, 200, 3))
x = layers.Conv2D(50, (3, 3), activation='relu',strides=4)(inp)
x = layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2) )(x)
x = layers.Conv2D(64, (5, 5), activation='relu')(x)
.
.
.
x = layers.Dense(4, activation='softmax')(x)
model = models.Model(inp, x)