[英]How to use a pre-trained TensorFlow net in Keras loss function
I have a pre-trained net I want to use in order to evaluate loss in my Keras net. 我有一个要使用的经过预先训练的网,以便评估我的Keras网中的损失。 The pre-trained network was trained using TensorFlow and I just want to use it as part of my loss calculation. 预训练网络是使用TensorFlow进行训练的,我只想将其用作损失计算的一部分。
The code of my custom loss function is currently: 我的自定义损失函数的代码当前为:
def custom_loss_func(y_true, y_pred):
# Get saliency of both true and pred
sal_true = deep_gaze.get_saliency_map(y_true)
sal_pred = deep_gaze.get_saliency_map(y_pred)
return K.mean(K.square(sal_true-sal_pred))
Where deep_gaze is an object that is ment to manage the access to the external pre-trained net I am using. 其中deep_gaze是一个对象,用于管理对我正在使用的外部预训练网络的访问。
It is defined this way: 它是这样定义的:
class DeepGaze(object):
CHECK_POINT = os.path.join(os.path.dirname(__file__), 'DeepGazeII.ckpt') # DeepGaze II
def __init__(self):
print('Loading Deep Gaze II...')
with tf.Graph().as_default() as deep_gaze_graph:
saver = tf.train.import_meta_graph('{}.meta'.format(self.CHECK_POINT))
self.input_tensor = tf.get_collection('input_tensor')[0]
self.log_density_wo_centerbias = tf.get_collection('log_density_wo_centerbias')[0]
self.tf_session = tf.Session(graph=deep_gaze_graph)
saver.restore(self.tf_session, self.CHECK_POINT)
print('Deep Gaze II Loaded')
'''
Returns the saliency map of the input data.
input format is a 4d array [batch_num, height, width, channel]
'''
def get_saliency_map(self, input_data):
log_density_prediction = self.tf_session.run(self.log_density_wo_centerbias,
{self.input_tensor: input_data})
return log_density_prediction
When I run this I get the error: 当我运行此错误时:
TypeError: The value of a feed cannot be a tf.Tensor object. TypeError:供稿的值不能是tf.Tensor对象。 Acceptable feed values include Python scalars, strings, lists, numpy ndarrays, or TensorHandles. 可接受的feed值包括Python标量,字符串,列表,numpy ndarrays或TensorHandles。
What am I doing wrong? 我究竟做错了什么? Is there a way to evaluate a net on a TensorFlow object coming for a different net (that was made by Keras with a TensorFlow backend). 有没有一种方法可以评估TensorFlow对象上的网络以获取其他网络(该网络由Keras与TensorFlow后端组成)。
Thanks in advance. 提前致谢。
There are two main problems: 主要有两个问题:
When you call get_saliency_map
with input_data=y_true
you are feeding a tensor input_data
to another tensor self.input_tensor
, and this is not valid. 当您使用input_data=y_true
调用get_saliency_map
,您正在将张量input_data
馈送到另一个张量self.input_tensor
,这是无效的。 Moreover, these tensors do not hold a value at graph creation time, but rather they define a computation that will eventually produce a value. 此外,这些张量在图创建时不保存值,而是定义最终将产生值的计算。
Even if you could get an output from get_saliency_map
, your code would still not work because this function disconnects your TensorFlow graph (it doesn't return a tensor), and all the logic must reside within the graph. 即使您可以从get_saliency_map
获得输出,您的代码仍然无法正常工作,因为此函数会断开TensorFlow图的连接(它不会返回张量),并且所有逻辑都必须驻留在图中。 Each tensor has to be computed based on the other available tensors in the graph. 必须根据图中的其他可用张量来计算每个张量。
The solution to this problem is to define the model producing self.log_density_wo_centerbias
within the graph where you define your loss function, using the tensors y_true
and y_pred
directly as input without disconnecting the graph. 解决此问题的方法是在直接定义张量y_true
和y_pred
作为输入而无需断开图形的情况下,在定义损失函数的图形中定义在模型中生成self.log_density_wo_centerbias
的模型。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.