python - NotImplementedError: Cannot convert a symbolic Tensor (2nd_target:0) to a numpy array

38
votes

I try to pass 2 loss functions to a model as Keras allows that.

loss: String (name of objective function) or objective function or Loss instance. See losses. If the model has multiple outputs, you can use a different loss on each output by passing a dictionary or a list of losses. The loss value that will be minimized by the model will then be the sum of all individual losses.

The two loss functions:

def l_2nd(beta):
    def loss_2nd(y_true, y_pred):
        ...
        return K.mean(t)

    return loss_2nd

and 

def l_1st(alpha):
    def loss_1st(y_true, y_pred):
        ...
        return alpha * 2 * tf.linalg.trace(tf.matmul(tf.matmul(Y, L, transpose_a=True), Y)) / batch_size

    return loss_1st

Then I build the model:

l2 = K.eval(l_2nd(self.beta))
l1 = K.eval(l_1st(self.alpha))
self.model.compile(opt, [l2, l1])

When I train, it produces the error:

1.15.0-rc3 WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow_core/python/ops/resource_variable_ops.py:1630: calling BaseResourceVariable.init (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version. Instructions for

updating: If using Keras pass *_constraint arguments to layers.

NotImplementedError Traceback (most recent call last) in () 47 create_using=nx.DiGraph(), nodetype=None, data=[('weight', int)]) 48 ---> 49 model = SDNE(G, hidden_size=[256, 128],) 50 model.train(batch_size=100, epochs=40, verbose=2) 51 embeddings = model.get_embeddings()

10 frames in init(self, graph, hidden_size, alpha, beta, nu1, nu2) 72 self.A, self.L = self._create_A_L( 73 self.graph, self.node2idx) # Adj Matrix,L Matrix ---> 74 self.reset_model() 75 self.inputs = [self.A, self.L] 76 self._embeddings = {}

in reset_model(self, opt)

---> 84 self.model.compile(opt, loss=[l2, l1]) 85 self.get_embeddings() 86

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/training/tracking/base.py in _method_wrapper(self, *args, **kwargs) 455 self._self_setattr_tracking = False # pylint: disable=protected-access 456 try: --> 457 result = method(self, *args, **kwargs) 458 finally: 459 self._self_setattr_tracking = previous_value # pylint: disable=protected-access

NotImplementedError: Cannot convert a symbolic Tensor (2nd_target:0) to a numpy array.

Please help, thanks!

3
pythontensorflowkerasloss-function
Even I am facing the same issue, and it works perfectly when I disable eager execution. – Siddhant
@Siddhant did you find an alternative without having to disable eager execution? Disabling it seems to fix the issue, but I am no longer benefiting from the other functionalities of eager execution. – Pleastry

3 Answers

66
votes

For me, the issue occurred when upgrading from numpy 1.19 to 1.20 and using ray's RLlib, which uses tensorflow 2.2 internally. Simply downgrading with

pip install numpy==1.19.5

solved the problem; the error did not occur anymore.

23
votes

I found the solution to this problem:

It was because I mixed symbolic tensor with a non-symbolic type, such as a numpy. For example. It is NOT recommended to have something like this:

def my_mse_loss_b(b):
     def mseb(y_true, y_pred):
         ...
         a = np.ones_like(y_true) #numpy array here is not recommended
         return K.mean(K.square(y_pred - y_true)) + a
     return mseb

Instead, you should convert all to symbolic tensors like this:

def my_mse_loss_b(b):
     def mseb(y_true, y_pred):
         ...
         a = K.ones_like(y_true) #use Keras instead so they are all symbolic
         return K.mean(K.square(y_pred - y_true)) + a
     return mseb

Hope this help!

15
votes

It might be the issue with numpy version. Try to use numpy less than 1.20

pip install numpy==1.19