Try in tensorflow2.6 by myself. But had some problems. Hope get some help!

[Liozizy]

When I try to define a layer as loss by myself and use the add_weight() function to declare the trainable return propagation variable，Threw an error：

ValueError: Variable <tf.Variable ‘eqn1_1/constant1:0’ shape=(1,) dtype=float32> has None for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.

My code is as follows：

class WbceLoss(KL.Layer):

  def __init__(self, **kwargs):
      super(WbceLoss, self).__init__(**kwargs)
      
  def build(self, input_shape):
      self.constant1 = self.add_weight(name = "constant1", shape[1],initializer='random_normal', trainable=True)
      self.constant2 = self.add_weight(name = "constant2", shape[1],initializer='random_normal', trainable=True)
  
  def call(self, inputs, **kwargs):
          
      tf.compat.v1.disable_eager_execution()
      out1, out2, out3, cur_time, cur_x_input, cur_y_input, cur_z_input, perm_input = inputs
      
      x_input = cur_x_input
      y_input = cur_y_input
      z_input = cur_z_input
      perm_input = perm_input
      
      constant1 = self.constant1
      constant2 = self.constant2
      print(constant1)
      print(constant2)
      
      gradient_with_time = tf.keras.backend.gradients(out1, cur_time)[0]
      constant1 = tf.convert_to_tensor(constant1)
      constant2 = tf.convert_to_tensor(constant2)
      a = tf.zeros((1,), dtype=tf.float32)
      bias = tf.convert_to_tensor([a, a, constant1])
      #bias = tf.expand_dims([0., 0., constant1], 0)
      bias = tf.expand_dims(bias, 2)
      
      pressure_grad_x = tf.keras.backend.gradients(out2, cur_x_input)[0]
      pressure_grad_y = tf.keras.backend.gradients(out2, cur_y_input)[0]
      pressure_grad_z = tf.keras.backend.gradients(out2, cur_z_input)[0]
        
      pressure_grad = tf.convert_to_tensor([pressure_grad_x, pressure_grad_y, pressure_grad_z])
      pressure_grad = tf.keras.backend.permute_dimensions(
      pressure_grad, (1, 0, 2))
      coeff = (1 - out1) / constant2
      
      m = tf.matmul(perm_input, (pressure_grad - bias))
      m_grad_x = tf.keras.backend.gradients(m, cur_x_input)[0]
      m_grad_y = tf.keras.backend.gradients(m, cur_y_input)[0]
      m_grad_z = tf.keras.backend.gradients(m, cur_z_input)[0]
      m_grad_1 = tf.add(m_grad_x, m_grad_y)
      m_grad = tf.add(m_grad_1, m_grad_z)
      
      m_final = tf.multiply(coeff, m_grad)
      eqn_1 = tf.add(gradient_with_time, m_final)
      eqn_2 = tf.add(eqn_1, out3)
      eqn = tf.negative(eqn_2)
      
      eqn = tf.compat.v1.to_float(eqn)
      
      self.add_loss(eqn, inputs=True)
      self.add_metric(eqn, aggregation="mean", name="eqn1")
      
      return eqn

The whole error when I train the model is as follows：

ValueError Traceback (most recent call last) in () 12 batch_size=241, 13 shuffle=True, ---> 14 verbose=1)

~\AppData\Roaming\Python\Python36\site-packages\keras\engine\training_v1.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs) 794 max_queue_size=max_queue_size, 795 workers=workers, --> 796 use_multiprocessing=use_multiprocessing) 797 798 def evaluate(self,

~\AppData\Roaming\Python\Python36\site-packages\keras\engine\training_arrays_v1.py in fit(self, model, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, **kwargs) 655 validation_steps=validation_steps, 656 validation_freq=validation_freq, --> 657 steps_name='steps_per_epoch') 658 659 def evaluate(self,

~\AppData\Roaming\Python\Python36\site-packages\keras\engine\training_arrays_v1.py in model_iteration(model, inputs, targets, sample_weights, batch_size, epochs, verbose, callbacks, val_inputs, val_targets, val_sample_weights, shuffle, initial_epoch, steps_per_epoch, validation_steps, validation_freq, mode, validation_in_fit, prepared_feed_values_from_dataset, steps_name, **kwargs) 175 # function we recompile the metrics based on the updated 176 # sample_weight_mode value. --> 177 f = _make_execution_function(model, mode) 178 179 # Prepare validation data. Hold references to the iterator and the input list

~\AppData\Roaming\Python\Python36\site-packages\keras\engine\training_arrays_v1.py in _make_execution_function(model, mode) 545 if model._distribution_strategy: 546 return distributed_training_utils_v1._make_execution_function(model, mode) --> 547 return model._make_execution_function(mode) 548 549

~\AppData\Roaming\Python\Python36\site-packages\keras\engine\training_v1.py in _make_execution_function(self, mode) 2077 def _make_execution_function(self, mode): 2078 if mode == ModeKeys.TRAIN: -> 2079 self._make_train_function() 2080 return self.train_function 2081 if mode == ModeKeys.TEST:

~\AppData\Roaming\Python\Python36\site-packages\keras\engine\training_v1.py in _make_train_function(self) 2009 # Training updates
2010 updates = self.optimizer.get_updates( -> 2011 params=self._collected_trainable_weights, loss=self.total_loss) 2012 # Unconditional updates
2013 updates += self.get_updates_for(None)

~\AppData\Roaming\Python\Python36\site-packages\keras\optimizer_v2\optimizer_v2.py in get_updates(self, loss, params) 757 758 def get_updates(self, loss, params): --> 759 grads = self.get_gradients(loss, params) 760 grads_and_vars = list(zip(grads, params)) 761 self._assert_valid_dtypes([

~\AppData\Roaming\Python\Python36\site-packages\keras\optimizer_v2\optimizer_v2.py in get_gradients(self, loss, params) 753 "gradient defined (i.e. are differentiable). " 754 "Common ops without gradient: " --> 755 "K.argmax, K.round, K.eval.".format(param)) 756 return grads 757

ValueError: Variable <tf.Variable 'constant1_6:0' shape=(1,) dtype=float32> has None for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.

Hope get some help. Thank you!

[nnnn123456789]

use tensorflow 1.x (like 1.15)

[manwu1994]

Try

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

[amiralizadeh1]

Try to do it in TensorFlow 1 because there are some major changes to the first version. Plus, I recommend using Google Colab instead of local processors because installing tf 1 is much easier (you don't have to change your python or IDE version. you can take a look at my repo for more details.

[hsks]

@amiralizadeh1 google colab no longer support tensor 1.x

[mingwei-yang-byte]

tensorflow 1.15.0
python 3.6

[justgoinggoxixi]

tensorflow 1.15.0 python 3.6

wow,thank y