自定义损失 Function 错误：ValueError：没有为任何变量提供梯度（TensorFlow）

Question

I am using a binary crossentropy model with non binary Y values & a sigmoid activation layer.

I have created my first custom loss function but when I execute it I get the error "ValueError: No gradients provided for any variable: [....]"

This is my loss function. It is used for cryptocurrency prediction. The y_true are the price change values and the y_pred values are rounded to 0/1 (sigmoid). It penalizes false positives with price_change * 3 and false negatives with price_change . I know my loss function is not like the regular loss functions but I wouldn't know how to achieve this goal with those functions.

def loss(y_true, y_pred):
    penalizer = 3
    loss_values = []
    y_true = y_true.numpy().tolist()
    y_pred = y_pred.numpy().tolist()
    for index in range(len(y_true)):
        pred = round(y_pred[index][0])
        lookup = y_true[index][0]
        if pred == 1:
            if lookup > 0:
                loss_values.append(0.0)
            else:
                loss_values.append(lookup * penalizer * -1)
        else:
            if lookup == 0:
                loss_values.append(0.0)
            elif lookup > 0:
                loss_values.append(lookup)
            else:
                loss_values.append(0.0)
    loss_values = K.constant(loss_values)
    return loss_values

And this is the tensor that the loss function returns.

tf.Tensor(
[ 0.          0.          0.          0.          2.76        0.
  2.16        3.75        0.          2.04        0.          0.03
  0.          0.          0.          2.8799999   1.41        0.
  0.          1.11        0.          2.79        1.2         0.
  0.69        1.92        0.          8.64        0.          6.2999997
  0.          0.          1.05        0.          4.08        0.84000003
  0.          0.          5.43        8.16        0.          0.
  0.6         3.87        0.          0.75        3.72        0.35999998
  1.74        8.07       13.92        1.74        4.41        0.
  1.23        0.          2.76        7.68        0.          0.63
  4.4700003   4.29        0.         10.59      ], shape=(64,), dtype=float32)

The error message:

Traceback (most recent call last):
  File "/vserver/storages///packages//trader/classes/neuralnet/numbers/categorical.py", line 1356, in <module>
    neuralnet.train(refresh="--refresh" in sys.argv)
  File "/vserver/storages///packages//trader/classes/neuralnet/numbers/categorical.py", line 783, in train
    history = self.model.model.fit(
  File "/home/administrator/venv/lib/python3.8/site-packages/keras/engine/training.py", line 1184, in fit
    tmp_logs = self.train_function(iterator)
  File "/home/administrator/venv/lib/python3.8/site-packages/keras/engine/training.py", line 853, in train_function
    return step_function(self, iterator)
  File "/home/administrator/venv/lib/python3.8/site-packages/keras/engine/training.py", line 842, in step_function
    outputs = model.distribute_strategy.run(run_step, args=(data,))
  File "/home/administrator/venv/lib/python3.8/site-packages/tensorflow/python/distribute/distribute_lib.py", line 1286, in run
    return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)
  File "/home/administrator/venv/lib/python3.8/site-packages/tensorflow/python/distribute/distribute_lib.py", line 2849, in call_for_each_replica
    return self._call_for_each_replica(fn, args, kwargs)
  File "/home/administrator/venv/lib/python3.8/site-packages/tensorflow/python/distribute/distribute_lib.py", line 3632, in _call_for_each_replica
    return fn(*args, **kwargs)
  File "/home/administrator/venv/lib/python3.8/site-packages/tensorflow/python/autograph/impl/api.py", line 597, in wrapper
    return func(*args, **kwargs)
  File "/home/administrator/venv/lib/python3.8/site-packages/keras/engine/training.py", line 835, in run_step
    outputs = model.train_step(data)
  File "/home/administrator/venv/lib/python3.8/site-packages/keras/engine/training.py", line 791, in train_step
    self.optimizer.minimize(loss, self.trainable_variables, tape=tape)
  File "/home/administrator/venv/lib/python3.8/site-packages/keras/optimizer_v2/optimizer_v2.py", line 522, in minimize
    return self.apply_gradients(grads_and_vars, name=name)
  File "/home/administrator/venv/lib/python3.8/site-packages/keras/optimizer_v2/optimizer_v2.py", line 622, in apply_gradients
    grads_and_vars = optimizer_utils.filter_empty_gradients(grads_and_vars)
  File "/home/administrator/venv/lib/python3.8/site-packages/keras/optimizer_v2/utils.py", line 72, in filter_empty_gradients
    raise ValueError("No gradients provided for any variable: %s." %
ValueError: No gradients provided for any variable: ['conv1d/kernel:0', 'conv1d_1/kernel:0', 'conv1d_2/kernel:0', 'conv1d_3/kernel:0', 'lstm/lstm_cell/kernel:0', 'lstm/lstm_cell/recurrent_kernel:0', 'lstm/lstm_cell/bias:0', 'lstm_1/lstm_cell_1/kernel:0', 'lstm_1/lstm_cell_1/recurrent_kernel:0', 'lstm_1/lstm_cell_1/bias:0', 'lstm_2/lstm_cell_2/kernel:0', 'lstm_2/lstm_cell_2/recurrent_kernel:0', 'lstm_2/lstm_cell_2/bias:0', 'lstm_3/lstm_cell_3/kernel:0', 'lstm_3/lstm_cell_3/recurrent_kernel:0', 'lstm_3/lstm_cell_3/bias:0', 'dense/kernel:0', 'dense/bias:0', 'dense_1/kernel:0', 'dense_1/bias:0', 'dense_2/kernel:0', 'dense_2/bias:0'].

Any ideas how to fix this?

Edit: New loss function.

def loss(y_true, y_pred):
    penalizer = 3
    batch_size = 64
    #loss_values = []
    loss_values = np.zeros(batch_size)
    y_true = y_true.numpy()
    y_pred = y_pred.numpy()
    for index in range(batch_size):
        pred = y_pred[index][0]
        #if pred >= 0.5:
        #   pred = 1
        #else:
        #   pred = 0
        #pred = tf.round(pred)
        differentiable_round = tf.maximum(pred - 0.499, 0)
        differentiable_round = differentiable_round * 10000
        pred = tf.minimum(differentiable_round, 1)
        lookup = y_true[index][0]
        if K.equal(pred, 1):
            if K.greater(lookup, 0):
                loss_values[index] = 0.0
            else:
                loss_values[index] = lookup * penalizer * -1
        else:
            if K.equal(lookup, 0):
                loss_values[index] = 0.0
            elif K.greater(lookup, 0):
                loss_values[index] = lookup
            else:
                loss_values[index] = 0.0
    loss_values = K.constant(loss_values)
    return loss_values

Answer 1

I think you are running into this issue because round is nondifferentiable. You might want to look into differentiable round functions (which really are approximations of them).

Answer 2

I found the correct differentiable code for the loss function i wanted to use.

def loss(y_true, y_pred):
    y_true_onehot = tf.where(
        tf.greater(y_true, 0.0),
        1.0,
        0.0
    )
    loss_values = keras.losses.BinaryCrossentropy()(y_true_onehot, y_pred)
    mask = tf.where(
        tf.logical_or(
            tf.logical_and(tf.greater(y_true, 0.0), tf.greater_equal(y_pred, 0.5)),
            tf.logical_and(tf.less(y_true, 0.0), tf.less(y_pred, 0.5)),
        ),
        0.0,
        1.0
    )[:,0]
    loss_values = tf.multiply(loss_values, mask)
    return loss_values