Denormalize Tensorflow

Question

I would like to denormalize my prediction from neural network. I first normalize my ground truth, and save its std and mean value in a numpy array:

def norm(x):
  return ((x - x.mean()) / x.std())

norm_y_train_n = norm(y_train_n)
mean_y_train_n = y_train_n.mean
std_y_train_n = y_train_n.std

Then I train my network:

history = model.fit(x_train_n, norm_y_train_n, batch_size=10, epochs=200, validation_split=0.1, shuffle=True, callbacks=[es])

I would like to denormalize my data to the original distribution so I can interpret the rmse in a range that makes sense as follows:

def rmse_denorm(y_true, y_pred):
    return backend.sqrt(backend.mean(backend.square((y_pred*std_y_train_n + mean_y_train_n) - (y_true*std_y_train_n+mean_y_train_n))))

but I can't, if I try to convert the numpy values to tensor values with tf.convert_to_tensor like this:

def rmse_denorm(y_true, y_pred):
return backend.sqrt(backend.mean(backend.square((y_pred*tf.convert_to_tensor(std_y_train_n) + tf.convert_to_tensor(mean_y_train_n)) - (y_true*tf.convert_to_tensor(std_y_train_n)+tf.convert_to_tensor(mean_y_train_n)))))

I will get this error:

Failed to convert object of type <class 'builtin_function_or_method'> to Tensor. Contents: <built-in method std of numpy.ndarray object at 0x7ff00811aee0>. Consider casting elements to a supported type.

and if I want to cast tensor values:

 def rmse_denorm(y_true, y_pred):
return backend.sqrt(backend.mean(backend.square((tf.dtypes.cast(y_pred)*std_y_train_n + mean_y_train_n) - (tf.dtypes.cast(y_true)*std_y_train_n+mean_y_train_n))))

I get this error:

TypeError: cast() missing 1 required positional argument: 'dtype'

how do I do this?

Answer 1

Try to create 2 tf.constant:

std_const = tf.constant(std_y_train_n)
mean_const = tf.constant(mean_y_train_n)
def rmse_denorm(y_true, y_pred):
    return backend.sqrt(backend.mean(backend.square((y_pred*std_const + mean_const) - (y_true*std_const+mean_const))))