numpy_and_tf_at_the_same_time.py

from keras import backend
from keras.backend import numpy_backend
import numpy as np
import tensorflow as tf


class NPTF(object):

  def __getattr__(self, name):
    if name in dir(numpy_backend) and name in dir(backend):
      k_symbol = getattr(backend, name)
      np_symbol = getattr(numpy_backend, name)
    else:
      raise ValueError('Unknown symbol:', name)

    def wrapped(*args, **kwargs):
      if args and isinstance(args[0], (np.ndarray, float, int)):
        return np_symbol(*args, **kwargs)
      return k_symbol(*args, **kwargs)
    return wrapped

X = NPTF()

# Use like you'd use a Keras backend

def dual_rmse(x, y):
  return X.sqrt(X.sum(X.square(x - y), axis=-1))

print(dual_rmse(tf.zeros((3, 4)), tf.ones((3, 4))))
print(dual_rmse(np.zeros((3, 4)), np.ones((3, 4))))

# Bonus points: also works with Theano, CNTK, MXNet