AttributeError: 'Tensor' object has no attribute 'numpy' when attempting to convert Tensor to nparray

Question

I have been looking for a solution for this for a while but I can't seem to find anything.

Some info about my env:

print(tf.__version__)
print(tf.executing_eagerly())
print(tf.config.list_physical_devices('GPU'))

Output:

2.7.0
True
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

I am using a tf.data.Dataset created with

train_ds = tf.data.Dataset.list_files(str(train_dir/'*'), shuffle=False)
train_ds = train_ds.shuffle(train_count, reshuffle_each_iteration=False)

val_ds = tf.data.Dataset.list_files(str(val_dir/'*'), shuffle=False)
val_ds = val_ds.shuffle(val_count, reshuffle_each_iteration=False)

I then created some helper functions

def process_path(file_path):
  img = tf.io.read_file(file_path)
  img = decode_img(img)
  return img

def bin_image(image, mask):
    mask = mask.numpy()
    bins = np.array([20, 40, 60, 80, 100, 120, 140, 160, 180, 200, 220, 240])
    new_mask = np.digitize(mask, bins)
    new_mask_tensor = tf.convert_to_tensor(new_mask, np.float32)
    return image, new_mask_tensor

def split_image(image):
    cityscape = tf.image.crop_to_bounding_box(image, 0, 0, 256, 256)
    mask = tf.image.crop_to_bounding_box(image, 0, 256, 256, 256)
    return bin_image(cityscape, mask)

Using the helper functions

train_ds = train_ds.map(process_path, num_parallel_calls=AUTOTUNE)
val_ds = val_ds.map(process_path, num_parallel_calls=AUTOTUNE)

train_images = train_ds.map(split_image, num_parallel_calls=tf.data.AUTOTUNE)
val_images = val_ds.map(split_image, num_parallel_calls=tf.data.AUTOTUNE)

Printing the type of the mask

print(type(mask))

Outputs:

<class 'tensorflow.python.framework.ops.Tensor'>

I get the error at mask = mask.numpy() . I believe that it is due to it being a different type of tensor not an EagerTensor , which supports .numpy() . The issue I have found is that most other ways of converting a tensor to a np.array were removed in TF2.0. How would one go about converting a tensor to a np.array given the scenario? Or should I just give up and attempt to find a different implementation?

Answer 1

Use tf.make_ndarray to convert Tensorflow tensor to numpy.

Sample code

import tensorflow as tf
a = tf.constant([[1,2,3],[4,5,6]])
proto_tensor = tf.make_tensor_proto(a)  # convert `tensor a` to a proto tensor
tf.make_ndarray(proto_tensor)

output

array([[1, 2, 3],
       [4, 5, 6]], dtype=int32)

In this case

mask = tf.make_ndarray(mask)