TypeError: 'Tensor' object cannot be interpreted as an integer when using tf.map_fn()

user_7375754 0
Viewed 1153

I am trying to construct label-dependent convolutional filters in keras/tensorflow. Therefore, the convolutional filter(s) depend on each example in the batch.

# function used for tf.map_fn
def single_conv(tupl):
    x, kernel = tupl
    return tf.nn.conv2d(x, kernel, strides=(1, 1, 1, 1), padding='SAME')

# first dimension is None (batch size)
input_img = tf.keras.layers.Input(shape=(28,28,1), dtype=tf.float32)
label = tf.keras.layers.Input(shape=(10,), dtype=tf.float32) 

# the network is learning a mapping for the label
label_encoded = tf.keras.layers.Dense(9, activation='relu')(label) 
# turn mapping into conv filter 
kernels = tf.keras.layers.Reshape((3,3,1,1))(label_encoded) 

# class dependent filter(s)
conditional_conv = tf.map_fn(single_conv, (tf.expand_dims(input_img, 1), kernels), fn_output_signature=tf.float32)

When I run this code snippet, I get a TypeError: 'Tensor' object cannot be interpreted as an integer for the last line. Since the last line uses tf.map_fn, I saw that tf.map_fn results in a TypeError if either the function used (single_conv in this case) is not callable or the structure of the output of function and fn_output_signature do not match: https://www.tensorflow.org/api_docs/python/tf/map_fn#raises.

However, I'm still not sure why this is happening? I feel like both of those reasons should not be an issue?

tf.keras conv-neural-network tensorflow mapping python
1 Answers

What worked for me was creating a custom layer, and doing the map_fn there.

class custom_layer(tf.keras.layers.Layer):
    def __init__(self):
        super(custom_layer, self).__init__()

    def call(self, inputs):
        return tf.map_fn(single_conv, inputs, fn_output_signature=tf.float32)

In your case,

conditional_conv = custom_layer()((tf.expand_dims(input_img, 1), kernels))
user_67676700
Related