Skip to content

Instantly share code, notes, and snippets.

Embed
What would you like to do?
Resize or interpolate tensors to any spatial dimension in Keras
import tensorflow as tf
import keras.backend as K
from keras.layers import Lambda
#
# Keras + TensorFlow
# Resize 'tensorA' to be of the same size as 'tensorB' using 'tf.image.resize_bilinear'
# Very useful for skip-connections and 'Concatenate' layers that might complain about being off by one column
# Only works when dimensions are provided since we use ' K.int_shape' that returns the static shape
#
def ResizeLayerLike(tensorA, tensorB):
sB = K.int_shape(tensorB)
def resize_like(tensor, sB): return tf.image.resize_bilinear(tensor, sB[1:3], align_corners=True)
return Lambda(resize_like, arguments={'sB':sB})(tensorA)
@AhmadMoussa

This comment has been minimized.

Copy link

@AhmadMoussa AhmadMoussa commented Jul 2, 2020

Thanks for this gist, only change I had to make was instead of tf.image.resize_bilinear I needed to use tf.compat.v1.image.resize_bilinear. Weird that this isn't already a standard layer in tensorflow/keras.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment