I basically want to have the option to feed input to the middle of the graph and compute the output going from there. One idea I had is to use tf.placeholder_with_default that defaults to a zero tensor. Then I could mix the optional inputs using addition, however addition on a large shape this seems to be a lot of unnecessary computation. Are there better ways of accomplishing that?
input_enabled = tf.placeholder_with_default(tf.constant(1.), [1])
input_shape = [None, in_size]
input = tf.placeholder_with_default(tf.zeros(input_shape), input_shape)
// ...
bottleneck_shape = [None, bottleneck_size]
bottleneck = input_enabled * f(prev_layer) + tf.placeholder_with_default(tf.zeros(bottleneck_shape), bottleneck_shape)
// ...
// Using graph with input at first layer:
sess.run([output], feed_dict={input: x})
// Using graph with input at bottleneck layer:
sess.run([output], feed_dict={bottleneck: b, input_enabled: 0.})
I understand better thanks to your code.
Basically the schema is:
input <- you can feed here
|
(encoder)
|
bottleneck <- you can also feed here instead
|
(decoder)
|
output
You want two use cases:
input, compute the outputYou don't need to create a placeholder for bottleneck, because sess.run() allows you to feed values to non placeholders in the Graph:
input_shape = [None, in_size]
input = tf.placeholder(tf.float32, input_shape)
# ...
bottleneck = f(prev_layer) # of shape [None, bottleneck_size]
# ...
# Using graph with input at first layer:
sess.run([output], feed_dict={input: x})
# Using graph with input at bottleneck layer:
sess.run([output], feed_dict={bottleneck: b})
From the documentation of sess.run():
The optional feed_dict argument allows the caller to override the value of tensors in the graph. Each key in feed_dict can be one of the following types:
If the key is a Tensor, the value may be a Python scalar, string, list, or numpy ndarray that can be converted to the same dtype as that tensor. Additionally, if the key is a placeholder, the shape of the value will be checked for compatibility with the placeholder.
Collected from the Internet
Please contact [email protected] to delete if infringement.
Comments