Input tensor is not saved in execute.py

[dvorotnev]

When I try to save an input tensor, which is randomly generated, during execution of simple example:

import tensorflow as tf
import nnef_tools.io.tf.graphdef as graphdef

def testnet_abs():
    x = tf.placeholder(tf.float32, shape=[6, 32, 32, 3], name='input')
    return tf.abs(x)

tf.reset_default_graph()
with tf.Session() as sess:
    result = testnet_abs()
    sess.run(tf.global_variables_initializer())
    graphdef.save_default_graph("model.pb", session=sess, outputs={result: "output"})

the script execute.py doesn't save it, because in TFExecutor::__call__ it is discarded by the condition if len(op.inputs) there. I think this condition should be if len(op.outputs), shouldn't it?

Steps to reproduse:

python3 ./test.py
python -m nnef_tools.convert --input-format=tf --output-format=nnef --input-model=./model.pb --optimize --keep-io-names --output-model=model.nnef
python -m nnef_tools.execute ./model.pb  --format=tf --random='uniform(0,1)' --output-names='{"input:0": "input", "output:0": "output"}' --output-path=./

python output:

Written ./output.dat

after fix python output is:

Written ./input.dat
Written ./output.dat

[gyenesvi]

No, that's not a mistake, that condition is exactly there to guard against outputting the inputs and constants. Not sure exactly why it was there, but it may have caused an error somewhere, let me think about it.

[gyenesvi]

I believe that the condition was a relic coming from the time before TFExecutor was refactored into a class, and got a separate init and call phase. The same condition is there in the init phase, when the implicit outputs of the model are enumerated, and it makes sense there. If I remove the condition altogether, the execution still succeeds, and I don't see a problem with it right now, so I have just removed it. The pull request can be discarded as it is fixed now. Let me know if it is saving fine on your side.

[dvorotnev]

It works. Thank you!