Closed chrhansen closed 8 years ago
At this point this PR only fixes the op-naming issues and adds a few extra specs, but I'm still working on implementing tf.zeros
and fixing tf.constant
. Merging wont do harm, should it happen.
@chrhansen I will make the constant work with the specifications soon. Thanks a lot for this.
@Arafatk well, wait off with that – as I wrote above I'm already working on that.
@chrhansen Just one thing, I have done something similar in image branch for this. Thanks
@Arafatk ?
@chrhansen In image recognition tutorial, I had made a few changes so that I could read scalars but it was not the best way to make scalars work.
I began implementing
tf.zeros
(a special kind of constant-op with all 0s – taking only shape as required argument) and found that ourconstant()
-method doesn't work with scalars and doesn't takeshape
as an argument. Writing specs for that, I then discovered that ourdefault_names
- method uses a character,:
, that makes it unretrievable from the TensorFlow graph.:
is not allowed in names on the actual (not the wrapper) TensorFlow graphconstant()
compliant (https://www.tensorflow.org/versions/r0.10/api_docs/python/constant_op.html#constant), e.g. work on scalars and takeshape
as argumentzeros()