Wercker fails

[dirkweissenborn]

I removed a redundant initialization statement (we were calling it once in reader.setup_from_data and in after that directly again in reader.train).

Now overfitting fails, which doesn't make a lot of sense, because I thought calling initializers multiple times is idempotent.

Any ideas @pminervini ?

[pminervini]

I'm not sure that calling initializers multiple times is idempotent .. you are now initializing the parameters with different values, and the training log can be different from the one it used to be.

I can fix this by updating the training log in the tests - give me a few minutes.

[dirkweissenborn]

Well it was at least in earlier TF versions. Anyway, this commit is necessary to not duplicate initialization especially if it is not idempotent.

Thanks!

[pminervini]

I'm not totally sure, let's quickly check:

In [7]: z = tf.get_variable(shape=[5], name='z')
In [8]: init_op = tf.global_variables_initializer()
In [9]: session = tf.InteractiveSession()
In [10]: session.run(init_op)
In [11]: session.run(z)
Out[11]: array([ 0.01121974, -0.20893741,  0.36991453,  0.48889792, -0.07987314], dtype=float32)
In [12]: session.run(init_op)
In [13]: session.run(z)
Out[13]: array([ 0.33691907, -0.73696464, -0.52026761, -0.72819018,  0.40745521], dtype=float32)

[dirkweissenborn]

Interesting, seems that this has changed :) Can you fix the tests?

[dirkweissenborn]

oh you did