Closed ghost closed 7 years ago
Possibly related:
If I change session initialisation to with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
, then instead of the above error it just hangs quietly for a bit after cl_driver DeviceAllocate
and then segfaults.
split
is missing currently. It's on my list of things to do https://github.com/hughperkins/tf-coriander/issues/33
thoughts on relative priority of tf.split
vs tf.random_normal
?
You'd know significantly more than I on their relative importance, I'm just getting started at TF. :)
Probably RNG is more important, because it better-enables operations in the first place. My poor understanding is that Split is used for multi-device learning? Which is great for those with multi-GPU setups but irrelevant to possibly-most users.
I don't know if this is related to the question about Random being needed to seed networks, but the 3_NeuralNetworks/multilayer_perceptron.py
example ends for me with Accuracy: 0.1135
... that seems suspiciously bad for an example network. Could that be a quiet bug all on its own? If so I can open another issue..
Nevermind, I just grepped the files and I see they're all explicitly seeded from tf.random_normal
, so if that's not working at all then it's clearly a priority to get things working again.
You can seed them using np.random.randn(3,4).astype(np.float32)
(obviously, update the shape as needed)
I'm going to handle random normal first, since:
In the meantime, I'm going to close this issue, 37, as a duplicate of 33 https://github.com/hughperkins/tf-coriander/issues/33 , ok?
Thanks! Sorry, my inexperience lead to me missing that this and Split were the same issue.
On 30 May 2017 12:50:32 GMT+01:00, Hugh Perkins notifications@github.com wrote:
I'm going to handle random normal first, since:
- I already started on it
- prefer consistency across platforms
- need to fix random normal soonish anyway, and
- Adam optimizer will give junky results without it
In the meantime, I'm going to close this issue, 37, as a duplicate of 33 https://github.com/hughperkins/tf-coriander/issues/33 , ok?
-- You are receiving this because you authored the thread. Reply to this email directly or view it on GitHub: https://github.com/hughperkins/tf-coriander/issues/37#issuecomment-304854605
-- Sent from my Android device with K-9 Mail. Please excuse my brevity.
No worries. Thank you very much for starting to investigate how much of tf-coriander is working for you, and what bits are missing. This is very helpful :-)
So, just decided to pull https://github.com/hughperkins/TensorFlow-Examples and run a few of the examples, to see how things are going since the fix to #34 and the addition of working ADAM.
The examples that specify a device always crash. Here's an example for 3_NeuralNetworks/dynamic_rnn.py:
If I change the specification to :1 instead of :0 (because ???) I get this instead:
My
clinfo
output: