Problem with AdamOptimizer? about tensorflow-seq2seq-tutorials HOT 5 CLOSED

Hi! Traceback you've shown indicates that you're running ops on GPU. I guess you've changed device manually (unless TF have some environment override for default device I'm not aware of). Could you post the code you modified?

Also, do you by any change run optimizer like tf.train.AdamOptimizer().minimize(loss, colocate_gradients_with_ops=True)?

from tensorflow-seq2seq-tutorials.

Hi,
I did not modify your code at all. Just a dry run gave me that error. So I don't think it was due to any changes in the code.
Also, no I didn't run the optimizer like the one you mentioned.

from tensorflow-seq2seq-tutorials.

This is extremely strange. Trace clearly shows you are running the tutorial on GPU, while nothing in it changes the device (and it doesn't change it for me).

Please try running with soft device placement: sess = tf.InteractiveSession(config=tf.ConfigProto(allow_soft_placement=True))

from tensorflow-seq2seq-tutorials.

I assume the issue was resolved. Closing it now.

from tensorflow-seq2seq-tutorials.

I had similar issue and after looking into the issue it seems the issue is with embedding. It seems there is no GPU implementation of embedding. One can declare the embedding variable to use cpu then it works fine, e.g.:

with tf.device("/cpu:0"):
    embeddings = tf.Variable(tf.random_uniform([vocab_size, input_embedding_size], -1.0, 1.0), dtype=tf.float32)

Ref: tensorflow/tensorflow#5117

from tensorflow-seq2seq-tutorials.