Building it on Cloud-9. Pywrap_tensorflow not found.

[Eulerianial]

Siraj,

While running on a Cloud-IDE, I see the following issue:
libdc1394 error: Failed to initialize libdc1394
Traceback (most recent call last):
File "atari.py", line 5, in
from train import Trainer
File "/home/ubuntu/workspace/Game-AI/train.py", line 2, in
import tensorflow as tf
File "/home/ubuntu/workspace/tensorflow/tensorflow/init.py", line 23, in
from tensorflow.python import *
File "/home/ubuntu/workspace/tensorflow/tensorflow/python/init.py", line 48, in
from tensorflow.python import pywrap_tensorflow
ImportError: cannot import name pywrap_tensorflow

Any clue/tips to help here?

Best,
Vijay

[llSourcell]

Hmm looks like you are running this on an Ubuntu machine. Have you checked to make sure you've installed the ubuntu version of tensorflow instead of the mac os x version?

https://www.tensorflow.org/versions/r0.8/get_started/os_setup.html

[Eulerianial]

Siraj, I installed the linux version only:
pip install --upgrade https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-0.5.0-cp27-none-linux_x86_64.whl

I just checked my other tensor flow code is working. For example, I was able to run this fine:

import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()
print(sess.run(hello))

a = tf.constant(10)
b = tf.constant(32)

print(sess.run(a + b))

[Eulerianial]

Siraj, quickly to share, I am able to start the training process for your code.
Just to ask, how much time it took you for training?

[llSourcell]

Awesome, glad to hear it! What was the problem? It took me a full day. Keep me updated with your results!

[Eulerianial]

I moved to the local build as opposed to cloud and it worked there. Now I saw this issues:

Traceback (most recent call last):
File "atari.py", line 61, in
env.gym.monitor.start(args.out, force=True)
File "/usr/local/lib/python2.7/dist-packages/gym/monitoring/monitor.py", line 113, in start
if not os.path.exists(directory):
File "/usr/lib/python2.7/genericpath.py", line 18, in exists
os.stat(path)
TypeError: coercing to Unicode: need string or buffer, NoneType found

So working on that. Any tips?

[llSourcell]

yup, i pushed a fix for that recently, pull the latest from master. It should work now, let me know if it doesn't.

[Eulerianial]

Sure, I will try again. And just to share I am running the default environment.
python atari.py

[Eulerianial]

Also, when I start it, it says:
libdc1394 error: Failed to initialize libdc1394
[2016-05-18 09:10:34,500] Making new env: SpaceInvaders-v0
start 50000 random plays to populate replay memory
What is that error?

[llSourcell]

Hmm libdc1394 is related to OpenCV. Some people have had success by disabling the driver http://stackoverflow.com/questions/12689304/ctypes-error-libdc1394-error-failed-to-initialize-libdc1394 another possible solution is installing OpenCV via Anaconda

download anaconda: https://www.continuum.io/downloads

conda install -c condo-forge opencv

[Eulerianial]

Looks like it moved ahead. The simulation ended with this:

Training step: 17120
memory size: 67120
Learning rate: 0.9845929
Successes: 916
Failures: 97
Sample successes: 1859
Sample failures: 214
Average batch loss: 0.0286087147397
Batch training time: 0.643544562161 s

Training step: 17152

memory size: 67152

Now do we have to push these to Open-AI evaluations to see the game and graphics? How to do that?

[Eulerianial]

Siraj, may I know at the end of the simulation the game will start automatically? Somehow it is not happening for me, though looks like Tensor flow is able to train the agent.

[llSourcell]

^ hey sorry for the late reply @Eulerianial i went ahead and fixed the readme, you just have to add '--display true' to the command line function when you run it