Compatibility Tips for TensorFlow-1.4.0 about thumt HOT 11 CLOSED

Sorry, the commit 7956f2a is the latest one that supports TensorFlow 1.4.0.

You can also use the following patch codes with the latest commit:

        # Create session, do not use default CheckpointSaverHook
        with tf.train.MonitoredTrainingSession(
                checkpoint_dir=params.output, hooks=train_hooks,
                save_checkpoint_secs=None, config=config) as sess:

            sess._tf_sess().run(restore_op)

            while not sess.should_stop():
                sess._tf_sess().run([init_op, ops["zero_op"]])
                for i in range(update_cycle - 1):
                    sess._tf_sess().run(ops["collect_op"])

                sess.run(ops["train_op"])

from thumt.

2018-04-12 13:10:28.825845: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
2018-04-12 13:10:29.011996: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Found device 0 with properties:
name: GeForce GTX 1080 major: 6 minor: 1 memoryClockRate(GHz): 1.7335
pciBusID: 0000:03:00.0
totalMemory: 7.92GiB freeMemory: 7.81GiB
2018-04-12 13:10:29.012047: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1120] Creating TensorFlow device (/device:GPU:0) -> (device: 2, name: GeForce GTX 1080, pci bus id: 0000:03:00.0, compute capability: 6.1)
Traceback (most recent call last):
File "thumt/bin/trainer.py", line 468, in
main(parse_args())
File "thumt/bin/trainer.py", line 461, in main
sess.run_step_fn(restore_fn)
AttributeError: 'MonitoredSession' object has no attribute 'run_step_fn'

I have encountered the same error .

from thumt.

@KelleyYin Have you solved it?

from thumt.

Upgrading TensorFlow to the latest version will eliminates this error.

from thumt.

@Playinf
Thank you for your reply.
1.The latest version of tensorflow needs cuda9.0, and my Ubuntu server is cuda8.0.
2.I saw the branch of github has updated , so what branch should I use?

from thumt.

You may find commit 2f5c288 helpful.

from thumt.

Thank you. But it failed to solve.
renhongkai at server01 in ~/project/THUMT/THUMT (2f5c288●) (py2.7)
.....
Traceback (most recent call last):
File "thumt/bin/trainer.py", line 468, in
main(parse_args())
File "thumt/bin/trainer.py", line 461, in main
sess.run_step_fn(restore_fn)
AttributeError: 'MonitoredSession' object has no attribute 'run_step_fn'

from thumt.

@Playinf Thank you very much. This problem has been solved.

from thumt.

@renhongkai Can you tell me how you solve this problem? I met the same thing 'AttributeError: 'MonitoredSession' object has no attribute 'run_step_fn''. After I use the patch code Playinf describes, it can only run 1 step and then another error ocuured.

from thumt.

@rgwt123 I have tested the patch codes under TensorFlow-1.4.0 and did not found any problems.

from thumt.

We recommend users to use the latest version of TensorFlow. We may drop the compatibility of old TensorFlow versions without notice.

from thumt.