Running the model completely on CPU gives -1 BLEU about unsupervisedmt HOT 8 CLOSED

Closing this for now. I will try running this for more epochs sometime later. If issues persists I will reopen.
Thanks for you help @glample appreciate it.

from unsupervisedmt.

Hi,

When you say you are getting very different results when training do you mean compared to when you run the code on GPU? Did you compare a CPU and GPU experiment? Can you provide the train.log as well? I'm not sure what is happening here, performance should be same on CPU and GPU so I'm curious to see the training losses.

from unsupervisedmt.

I could not find the train.log. As I noticed my dumped folder was not getting created correctly. I'm guessing it is because of Windows/Linux directory structure differences.
Let me fix this in sometime and get back to you.
Thanks !

from unsupervisedmt.

@glample I'm assuming GPU runs perfectly fine as no changes are required. Here is my train log for CPU run. It's only for couple of steps but it does gives the same error as before.
train.log
On GPU it gives me the correct output, like this:

INFO - 10/11/18 02:23:10 - 0:03:24 - Creating new training otf,fr iterator ...
INFO - 10/11/18 02:23:17 - 0:03:32 - Creating new training otf,en iterator ...
INFO - 10/11/18 02:29:41 - 0:09:56 - 50 - 11.87 sent/s - 302.00 words/s - XE-en-en: 9.0458 || XE-fr-fr: 9.3363 || XE-fr-en-fr: 8.8043 || XE-en-fr-en: 9.5765 || ENC-L2-en: 4.4486 || ENC-L2-fr: 4.4885 - LR enc=1.0000e-04,dec=1.0000e-04 - Sentences generation time: 128.76s (23.88%)
INFO - 10/11/18 02:34:47 - 0:15:01 - 100 - 20.95 sent/s - 537.00 words/s - XE-en-en: 6.7902 || XE-fr-fr: 6.8785 || XE-fr-en-fr: 6.5884 || XE-en-fr-en: 7.1309 || ENC-L2-en: 4.1948 || ENC-L2-fr: 4.1622 - LR enc=1.0000e-04,dec=1.0000e-04 - Sentences generation time: 27.85s (9.12%)
INFO - 10/11/18 02:39:52 - 0:20:06 - 150 - 20.99 sent/s - 576.00 words/s - XE-en-en: 6.3104 || XE-fr-fr: 6.1536 || XE-fr-en-fr: 6.1312 || XE-en-fr-en: 6.5265 || ENC-L2-en: 4.2309 || ENC-L2-fr: 4.1930 - LR enc=1.0000e-04,dec=1.0000e-04 - Sentences generation time: 25.22s (8.27%)
INFO - 10/11/18 02:44:44 - 0:24:59 - 200 - 21.90 sent/s - 585.00 words/s - XE-en-en: 6.0423 || XE-fr-fr: 5.9185 || XE-fr-en-fr: 5.9556 || XE-en-fr-en: 6.4293 || ENC-L2-en: 4.2429 || ENC-L2-fr: 4.2192 - LR enc=1.0000e-04,dec=1.0000e-04 - Sentences generation time: 15.95s (5.46%)

from unsupervisedmt.

How do you know this is the correct output? For CPU it looks like you trained for 6 epochs and still get -1 BLEU. How much do you get if you train for 6 epochs on GPU?

from unsupervisedmt.

@glample
Here is the train.log for the run on GPU (not complete just 8 epochs). Blues scores show 0.0 , it doesn't give the warning message like the CPU run. Is this expected ?
FAIR_train_GPU.log

from unsupervisedmt.

Hey,
You are using --epoch_size 500, why? With this, the model only trains on a few batches per epoch. Default value for this is --epoch_size 500000.
Maybe you can try something in between --epoch_size 100000 since CPU will be very slow, and check that you have the same performance at the end of one GPU and one CPU epoch.

from unsupervisedmt.

@glample 500 , Just because I wanted to see if my CPU level changes to all files were correct and does it even start to run, which I'm guessing that all all changes were positive and the model runs fine. It just need to run more epochs to give good BLEU score.

from unsupervisedmt.