Discussion and a few observations about easynmt HOT 2 CLOSED

Hi,

1. Is there any error?

2. This can be ignored

3. Yes, the opus-mt models has some issues. It was trained on a rather small dataset consisting of sentences. If you translated words or short phrases with it, it can lead to these strange behavior. The M2M models are more robust (im my impression). Sadly there is not easy fix for the opus-mt models

4. Running on a GPU significantly improves the performance. But otherwise, there is sadly not much tuning possible as transformers are quite compute expensive on CPU. You can use Google Colab to get a free GPU. See the links in the Readme for some examples how to run the code or a REST API on Google Colab.

from easynmt.

regarding:

1. I originally started out trying to build my own image using pip3 to install easynmt. Build failed a number of times. I would try to figure out what would satisfy the issue and pip3 install that before the pip3 install easynmt so the requirement would be met. I ran out of the allocated time before I got it working. The last issue that got me red text was sentencepiece. Next day I pulled your image and have been using it since.
2. Consider it done.
3. I tried to load m2m_100_1.2B but it went into a loop tripping a watchdog timer. Then I loaded m2m_100_418M, but it translated "eighteen" to "18" and that isn't much use when building word match flashcards. I went back to opus-mt. We are doing a few manual corrections, as needed. Most of the translating that we are doing is for one time use. Most TV isn't worth watching a second time.
4. I understand the need for a gpu. I'm running on a Dell PowerEdge R610 and there really isn't an easy way to get a gpu on board. I translated the Swedish CEFR Kelly list yesterday (8500 words, one at a time) and that went quickly enough. When we are doing a TV program, there are about 2500 lines of closed captions in a 90 minute program and so far that has been about 1700 unique words. Some are, of course, people and place names. I am getting good enough performance to just keep chugging along.
5. Again.

Feel free to close this issue when you tire of discussion.

from easynmt.