I was trying to make my BBQ cook recipes with machine for a week. Here is the experience what I learned from trying to implement it.
First Lesson: Always test training process with small circle before running entire session.
Test your script with 1 epoch. It can save your 30 hours work without sigh.
Let us take a look at code snippet after sigh.
If you follow the tutorial to train a LSTM, you will copy below command and change “my_own_data” in train.py file.
However, it will return an empty string.
Second Lesson: Run your training program on a GPU machine if possible.
Machine learning is really time consuming.
and TensorFlow do not support GPU mode on Mac OS.
This training was crushed on my laptop. The actual steps I need is 1,320,000.
Koji helped me run my training in NYU HPC machine. However, the training still got crashed after one day for uncertain issue. sigh
Third Lesson: Never train your data during the finals on a shared machine.
or it will crush with Unreal…..
Fourth Lesson: Checkpoint is my lifesaver.
Finally, I gave up training my own BBQ receipts data and feed myself TensorFlow documents with learned helplessness. Wait, what is Checkpoint?
It seems possible to restart training process. However, I did not want to continue my data training anymore. But, I still need to show something for my NOC final. Could I use the unfinished training data to generator a model?
After modified the main function in train.py file, I got my 66% unfinished BBQ receipts LSTM model. I can not wait to test it.
(Note: Only works on Firefox)
seed text “- – – – – – – – – ” with temperature 1 will come with the best result.
Bonus: chicken wings receipts model.
Try start with ———————
seed text “—— ————-” with temperature 0.5 will come with the best result.