I started off by trying to get tensorflow to run on my gpu with cuda and whatnot, but things got way too complex real fast. It’s now on my to-do list. So I did my testing by just running the example on my CPU. I played around with epochs and batchsizes.
Here are the most significant results:
| BATCH SIZE | EPOCHS | LOSS | ACCURACY |
| 32 | 3 | 1.24 | 0.56 |
| 128 | 15 | 0.94 | 0.67 |
| 128 | 3 | 1.49 | 0.46 |
| 16 | 15 | 0.75 | 0.74 |
It’d definitely be interesting to train the model using the intel server, which I assume would be much faster, allowing to try and run for more epochs. As I was curious about epochs and why increasing them increases the accuracy, I did a little reading. Each epoch is basically an iteration over the data set. The weights get updated with each epoch and running through the data more times gives more accurate results (naturally). The batch size on the other hand, is how many training examples to run through before making an update. Making the batch size too large would thus cause a lower accuracy because it would seem as though you’re averaging out more. A batch size that’s too low will potentially be more accurate, but then again, would be super slow. I think 16/32 seemed like the ideal batch size for this particular example. Once I can get my intel cloud server running, I would be interested to even plot some graphs comparing accuracy, epochs, and batch size for larger number of epochs.