search

JiahuiYu / neuralgym

Deep Learning Toolkit
MIT License
101 stars 97 forks source link

epoch, step, pse setting #3

Closed zengyh1900 closed 6 years ago

zengyh1900 commented 6 years ago

Hi Jiahui,

Thanks for your toolkit, it's really helpful !

In my world, an epoch means running through the training dataset once. So I'm not sure why you set an epoch as TRAIN_SPE steps ignoring the size of training dataset.

see progress_logger

JiahuiYu commented 6 years ago

Hi it is a good question. You are right, basically one epoch means running through the training dataset once. However in some cases, like if you sample a small patch in an image (for tasks like inpainting, super-resolution, etc.), you don't know how to measure one epoch. Also sometimes you just want to save the model frequently (like for every 0.5 epoch), it is better to define an epoch as how many steps to run as a whole unit.

Thus I use SPE as steps per epoch so that you can define how many iterations you want as a running unit. Just some preferences in developing.

Again, you are right. An epoch means running through the training dataset once.

zengyh1900 commented 6 years ago

@JiahuiYu Thank you for your fast reply very much!

muaxiu commented 5 years ago

Does this mean that the train always covers all train dataset? Just like this: 3209A94E-828E-4D77-A5C4-D7FF98DE2A4E

JiahuiYu commented 5 years ago

Yes, it is also possible that one SPE is larger than the length of train dataset.

muaxiu commented 5 years ago

Yes, it is also possible that one SPE is larger than the length of train dataset.

Thank you very much!