Mercurial > pylearn
comparison doc/v2_planning/dataset.txt @ 1104:5e6d7d9e803a
a comment on the GPU issue for datasets
| author | Razvan Pascanu <r.pascanu@gmail.com> |
|---|---|
| date | Mon, 13 Sep 2010 20:21:23 -0400 |
| parents | 75175e2e697d |
| children | 546bd0ccb0e4 |
comparison
equal
deleted
inserted
replaced
| 1103:56c5f0990869 | 1104:5e6d7d9e803a |
|---|---|
| 322 OD (continued): After thinking a bit more about it, I am not sure that would | 322 OD (continued): After thinking a bit more about it, I am not sure that would |
| 323 work. I definitely need to look at some code doing it to get a better | 323 work. I definitely need to look at some code doing it to get a better |
| 324 understanding of it, but my feeling is that you need your learner to be | 324 understanding of it, but my feeling is that you need your learner to be |
| 325 written in a specific way to achieve this, in which case it may be up to the | 325 written in a specific way to achieve this, in which case it may be up to the |
| 326 learner to take its input data and store it into a shared variable. | 326 learner to take its input data and store it into a shared variable. |
| 327 | |
| 328 RP comment: Yes, the dataset object alone can not handle this, the issue is somewhere | |
| 329 between the dataset and the learner. Or in other words, everytime you change | |
| 330 the data you need to recompile your theano function. So the learner can not | |
| 331 only get data from the dataset, it needs to get a shared variable. The learner | |
| 332 should also be aware when the dataset is changed, to recompile its internal | |
| 333 functions. I'm not sure which is the best wa to do this. My personal feeling | |
| 334 is that the dataset should be part of the learner. The lerner should provide | |
| 335 a function use_dataset ( or replace_dataset). When this function is called, | |
| 336 all the theano functions in the learner get recompiled based on shared | |
| 337 variables that the dataset object provides. It sort of fits very well in the | |
| 338 framework that I have in mind, which was spattered around in the learner.txt | |
| 339 and some of my previous emails. I think it shares a lot with James concepts, | |
| 340 since it follows quite closely the concepts behind Theano. |