Mercurial > pylearn
comparison doc/v2_planning/dataset.txt @ 1084:7e6e77d50eeb
dataset: I say the learner committee should take care of dataset as well
| author | Olivier Delalleau <delallea@iro> |
|---|---|
| date | Fri, 10 Sep 2010 17:06:38 -0400 |
| parents | 4c00af69c164 |
| children | de456561ec40 |
comparison
equal
deleted
inserted
replaced
| 1083:4c00af69c164 | 1084:7e6e77d50eeb |
|---|---|
| 255 we may need to revise our idea of what 'mini' means) Hopefully the answer to | 255 we may need to revise our idea of what 'mini' means) Hopefully the answer to |
| 256 that last question is no, as I think it would definitely keep things simpler, | 256 that last question is no, as I think it would definitely keep things simpler, |
| 257 since we could simply use numpy arrays (for numeric data) or lists (for | 257 since we could simply use numpy arrays (for numeric data) or lists (for |
| 258 anything else) to store mini-batches' data. So I vote for 'no'. | 258 anything else) to store mini-batches' data. So I vote for 'no'. |
| 259 | 259 |
| 260 A dataset is a learner | |
| 261 ~~~~~~~~~~~~~~~~~~~~~~ | |
| 262 | |
| 263 OD: This is more a high-level comment that may or may not be relevant | |
| 264 depending on how we get to plug our different classes together. | |
| 265 In PLearn (old C++ lisa ML library) we had *lots* of dataset subclasses doing | |
| 266 all sorts of fancy things, the majority of these classes taking as input | |
| 267 another dataset, and transforming it in some way (e.g. taking a subset of | |
| 268 samples, a subset of features, normalizing features, computing extra fields | |
| 269 given existing fields, etc.). I think right now our interface is heading in a | |
| 270 similar direction. | |
| 271 When you think about it, this kind of operation is equivalent to writing a | |
| 272 learner class that is trained on the input dataset, and whose output on this | |
| 273 same dataset is used to obtain an output dataset (note that the training phase | |
| 274 may do nothing, e.g. if the goal is only to filter out a predefined set of | |
| 275 samples). | |
| 276 If you push it even further, even a dataset that has no input dataset, say | |
| 277 e.g. a dataset view of a 2D numpy matrix, can be seen as the output of a | |
| 278 learner that was trained on nothing and whose output is computed on nothing | |
| 279 (but still outputs this 2D matrix). | |
| 280 In the small ML library I have been using at Ubisoft, the dataset class | |
| 281 actually inherits from learner, based on this point of view. Actually pretty | |
| 282 much all objects that are plugged together to make an experiment are learners. | |
| 283 The main advantage is everything has the same interface and the "plugging" of | |
| 284 the different parts can remain very simple. Confusion is avoided by the module | |
| 285 hierarchy to ensure objects with different behavior have different names. | |
| 286 Something like dataset.MatrixDataset would create a dataset from scratch (i.e. | |
| 287 a numpy matrix), process.FilterSamples would be something that does not need | |
| 288 to be trained, but needs an input dataset, and learner.NNet would be a usual | |
| 289 learning algorithm that must be trained on an input dataset, and computes an | |
| 290 output (possibly on the same dataset, possibly on another one). | |
| 291 | |
| 292 Ok, this is getting too long, I am definitely not saying we should do this, | |
| 293 but I think there is some close relationship between the usual data processing | |
| 294 we do and the learning process, so it may be worth thinking how to put them | |
| 295 together in a coherent framework. For instance, in PLearn there was (something | |
| 296 like) a NormalizeVMatrix (think of it as a dataset subclass), but it could | |
| 297 not be used in a natural way to learn the normalization parameters on a | |
| 298 training set (e.g. mean and std of features) and normalize another dataset. | |
| 299 Instead you could use (something like) a | |
| 300 PLearnerOutputVMatrix(learner=NormalizeLearner(train_on=....)). Having both | |
| 301 ways to do (almost the) same thing can be confusing. | |
| 302 |