Datamodule
PickledDataWDS
Bases: WebDataModule
A LightningDataModule to process pickled data into webdataset tar files.
PickledDataWDS
is a LightningDataModule to process pickled data into webdataset tar files
and setup dataset and dataloader. This inherits the webdataset setup from its parent module
WebDataModule
. This data module takes a directory of pickled data files, data filename
prefixes for train/val/test splits, data filename suffixes and prepare webdataset tar files
by globbing the specific pickle data files {dir_pickles}/{name_subset[split]}.{suffix_pickles}
and outputing to webdataset tar file with the dict structure:
{"__key__" : name.replace(".", "-"),
suffix_pickles : pickled.dumps(data) }
pipeline_wds
workflow. In its train/val/test_dataloader(), it creates the
WebLoader object chaining up the pipeline_prebatch_wld
workflow.
Examples:
- create the data module with a directory of pickle files and the file name
prefix thereof for different splits to used by
Lightning.Trainer.fit()
>>> from bionemo.core.data.datamodule import Split, PickledDataWDS
>>> dir_pickles = "/path/to/my/pickles/dir"
>>> # the following will use `sample1.mydata.pt` and `sample2.mydata.pt` as the
>>> # training dataset and `sample4.mydata.pt` and `sample5.mydata.pt` as the
>>> # validation dataset
>>> suffix_pickles = "mydata.pt"
>>> names_subset = {
>>> Split.train: [sample1, sample2],
>>> Split.val: [sample4, sample5],
>>> }
>>> # the following setting will attempt to create at least 5 tar files in
>>> # `/path/to/output/tars/dir/myshards-00000{0-5}.tar`
>>> n_tars_wds = 5
>>> prefix_tars_wds = "myshards"
>>> output_dir_tar_files = {
Split.train : "/path/to/output/tars/dir-train",
Split.val : "/path/to/output/tars/dir-val",
Split.test : "/path/to/output/tars/dir-test",
}
>>> # see the `WebDataModule` API doc for the definition of global_batch_size
>>> global_batch_size = 16
>>> # user can optionally customize the data processing routines and kwargs used
>>> # in the WebDataset and WebLoader (see the examples in `WebDataModule`)
>>> pipeline_wds = { Split.train: ... }
>>> pipeline_prebatch_wld = { Split.train: ... }
>>> kwargs_wds = { Split.train: ..., Split.val: ... }
>>> kwargs_wld = { Split.train: ..., Split.val: ... }
>>> # create the data module
>>> data_module = PickledDataWDS(
>>> dir_pickles,
>>> names_subset,
>>> suffix_pickles, # `WebDataModule` args
>>> output_dir_tar_files, # `WebDataModule` args
>>> global_batch_size, # `WebDataModule` args
>>> n_tars_wds=n_tars_wds,
>>> prefix_tars_wds=prefix_tars_wds, # `WebDataModule` kwargs
>>> pipeline_wds=pipeline_wds, # `WebDataModule` kwargs
>>> pipeline_prebatch_wld=pipelines_wdl_batch, # `WebDataModule` kwargs
>>> kwargs_wds=kwargs_wds, # `WebDataModule` kwargs
>>> kwargs_wld=kwargs_wld, # `WebDataModule` kwargs
>>> )
Source code in bionemo/webdatamodule/datamodule.py
326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 |
|
__init__(dir_pickles, names_subset, *args, n_tars_wds=None, **kwargs)
Constructor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
dir_pickles
|
str
|
input directory of pickled data files |
required |
names_subset
|
Dict[Split, List[str]]
|
list of filename prefix of the data samples to be loaded in the dataset and dataloader for each of the split |
required |
*args
|
arguments passed to the parent WebDataModule after its |
()
|
|
n_tars_wds
|
Optional[int]
|
attempt to create at least this number of webdataset shards |
None
|
**kwargs
|
arguments passed to the parent WebDataModule |
{}
|
Source code in bionemo/webdatamodule/datamodule.py
407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 |
|
prepare_data()
This is called only by the main process by the Lightning workflow.
Do not rely on this data module object's state update here as there is no
way to communicate the state update to other subprocesses. The nesting
pickles_to_tars
function goes through the data name prefixes in the
different splits, read the corresponding pickled file and output a
webdataset tar archive with the dict structure: {"key" :
name.replace(".", "-"), suffix_pickles : pickled.dumps(data) }.
Source code in bionemo/webdatamodule/datamodule.py
441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 |
|
Split
Bases: Enum
Names for each data split.
Source code in bionemo/webdatamodule/datamodule.py
27 28 29 30 31 32 |
|
WebDataModule
Bases: LightningDataModule
A LightningDataModule for using webdataset tar files.
WebDataModule
is a LightningDataModule
for using webdataset tar files to setup PyTorch
datasets and dataloaders. This data module takes as input a dictionary: Split -> tar file
directory and vaiours webdataset config settings. In its setup() function, it creates the
webdataset object chaining up the input pipeline_wds
workflow. In its train/val/test_dataloader(),
it creates the WebLoader object chaining up the pipeline_prebatch_wld
workflow.
Examples:
-
create the data module with input directory to webdataset tar files. Depending on which of the downstream Lightning.Trainer methods are called, e.g.,
Trainer.fit()
,Trainer.validate()
,Trainer.test()
orTrainer.predict()
, only a subset of the train, val and test splits need to be specified in the various input options to the data module: -
Trainer.fit()
requires thetrain
andval
splits Trainer.validate()
requires theval
splitTrainer.test()
requires thetest
splitsTrainer.predict()
requires thetest
splits
Here is an example of constructing the data module for Trainer.fit()
:
>>> from bionemo.webdatamodule.datamodule import Split, WebDataModule
>>>
>>> tar_file_prefix = "shards"
>>>
>>> dirs_of_tar_files = {
>>> Split.train: "/path/to/train/split/tars",
>>> Split.val: "/path/to/val/split/tars",
>>> }
>>>
>>> n_samples {
>>> Split.train: 1000,
>>> Split.val: 100,
>>> }
>>>
>>> # this is the string to retrieve the corresponding data object from the
>>> # webdataset file (see
>>> # https://github.com/webdataset/webdataset?tab=readme-ov-file#the-webdataset-format
>>> # for details)
>>> suffix_keys_wds = "tensor.pyd"
>>>
>>> # see the API doc for the definition of global_batch_size
>>> global_batch_size = 16
>>>
>>> seed = 27193781
>>>
>>> # Specify the routines to process the samples in the WebDataset object.
>>> # The routine is a generator of an Iterable of generators that are chained
>>> # together by nested function calling. The following is equivalent of
>>> # defining a overall generator of `shuffle(untuple(...))` which
>>> # untuples the samples and shuffles them. See webdataset's Documentation
>>> # for details.
>>> # NOTE: the `untuple` is almost always necessary due to the webdataset's
>>> # file parsing rule.
>>>
>>> untuple = lambda source : (sample for (sample,) in source)
>>>
>>> from webdatast import shuffle
>>> pipeline_wds = {
>>> Split.train : [untuple, shuffle(n_samples[Split.train],
>>> rng=random.Random(seed_rng_shfl))],
>>> Split.val: untuple
>>> }
>>>
>>> # Similarly the user can optionally define the processing routine on the
>>> # WebLoader (the dataloader of webdataset).
>>> # NOTE: these routines by default take unbatched sample as input so the
>>> # user can customize their batching routines here
>>>
>>> batch = batched(local_batch_size, collation_fn=lambda
list_samples : torch.vstack(list_samples))
>>> pipeline_prebatch_wld = {
Split.train: [shuffle(n_samples[Split.train],
rng=random.Random(seed_rng_shfl)), batch],
Split.val : batch,
Split.test : batch
}
>>>
>>> # the user can optionally specify the kwargs for WebDataset and
>>> # WebLoader
>>>
>>> kwargs_wds = {
>>> split : {'shardshuffle' : split == Split.train,
>>> 'nodesplitter' : wds.split_by_node,
>>> 'seed' : seed_rng_shfl}
>>> for split in Split
>>> }
>>>
>>> kwargs_wld = {
>>> split : {"num_workers": 2} for split in Split
>>> }
>>>
>>> # construct the data module
>>> data_module = WebDataModule(n_samples, suffix_keys_wds,
dirs_of_tar_files, global_batch_size,
prefix_tars_wds=tar_file_prefix,
pipeline_wds=pipeline_wds,
pipeline_prebatch_wld=pipeline_prebatch_wld,
kwargs_wds=kwargs_wds,
kwargs_wld=kwargs_wld)
Source code in bionemo/webdatamodule/datamodule.py
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 |
|
__init__(n_samples, suffix_keys_wds, dirs_tars_wds, global_batch_size, prefix_tars_wds='wdshards', pipeline_wds=None, pipeline_prebatch_wld=None, kwargs_wds=None, kwargs_wld=None)
Constructor.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
n_samples
|
Dict[Split, int]
|
input dictionary: Split -> number of data samples for each split |
required |
suffix_keys_wds
|
Union[str, Iterable[str]]
|
a set of keys each corresponding to a data object in the webdataset tar file dictionary. The data objects of these keys will be extracted and tupled for each sample in the tar files |
required |
dirs_tars_wds
|
Dict[Split, str]
|
input dictionary: Split -> tar file directory that contains the webdataset tar files for each split |
required |
global_batch_size
|
int
|
size of batch summing across nodes in Data
Distributed Parallel, i.e., local_batch_size * n_nodes. NOTE:
this data module doesn't rely on the input |
required |
Kwargs: prefix_tars_wds: name prefix of the input webdataset tar files. The input tar files are globbed by "{dirs_tars_wds[split]}/{prefix_tars_wds}-*.tar" pipeline_wds: a dictionary of webdatast composable, i.e., functor that maps a iterator to another iterator that transforms the data sample yield from the dataset object, for different splits, or an iterable to such a sequence of such iterators. For example, this can be used to transform the sample in the worker before sending it to the main process of the dataloader pipeline_prebatch_wld: a dictionary of webloader composable, i.e., functor that maps a iterator to another iterator that transforms the data sample yield from the WebLoader object, for different splits, or an iterable to a seuqnence of such iterators. For example, this can be used for batching the samples. NOTE: this is applied before batching is yield from the WebLoader kwargs_wds: kwargs for the WebDataset.init() kwargs_wld : kwargs for the WebLoader.init(), e.g., num_workers, of each split
Source code in bionemo/webdatamodule/datamodule.py
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 |
|
predict_dataloader()
Alias for :func:test_dataloader
.
Source code in bionemo/webdatamodule/datamodule.py
321 322 323 |
|
prepare_data()
This is called only by the main process by the Lightning workflow.
Do not rely on this data module object's state update here as there is no way to communicate the state update to other subprocesses. Is a no-op.
Source code in bionemo/webdatamodule/datamodule.py
224 225 226 227 228 229 230 |
|
setup(stage)
This is called on all Lightning-managed nodes in a multi-node training session.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
stage
|
str
|
"fit", "test" or "predict" |
required |
Source code in bionemo/webdatamodule/datamodule.py
259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 |
|
test_dataloader()
Webdataset for the test data.
Source code in bionemo/webdatamodule/datamodule.py
317 318 319 |
|
train_dataloader()
Webdataset for the training data.
Source code in bionemo/webdatamodule/datamodule.py
309 310 311 |
|
val_dataloader()
Webdataset for the validation data.
Source code in bionemo/webdatamodule/datamodule.py
313 314 315 |
|