o 7wiC#@spdZddlZddlZddlmZddlmZddlm Z m Z m Z e dddgZ Gd d d ZGd d d ZdS) z,Batch collation Authors * Aku Rouhe 2020 N)default_convert) pin_memory)batch_pad_rightmod_default_collate recursive_to PaddedDatadatalengthsc@sbeZdZdZddeiddfddZddZdd Zd d Zd d Z ddZ ddZ e ddZ dS) PaddedBatcha+ Collate_fn when examples are dicts and have variable-length sequences. Different elements in the examples get matched by key. All numpy tensors get converted to Torch (PyTorch default_convert) Then, by default, all torch.Tensor valued elements get padded and support collective pin_memory() and to() calls. Regular Python data types are just collected in a list. Arguments --------- examples : list List of example dicts, as produced by Dataloader. padded_keys : list, None (Optional) List of keys to pad on. If None, pad all torch.Tensors device_prep_keys : list, None (Optional) Only these keys participate in collective memory pinning and moving with to(). If None, defaults to all items with torch.Tensor values. padding_func : callable, optional Called with a list of tensors to be padded together. Needs to return two tensors: the padded data, and another tensor for the data lengths. padding_kwargs : dict (Optional) Extra kwargs to pass to padding_func. E.G. mode, value apply_default_convert : bool Whether to apply PyTorch default_convert (numpy to torch recursively, etc.) on all data. Default:True, usually does the right thing. nonpadded_stack : bool Whether to apply PyTorch-default_collate-like stacking on values that didn't get padded. This stacks if it can, but doesn't error out if it cannot. Default:True, usually does the right thing. Example ------- >>> batch = PaddedBatch([ ... {"id": "ex1", "foo": torch.Tensor([1.])}, ... {"id": "ex2", "foo": torch.Tensor([2., 1.])}]) >>> # Attribute or key-based access: >>> batch.id ['ex1', 'ex2'] >>> batch["id"] ['ex1', 'ex2'] >>> # torch.Tensors get padded >>> type(batch.foo) >>> batch.foo.data tensor([[1., 0.], [2., 1.]]) >>> batch.foo.lengths tensor([0.5000, 1.0000]) >>> # Batch supports collective operations: >>> _ = batch.to(dtype=torch.half) >>> batch.foo.data tensor([[1., 0.], [2., 1.]], dtype=torch.float16) >>> batch.foo.lengths tensor([0.5000, 1.0000], dtype=torch.float16) >>> # Numpy tensors get converted to torch and padded as well: >>> import numpy as np >>> batch = PaddedBatch([ ... {"wav": np.asarray([1,2,3,4])}, ... {"wav": np.asarray([1,2,3])}]) >>> batch.wav # +ELLIPSIS PaddedData(data=tensor([[1, 2,... >>> # Basic stacking collation deals with non padded data: >>> batch = PaddedBatch([ ... {"spk_id": torch.tensor([1]), "wav": torch.tensor([.1,.0,.3])}, ... {"spk_id": torch.tensor([2]), "wav": torch.tensor([.2,.3,-.1])}], ... padded_keys=["wav"]) >>> batch.spk_id tensor([[1], [2]]) >>> # And some data is left alone: >>> batch = PaddedBatch([ ... {"text": ["Hello"]}, ... {"text": ["How", "are", "you?"]}]) >>> batch.text [['Hello'], ['How', 'are', 'you?']] NTc st||_t|d|_g|_g|_|jD]bfdd|D}|r(t|}|dur0|vs<|durSt|dt j rS|j t ||fi|} t || n |rYt|}t |||durg|vss|duryt|dt j ry|j qdS)Nrcsg|]}|qSr ).0examplekeyr U/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/speechbrain/dataio/batch.py xsz(PaddedBatch.__init__..)len_PaddedBatch__lengthlistkeys_PaddedBatch__keys_PaddedBatch__padded_keys_PaddedBatch__device_prep_keysr isinstancetorchTensorappendrsetattrr) selfexamples padded_keysdevice_prep_keys padding_funcpadding_kwargsapply_default_convertnonpadded_stackvaluespaddedr rr__init__is*     zPaddedBatch.__init__cC|jSNrrr r r__len__szPaddedBatch.__len__cCs"||jvr t||Std|)NzBatch doesn't have key: )rgetattrKeyError)rrr r r __getitem__s  zPaddedBatch.__getitem__cstfddjDS)a~Iterates over the different elements of the batch. Returns ------- Iterator over the batch. Example ------- >>> batch = PaddedBatch([ ... {"id": "ex1", "val": torch.Tensor([1.])}, ... {"id": "ex2", "val": torch.Tensor([2., 1.])}]) >>> ids, vals = batch >>> ids ['ex1', 'ex2'] c3s|]}t|VqdSr*)r.)r rr,r r sz'PaddedBatch.__iter__..)iterrr,r r,r__iter__szPaddedBatch.__iter__cCs.|jD]}t||}t|}t|||q|S)z3In-place, moves relevant elements to pinned memory.)rr.recursive_pin_memoryr)rrvaluepinnedr r rrs  zPaddedBatch.pin_memorycOs<|jD]}t||}t|g|Ri|}t|||q|S)zwIn-place move/cast relevant elements. Passes all arguments to torch.Tensor.to, see its documentation. )rr.rr)rargskwargsrr5movedr r rtos  zPaddedBatch.tocCs|j|}t||S)zGets the position.)rr.)rposrr r r at_positions  zPaddedBatch.at_positioncCr))zReturns the bach sizer+r,r r r batchsizeszPaddedBatch.batchsize)__name__ __module__ __qualname____doc__rr(r-r0r3rr:r<propertyr=r r r rr s"S % r c@sPeZdZdZddZddZddZdd Zd d Zd d Z ddZ ddZ dS)BatchsizeGuessera;Try to figure out the batchsize, but never error out If this cannot figure out anything else, will fallback to guessing 1 Example ------- >>> guesser = BatchsizeGuesser() >>> # Works with simple tensors: >>> guesser(torch.randn((2,3))) 2 >>> # Works with sequences of tensors: >>> guesser((torch.randn((2,3)), torch.randint(high=5, size=(2,)))) 2 >>> # Works with PaddedBatch: >>> guesser(PaddedBatch([{"wav": [1.,2.,3.]}, {"wav": [4.,5.,6.]}])) 2 >>> guesser("Even weird non-batches have a fallback") 1 cCs d|_dSr*)methodr,r r rr(s zBatchsizeGuesser.__init__cCs"z||WS||YSr*)rDfind_suitable_methodrbatchr r r__call__s zBatchsizeGuesser.__call__cCsz ||}|j|_|WSYz ||}|j|_|WSYz ||}|j|_|WSYz ||}|j|_|WSY||}|||_|S)z/Try the different methods and note which worked) attr_basedrDtorch_tensor_bs len_of_firstlen_of_iter_firstfallback)rrGbsr r rrEs6      z%BatchsizeGuesser.find_suitable_methodcCs|jS)zImplementation of attr_based.)r=rFr r rrIszBatchsizeGuesser.attr_basedcCs |jdS)z"Implementation of torch_tensor_bs.r)shaperFr r rrJ s z BatchsizeGuesser.torch_tensor_bscCs t|dS)zImplementation of len_of_first.r)rrFr r rrK s zBatchsizeGuesser.len_of_firstcCsttt|S)z$Implementation of len_of_iter_first.)rnextr2rFr r rrLsz"BatchsizeGuesser.len_of_iter_firstcCsdS)zImplementation of fallback.r rFr r rrMszBatchsizeGuesser.fallbackN) r>r?r@rAr(rHrErIrJrKrLrMr r r rrCs rC)rA collectionsrtorch.utils.data._utils.collater"torch.utils.data._utils.pin_memoryrr4speechbrain.utils.data_utilsrrr namedtuplerr rCr r r rs  0