>:j#SSKJr SSKJrJr SSKrSSKrSSKJ r SSKJ r J r SSjr SSjr SSjrSS jr\SS j5r\SS j5rSS jrSS jrSSjrSSjrSSjrg)) annotations)AnyoverloadN) coo_matrix)Tensordevicec [U[5(a}[SU55(aO[R"UVs/sH-oR 5R [RS9PM/ sn5$[R"U5nO+[U[5(d[R"U5nUR(aUR [RS9$U$s snf)z Converts the input `a` to a PyTorch tensor if it is not already a tensor. Handles lists of sparse tensors by stacking them. Args: a (Union[list, np.ndarray, Tensor]): The input array or tensor. Returns: Tensor: The converted tensor. c3h# UH(n[U[5=(a URv M* g7fN) isinstancer is_sparse).0xs ^/home/mande/repo/quber/.venv/lib/python3.13/site-packages/sentence_transformers/util/tensor.py %_convert_to_tensor..s#@az!V$44as02dtype) r listalltorchstackcoalescetofloat32tensorrr )ars r_convert_to_tensorr s!T @a@ @ @;;aPa emm DaPQ Q QA 6 " " LLO{{tt%--t(( H Qs4C.cPUR5S:XaURS5nU$)z If the tensor `a` is 1-dimensional, it is unsqueezed to add a batch dimension. Args: a (Tensor): The input tensor. Returns: Tensor: The tensor with a batch dimension. r)dim unsqueezers r_convert_to_batchr$$s# uuw!| KKN Hcf[U5nUR5S:XaURS5nU$)a Converts the input data to a tensor with a batch dimension. Handles lists of sparse tensors by stacking them. Args: a (Union[list, np.ndarray, Tensor]): The input data to be converted. Returns: Tensor: The converted tensor with a batch dimension. r r)rr!r"r#s r_convert_to_batch_tensorr'3s- 1Auuw!| KKN Hr%cDUR(d)[RRR USSS9$UR 5nUR 5UR5p![R"URS5URS9nURSUSUS-5 [R"U5RSUS5nUS:nUR5nXT==X4-ss'[R"XUR55$)z Normalizes the embeddings matrix, so that each sentence embedding has unit length. Args: embeddings (Tensor): The input embeddings matrix. Returns: Tensor: The normalized embeddings matrix. r )pr!rr)r rnn functional normalizerindicesvalueszerossizer index_add_sqrt index_selectclonesparse_coo_tensor) embeddingsr/r0 row_normsmasknormalized_valuess rnormalize_embeddingsr<Ds   xx"",,Z1!,DD$$&J ((*J,=,=,?V JOOA.z7H7HII GAJ 2 9%221gajAI q=D y.  " "7z?P QQr%cgr r8 truncate_dims rtruncate_embeddingsrAasY\r%cgr r>r?s rrArAes]`r%cUSSU24$)a\ Truncates the embeddings matrix. Args: embeddings (Union[np.ndarray, torch.Tensor]): Embeddings to truncate. truncate_dim (Optional[int]): The dimension to truncate sentence embeddings to. `None` does no truncation. Example: >>> from sentence_transformers import SentenceTransformer >>> from sentence_transformers.util import truncate_embeddings >>> model = SentenceTransformer("tomaarsen/mpnet-base-nli-matryoshka") >>> embeddings = model.encode(["It's so nice outside!", "Today is a beautiful day.", "He drove to work earlier"]) >>> embeddings.shape (3, 768) >>> model.similarity(embeddings, embeddings) tensor([[1.0000, 0.8100, 0.1426], [0.8100, 1.0000, 0.2121], [0.1426, 0.2121, 1.0000]]) >>> truncated_embeddings = truncate_embeddings(embeddings, 128) >>> truncated_embeddings.shape >>> model.similarity(truncated_embeddings, truncated_embeddings) tensor([[1.0000, 0.8092, 0.1987], [0.8092, 1.0000, 0.2716], [0.1987, 0.2716, 1.0000]]) Returns: Union[np.ndarray, torch.Tensor]: Truncated embeddings. .Nr>r?s rrArAis: c=L=( ))r%c$UcU$[U[R5(a[R"U5nUR up#UR n[R"[R"U5[X5SS9upV[R"U[RS9n[R"X$S9RS5RS[X55nSXxR5UR54'SX)'U$)a{ Keeps only the top-k values (in absolute terms) for each embedding and creates a sparse tensor. Args: embeddings (Union[np.ndarray, torch.Tensor]): Embeddings to sparsify by keeping only top_k values. max_active_dims (int): Number of values to keep as non-zeros per embedding. Returns: torch.Tensor: A sparse tensor containing only the top-k values per embedding. r )kr!rr+Tr)r npndarrayrrshapertopkabsmin zeros_likeboolaranger"expandflatten) r8max_active_dims batch_sizer!r_ top_indicesr: batch_indicess rselect_max_active_dimsrWs*bjj))\\*-  &&OJ   FZZ * 5_9RXYZNA   Jejj 9DLL;EEaHOOPRTWXgTmnM;?D   +"5"5"7 78Ju r%ctUH1n[X[5(dMXRU5X'M3 U$)aY Send a PyTorch batch (i.e., a dictionary of string keys to Tensors) to a device (e.g. "cpu", "cuda", "mps"). Args: batch (Dict[str, Tensor]): The batch to send to the device. target_device (torch.device): The target device (e.g. "cpu", "cuda", "mps"). Returns: Dict[str, Tensor]: The batch with tensors sent to the target device. )r rr)batch target_devicekeys rbatch_to_devicer\s6 ej& ) )}5EJ Lr%c UR5nUR5R5R5nUR 5R5R5n[ X!SUS44UR S9$)Nrr )rI)rr/cpunumpyr0rrI)rr/r0s r to_scipy_coor`sd Aiikoo%%'G XXZ^^  # # %F v GAJ78 HHr%clUR(dUR5nUR5n[R"UR S5UR [RS9nUR5S:Xa"SXR5R5'U$UR5S:Xa`UR5R5S:a<UR5n[R"USSS9up4UR5X'U$[SUR5S 35e) z Compute count vector from sparse embeddings indicating how many samples have non-zero values in each dimension. Args: embeddings: Sparse tensor of shape (batch_size, vocab_size) or (vocab_size,) Returns: Count vector of shape (vocab_size,) rF)rrr r)rT) return_countszExpected 1D or 2D tensor, got D)r to_sparserrr1r2rint32r!r/squeezer0numeluniqueint ValueError)r8 count_vectorr/ unique_dimscountss rcompute_count_vectorrns   ))+ $$&J;;zr2:;L;LTYT_T_`L~~178 '')1134  Q      $ $ & * ((*G"',,wqz"N K(. L %9*..:J9K1MNNr%)rzlist | np.ndarray | Tensorreturnr)rrror)r8rror)r8 np.ndarrayr@ int | Nonerorp)r8 torch.Tensorr@rqrorr)r8np.ndarray | torch.Tensorr@rqrors)r8rsrRrqrorr)rYdict[str, Any]rZrrort)rrror)r8rrrorr) __future__rtypingrrr_rGr scipy.sparserrrrr$r'r<rArWr\r`rnr>r%rrxsi"  #  2   "R: \ \ ` `*@D"IOr%