Skip to main content
torch.js has not been released yet.
torch.js logotorch.js logotorch.js
PlaygroundContact
Login
Documentation
IntroductionType SafetyTensor ExpressionsTensor IndexingEinsumEinopsAutogradTraining a ModelProfiling & MemoryPyTorch MigrationBest PracticesRuntimesPerformancePyTorch CompatibilityBenchmarksDType Coverage
ActivationOptionsAdaptiveAvgPool1dAdaptiveAvgPool2dAdaptiveAvgPool3dAdaptiveLogSoftmaxOptionsAdaptiveLogSoftmaxWithLossAdaptiveMaxPool1dAdaptiveMaxPool1dOptionsAdaptiveMaxPool2dAdaptiveMaxPool2dOptionsAdaptiveMaxPool3dAdaptiveMaxPool3dOptionsadd_moduleAlphaDropoutappendappendapplyAvgPool1dAvgPool1dOptionsAvgPool2dAvgPool2dOptionsAvgPool3dAvgPool3dOptionsBackwardHookBackwardPreHookBatchNorm1dBatchNorm2dBatchNorm3dBatchNormOptionsBCELossBCEWithLogitsLossBilinearBilinearOptionsBufferBufferOptionsBufferRegistrationHookbufferscallCELUCELUOptionsChannelShufflechildrenCircularPad1dCircularPad2dCircularPad3dclearConstantPad1dConstantPad2dConstantPad3dConv1dConv2dConv3dConvOptionsConvTranspose1dConvTranspose2dConvTranspose3dConvTransposeOptionsCosineEmbeddingLossCosineEmbeddingLossOptionsCosineSimilarityCosineSimilarityOptionscreatecreateCrossEntropyLossCTCLossdecodedecodedeleteDropoutDropout1dDropout2dDropout3dDropoutOptionsELUELUOptionsEmbeddingEmbeddingBagEmbeddingBagForwardOptionsEmbeddingBagFromPretrainedOptionsEmbeddingBagOptionsEmbeddingFromPretrainedOptionsEmbeddingOptionsencodeencodeentriesentriesevalextendFeatureAlphaDropoutFlattenFlattenOptionsFoldFoldOptionsforwardforwardforwardforwardforwardforwardforwardforwardforwardforwardforwardforwardforward_with_targetForwardHookForwardPreHookFractionalMaxPool2dFractionalMaxPool3dFractionalMaxPoolOptionsfrom_pretrainedfrom_pretrainedGaussianNLLLossGELUGELUOptionsgenerate_square_subsequent_maskgetgetgetgetgetget_bufferget_parameterget_submoduleGLUGLUOptionsGroupNormGroupNormOptionsGRUGRUCellHardshrinkHardshrinkOptionsHardsigmoidHardswishHardtanhHardtanhOptionshashasHingeEmbeddingLossHingeEmbeddingLossOptionsHuberLossHuberLossOptionsIdentityInstanceNorm1dInstanceNorm2dInstanceNorm3dInstanceNormOptionsis_uninitialized_bufferis_uninitialized_parameteriterator]iterator]iterator]iterator]keyskeysKLDivLossL1LossL1LossOptionsLayerNormLayerNormOptionsLazyBatchNorm1dLazyBatchNorm2dLazyBatchNorm3dLazyConv1dLazyConv2dLazyConv3dLazyConvOptionsLazyConvTranspose1dLazyConvTranspose2dLazyConvTranspose3dLazyConvTransposeOptionsLazyInstanceNorm1dLazyInstanceNorm2dLazyInstanceNorm3dLazyLinearLeakyReLULeakyReLUOptionsLinearLinearOptionsload_state_dictload_state_dictLocalResponseNormLocalResponseNormOptionslog_probLogSigmoidLogSoftmaxLogSoftmaxOptionsLPPool1dLPPool1dOptionsLPPool2dLPPool2dOptionsLPPool3dLPPool3dOptionsLSTMLSTMCellLSTMCellOptionsMarginRankingLossMarginRankingLossOptionsmaterializematerializematerialize_uninitializedmaterialize_uninitializedMaxPool1dMaxPool1dOptionsMaxPool2dMaxPool2dOptionsMaxPool3dMaxPool3dOptionsMaxUnpool1dMaxUnpool1dOptionsMaxUnpool2dMaxUnpool2dOptionsMaxUnpool3dMaxUnpool3dOptionsMishModuleModuleBuffersModuleChildrenModuleDictModuleDictOptionsModuleListModuleListOptionsModuleParametersModuleRegistrationHookmodulesMSELossMSELossOptionsmultihead_attnMultiheadAttentionMultiheadAttentionOptionsMultiheadAttnOptionsMultiLabelMarginLossMultiLabelMarginLossOptionsMultiLabelSoftMarginLossMultiMarginLossnamed_buffersnamed_childrennamed_modulesnamed_parametersNamedModulesOptionsNamedRecurseOptionsNLLLossnum_parametersNumParametersOptionsPairwiseDistancePairwiseDistanceOptionsParameterParameterDictParameterDictOptionsParameterListParameterListOptionsParameterOptionsParameterRegistrationHookparametersPixelShufflePixelUnshufflePoissonNLLLosspoppopPReLUPReLUOptionsRecurseOptionsReflectionPad1dReflectionPad2dReflectionPad3dregister_backward_hookregister_bufferregister_forward_hookregister_forward_pre_hookregister_full_backward_hookregister_full_backward_pre_hookregister_module_backward_hookregister_module_buffer_registration_hookregister_module_forward_hookregister_module_forward_pre_hookregister_module_full_backward_hookregister_module_full_backward_pre_hookregister_module_module_registration_hookregister_module_parameter_registration_hookregister_parameterReLUReLU6RemovableHandleremoveReplicationPad1dReplicationPad2dReplicationPad3dRMSNormRMSNormOptionsRNNRNNBaseRNNBaseOptionsRNNCellRNNCellOptionsRReLURReLUOptionsrunrunSELUSequentialsetsetsetSigmoidSiLUSmoothL1LossSmoothL1LossOptionsSoftMarginLossSoftMarginLossOptionsSoftmaxSoftmax2dSoftmaxOptionsSoftminSoftminOptionsSoftplusSoftplusOptionsSoftshrinkSoftshrinkOptionsSoftsignstate_dictstate_dictStateDictOptionsstepSyncBatchNormTanhTanhshrinkThresholdThresholdOptionstotototrainTrainOptionsTransformerTransformerDecoderTransformerDecoderDecodeOptionsTransformerDecoderLayerTransformerDecoderLayerDecodeOptionsTransformerDecoderLayerOptionsTransformerDecoderOptionsTransformerEncoderTransformerEncoderEncodeOptionsTransformerEncoderLayerTransformerEncoderLayerEncodeOptionsTransformerEncoderLayerOptionsTransformerEncoderOptionsTransformerOptionsTransformerRunOptionsTripletMarginLossTripletMarginWithDistanceLossUnflattenUnfoldUnfoldOptionsUninitializedBufferUninitializedOptionsUninitializedParameterupdateUpsampleUpsamplingBilinear2dUpsamplingNearest2dvaluesvalueszero_gradZeroPad1dZeroPad2dZeroPad3d
absacosacoshAdaptivePool1dShapeAdaptivePool2dShapeaddaddbmmAddbmmOptionsaddcdivAddcdivOptionsaddcmulAddcmulOptionsaddmmAddmmOptionsaddmvAddmvOptionsaddrAddrOptionsadjointallallcloseAllcloseOptionsAlphaBetaOptionsamaxaminaminmaxAminmaxOptionsangleanyapplyOutarangeare_deterministic_algorithms_enabledargmaxargminargsortargwhereas_stridedas_tensorasinasinhAssertNoShapeErrorAssertNotErrorAsStridedOptionsAtat_error_index_out_of_boundsatanatan2atanhatleast_1datleast_2datleast_3dAtShapeautocast_decrement_nestingautocast_increment_nestingautograd_gradient_mismatch_errorautograd_not_registered_errorAutogradConfigAutogradDeviceAutogradDTypeAutogradEntryAutogradHandleAutogradHandleImplAxesRecordBackwardFnbaddbmmBaddbmmOptionsbartlett_windowBaseKernelConfigbatch_dimensions_do_not_match_errorbernoulliBernoulliOptionsBinaryBackwardFnBinaryBroadcastResultBinaryDTypeBinaryKernelConfigCPUBinaryKernelCPUBinaryOpConfigBinaryOpNamesBinaryOpSchemaBinaryOptionsbincountBincountOptionsbitwise_andbitwise_left_shiftbitwise_notbitwise_orbitwise_right_shiftbitwise_xorblackman_windowblock_diagbmmBooleanDTypeRulebroadcast_error_incompatible_dimensionsbroadcast_shapesbroadcast_tensorsbroadcast_toBroadcastShapeBroadcastShapeRulebroadcastShapesbucketizeBucketizeOptionsBufferUsagebuildEinopsErrorbuildErrorMessagecanBroadcastTocartesian_prodcatCatOptionsCatShapeCauchyOptionscdistCdistOptionsceilceluCeluFunctionalOptionschain_matmulCheckShapeErrorCholeskyShapechunkchunk_error_dim_out_of_rangeChunkOptionsclampClampOptionsclear_autocast_cacheclearEinopsCacheclearEinsumCacheclonecolumn_stackcombinationsCombinationsOptionscompiled_with_cxx11_abicomplexconjconj_physicalcontiguousConv1dShapeConv2dShapeConv3dShapeConvTranspose2dShapecopysigncorrcoefcoscoshcount_nonzeroCountNonzeroOptionscovcoverage_reportcoverageReportCoverageReportCovOptionsCPUForwardFnCPUKernelConfigCPUKernelEntryCPUOnlyResultCPUTensorDatacreateCumExtremeResultcreateTorchCreationOpSchemaCumExtremeResultcummaxcummincumprodCumShapecumsumcumulative_trapezoidCumulativeOptionsCumulativeOptionsWithDimdeg2raddetachDeterministicOptionsDetShapeDevicedevice_error_requiresDeviceBufferDeviceCapabilitiesDeviceCheckedResultDeviceConfigDeviceContextDeviceEntryDeviceHandleDeviceInputDeviceOptionsDeviceRegistryDeviceTypediagdiag_embedDiagEmbedOptionsdiagflatDiagflatOptionsDiagFlatOptionsdiagonal_scatterDiagonalOptionsDiagonalScatterOptionsDiagOptionsDiagShapediffDiffOptionsdigammadimension_error_out_of_rangeDispatchConfigdistDistOptionsdivdotDotShapeRuleDoubleDoubleDimdropoutDropoutFunctionalOptionsdsplitdstackDTypedtype_already_registered_errordtype_components_mismatch_errordtype_not_found_errorDTypeComponentsDTypeConfigDTypeCoverageReportDTypeDisplayConfigDTypeEntryDTypeHandleDTypeHandleImplDTypeInfoDTypeRegistryDTypeRuleDTypeSerializationConfigDynamicShapeEigShapeeinops_error_ambiguous_decompositioneinops_error_anonymous_in_outputeinops_error_dimension_mismatcheinops_error_invalid_patterneinops_error_reduce_undefined_outputeinops_error_repeat_missing_sizeeinops_error_undefined_axiseinsumeinsum_error_dimension_mismatcheinsum_error_index_out_of_rangeeinsum_error_invalid_equationeinsum_error_invalid_sublist_elementeinsum_error_operand_count_mismatcheinsum_error_subscript_rank_mismatcheinsum_error_unknown_output_indexEinsumOptionsEinsumOutputShapeEllipsiseluelu_EluFunctionalOptionsembedding_bag_error_requires_2d_inputemptyempty_cacheempty_likeeqequalerferfcerfinvexpexp2expandexpand_asexpand_error_incompatibleExpandShapeexpm1ExponentialOptionseyeEyeOptionsfftFFTOptionsfindKernelWithPredicatefindSimilarPatternsflattenFlattenOptionsFlattenShapeflipflip_error_dim_out_of_rangefliplrFlipShapeflipudfloat_powerFloatDTypeRulefloorfloor_dividefmaxfminfmodformatEquationErrorformatShapefracfrexpfrombufferfullfull_likefunction_already_registered_errorFunctionConfigFunctionEntryFunctionHandlegathergather_error_dim_out_of_rangeGatherShapegcdgegeluGeometricOptionsget_autocast_cpu_dtypeget_autocast_gpu_dtypeget_autocast_ipu_dtypeget_autocast_xla_dtypeget_default_deviceget_default_dtypeget_deterministic_debug_modeget_device_configget_device_contextget_device_moduleget_dtype_infoget_file_pathget_float32_matmul_precisionget_num_interop_threadsget_num_threadsget_op_infoget_printoptionsget_real_dtypeget_rng_stategetAutogradgetDTypegetEinopsCacheSizegetEinsumCacheSizegetFunctiongetKernelgetMethodgetOpInfoGetOpKindGetOpSchemagetScalarKernelgluGluFunctionalOptionsGradContextGradFnGradientsForgtHalfHalfDimhamming_windowhann_windowhardshrinkhardsigmoidhardswishhardtanhhardtanh_HardtanhFunctionalOptionshas_autogradhas_devicehas_dtypehas_kernelhasAutogradhasDTypehasFunctionhasKernelhasMethodhasScalarKernelHasShapeErrorheavisidehistcHistcOptionshistogramHistogramOptionsHistogramResulthsplithstackhypoti0IdentityShapeifftimagindex_addindex_copyindex_fillindex_putindex_reduceindex_selectindex_select_error_dim_out_of_rangeIndexPutOptionsIndexSelectShapeIndexSpecIndicesOptionsIndicesSpecinitialize_deviceInputsForInsertDiminvalid_config_errorinverseInverseShapeirfftis_anomaly_check_nan_enabledis_anomaly_enabledis_autocast_cache_enabledis_autocast_cpu_enabledis_autocast_ipu_enabledis_autocast_xla_enabledis_complexis_complex_dtypeis_cpu_only_modeis_deterministic_algorithms_warn_only_enabledis_floating_pointis_floating_point_dtypeis_inference_mode_enabledis_nonzerois_tensoris_warn_always_enabledis_webgpu_availableIs2DIsAtLeast1DIsBinaryOpIsBinaryOpNameiscloseIscloseOptionsisfiniteisinisinfisnanisneginfisposinfisrealIsReductionOpIsReductionOpNameIsRegistryErrorIsShapeErroristftISTFTOptionsIsUnaryOpIsUnaryOpNameitem_error_not_scalarItemResultkaiser_windowKaiserWindowOptionskernel_not_registered_errorkernel_signature_mismatch_errorKernelConfigKernelConfigWebGPUKernelEntryKernelHandleKernelInfoKernelPredicateKernelRegistryKernelWebGPUkronkthvalueKthvalueOptionslcmldexpleleaky_reluleaky_relu_LeakyReluFunctionalOptionslerplevenshteinDistancelgammalinalg_error_not_square_matrixlinalg_error_requires_2dlinalg_error_requires_at_least_2dlinearlinspacelist_custom_deviceslist_custom_dtypeslist_deviceslist_dtypeslist_functionslist_kernelslist_methodslist_opslistCustomDTypeslistDTypeslistFunctionslistKernelsListKernelsOptionslistMethodslistOpsListOpsOptionsloglog_softmaxlog10log1plog2logaddexplogaddexp2logcumsumexplogical_andlogical_notlogical_orlogical_xorLogitOptionsLogNormalOptionsLogOptionslogsigmoidlogspacelogsumexpLogsumexpOptionsltLUShapeLuSolveOptionsmasked_fillmasked_selectmasked_select_asyncMaskSpecmatmulmatmul_error_inner_dimensions_do_not_matchMatmul2DShapeMatmulShapeMatmulShapeRuleMatrixTransposeShapemaxmaximummeanmedianmemory_statsmemory_summarymeshgridmethod_already_registered_errormethod_dtype_not_supported_errorMethodConfigMethodEntryMethodHandleminminimummishmmMMShapeRulemodemovedimmsortmulmultinomialmultinomial_asyncMultinomialAsyncOptionsMultinomialOptionsMultiplyBymvMVShapeRulenan_to_numnanmeannanmediannanquantileNanReductionOptionsnansumNanToNumOptionsnarrownarrow_copynarrow_error_length_exceeds_boundsnarrow_error_start_out_of_boundsNarrowShapeneneedsBroadcastnegNegativeDimnextafternonzeroNonzeroOptionsnormnormalNormalOptionsNormOptionsnumelonesones_likeop_kind_mismatch_errorop_not_found_errorOpCoverageEntryOpInfoOpKindOpNameOpSchemaOpSchemasouterOuterShapepackPackShapepermutepermute_error_dimension_count_mismatchPermuteShapepoissonpolarPool1dShapePool2dShapePool3dShapepositivepowpreluPrintOptionsprodprofiler_allow_cudagraph_cupti_lazy_reinit_cuda12promote_typesPromoteDTypeRulePutOptionsquantileQuantileOptionsrad2degrandrand_likerandintrandint_likeRandintLikeOptionsRandintOptionsrandnrandn_likeRandomLikeOptionsRandomOptionsrandpermRangeSpecRankravelrealrearrangeRearrangeOptionsRearrangeShapereciprocalreduceReduceOperationReduceOptionsReduceShapeReductionKernelConfigCPUReductionKernelCPUReductionOpNamesReductionOpSchemaReductionOptionsReductionShapeRuleregister_backwardregister_deviceregister_dtyperegister_forwardregister_functionregister_methodregister_scalar_forwardregisterAutogradRegisterBackwardOptionsregisterBinaryOpregisterDTypeRegisterDTypeOptionsRegisteredDTyperegisterFunctionRegisterFunctionOptionsregisterKernelRegisterKernelOptionsregisterMethodRegisterMethodOptionsregisterScalarKernelregisterUnaryOpregistration_failed_errorrelurelu_relu6ReluFunctionalOptionsremainderRemoveDimrepeatrepeat_interleaveRepeatInterleaveOptionsRepeatOptionsRepeatShapeReplaceDimrequireWebGPUreset_peak_memory_statsreshapeReshapeShaperesult_typerfftrollRollOptionsrot90Rot90Optionsroundrrelurrelu_RreluFunctionalOptionsrsqrtSafeExpandShapeSameDTypeRuleSameShapeRuleSaveForBackwardScalarCPUForwardFnScalarCPUKernelConfigScalarKernelEntryScalarKernelHandleScalarWebGPUKernelConfigScaleDimscatterscatter_addscatter_add_scatter_error_dim_out_of_rangescatter_reducescatter_reduce_ScatterReduceOptionsScatterShapesearchsortedSearchSortedOptionsselectselect_error_index_out_of_boundsselect_scatterSelectShapeseluset_default_deviceset_default_tensor_typeset_deterministic_debug_modeset_float32_matmul_precisionset_printoptionsset_warn_alwaysSetupContextFnShapeShapeCheckedResultShapedTensorShapeErrorMessageShapeOpSchemaShapeRulesigmoidsignsignbitsilusinsincsinhSizeOptionsslice_error_out_of_boundsslice_scatterSliceOptionsSliceScatterOptionsSliceShapeSliceSpecsoftmaxsoftmax_error_dim_out_of_rangeSoftmaxShapesoftminSoftminFunctionalOptionssoftplusSoftplusFunctionalOptionssoftshrinksoftsignsortSortOptionssplitsplit_error_dim_out_of_rangeSplitOptionssqrtsquaresqueezeSqueezeOptionsSqueezeShapestackStackOptionsStackShapestdstd_meanStdVarMeanOptionsStdVarOptionsstftSTFTOptionsStrideOptionssubSublistSublistElementSubscriptIndexsumSVDShapeswapaxessym_floatsym_intsym_notttaketake_along_dimTakeAlongDimOptionstantanhtanhshrinktensortensor_splitTensorCreatorTensorDatatensordotTensordotOptionsTensorLikeTensorMetaTensorOptionsTensorStoragethresholdthreshold_tileTileShapeToOptionstopkTopkOptionsTorchtraceTraceShapetransposetranspose_dims_error_out_of_rangetranspose_error_requires_2d_tensorTransposeDimsShapeTransposeDimsShapeCheckedTransposeShapetrapezoidTrapezoidOptionsTriangularOptionstriltril_indicesTriOptionsTripletriutriu_indicestrue_dividetruncTupleOfLengthTypedArrayTypedArrayForTypedStorageTypeOptionsUnaryBackwardFnUnaryDTypeUnaryKernelConfigCPUUnaryKernelCPUUnaryOpConfigUnaryOpFnUnaryOpNamesUnaryOpParamsUnaryOpSchemaUnaryOptionsunbindunbind_error_dim_out_of_rangeUnbindOptionsunflattenUniformOptionsuniqueunique_consecutiveUniqueConsecutiveOptionsUniqueOptionsunpackUnpackShapeunravel_indexunregister_deviceunsqueezeUnsqueezeOptionsUnsqueezeShapeuse_deterministic_algorithmsValidateBatchedSquareMatrixValidateChunkDimValidatedEinsumShapevalidateDeviceValidateDeviceValidatedRearrangeShapeValidatedReduceShapeValidatedRepeatShapevalidateDTypeValidateEinsumValidateOperandCountValidateRanksValidateScalarValidateSplitDimValidateSquareMatrixValidateUnbindDimValueOptionsvar_var_meanvdotviewview_as_complexview_as_realvmapvsplitvstackWebGPUKernelConfigWebGPUOnlyResultWebGPUTensorDatawhereWindowOptionsxlogyzeroszeros_like
torch.js· 2026
LegalTerms of UsePrivacy Policy
/
/
  1. docs
  2. torch.js
  3. torch
  4. nn
  5. Embedding

torch.nn.Embedding

class Embedding extends Module
new Embedding(num_embeddings: number, embedding_dim: number, options?: EmbeddingOptions)
readonlynum_embeddings(number)
readonlyembedding_dim(number)
readonlypadding_idx(number | null)
readonlymax_norm(number | null)
readonlynorm_type(number)
readonlyscale_grad_by_freq(boolean)
readonlysparse(boolean)
weight(Parameter)

Learnable embedding lookup table: converts token indices to dense vectors.

A fundamental layer that transforms discrete token IDs into continuous vector representations. Essential for:

  • Natural language processing (word embeddings for any vocabulary)
  • Sequence modeling (tokens → dense representations → transformers)
  • Recommendation systems (item/user IDs → embedding vectors)
  • Graph neural networks (node IDs → embeddings)
  • Categorical feature embeddings in deep learning
  • Position embeddings in attention-based models (combined with positional encoding)

Creates a weight matrix of shape [num_embeddings, embedding_dim] where each row is the learnable embedding vector for a vocabulary item. During forward pass, looks up embeddings for given token indices, enabling end-to-end training with gradient flow back to embedding weights.

output[i]=W[input[i],:](simple lookup)With max_norm: if ∥embedding∥2>max_norm, scale by max_norm∥embedding∥2\begin{aligned} \text{output}[i] = W[\text{input}[i], :] \quad \text{(simple lookup)} \\ \text{With max\_norm: if } \|\text{embedding}\|_2 > \text{max\_norm}, \text{ scale by } \frac{\text{max\_norm}}{\|\text{embedding}\|_2} \end{aligned}output[i]=W[input[i],:](simple lookup)With max_norm: if ∥embedding∥2​>max_norm, scale by ∥embedding∥2​max_norm​​
  • Vocabulary size matters: Large vocabularies need large embedding matrices (memory-intensive)
  • Embedding dimension selection: Typical range: 64-768 for general tasks, 768-2048 for BERT-scale
  • Initialization: Random normal initialization (N(0,1)) by default
  • Padding token: Set padding_idx to have zero embedding vector (important for masking)
  • Frozen embeddings: Use from_pretrained with freeze=true for transfer learning
  • Memory usage: Parameters = num_embeddings × embedding_dim × 4 bytes (float32)
  • Sparse vs dense gradients: Sparse updates only for accessed embeddings (not implemented)
  • GPU acceleration: Embedding lookup is efficient on GPU (parallel table lookup)
  • Typical architecture pattern: Token IDs → Embedding → Transformer/LSTM/CNN → Output
  • Combined with position encoding: Position embeddings added/concatenated to token embeddings

Examples

// Natural language processing: word embedding lookup
const vocab_size = 10000;  // Dictionary with 10k words
const embed_dim = 300;     // GloVe-like embedding dimension
const embed = new torch.nn.Embedding(vocab_size, embed_dim);

// Token IDs (e.g., from tokenizer)
const token_ids = torch.tensor([2, 145, 8, 9999], { dtype: 'int32' });
// [batch, seq_len] shape
const token_ids_batch = torch.tensor(
  [[101, 2054, 2003, 1045, 102],  // "what is i"
   [101, 2234, 3431, 102, 0]],     // "good day" with padding
  { dtype: 'int32' }
);

const embeddings = embed.forward(token_ids_batch);  // [2, 5, 300]
// Transformer-style architecture with word + position embeddings
class PositionalEmbedding extends torch.nn.Module {
  word_embed: torch.nn.Embedding;
  pos_embed: torch.nn.Embedding;
  max_seq_len: number = 512;

  constructor(vocab_size: number, embed_dim: number) {
    super();
    this.word_embed = new torch.nn.Embedding(vocab_size, embed_dim);
    this.pos_embed = new torch.nn.Embedding(this.max_seq_len, embed_dim);
  }

  forward(token_ids: torch.Tensor): torch.Tensor {
    const seq_len = token_ids.shape[-1];
    // Create position IDs
    const pos_ids = torch.arange(0, seq_len, { dtype: 'int32' });
    if (token_ids.shape.length > 1) {
      // Expand for batch dimension
      const pos_ids_batch = pos_ids.unsqueeze(0).expand_as(token_ids);
    }
    // Combine word and position embeddings
    const word_emb = this.word_embed.forward(token_ids);       // [B, L, D]
    const pos_emb = this.pos_embed.forward(pos_ids_batch);     // [B, L, D]
    return word_emb.add(pos_emb);  // [B, L, D] - summed embeddings
  }
}
// Recommendation system: item embeddings
const num_items = 100000;      // 100k item catalog
const embed_dim = 128;         // Embedding dimension
const item_embed = new torch.nn.Embedding(num_items, embed_dim);

// User interactions as item IDs
const user_history = torch.tensor(
  [[123, 456, 789, 0],        // User 1's item history
   [234, 567, 0, 0]],         // User 2's item history (padded)
  { dtype: 'int32' }
);

// Get embeddings for items
const history_embeddings = item_embed.forward(user_history);  // [2, 4, 128]
// Now can aggregate embeddings (mean, attention, etc.) for user representation
const user_reps = history_embeddings.mean(1);  // [2, 128]
// Using pretrained embeddings (e.g., GloVe, FastText)
const pretrained_weights = torch.randn([10000, 300]);  // Pre-computed embeddings
const embed = torch.nn.Embedding.from_pretrained(pretrained_weights, { freeze: true });

// Frozen embeddings (no gradient updates)
embed.weight.requires_grad = false;
const token_ids = torch.tensor([1, 2, 3, 4], { dtype: 'int32' });
const embeddings = embed.forward(token_ids);  // Uses fixed pretrained values

// Or allow fine-tuning
const embed_finetune = torch.nn.Embedding.from_pretrained(pretrained_weights, { freeze: false });
const embeddings2 = embed_finetune.forward(token_ids);  // Can be updated via backprop
// Padding token handling
const PAD_ID = 0;
const embed = new torch.nn.Embedding(
  5000,
  128,
  { padding_idx: PAD_ID }  // Padding token embedding always zero
);

// Variable length sequences with padding
const sequences = torch.tensor(
  [[1, 2, 3, 0, 0],     // Length 3, then padded
   [4, 5, 0, 0, 0],     // Length 2, then padded
   [6, 7, 8, 9, 10]],   // Length 5, no padding
  { dtype: 'int32' }
);

const embeddings = embed.forward(sequences);  // [3, 5, 128]
// Padding token embeddings are guaranteed to be zero vectors
// Gradient flow through embeddings
const embed = new torch.nn.Embedding(100, 50);
const token_ids = torch.tensor([1, 2, 3], { dtype: 'int32' });
const embed_out = embed.forward(token_ids);  // [3, 50]

// Gradients flow back to embedding weights
const loss = embed_out.sum();  // Dummy loss
// loss.backward();  // Gradients computed for embed.weight

// Only accessed embeddings get gradient updates (sparse gradients concept)

See Also

  • PyTorch torch.nn.Embedding
Previous
ELUOptions
Next
Embedding.forward