U Kc@sbddlZddlmZddlmZddlmmZddgZGdddejZGdddej Z dS)N)Tensor Embedding EmbeddingBagc sPeZdZdZejZdddfdd Zeddd Z e d d Z d d Z Z S)ra An embedding bag module attached with FakeQuantize modules for weight, used for quantization aware training. We adopt the same interface as `torch.nn.Embedding`, please see https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html#torch.nn.Embedding for documentation. Similar to `torch.nn.Embedding`, with FakeQuantize modules initialized to default. Attributes: weight: fake quant module for weight N@Freturnc sr| | d} tj||||||||f| | s4td| jtjksZtdt| j| |_| j| d|_ dS)Ndevicedtype'qconfig must be provided for QAT moduleSEmbedding weights requires a qscheme of torch.per_channel_affine_float_qparams Got factory_kwargs super__init__AssertionErrorweightqschemetorch per_channel_affine_float_qparamsstrqconfigweight_fake_quant) selfnum_embeddings embedding_dim padding_idxmax_norm norm_typescale_grad_by_freqsparse_weightr r rr __class__I/tmp/pip-unpacked-wheel-gikjz4vx/torch/ao/nn/qat/modules/embedding_ops.pyrs"   zEmbedding.__init__c Cs(t|||j|j|j|j|j|jSN) FZ embeddingrrrrrr r!)rinputr%r%r&forward's zEmbedding.forwardc Cst||jks(td|jd|jjt|ds:td|jsHtd|jj}|tj ksntdt ||j}||j |j |j |j|j|j|j|j|d }|S)Create a qat module from a float module Args: `mod` a float module, either produced by torch.ao.quantization utilities or directly from user  qat..from_float only works for r,Input float module must have qconfig defined,Input float module must have a valid qconfigr r)type _FLOAT_MODULEr__name__hasattrrrrrrrrrrrrr r!clsmodZweight_qschemerZqat_embedding_bagr%r%r& from_float,s*  zEmbedding.from_floatc CsLtj|j|j|j|j|j|j|j d}tj |j |_ | |j|Sr')rnnrrrrrrr r! Parameterrdetachtraintrainingr embedding_bagr%r%r&to_floatCs zEmbedding.to_float) NNrFFNNNN)r3 __module__ __qualname____doc__r9rr2rrr* classmethodr8r@ __classcell__r%r%r#r&rs$ c sReZdZdZejZdddfdd Zdedd d Z e d d Z d dZ Z S)ra An embedding bag module attached with FakeQuantize modules for weight, used for quantization aware training. We adopt the same interface as `torch.nn.EmbeddingBag`, please see https://pytorch.org/docs/stable/generated/torch.nn.EmbeddingBag.html#torch.nn.EmbeddingBag for documentation. Similar to `torch.nn.EmbeddingBag`, with FakeQuantize modules initialized to default. Attributes: weight: fake quant module for weight NrFmeanrc sv| | d}tj||||||||| | f || s8td| jtjks^tdt| j| |_| j|d|_ dS)Nrr WEmbedding Bag weights requires a qscheme of torch.per_channel_affine_float_qparams Got r r)rrrrrr moder!r"include_last_offsetrrr r rr#r%r&r\s&   zEmbeddingBag.__init__c Cs4t|||j||j|j|j|j|j||j |j Sr') r(r?rrrrr rHr!rIr)rr)offsetsZper_sample_weightsr%r%r&r*kszEmbeddingBag.forwardc Cst||jks(td|jd|jjt|ds:td|jsHtd|jj}|tj ksntdt ||j}||j |j |j |j|j|j|j|j|j|j|d }|S)r+r,r-rr.r/rGr0)r1r2rr3r4rrrrrrrrrrr rHr!rIrr5r%r%r&r8rs,  zEmbeddingBag.from_floatc CsTtj|j|j|j|j|j|j|j d|j |j }tj |j |_ ||j|Sr')rr9rrrrrr rHr!rIrr:rr;r<r=r>r%r%r&r@s zEmbeddingBag.to_float) NrFrFFNFNNNN)NN)r3rArBrCr9rr2rrr*rDr8r@rEr%r%r#r&rKs( ) rrZtorch.nnr9Ztorch.nn.functionalZ functionalr(__all__rrr%r%r%r&s   C