o Gi@sddlmZddlmZddlZddlmZddlmZm Z ddl m Z ddl m Z mZeeZGdd d ejZeGd d d e ZGd d d ejZGddde eZdS)) dataclass)piN) ConfigMixinregister_to_config) ModelMixin) BaseOutputloggingcs<eZdZdZdeffdd ZdejdejfddZZ S) StableAudioPositionalEmbeddingzUsed for continuous timedimcs8t|ddks J|d}tt||_dS)Nr)super__init__nn Parametertorchrandnweights)selfr half_dim __class__j/home/ubuntu/.local/lib/python3.10/site-packages/diffusers/pipelines/stable_audio/modeling_stable_audio.pyr s z'StableAudioPositionalEmbedding.__init__timesreturncCsN|d}||jddt}tj||fdd}tj||fdd}|S)N).Nr )r )rrrcatsincos)rrfreqs fourieredrrrforward&s z&StableAudioPositionalEmbedding.forward __name__ __module__ __qualname____doc__intrrTensorr" __classcell__rrrrr sr c@sHeZdZUdZdZejdBed<dZejdBed<dZ ejdBed<dS) StableAudioProjectionModelOutputa Args: Class for StableAudio projection layer's outputs. text_hidden_states (`torch.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states obtained by linearly projecting the hidden-states for the text encoder. seconds_start_hidden_states (`torch.Tensor` of shape `(batch_size, 1, hidden_size)`, *optional*): Sequence of hidden-states obtained by linearly projecting the audio start hidden states. seconds_end_hidden_states (`torch.Tensor` of shape `(batch_size, 1, hidden_size)`, *optional*): Sequence of hidden-states obtained by linearly projecting the audio end hidden states. Ntext_hidden_statesseconds_start_hidden_statesseconds_end_hidden_states) r$r%r&r'r,rr)__annotations__r-r.rrrrr+.s  r+cs>eZdZdZ d dedBffdd Zdejfdd ZZ S) StableAudioNumberConditionera A simple linear projection model to map numbers to a latent space. Args: number_embedding_dim (`int`): Dimensionality of the number embeddings. min_value (`int`): The minimum value of the seconds number conditioning modules. max_value (`int`): The maximum value of the seconds number conditioning modules internal_dim (`int`): Dimensionality of the intermediate number hidden states.  internal_dimNcs@ttt|tj|d|d|_||_||_||_ dS)N) in_features out_features) r rr Sequentialr Lineartime_positional_embeddingnumber_embedding_dim min_value max_value)rr9r:r;r2rrrrOs  z%StableAudioNumberConditioner.__init__floatscCs^||j|j}||j|j|j}t|jj}||}||}|dd|j }|S)Nrr3) clampr:r;nextr8 parametersdtypetoviewr9)rr<normalized_floatsembedder_dtype embedding float_embedsrrrr"`s  z$StableAudioNumberConditioner.forward)r1r#rrrrr0@sr0csTeZdZdZefddZ   d dejdBdejdBdejdBfdd ZZ S) StableAudioProjectionModela A simple linear projection model to map the conditioning values to a shared latent space. Args: text_encoder_dim (`int`): Dimensionality of the text embeddings from the text encoder (T5). conditioning_dim (`int`): Dimensionality of the output conditioning tensors. min_value (`int`): The minimum value of the seconds number conditioning modules. max_value (`int`): The maximum value of the seconds number conditioning modules csHt||kr tnt|||_t||||_t||||_dS)N) r rrIdentityr7text_projectionr0start_number_conditionerend_number_conditioner)rtext_encoder_dimconditioning_dimr:r;rrrrs z#StableAudioProjectionModel.__init__Nr, start_seconds end_secondscCsP|dur|n||}|dur|n||}|dur|n||}t|||dS)N)r,r-r.)rIrJrKr+)rr,rNrOr-r.rrrr"sz"StableAudioProjectionModel.forward)NNN) r$r%r&r'rrrr)r"r*rrrrrGrs rG) dataclassesrmathrrtorch.nnrconfiguration_utilsrrmodels.modeling_utilsrutilsrr get_loggerr$loggerModuler r+r0rGrrrrs     2