o piu @sddlZddlmZddlmZ     ddejdede d e d e d e d e d ejfddZ Gdddej Z Gdddej ZdS)N@F? timesteps embedding_dim freq_shift min_timescale max_timescaleflip_sin_to_cosscalereturnc Cs|jdks Jd|ddksJd|dt|d}t||||}|ttj|tjd| } t|dt| d} || } |rZtj t | t | gdd} ntj t | t | gdd} t | t |d|g} | S) aReturns the positional encoding (same as Tensor2Tensor). Args: timesteps: a 1-D Tensor of N indices, one per batch element. These may be fractional. embedding_dim: The number of output channels. min_timescale: The smallest time unit (should probably be 0.0). max_timescale: The largest time unit. Returns: a Tensor of timing signals [N, num_channels] rzTimesteps should be a 1d-arrayrzEmbedding dimension z should be even)dtype)axis)ndimfloatmathlogjnpexparangefloat32 expand_dims concatenatecossinreshapeshape) rrrrr r r num_timescaleslog_timescale_incrementinv_timescalesemb scaled_timesignalr$^/home/ubuntu/SoloSpeech/.venv/lib/python3.10/site-packages/diffusers/models/embeddings_flax.pyget_sinusoidal_embeddingss   r&c@s<eZdZUdZdZeed<ejZ ej ed<e j ddZ dS)FlaxTimestepEmbeddinga. Time step Embedding Module. Learns embeddings for input time steps. Args: time_embed_dim (`int`, *optional*, defaults to `32`): Time step embedding dimension dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32): Parameters `dtype` time_embed_dimrcCs>tj|j|jdd|}t|}tj|j|jdd|}|S)Nlinear_1)rnamelinear_2)nnDenser)rsilu)selftembr$r$r%__call__Hs zFlaxTimestepEmbedding.__call__N) __name__ __module__ __qualname____doc__r)int__annotations__rrrr-compactr2r$r$r$r%r':s  r'c@sDeZdZUdZdZeed<dZeed<dZ e ed<e j dd Z d S) FlaxTimestepsz Wrapper Module for sinusoidal Time step Embeddings as described in https://arxiv.org/abs/2006.11239 Args: dim (`int`, *optional*, defaults to `32`): Time step embedding dimension r(dimFr rrcCst||j|j|jdS)N)rr r)r&r;r r)r0rr$r$r%r2]szFlaxTimesteps.__call__N)r3r4r5r6r;r7r8r boolrrr-r9r2r$r$r$r%r:Ps    r:)rrrFr)r flax.linenlinenr- jax.numpynumpyrndarrayr7rr<r&Moduler'r:r$r$r$r%s6   &