o }oiK @sPddlZddlZddlmZddlmZGdddejZGdddejZdS)N) rearrangecs*eZdZdZ fddZddZZS)LearnedSinusoidalPosEmbz> following @crowsonkb 's lead with learned sinusoidal pos emb cs8t|ddks J|d}tt||_dS)Nr)super__init__nn Parametertorchrandnweights)selfdimhalf_dim __class__t/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/multimodal/modules/imagen/diffusionmodules/embs.pyrs z LearnedSinusoidalPosEmb.__init__cCsTt|d}|t|jddtj}tj||fdd}tj||fdd}|S)Nzb -> b 1zd -> 1 drr )rr mathpir catsincos)r xfreqs fourieredrrrforward s zLearnedSinusoidalPosEmb.forward)__name__ __module__ __qualname____doc__rr __classcell__rrrrrs  rcs&eZdZdfdd ZddZZS)UnLearnedSinusoidalPosEmb'cs.t||_||_td|d|dS)as Create sinusoidal timestep embeddings. :param timesteps: a 1-D Tensor of N indices, one per batch element. These may be fractional. :param dim: the dimension of the output. :param max_period: controls the minimum frequency of the embeddings. :return: an [N x dim] Tensor of positional embeddings. z+Unlearned Timestep Embedding Schedule: dim=z , max_period=N)rrr max_periodprint)r r r%rrrr)s z"UnLearnedSinusoidalPosEmb.__init__c Cs|j}|d}|j}|j}tt| tjd|tjd|j |j d}|dddf |d}|j |d}tj t |t|gdd}|dratj |t|ddddfgdd}|S) Nrr)startenddtype)device)r)rr)r r%r)r exprlogarangefloat32tor*floatrrr zeros_like) r timestepsr halfr%r)rargs embeddingrrrr8s( (z!UnLearnedSinusoidalPosEmb.forward)r$)rrr rrr"rrrrr#(sr#) rr torch.nnreinopsrModulerr#rrrrs