o 9wia @sjdZddlZddlmZddlmmZddZGdddejZ GdddejZ d d Z d d Z dS) zThis pytorch_utils.py contains functions from: https://github.com/qiuqiangkong/audioset_tagging_cnn/blob/master/pytorch/pytorch_utils.py NcCsBdt|jvr t|}ndt|jvrt|}n|S||S)Nfloatint)strdtypetorchTensor LongTensorto)xdevicer Z/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/panns_inference/pytorch_utils.pymove_data_to_device s    rcs&eZdZdfdd ZddZZS) Interpolatornearestcs(tt||dkrt||_dSdS)zInterpolate the sound event detection result along the time axis. Args: ratio: int interpolate_mode: str rN)superr__init__NearestInterpolator interpolator)selfratiointerpolate_mode __class__r r rszInterpolator.__init__cCs ||S)zInterpolate the sound event detection result along the time axis. Args: x: (batch_size, time_steps, classes_num) Returns: (batch_size, new_time_steps, classes_num) )r)rr r r r forward"s zInterpolator.forward)r__name__ __module__ __qualname__rr __classcell__r r rr rs rcs$eZdZfddZddZZS)rcstt|||_dS)zxNearest interpolate the sound event detection result along the time axis. Args: ratio: int N)rrrr)rrrr r r/s zNearestInterpolator.__init__cCsN|j\}}}|dddddddfdd|jd}||||j|}|S)zInterpolate the sound event detection result along the time axis. Args: x: (batch_size, time_steps, classes_num) Returns: upsampled: (batch_size, new_time_steps, classes_num) N)shaperepeatrreshape)rr batch_size time_steps classes_num upsampledr r r r9s *zNearestInterpolator.forwardrr r rr r.s  rcCsH|ddddddfd||jdd} tj||fdd} |S)zPad framewise_output to the same length as input frames. Args: framewise_output: (batch_size, frames_num, classes_num) frames_num: int, number of frames to pad Outputs: output: (batch_size, frames_num, classes_num) Nr )dim)r"r!rcat)framewise_output frames_numpadoutputr r r pad_framewise_outputHs .r/cCsT|ddddd|ddd|ddddd|ddd}|ddS)Nrr(r ) transpose)r mixup_lambdaoutr r r do_mixupYs"" r4) __doc__rtorch.nnnntorch.nn.functional functionalFrModulerrr/r4r r r r s