o 1i1@sfdZddlmZddlmZmZddlmZmZm Z ddl Z eGdddZ e Gdd d eZ dS) zEUnified codec interface: TokenBatch dataclass + NeuralCodec protocol.) annotations) dataclassfield)AnyProtocolruntime_checkableNc@sleZdZUdZded<ded<ded<eedZd ed <edd d Z edddZ dddZ dddZ dS) TokenBatchaXContainer for codec token output, supporting diverse token structures. tokens can be: - torch.LongTensor [B, T] single-stream (XCodec2, WavTokenizer) - list[torch.LongTensor] multi-scale (SNAC: different T per level) - dict[str, torch.LongTensor] named streams (BiCodec: semantic + global) str codec_nameint sample_ratertokens)default_factoryzdict[str, Any]auxreturncCsrt|jtjr |jjdSt|jtrtt|jjdSt|jt t fr/|jdjdSt dt |j)Nrz)Cannot infer batch_size from tokens type ) isinstancer torchTensorshapedictnextitervalueslisttuple ValueErrortypeselfr3/home/ubuntu/bench-codecs/codecbench/codecs/base.py batch_sizes  zTokenBatch.batch_sizecCsdt|jtjr |jSt|jtrtdd|jDSt|jtt fr0tdd|jDSdS)z9Total number of token elements across all streams/levels.cs|]}|VqdSNnumel).0vrrr *z)TokenBatch.token_count..csr"r#r$r&trrr r(,r)r) rr rrr%rsumrrrrrrr token_count$s  zTokenBatch.token_counttuple[int, int]cCsg}t|jtjr|jg}nt|jtrt|j}n t|jttfr)t|j}tdd|D}t dd|D}t |t |fS)z@Return (min_token, max_token) observed across all token tensors.cs|] }|VqdSr#)minitemr*rrr r(8z,TokenBatch.observed_vocab..csr/r#)maxr1r*rrr r(9r2) rr rrrrrrr0r3r )rtensorsall_minall_maxrrr observed_vocab/s   zTokenBatch.observed_vocabcCsjt|jtjrtt|jjSt|jtr!tdd|jDSt|jtt fr3tdd|jDSdS)NcSsi|] \}}|t|jqSrrr)r&kr'rrr @sz-TokenBatch.shapes_summary..cSsg|]}t|jqSrr8r*rrr Bsz-TokenBatch.shapes_summary..unknown) rr rrr rrritemsrrrrr shapes_summary<s zTokenBatch.shapes_summaryN)rr )rr.)rr ) __name__ __module__ __qualname____doc____annotations__rrrpropertyr!r-r7r>rrrr r s     rc@sTeZdZUdZded<ded<dd d Zd ddZd!ddZd"ddZd#ddZ dS)$ NeuralCodecz3Unified interface every codec wrapper must satisfy.r namer native_srdevicedtype torch.dtyperNonecCdS)z1Load model weights and move to device. Call once.Nr)rrHrIrrr loadMzNeuralCodec.load batch_secondsfloatr!cCrL)z9Run throwaway forward passes so JIT/CUDA caches are warm.Nr)rrOr!rrr warmupQrNzNeuralCodec.warmupwav torch.TensorsrrcCrL)zEncode waveform to tokens. Args: wav: float32 [-1, 1], shape [B, 1, T] sr: sample rate of wav (wrapper resamples if needed) Nr)rrRrTrrr encodeUszNeuralCodec.encodetbcCrL)z3Decode tokens back to waveform. Returns [B, 1, T'].NrrrVrrr decode^rNzNeuralCodec.decodetorch.LongTensorcCrL)zFlatten token structure to a single 1-D sequence per batch element. Returns [B, T_flat]. Used to estimate modeling complexity. NrrWrrr flatten_for_lmbszNeuralCodec.flatten_for_lmN)rHr rIrJrrK)rOrPr!r rrK)rRrSrTr rr)rVrrrS)rVrrrY) r?r@rArBrCrMrQrUrXrZrrrr rEFs     rE) rB __future__r dataclassesrrtypingrrrrrrErrrr s :