o 3NPi- @s(dZddlZddlZddlmZmZmZddlZddl Z ddl Z ddl m Z e dZ ddejded ed e jd ed ejf d dZ d dedeed eeejffddZdeded efddZdeedee jded e jdeeejff ddZdeedee jd e jdeeejffddZdS)!a Speaker embedding extraction with OOM protection and compute monitoring. === OPTIMIZATION (v6.2) === Key improvements: - Accept audio buffer: No file re-read when buffer provided - GPU memory-aware batching: Query torch.cuda.mem_get_info() for dynamic batch sizing - Adaptive batching based on segment length - OOM protection with fallback to individual processing - Memory-efficient processing order (short segments first) N)ListDictOptional)MODELSzFastPipelineV6.Embeddings? audio_chunksrmax_window_samplesdevicewindow_overlapreturnc Cs(t|}t|d|}g}td||D]t} t| ||} || | } t| |kr*qzDt| d|} tdg|} t | | |  }|j dkrW| }||Wdn1sfwY~ ~ Wqty}ztd|WYd}~qd}~ww|sdStj|ddS)a Extract embedding from a long segment by chunking and averaging. === OPTIMIZATION (v6.1) === Instead of discarding long segments (>10s), we: 1. Split into overlapping windows 2. Extract embedding from each window 3. Average the embeddings This preserves ALL your data - longer samples are valuable for TTS! Args: audio_chunk: Audio samples (numpy array) sr: Sample rate max_window_samples: Maximum samples per window device: Torch device embedding_model: Embedding extraction model window_overlap: Overlap ratio between windows (0.5 = 50%) Returns: Averaged embedding vector r?NzWindow embedding failed: )axis)lenintrangemintorchtensor unsqueezetono_grad encode_batchcpunumpyndimsqueezeappend Exceptionloggerdebugnpmean)rrr r embedding_modelr total_samples step_sampleswindow_embeddings start_idxend_idxwindowpaddedwav_lensember0T/home/ubuntu/.cursor/worktrees/maya3data__SSH__216.81.248.184_/zxg/src/embeddings.py_extract_embedding_chunkeds4      r2 audio_pathsegmentsc# Cstdt|dt}tjddt}tj}|dur,|j}|j }tdnt |\} }| d }i} td|} |j} g} g}d}t|D]E\}}|d d vr[qOt|d |}t|d |}|||}t|| kr||d 7}qOt|| kr|||fqO| |t|fqO|dkrtd|d|rtdt|d|D]\}}t||| ||}|dur|| |<qtjdd| }|s| rtd| Std| S|jdddt||j}t||d |}tdt|d|d|dttdt||D]\}}t||t|}|||}dd|D}dd|D}tdd|D}t||d} | d krktd!|d d"| d#d$d%t||||| tjddqz t|||||| Wn)ty}!zt d!|d d&|!d't||||| WYd}!~!nd}!~!ww|d d(dkrtjddqtjddt|}"td)|"d*d+t| d,t|d-| S).u Extract speaker embeddings with adaptive batching to prevent OOM. === OPTIMIZATION (v6.2) === - Accept audio_buffer: No file re-read when buffer provided - GPU memory-aware batching: Use torch.cuda.mem_get_info() for dynamic batch sizing - Long segments (>10s) are chunked and averaged - preserving valuable data Strategy to prevent OOM while preserving ALL data: 1. Query available VRAM before each batch 2. Short segments: batch processing for efficiency 3. Long segments: chunk into windows → extract → average 4. Adaptive batch size based on VRAM + segment lengths 5. Aggressive cache clearing between batches Args: audio_path: Path to audio file segments: List of segments to extract embeddings for config: Pipeline configuration audio_buffer: Optional AudioBuffer (avoids file re-read) Returns: Dict mapping segment index to embedding array u🔢 Extracting embeddings (z segments)...T) aggressiveNz. Using pre-loaded audio buffer (no file I/O)rg333333?speaker)OVERLAP NON_SPEECHstartendr z Skipped z segments (too short <0.3s)z Processing z+ long segments via chunking (no data loss!)z2 All segments were long - processed via chunkingz- No valid segments for embedding extractioncSs t|dS)Nr r)xr0r0r1s z,extract_embeddings_batched..)keyz segments in z batches (size=)cSg|]}|dqS)rr0.0itemr0r0r1 z.extract_embeddings_batched..cSr@)r r0rAr0r0r1rDrEcss|]}t|VqdSNr;)rBcr0r0r1 sz-extract_embeddings_batched..iez Batch z too large (geAz.2fzGB), processing individuallyz failed: z, trying individual processingu✅ Embeddings: z.1fzs | /z extracted)!r!infortimer clear_cache get_devicer% waveform_np sample_rate torchaudioloadrrrmax_embedding_length enumerategetrrrr2warningsort_compute_adaptive_batch_sizeembedding_batch_sizerrmax_process_individually_process_batchr error)#r3r4config audio_bufferr9r r%rPrwaveform embeddings min_samples max_samples short_items long_itemsskipped_too_shortiseg start_sample end_samplechunkr. valid_items batch_size total_batches batch_num batch_start batch_end batch_itemsindices audio_chunksmax_lenestimated_sizer/elapsedr0r0r1extract_embeddings_batched]s           "  "    (ryrm base_batchc Cs|s|Stdd|Dt|}zddlm}|t|t||dddd}Wn ttfy4|}Ynw|d krAt|dd}n|d krH|}nt|dd }t||S) z Compute optimal batch size based on segment lengths AND available VRAM. === OPTIMIZATION (v6.2) === Query torch.cuda.mem_get_info() for dynamic batch sizing instead of relying solely on average length heuristics. css|] }t|dVqdS)r Nr;rAr0r0r1rHsz/_compute_adaptive_batch_size..r)compute_optimal_batch_size@g@) num_itemsavg_item_samplesrz min_batch max_batchvram_headroom_gbi>irI) sumrsrc.audio_bufferr{r ImportErrorr rr[)rmrzavg_lenr{ vram_batch length_batchr0r0r1rYs,   rYrtrurvrbc Cstt||}tt|}t|D]\}} | ||dt| f<t| |||<q||}||}t2|||} | j dkrU| j ddkrU| d} n| j dkrc| | j dd} Wdn1smwYt|D]\}} | j dkr| |n| || <qv~~~ dS)zProcess a batch of segments.Nrr r) rzerosrrUrrrrrrshaperreshape) rtrurvr r%rbr,r-jrlembsidxr0r0r1r]s$       r]c Cst||D]a\}}z@|d|}tdg|}t|||} | j dkr3| } | ||<Wdn1sAwY~~Wqt yf} zt d|d| WYd} ~ qd} ~ wwdS)z*Process segments one at a time (fallback).rrrNz# Individual embedding failed for z: )ziprrrrrrrrrrr r!r^) rtrur r%rbrrl padded_singlewav_lens_singler.r/r0r0r1r\<s    "r\)rrF)__doc__rMloggingtypingrrrrr#rrR src.modelsr getLoggerr!ndarrayrr floatr2strryrYTensorr]r\r0r0r0r1sj     H  )  !