^j->UdZddlmZddlZddlZddlmZddlmZddl Z ddl m Z ddl mZddlmZdd lmZeZeed ej,Zgd Zd ed <gdZd ed<ddZddZddZ d ddZ d ddZy)a`TFLite inference helpers for RF-DETR exported models. These functions handle interpreter creation, image preprocessing, and decoding of detection and segmentation-mask outputs without requiring PyTorch or the RF-DETR training stack: only ``tflite-runtime`` (or ``tensorflow``), ``numpy``, ``supervision``, and ``Pillow`` are needed at inference time. ) annotationsN)Path)Any)NDArray)Image) Detections) get_logger Resampling)g ףp= ?gv/?gCl?z list[float]_IMAGENET_MEAN)gZd;O?gy&1?g? _IMAGENET_STDc d}g}dD]H\}}tjt5tt j ||} dddn|'djd|D}td|d|t| }|j|j}|j}tjd |dd |dd j|D],} tjd | d | j!dd.|S#1swYnxYw|j |jddG)awLoad a TFLite model, allocate tensors, and log I/O shapes. Tries ``tflite_runtime`` first (lightweight; preferred on edge devices), then falls back to ``tensorflow.lite`` (pre-installed on Colab / full TF environments). Args: model_path: Path to the ``.tflite`` model file. Returns: An allocated TFLite interpreter ready for inference. N))zai_edge_litert.interpreter Interpreter)ztflite_runtime.interpreterr)ztensorflow.liter.rz, c3(K|] }d|d yw)'N).0ps j/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/export/_tflite/inference.py z&_create_interpreter..<s8A1#Qx8szVTFLite inference requires 'ai_edge_litert', 'tflite-runtime', or 'tensorflow' (tried: zN). Install one: `pip install ai_edge_litert` OR `pip install tflite-runtime`) model_pathzInput : %s %sshapedtypezOutput : %s name=%sname ) contextlibsuppress ImportErrorgetattr importlib import_moduleappendsplitjoinstrallocate_tensorsget_input_detailsget_output_detailsloggerdebug__name__get) r _Interpreter_tried_pkg_attr _tried_strinterpinp_detout_detods r_create_interpreterr6$sXLF* e   - "9#:#:4#@%HL    *YY888 !l#Z Z  S_ 5F &&(G'')G LL"GAJw$7G9L9U9UVW +R["&&:UVW M'    djjoa()s !DD& c|jd|jd}}tj|tjdz||z zdz }tj|tjdz||z zdz }tj|d|dz }tj|d|dz }tj |j tj}tj |j tj}tj|dz|dz } tj|dz|dz } ||z dddf} ||z dddf} |d|dddf|dddff} |d|dddf| dddff}|d| dddf|dddff}|d| dddf| dddff}d| z d| z | z| |zzz| d| z |z| |zzzz}tj|tjS) aLNumpy bilinear resize matching ``F.interpolate(mode="bilinear", align_corners=False)``. Half-pixel center convention. Used by ``_decode_masks`` only when ``torch`` is not importable. Args: src: Source array of shape ``(K, src_h, src_w)``. out_h: Target height in pixels. out_w: Target width in pixels. Returns: Float32 array of shape ``(K, out_h, out_w)``. Note: Replaces ``PIL.Image.resize(BILINEAR)``, which uses a corner-aligned half-pixel convention and produced border-pixel discrepancies vs ``F.interpolate``. rg?N.) rnparangefloat32clipfloorastypeint64minimumasarray)srcout_hout_wsrc_hsrc_wsrc_ysrc_xy0x0y1x1dydxabcdouts r_bilinear_resize_half_pixelrXMs"99R=#))B-5E YYuBJJ /# 5%%- H3 NE YYuBJJ /# 5%%- H3 NE GGE3 *E GGE3 *E %   )B %   )B BFEAI &B BFEAI &B "*ag B "*dAg B CAtGbqk )*A CAtGbqk )*A CAtGbqk )*A CAtGbqk )*A r6q2vlR!V+ ,ra"f\BF5J/K KC ::c ,,c|jdk7rtd|jd|\}}|jddk(r(tjd||ftj S ddl}ddlmcm }|j5|j|jtjjd}|j|||fdd }dddtj j#dj%tj}|d kDS#1swYPxYw#t&$r1t)|jtj||}Y|d kDSwxYw) aUpsample mask logits to image size and threshold at zero. Matches ``PostProcess.forward``: bilinear upsample with ``align_corners=False`` followed by ``> 0``. Uses ``torch.nn.functional.interpolate`` when torch is importable for bit-exact parity, and falls back to the pure-NumPy ``_bilinear_resize_half_pixel`` otherwise. Args: mask_logits: Raw mask logits of shape ``(K, Hm, Wm)``. out_size: Target ``(width, height)`` in pixels. Returns: Boolean mask array of shape ``(K, height, width)``. Raises: ValueError: If *mask_logits* is not rank-3. Note: ``out_size`` follows PIL convention ``(width, height)``; the returned array uses NumPy/PyTorch convention ``(K, height, width)``. z4_decode_masks expects rank-3 (K, Hm, Wm); got shape za. This usually means the rank-4 mask-output heuristic in _run_inference matched the wrong tensor.rr:NbilinearF)sizemode align_cornersr;)ndim ValueErrorrr=zerosbool_torchtorch.nn.functionalnn functionalno_grad from_numpyrBr? unsqueeze interpolaterEsqueezenumpyrrX) mask_logitsout_sizewidthheightrd_Ftresizeds r _decode_masksruqsW*1B;CTCTBUVn n  ME6q xxFE*"((;; ](( ]]_ ^  !3!3BJJ!?@JJ1MAqZW\]A ^(*zz!))A,2D2D2Fbjj'Y S=  ^ ^ ]-k.@.@.LfV[\ S=]s,)D6AD*A D6*D3/D663E0/E0c$|\}}|dk(rdnd}|j|}d} ddl}ddlmcm} |j 5| j |} | j| t|} t|D cgc]} t| dz} } t|D cgc]} t| dz} } | j| | | } dddtj jdj!j#tj$}|7tj|j)ddddtj$Stj*|j||ft,tj$d z }|j.dk(r|ddddtj0f}tj*t|D cgc]} t| dzc} tj$}tj*t|D cgc]} t| dzc} tj$}||z |z tj0Scc} wcc} w#1swYxYw#t&$rYhwxYwcc} wcc} w) a!Resize and ImageNet-normalise an image to match ``RFDETR.predict()``. Uses ``torchvision.transforms.functional`` when importable for bit-exact parity, and falls back to ``PIL.Image.resize`` with BILINEAR for torch-free deployments. Args: pil_img: Source PIL image at native resolution. hw: Target ``(height, width)`` from the interpreter's input shape. channels: Channel count (3 for RGB, 1 for grayscale). Returns: Float32 array of shape ``(1, height, width, channels)`` in NHWC. Note: The PIL fallback uses BILINEAR resize, which does not perfectly match PyTorch's ``F.resize`` (different coordinate conventions). For bit-exact parity with ``RFDETR.predict()``, ensure ``torch`` and ``torchvision`` are importable. r<LRGBNrr[r:go@)convertrd!torchvision.transforms.functional transformsrgrh to_tensorresizelistranger r normalizer=rErjcpurmr?r transposearray _PIL_BILINEARr`newaxis)pil_imghwchannelsrqrppil_modepil_rgb nchw_floatrdrrrsi mean_liststd_listarrmeanstds r_preprocess_imagers .MFE!msHooh'G-1J  66 ]]_ 5 W%A !T"X&A8=hH1A.HIH6;HoF a!e,FHF Q 84A  5 ZZ A 2 2 4 : : >5&/=A TW\ \C xx1}!Q "# 88E(ODq^AE*DBJJ WD ((%/BQM!a%(B"** UC 4Z3  ++'IF  5 5    EBsTI8:I+:I! I+I&/I+AI8JJ ! I++I50I88 JJc  |j}|j}|dd\}}}}tj} |dd} | | k7rt d| j dt j|} t| t|t|ft|} |j|dd| |j|D cgc]} t| jdd}} td t|Dd }td t|Dd }||t j#d |t|D cgc]%\}} t%| dd k(s| dddk(s$|'}}} t|D cgc]%\}} t%| dd k(s| dddk7s$|'}}} t%|dk(rt%|dk(r |d}|d}nQt%|dk(rt j#dd}d}n)|D cgc]} t'| d}} t d||j)||dd}t+|j-dkDst+|j/dkrqt j#dt+|j-|j)||dj0ddz ||}}|j)||dd}|j)||ddd d d df}t j#d|j0t+|j/t+|j-t+|j3tj4d|j6}||tj8|j;dd zz }|j-d}|j=d}t j#dt+|j/t+|j-|t||kDj?||kD}||j@\}}}}| jB\} }!tjD||dz z ||dz z ||dz z||dz zgd}"|"tjF| |!| |!gtjz}"tdt|Dd }#|#pt|D cgc]\}} t%| ddk(s|}$}} t%|$dk(r|$d}#n-t%|$dk\rt jIdt%|$d }%|#;|jKr+|j)||#dd}&tM|&|| |!f}%tO|"||||jQt|%}'|'| fScc} wcc} }wcc} }wcc} wcc} }w) aPreprocess one image, run TFLite inference, and decode detections. Reads input shape from the interpreter (NHWC ``float32``), resizes and normalises the image with ImageNet statistics, invokes the model, then decodes the ``dets`` / ``labels`` output tensors into a :class:`supervision.Detections` object with pixel-space ``xyxy`` boxes. For segmentation exports the ``masks`` output is also decoded into ``Detections.mask``. Args: interp: Allocated TFLite interpreter returned by ``_create_interpreter``. image_path: Path to the input image (any format supported by Pillow). threshold: Confidence threshold; detections below this are discarded. Returns: A tuple of ``(detections, pil_img)`` where ``detections`` contains pixel-space ``xyxy`` boxes (and ``mask`` for segmentation models) and ``pil_img`` is the original PIL image at its original resolution. rrrzF_run_inference only supports float32 input tensors, but model expects zV. Export the model with float32 quantization or implement input quantization manually.indexrrc3bK|]'\}}dt|jddvs$|)yw)detsrNr%r,rrr5s rrz!_run_inference..s-]EAr6SPVXZI[E\;\a]%//Nc3bK|]'\}}dt|jddvs$|)yw)labelsrrNrrs rrz!_run_inference..s-`UQHBFFSY[]L^H_<_q`rzXName-based output matching failed (available: %s). Falling back to shape-based matching.r[r9r<ryzKShape-based matching ambiguous. Using positional order (0=boxes, 1=logits).zShape-based TFLite output matching failed. Expected exactly one rank-3 tensor with last dim == 4 (boxes) and one rank-3 tensor with last dim != 4 (logits). Available output shapes: g@guBox tensor max=%.2f exceeds [0,1] — swapping boxes/logits assignment (num_classes==%d likely caused ambiguous positional fallback).z2Logits stats: shape=%s min=%.3f max=%.3f mean=%.3fr:iX)axisuGScores stats: min=%.3f max=%.3f — detections above threshold %.2f: %dc3bK|]'\}}dt|jddvs$|)yw)masksrrNrrs rrz!_run_inference..Os-]51b'SPVXZI[E\:\Q]rzyAmbiguous rank-4 outputs (%d candidates); skipping mask decode. Name your mask output to contain 'masks' to disambiguate.)xyxy confidenceclass_idmask))r'r(r=r?rar+PILImageopenrint set_tensorinvoker%r,next enumerater)r*lenr get_tensorfloatmaxminrrrErexpr@argmaxsumTr]stackrwarninganyrurrB)(r2 image_path thresholdr3r4_rqrprexpected_dtype actual_dtyper inp_tensorr5available_output_names boxes_idx logits_idxrshape_boxes_candidatesshape_logits_candidatesavailable_shapes boxes_cwhlogitsone scores_allscoresclskeepcxcybwbhowohrmask_idxrank4_candidatesr raw_masks detectionss( r_run_inferencers*&&(G'')G!(G!4AvuhZZN1:g&L~%TUaUjUjTklc c  mmJ'G"7S[#e*,Es8}UJ gaj):6 MMOJQQ2c"&&"=>QQ]Yw%7]_cdI`i&8`bfgJJ.  f " 2;71C!v2s2g;GW[\G\acdkalmoaptuau!!v!v2;G2D"wBBwKHX\]H]bdelbmnpbquvbv1"w"w % &! +4K0LPQ0Q.q1I03J \Q  LLf gIJqAI Y]]_#uY]]_'='D  M )--/ "   gj1': ; A A" E I  !+I: %%gi&8&AB1E   wz27; rs # !". ,9BB 3 32 {2&R!-H'Z7, 7,7,7, 7,zK KKK' KrY