§
    ­¸jË;  ã                  ó  — U d Z ddlmZ ddlZddlZddlmZ ddlmZm	Z	 ddl
ZddlmZ ddlmZ ddlmZ dd	lmZ dd
lmZ ddlmZ ddlmZ  e¦   «         Zg d¢Zded<   g d¢Zded<   dZded<   d5d„Z d6d„Z!	 d7d8d'„Z"	 	 	 	 d9d:d4„Z#dS );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.
é    )ÚannotationsN)ÚPath)ÚAnyÚLiteral)ÚNDArray)ÚImage)Ú
Detections)Ú_bilinear_resize_half_pixel)Údecode_detections)Úpreprocess_to_nchw)Ú
get_logger)g
×£p=
ß?gÉv¾Ÿ/Ý?g–C‹lçûÙ?zlist[float]Ú_IMAGENET_MEAN)gZd;ßOÍ?gyé&1¬Ì?gÍÌÌÌÌÌÌ?Ú_IMAGENET_STD)ÚmasksÚ	keypointsztuple[str, ...]Ú_RANK4_OUTPUT_KINDSÚ
model_pathú
str | PathÚreturnr   c           	     ó  — d}g }dD ]ˆ\  }}t          j        t          ¦  «        5  t          t	          j        |¦  «        |¦  «        }	 ddd¦  «          n:# 1 swxY w Y   |                     |                     d¦  «        d         ¦  «         Œ‰|€2d                     d„ |D ¦   «         ¦  «        }t          d|› d�¦  «        ‚ |t          | ¦  «        ¬	¦  «        }| 
                    ¦   «          |                     ¦   «         }|                     ¦   «         }t                               d
|d         d         |d         d         j        ¦  «         |D ]8}	t                               d|	d         |	                     dd¦  «        ¦  «         Œ9|S )aw  Load 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))zai_edge_litert.interpreterÚInterpreter)ztflite_runtime.interpreterr   )ztensorflow.liter   ú.r   z, c              3  ó"   K  — | ]
}d |› d �V — ŒdS )ú'N© )Ú.0Úps     úo/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/rfdetr/export/_tflite/inference.pyú	<genexpr>z&_create_interpreter.<locals>.<genexpr>=   s*   è è € Ð8Ð8¨A˜x 1˜x˜x˜xÐ8Ð8Ð8Ð8Ð8Ð8ó    zVTFLite inference requires 'ai_edge_litert', 'tflite-runtime', or 'tensorflow' (tried: zN). Install one: `pip install ai_edge_litert`  OR  `pip install tflite-runtime`)r   zInput  : %s  %sÚshapeÚdtypezOutput : %s  name=%sÚnameú	<unnamed>)Ú
contextlibÚsuppressÚImportErrorÚgetattrÚ	importlibÚimport_moduleÚappendÚsplitÚjoinÚstrÚallocate_tensorsÚget_input_detailsÚget_output_detailsÚloggerÚdebugÚ__name__Úget)
r   Ú_InterpreterÚ_triedÚ_pkgÚ_attrÚ
_tried_strÚinterpÚinp_detÚout_detÚods
             r   Ú_create_interpreterr?   %   sâ  € ð €LØ€Fðð *ð *‰ˆˆeõ
 Ô ¥Ñ-Ô-ð 	ð 	Ý"¥9Ô#:¸4Ñ#@Ô#@À%ÑHÔHˆLØð	ð 	ð 	ñ 	ô 	ð 	ð 	ð 	ð 	ð 	ð 	ð 	øøøð 	ð 	ð 	ð 	ð 	�Š�d—j’j ‘o”o aÔ(Ñ)Ô)Ð)Ð)ØÐØ—Y’YÐ8Ð8°Ð8Ñ8Ô8Ñ8Ô8ˆ
ÝðZØ!ðZð Zð Zñ
ô 
ð 	
ð ˆ\¥S¨¡_¤_Ð5Ñ5Ô5€FØ
×ÒÑÔÐØ×&Ò&Ñ(Ô(€GØ×'Ò'Ñ)Ô)€GÝ
‡L‚LÐ" G¨A¤J¨wÔ$7¸À¼ÀGÔ9LÔ9UÑVÔVÐVØð Wð WˆÝ�ŠÐ+¨R°¬[¸"¿&º&ÀÈÑ:UÔ:UÑVÔVÐVÐVØ€Ms   ¥$AÁA	ÁA	Úmask_logitsúNDArray[np.floating[Any]]Úout_sizeútuple[int, int]úNDArray[np.bool_]c                ó  — | j         dk    rt          d| j        › d�¦  «        ‚|\  }}| j        d         dk    r#t          j        d||ft          j        ¬¦  «        S 	 ddl}ddlmc m	} | 
                    ¦   «         5  |                     |                      t          j        ¦  «        ¦  «                             d¦  «        }|                     |||fdd¬	¦  «        }ddd¦  «         n# 1 swxY w Y   t          j        |                     d¦  «                             ¦   «         t          j        ¬¦  «        }n># t&          $ r1 t)          |                      t          j        ¦  «        ||¦  «        }Y nw xY w|d
k    S )aþ  Upsample 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)``.
    é   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.r   ©r"   NÚbilinearF)ÚsizeÚmodeÚalign_cornersg        )ÚndimÚ
ValueErrorr!   ÚnpÚzerosÚbool_ÚtorchÚtorch.nn.functionalÚnnÚ
functionalÚno_gradÚ
from_numpyÚastypeÚfloat32Ú	unsqueezeÚinterpolateÚasarrayÚsqueezeÚnumpyr'   r
   )r@   rB   ÚwidthÚheightrQ   Ú_FÚtÚresizeds           r   Ú_decode_masksrc   N   sé  € ð* Ô˜1ÒÐÝðnÀ;ÔCTð nð nð nñ
ô 
ð 	
ð �M€Eˆ6ØÔ˜Ô˜qÒ Ð ÝŒx˜˜F EÐ*µ"´(Ð;Ñ;Ô;Ð;ð	]ØˆˆˆØ(Ð(Ð(Ð(Ð(Ð(Ð(Ð(Ð(à�]Š]‰_Œ_ð 	^ð 	^Ø× Ò  ×!3Ò!3µB´JÑ!?Ô!?Ñ@Ô@×JÒJÈ1ÑMÔMˆAØ—’˜q¨° ¸ZÐW\�Ñ]Ô]ˆAð	^ð 	^ð 	^ñ 	^ô 	^ð 	^ð 	^ð 	^ð 	^ð 	^ð 	^øøøð 	^ð 	^ð 	^ð 	^õ (*¤z°!·)²)¸A±,´,×2DÒ2DÑ2FÔ2FÍbÌjÐ'YÑ'YÔ'YˆˆøÝð ]ð ]ð ]Ý-¨k×.@Ò.@ÅÄÑ.LÔ.LÈfÐV[Ñ\Ô\ˆˆˆð]øøøà�SŠ=Ðs8   Á!D= Á?A!C,Ã D= Ã,C0Ã0D= Ã3C0Ã4AD= Ä=8E8Å7E8rF   Úpil_imgúPILImage.ImageÚhwÚchannelsÚintúNDArray[np.float32]c                óœ   — |\  }}t          | |||¦  «        }t          j        |                     dddd¦  «        t          j        ¬¦  «        S )aÛ  Resize and ImageNet-normalise an image to match ``RFDETR.predict()``.

    Thin NHWC adapter over :func:`~rfdetr.export._runtime.preprocess.preprocess_to_nchw`, which uses
    ``torchvision.transforms.functional`` when importable for bit-exact parity and falls back to the pure-NumPy
    ``_bilinear_resize_half_pixel`` for torch-free deployments. Both paths resize with predict()'s convention:
    bilinear, half-pixel centers, ``antialias=False``.

    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 NumPy fallback matches the torchvision path up to float32 op-order noise (~5e-5 in
        normalised space). For bit-exact parity with ``RFDETR.predict()``, ensure ``torch`` and
        ``torchvision`` are importable.
    r   é   rF   é   rG   )r   rN   r[   Ú	transposerX   )rd   rf   rg   r_   r^   Únchws         r   Ú_preprocess_imagero   x   sI   € ð2 �M€FˆEÝ˜g v¨u°hÑ?Ô?€DåŒ:�d—n’n Q¨¨1¨aÑ0Ô0½¼
ÐCÑCÔCÐCr    ç333333Ó?éÿÿÿÿr;   Ú
image_pathÚ	thresholdÚfloatÚ
num_selectú
int | NoneÚbackground_class_idÚrank4_outputú$Literal['masks', 'keypoints'] | Noneú!tuple[Detections, PILImage.Image]c           	     óR
  — |�#|t           vrt          dt           › d|›�¦  «        ‚|                      ¦   «         }|                      ¦   «         }|d         d         \  }}	}
}t          j        }|d         d         }||k    rt          d|j        › d�¦  «        ‚t          j        |¦  «        5 }t          |t          |	¦  «        t          |
¦  «        ft          |¦  «        ¦  «        }ddd¦  «         n# 1 swxY w Y   |                      |d         d	         |¦  «         |                      ¦   «          d
„ |D ¦   «         }t          d„ t          |¦  «        D ¦   «         d¦  «        }t          d„ t          |¦  «        D ¦   «         d¦  «        }|�|€Ôt                               d|¦  «         d„ t          |¦  «        D ¦   «         }d„ t          |¦  «        D ¦   «         }t#          |¦  «        dk    r$t#          |¦  «        dk    r|d         }|d         }nPt#          |¦  «        dk    rt                               d¦  «         d}d}nd„ |D ¦   «         }t          d|› �¦  «        ‚|                      ||         d	         ¦  «        d         }t'          |                     ¦   «         ¦  «        dk    s%t'          |                     ¦   «         ¦  «        dk     r“t                               dt'          |                     ¦   «         ¦  «        |                      ||         d	         ¦  «        j        d         dz
  ¦  «         ||}}|                      ||         d	         ¦  «        d         }|                      ||         d	         ¦  «        d         }t/          |||j        |||¬¦  «        }|j        }t          d„ t          |¦  «        D ¦   «         d¦  «        }|€¸|dk    r²d„ t          |¦  «        D ¦   «         }t#          |¦  «        dk    rK|d         }t                               dt5          ||                              dd¦  «        ¦  «        ¦  «         n;t#          |¦  «        dk    r(t                               d t#          |¦  «        ¦  «         d}|�S|j        d         dk    rB|                      ||         d	         ¦  «        d         }t;          ||         |j        ¦  «        }t=          |j        |j         |j!         "                    t          ¦  «        |¬!¦  «        }||fS )"a	  Preprocess 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.
        num_select: Maximum query/class pairs selected before thresholding. ``None`` uses the exported model's query
            count, matching shipped RF-DETR configurations; pass an explicit value for custom exports.
        background_class_id: Exported class slot to exclude before selection. The default ``-1`` preserves the common
            final-slot background convention. Pass ``None`` for sparse COCO checkpoints, whose final slot is class 90,
            or ``0`` for legacy background-first keypoint checkpoints.
        rank4_output: What a rank-4 output holds when no output names itself ``masks``. Segmentation and keypoint
            exports each add exactly one rank-4 tensor, and RF-DETR's own TFLite files reach the interpreter with
            the ONNX output names replaced by ``StatefulPartitionedCall:N``, so the kind is usually not readable
            from the graph. The default ``None`` decodes a mask only from an output that names itself. Pass
            ``"masks"`` for a name-stripped segmentation export or ``"keypoints"`` to suppress anonymous-mask
            decoding for a keypoint export.

    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.

    Raises:
        ValueError: If *rank4_output* is neither ``None`` nor one of ``"masks"``/``"keypoints"``, if the model's
            input tensor is not ``float32``, or if the ``dets``/``labels`` outputs cannot be matched by name or
            shape.
    Nzrank4_output must be one of z or None; got r   r!   r"   zF_run_inference only supports float32 input tensors, but model expects zV. Export the model with float32 quantization or implement input quantization manually.Úindexc                óT   — g | ]%}t          |                     d d¦  «        ¦  «        ‘Œ&S )r#   r$   ©r.   r5   ©r   r>   s     r   ú
<listcomp>z"_run_inference.<locals>.<listcomp>Ö   s.   € ÐQÐQÐQÀ2�c "§&¢&¨°Ñ"=Ô"=Ñ>Ô>ÐQÐQÐQr    c              3  ój   K  — | ].\  }}d t          |                     dd¦  «        ¦  «        v ¯*|V — Œ/dS )Údetsr#   Ú Nr~   ©r   Úir>   s      r   r   z!_run_inference.<locals>.<genexpr>×   sH   è è € Ð]Ð]™E˜A˜r¸6ÅSÈÏÊÐPVÐXZÑI[ÔI[ÑE\ÔE\Ð;\Ð;\�aÐ;\Ð;\Ð;\Ð;\Ð]Ð]r    c              3  ój   K  — | ].\  }}d t          |                     dd¦  «        ¦  «        v ¯*|V — Œ/dS )Úlabelsr#   rƒ   Nr~   r„   s      r   r   z!_run_inference.<locals>.<genexpr>Ø   sH   è è € Ð`Ð`™U˜Q ¸HÍÈBÏFÊFÐSYÐ[]ÑL^ÔL^ÑH_ÔH_Ð<_Ð<_�qÐ<_Ð<_Ð<_Ð<_Ð`Ð`r    zXName-based output matching failed (available: %s). Falling back to shape-based matching.c                ón   — g | ]2\  }}t          |d          ¦  «        dk    ¯|d          d         dk    ¯0|‘Œ3S ©r!   rF   rq   é   ©Úlenr„   s      r   r€   z"_run_inference.<locals>.<listcomp>ã   sK   € Ð!vÐ!vÐ!v©¨¨2ÅsÈ2ÈgÌ;ÑGWÔGWÐ[\ÒG\ÐG\ÐacÐdkÔalÐmoÔapÐtuÒauÐau !ÐauÐauÐaur    c                ón   — g | ]2\  }}t          |d          ¦  «        dk    ¯|d          d         dk    ¯0|‘Œ3S r‰   r‹   r„   s      r   r€   z"_run_inference.<locals>.<listcomp>ä   sK   € Ð"wÐ"wÐ"w©¨¨BÍÈBÈwÌKÑHXÔHXÐ\]ÒH]ÐH]ÐbdÐelÔbmÐnpÔbqÐuvÒbvÐbv 1ÐbvÐbvÐbvr    rl   rk   zKShape-based matching ambiguous. Using positional order (0=boxes, 1=logits).c                ó8   — g | ]}t          |d          ¦  «        ‘ŒS )r!   )Úlistr   s     r   r€   z"_run_inference.<locals>.<listcomp>ï   s$   € ÐDÐDÐD°b¥ R¨¤[Ñ 1Ô 1ÐDÐDÐDr    zµShape-based TFLite output matching failed. Expected exactly one rank-3 tensor with last dim == 4 (boxes) and one rank-3 tensor with last dim != 4 (logits). Available output shapes: g       @g       Àu…   Box tensor max=%.2f exceeds [0,1] â€” swapping boxes/logits assignment (num_classes==%d likely caused ambiguous positional fallback).rq   )rs   ru   rw   c              3  ój   K  — | ].\  }}d t          |                     dd¦  «        ¦  «        v ¯*|V — Œ/dS )r   r#   rƒ   Nr~   r„   s      r   r   z!_run_inference.<locals>.<genexpr>  sH   è è € Ð]Ð]™5˜1˜b¸'ÅSÈÏÊÐPVÐXZÑI[ÔI[ÑE\ÔE\Ð:\Ð:\�QÐ:\Ð:\Ð:\Ð:\Ð]Ð]r    r   c           	     ó”   — g | ]E\  }}t          |d          ¦  «        dk    ¯dt          |                     dd¦  «        ¦  «        v¯C|‘ŒFS )r!   rŠ   r   r#   rƒ   )rŒ   r.   r5   r„   s      r   r€   z"_run_inference.<locals>.<listcomp>  s_   € ð 
ð 
ð 
Ù�!�Rµ°R¸´[Ñ1AÔ1AÀQÒ1FÐ1FÈ;Õ^aÐbd×bhÒbhÐioÐqsÑbtÔbtÑ^uÔ^uÐKuÐKuˆAÐKuÐKuÐKur    zaRank-4 output %s carries no kind in its name; decoding it as a caller-declared segmentation mask.r#   r$   zyAmbiguous rank-4 outputs (%d candidates); skipping mask decode. Name your mask output to contain 'masks' to disambiguate.)ÚxyxyÚ
confidenceÚclass_idÚmask)#r   rM   r0   r1   rN   rX   r4   ÚPILImageÚopenro   rh   Ú
set_tensorÚinvokeÚnextÚ	enumerater2   r3   rŒ   Ú
get_tensorrt   ÚmaxÚminr!   r   rI   Úquery_indexr.   r5   Úwarningrc   r	   r’   r“   r”   rW   )r;   rr   rs   ru   rw   rx   r<   r=   Ú_r_   r^   rg   Úexpected_dtypeÚactual_dtyperd   Ú
inp_tensorÚavailable_output_namesÚ	boxes_idxÚ
logits_idxÚshape_boxes_candidatesÚshape_logits_candidatesÚavailable_shapesÚ	boxes_cwhÚlogitsÚdecodedÚ	query_idxÚmask_idxÚrank4_candidatesr   Ú	raw_masksÚ
detectionss                                  r   Ú_run_inferencer³   —   sÆ  € ðP Ð LÕ8KÐ$KÐ$KÝÐkÕ8KÐkÐkÐ[gÐkÐkÑlÔlÐlà×&Ò&Ñ(Ô(€GØ×'Ò'Ñ)Ô)€GØ!(¨¤¨GÔ!4Ñ€A€vˆu�hå”Z€NØ˜1”:˜gÔ&€LØ�~Ò%Ð%ÝðcÐUaÔUjð cð cð cñ
ô 
ð 	
õ
 
Œ�zÑ	"Ô	"ð Z gÝ& wµ°V±´½cÀ%¹j¼jÐ0IÍ3ÈxÉ=Ì=ÑYÔYˆ
ðZð Zð Zñ Zô Zð Zð Zð Zð Zð Zð Zøøøð Zð Zð Zð Zð ×Ò�g˜a”j Ô)¨:Ñ6Ô6Ð6Ø
‡M‚M�O„O€Oð RÐQÈÐQÑQÔQÐÝÐ]Ð]¥Y¨wÑ%7Ô%7Ð]Ñ]Ô]Ð_cÑdÔd€IÝÐ`Ð`¥i°Ñ&8Ô&8Ð`Ñ`Ô`ÐbfÑgÔg€JØÐ˜JÐ.õ 	�ŠØfØ"ñ	
ô 	
ð 	
ð "wÐ!vµ¸7Ñ1CÔ1CÐ!vÑ!vÔ!vÐØ"wÐ"wµ)¸GÑ2DÔ2DÐ"wÑ"wÔ"wÐÝÐ%Ñ&Ô&¨!Ò+Ð+µÐ4KÑ0LÔ0LÐPQÒ0QÐ0QØ.¨qÔ1ˆIØ0°Ô3ˆJˆJÝ�‰\Œ\˜QÒÐõ �LŠLÐfÑgÔgÐgØˆIØˆJˆJàDÐD¸GÐDÑDÔDÐÝð?à,<ð?ð ?ñô ð ð
 ×!Ò! '¨)Ô"4°WÔ"=Ñ>Ô>¸qÔA€Iõ ˆY�]Š]‰_Œ_ÑÔ Ò#Ð#¥u¨Y¯]ª]©_¬_Ñ'=Ô'=ÀÒ'DÐ'DÝ�ŠðMå�)—-’-‘/”/Ñ"Ô"Ø×Ò˜g jÔ1°'Ô:Ñ;Ô;ÔAÀ"ÔEÈÑIñ		
ô 	
ð 	
ð !+¨I�:ˆ	Ø×%Ò% g¨iÔ&8¸Ô&AÑBÔBÀ1ÔEˆ	ð ×Ò˜w zÔ2°7Ô;Ñ<Ô<¸QÔ?€FåØØØŒØØØ/ðñ ô €Gð Ô#€Iõ
 Ð]Ð]¥I¨gÑ$6Ô$6Ð]Ñ]Ô]Ð_cÑdÔd€HØÐ˜L¨GÒ3Ð3ð
ð 
Ý$ WÑ-Ô-ð
ñ 
ô 
Ðõ ÐÑ Ô  AÒ%Ð%Ø'¨Ô*ˆHÝ�LŠLØsÝ�G˜HÔ%×)Ò)¨&°+Ñ>Ô>Ñ?Ô?ñô ð ð õ Ð!Ñ"Ô" aÒ'Ð'Ý�NŠNðLåÐ$Ñ%Ô%ñô ð ð
 €EØÐ 	¤°Ô 2°QÒ 6Ð 6Ø×%Ò% g¨hÔ&7¸Ô&@ÑAÔAÀ!ÔDˆ	õ ˜i¨	Ô2°G´LÑAÔAˆåØŒ\ gÔ&8À7ÔCS×CZÒCZÕ[^ÑC_ÔC_Ðfkðñ ô €Jð �wÐÐs   Â-;C4Ã4C8Ã;C8)r   r   r   r   )r@   rA   rB   rC   r   rD   )rF   )rd   re   rf   rC   rg   rh   r   ri   )rp   Nrq   N)r;   r   rr   r   rs   rt   ru   rv   rw   rv   rx   ry   r   rz   )$Ú__doc__Ú
__future__r   r%   r)   Úpathlibr   Útypingr   r   r]   rN   Únumpy.typingr   ÚPILr   r–   Úsupervisionr	   Úrfdetr.export._resizer
   Úrfdetr.export._runtime.decoder   Ú!rfdetr.export._runtime.preprocessr   Úrfdetr.utilities.loggerr   r2   r   Ú__annotations__r   r   r?   rc   ro   r³   r   r    r   ú<module>rÀ      sÅ  ððð ð ð #Ð "Ð "Ð "Ð "Ð "à Ð Ð Ð Ø Ð Ð Ð Ø Ð Ð Ð Ð Ð Ø Ð Ð Ð Ð Ð Ð Ð à Ð Ð Ð Ø  Ð  Ð  Ð  Ð  Ð  Ø !Ð !Ð !Ð !Ð !Ð !Ø "Ð "Ð "Ð "Ð "Ð "à =Ð =Ð =Ð =Ð =Ð =Ø ;Ð ;Ð ;Ð ;Ð ;Ð ;Ø @Ð @Ð @Ð @Ð @Ð @Ø .Ð .Ð .Ð .Ð .Ð .à	ˆ‰Œ€à3Ð3Ð3€Ð 3Ð 3Ð 3Ñ 3Ø2Ð2Ð2€Ð 2Ð 2Ð 2Ñ 2Ø'=Ð Ð =Ð =Ð =Ñ =ð&ð &ð &ð &ðR'ð 'ð 'ð 'ðZ ðDð Dð Dð Dð DðD Ø!Ø&(Ø9=ð]ð ]ð ]ð ]ð ]ð ]ð ]r    