i ^dZddlZddlmZmZddej deeefdeeefdej fdZ dd ej d eed fdeeefdeeefdej f d Z dd e dede fdZ dedefdZ de deed fdefdZ GddZddej dedefdZy)zQ ConvMLP Utilities ================= Vedic-inspired optimized utility functions. N)TupleUnionx kernel_sizestridesreturnc |j\}}}}|\}}|\} } ||z | zdz} ||z | zdz} tj||||| | f|j} t | D]D}t | D]4}|| z}|| z}|dddd|||z|||zf| dddddddd||f<6F| j ddddddj || | z||z|z} | j dddS) a Image to Column conversion (Vedic Block Multiplication). Inspired by Urdhva-Tiryagbyham (vertically and crosswise). Converts 2D convolution to efficient matrix multiplication. Parameters ---------- x : np.ndarray Input tensor of shape (batch, channels, height, width) kernel_size : tuple (kernel_height, kernel_width) strides : tuple (stride_h, stride_w) Returns ------- col : np.ndarray Unfolded patches of shape (batch, channels * kh * kw, out_h * out_w) Example ------- >>> patches = im2col(x, kernel_size=(3, 3), strides=(1, 1)) >>> # patches[:, :, 0] = first patch flattened dtypeNr)shapenpzerosr range transposereshape)rrrbatchchannelshwkhkwshswout_hout_wcolijh_startw_starts ;/home/per/Documents/sklearn ConvMLP module/convmlp/utils.pyim2colr' s.4GGE8Q FB FB VNQ E VNQ E ((E8RUE:!'' JC 5\Tu TA"fG"fG$%aGGBJ,>PR @R&R$SC1aAq ! TT --1aAq ) 1 1%SU XZHZ [C ==Aq !!r!x_shape.c |\}}}}|\}} |\} } ||z | zdz} || z | zdz} |jdddj|| | ||| }tj||j}t | D]C}t | D]3}|| z}|| z}|dddd|||z||| zfxx|dd||fz cc<5E|S)a Column to Image conversion (Reverse of im2col). Parameters ---------- col : np.ndarray Column data of shape (batch, channels * kh * kw, out_h * out_w) x_shape : tuple Original input shape (batch, channels, height, width) kernel_size : tuple (kernel_height, kernel_width) strides : tuple (stride_h, stride_w) Returns ------- x : np.ndarray Reconstructed image of shape x_shape r rrr N)rrrrr r)r!r)rrrrrrrrrrrr rr"r#r$r%s r&col2imr+;s*$E8Q FB FB VNQ E VNQ E --1a ( (uhB OC  *A5\Lu LA"fG"fG aGGBJ& (:: ;s1a7| K ; LL Hr(layersstart_rfc x|g}|}|D])}t|drt|dr|j\}}t|jtr |jn|j|jf\}}||dz t j |D cgc] } t| tr | jn| "c} zz}|j|t|dst|jtr |jn|j|jf\} } || z}|j|,|Scc} w)a Calculate receptive field at each layer. Based on the "coconut sellers" concept - each layer adds to the receptive field. Parameters ---------- layers : list List of layer objects with kernel_size and strides attributes start_rf : int Initial receptive field size Returns ------- receptive_fields : list Receptive field size at each layer rrr pool_size) hasattrr isinstancertuplerprodappendr/) r,r-receptive_fields current_rflayerrrrrlphpws r&calculate_receptive_fieldr;gs,$!zJ 0 5- (WUI-F&&FB&0&FU]]U]]\a\i\iLjFB#rAvCS:U>?HRRSUZG[!))ab:b:U2V'VVJ  # #J / UK ((25??E(JU__QVQ`Q`bgbqbqPrFB#bJ  # #J / 0 :Us %D7ncjdttjtj|zS)z Return the next power of 2 >= n.r)intrceillog2)r<s r&next_power_of_2rAs# BGGBGGAJ'( ((r( conv_layers input_shapecd}|}|D]}t|tr[|j\}}|j|dd\}}||z|z|z|dz|jz} || z }|j|}nt|t t fst|jtr |jn|j|jf\} } |d|d|d| z|d| zf}|S)z Calculate total FLOPs for a convolutional network. Inspired by Bhaskara's wheel method - counting all operations. rNr rr) r1Conv2Dr_compute_output_shapefilters MaxPool2D AvgPool2Dr/r2) rBrC total_flops current_shaper7rrrr flopsr9r:s r&calculate_flopsrOs  KM K eV $&&FB 66}EbcJLE5EMB&+mB.??%--OE 5 K!77 FM  95 6(25??E(JU__QVQ`Q`bgbqbqPrFB*1-}Q/?(+r1=3Cr3IKM K r(ceZdZdZd defdZdejfdZdejdejfdZ dejdejfdZ y ) NikhilamQuantizerz Nikhilam Quantization for fast inference. Based on the Vedic "deficiency from base" method. Quantizes weights to low-precision for faster computation. num_bitsc.||_d|_d|_yN)rRscale zero_point)selfrRs r&__init__zNikhilamQuantizer.__init__s   r(weightsc|j|jz d|jzdz z |_|j|_|S)z"Calculate quantization parameters.rr )maxminrRrUrVrWrYs r&fitzNikhilamQuantizer.fits?kkmgkkm3T]]8JQ8NO !++- r(rc|j|j|tj||jz |jz j tj S)z"Quantize weights to low precision.)rUr^rroundrVastypeint8r]s r&quantizezNikhilamQuantizer.quantizesH ::  HHW xx4??2djj@AHHQQr(ct|jtj|jz|jzS)z#Restore weights from low precision.)rarfloat32rUrVr]s r& dequantizezNikhilamQuantizer.dequantizes'~~bjj)DJJ6HHr(N)) __name__ __module__ __qualname____doc__r>rXrndarrayr^rcrfr(r&rQrQs_ 2:: R RrzzR I"**IIr(rQexpected_channelsc,t|jdk(ryt|jdk(ryt|jdk(r3|r0|jd|k7rtd|d|jdytd |j) zValidate input tensor shape.rFrr r z Expected z channels, got TzInvalid input shape: )lenr ValueError)rrns r&validate_input_shaperrs 177|q QWW  QWW  /@!@y):(;?177ST:,WX X0 :;;r())r r )r rT)rknumpyrtypingrrrlr>r'r+listr;rArOrQboolrrrmr(r&rws -"bjj-"uS#X-"sCx-"^`^h^h-"b'-)  ) U38_) 5c?) #s(O) 13) X"d"c"$"J)s)s) E#s(O6II<