o ߥimN@sddlZddlZddlZddlmZddlZddlmZddlmZ ddl m Z e dgdZ e dgd Zd ee je j_d eejej_eed rSejZn Gd d d ejZGdddejjZGdddejZddZddZddZddZddZd ddZ Gdddej!Z"Gd d!d!ej!Z#d d"d#Z$Gd$d%d%ej%Z&Gd&d'd'ej%Z'Gd(d)d)e(Z)d*d+Z*    , , - .d6d/d0Z+d1d2Z,  . 3 .d7d4d5Z-dS)8N)partial)nn) functional) model_zoo GlobalParams width_coefficientdepth_coefficient image_size dropout_rate num_classesbatch_norm_momentumbatch_norm_epsilondrop_connect_rate depth_divisor min_depth include_top BlockArgs num_repeat kernel_sizestride expand_ratio input_filtersoutput_filtersse_ratioid_skipNSiLUc@eZdZddZdS)SwishcCs|t|Sr)torchsigmoidselfxr&e/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/cv/face_emotion/efficient/utils.pyforward"sz Swish.forwardN__name__ __module__ __qualname__r(r&r&r&r'r  r c@s$eZdZeddZeddZdS)SwishImplementationcCs|t|}|||Sr)r!r"save_for_backward)ctxiresultr&r&r'r((s zSwishImplementation.forwardcCs,|jd}t|}||d|d|S)Nr) saved_tensorsr!r")r0 grad_outputr1 sigmoid_ir&r&r'backward.s  zSwishImplementation.backwardN)r*r+r, staticmethodr(r7r&r&r&r'r.&s  r.c@r)MemoryEfficientSwishcCs t|Sr)r.applyr#r&r&r'r(7s zMemoryEfficientSwish.forwardNr)r&r&r&r'r95r-r9cCsd|j}|s|S|j}|j}||9}|p|}t|t||d||}|d|kr.||7}t|S)a[Calculate and round number of filters based on width multiplier. Use width_coefficient, depth_divisor and min_depth of global_params. Args: filters (int): Filters number to be calculated. global_params (namedtuple): Global params of the model. Returns: new_filters: New filters number after calculating. g?)rrrmaxint)filters global_params multiplierdivisorr new_filtersr&r&r' round_filters;s  rCcCs |j}|s|Stt||S)a>Calculate module's repeat number of a block based on depth multiplier. Use depth_coefficient of global_params. Args: repeats (int): num_repeat to be calculated. global_params (namedtuple): Global params of the model. Returns: new repeat: New repeat number after calculating. )r r=mathceil)repeatsr?r@r&r&r' round_repeatsSs rGcCs|d|kr dksJdJd|s|S|jd}d|}|}|tj|dddg|j|jd7}t|}|||}|S)zDrop connect. Args: input (tensor: BCWH): Input of this structure. p (float: 0.0~1.0): Probability of drop connection. training (bool): The running mode. Returns: output: Output after drop connection. rr3zp must be in range of [0,1])dtypedevice)shaper!randrHrIfloor)inputsptraining batch_size keep_prob random_tensor binary_tensoroutputr&r&r' drop_connectbs$    rUcCs0t|tr ||fSt|tst|tr|St)zObtain height and width from x. Args: x (int, tuple or list): Data size. Returns: size: A tuple or list (H,W). ) isinstancer=listtuple TypeError)r%r&r&r'get_width_and_height_from_size}s rZcCsZ|durdSt|\}}t|tr|n|d}tt||}tt||}||gS)agCalculates the output image size when using Conv2dSamePadding with a stride. Necessary for static padding. Thanks to mannatsingh for pointing this out. Args: input_image_size (int, tuple or list): Size of input image. stride (int, tuple or list): Conv2d operation's stride. Returns: output_image_size: A list [H,W]. Nr)rZrVr=rDrE)input_image_sizer image_height image_widthr&r&r'calculate_output_image_sizes r^cC|durtStt|dS)a*Chooses static padding if you have specified an image size, and dynamic padding otherwise. Static padding is necessary for ONNX exporting of models. Args: image_size (int or tuple): Size of the image. Returns: Conv2dDynamicSamePadding or Conv2dStaticSamePadding. Nr )Conv2dDynamicSamePaddingrConv2dStaticSamePaddingr`r&r&r'get_same_padding_conv2d rccs2eZdZdZ    dfdd ZddZZS) raz2D Convolutions like TensorFlow, for a dynamic image size. The padding is operated in forward function by calculating dynamically. r3Tc sJt||||d|||t|jdkr|j|_dS|jdgd|_dS)Nrr;)super__init__lenr)r$ in_channels out_channelsrrdilationgroupsbias __class__r&r'rfsz!Conv2dDynamicSamePadding.__init__c Cs|dd\}}|jdd\}}|j\}}t||t||}} |d|jd} t| |d|jdd|d} | d|jd} t| |d|jdd|d} | dksg| dkr}t|| d| | d| d| | dg}t ||j|j |j|j |j|j SNr3rr;) sizeweightrrDrEr<rjFpadconv2drlpaddingrk)r$r%ihiwkhkwshswohowa1pad_ha2pad_wr&r&r'r(s $$ z Conv2dDynamicSamePadding.forward)r3r3r3Tr*r+r,__doc__rfr( __classcell__r&r&rmr'ras racs.eZdZdZ  dfdd ZddZZS) rbz2D Convolutions like TensorFlow's 'SAME' mode, with the given input image size. The padding module is calculated in construction function, then used in forward. r3Nc s^tj||||fi|t|jdkr|jn|jdgd|_|dus&Jt|tr/||fn|\}}|jdd\} } |j\} } t || t || } }| d|jd}t || d|j dd|d}|d|jd}t || d|j dd|d}|dks|dkrt |d||d|d||df|_dSt |_dS)Nr;rrpr3)rerfrgrrVr=rrrqrDrEr<rjr ZeroPad2dstatic_paddingIdentity)r$rhrirrr kwargsrwrxryrzr{r|r}r~b1rb2rrmr&r'rfs>   $$  z Conv2dStaticSamePadding.__init__c C0||}t||j|j|j|j|j|j}|Sr) rrsrurrrlrrvrjrkr#r&r&r'r(s zConv2dStaticSamePadding.forward)r3Nrr&r&rmr'rbs rbcCr_)a0Chooses static padding if you have specified an image size, and dynamic padding otherwise. Static padding is necessary for ONNX exporting of models. Args: image_size (int or tuple): Size of the image. Returns: MaxPool2dDynamicSamePadding or MaxPool2dStaticSamePadding. Nr`)MaxPool2dDynamicSamePaddingrMaxPool2dStaticSamePaddingr`r&r&r'get_same_padding_maxPool2drdrcs2eZdZdZ    d fdd ZddZZS) rz2D MaxPooling like TensorFlow's 'SAME' mode, with a dynamic image size. The padding is operated in forward function by calculating dynamically. rr3Fcst||||||t|jtr|jgdn|j|_t|jtr'|jgdn|j|_t|jtr:|jgd|_dS|j|_dS)Nr;)rerfrVrr=rrj)r$rrrvrjreturn_indices ceil_modermr&r'rf s$z$MaxPool2dDynamicSamePadding.__init__c Cs|dd\}}|j\}}|j\}}t||t||}} |d|jd} t| |d|jdd|d} | d|jd} t| |d|jdd|d} | dksa| dkrwt|| d| | d| d| | dg}t ||j|j|j |j|j |j Sro) rqrrrDrEr<rjrsrt max_pool2drvrr)r$r%rwrxryrzr{r|r}r~c1rc2rr&r&r'r(s  $$  z#MaxPool2dDynamicSamePadding.forward)rr3FFrr&r&rmr'rsrcs*eZdZdZdfdd ZddZZS)rz2D MaxPooling like TensorFlow's 'SAME' mode, with the given input image size. The padding mudule is calculated in construction function, then used in forward. Nc stj||fi|t|jtr|jgdn|j|_t|jtr'|jgdn|j|_t|jtr7|jgdn|j|_|dusAJt|trJ||fn|\}}|j\}}|j\} } t|| t|| } } | d|jd} t | |d|jdd|d}| d|jd}t ||d|jdd|d}|dks|dkrt |d||d|d||df|_ dSt |_ dS)Nr;r3r)rerfrVrr=rrjrDrEr<rrrr)r$rrr rrwrxryrzr{r|r}r~d1rd2rrmr&r'rf2sJ   $$  z#MaxPool2dStaticSamePadding.__init__c Crr) rrsrrrrvrjrrr#r&r&r'r(Ls  z"MaxPool2dStaticSamePadding.forwardrrr&r&rmr'r-src@s@eZdZdZeddZeddZeddZedd Zd S) BlockDecoderz\Block Decoder for readability, straight from the official TensorFlow repository. c Cst|tsJ|d}i}|D]}td|}t|dkr*|dd\}}|||<qd|vr7t|ddksMt|ddkrK|dd|ddksMJtt|dt|d t|ddgt|d t|d t|d d |vryt|d ndd|vdS)a2Get a block through a string notation of arguments. Args: block_string (str): A string notation of arguments. Examples: 'r1_k3_s11_e1_i32_o16_se0.25_noskip'. Returns: BlockArgs: The namedtuple defined at the top of this file. _z(\d.*)r;Nsr3rrker1osenoskipr)rVstrsplitrergrr=float) block_stringopsoptionsopsplitskeyvaluer&r&r'_decode_block_stringXs,        z!BlockDecoder._decode_block_stringcCsd|jd|jd|jd|jdfd|jd|jd|jg}d|jkr,dkr6nn|d |j|jd ur@|d d |S) zEncode a block to a string. Args: block (namedtuple): A BlockArgs type argument. Returns: block_string: A String form of BlockArgs. zr%dzk%dzs%d%drr3ze%szi%dzo%dzse%sFrr) rrstridesrrrrappendrjoin)blockargsr&r&r'_encode_block_stringzs    z!BlockDecoder._encode_block_stringcCs0t|tsJg}|D] }|t|q |S)aDecode a list of string notations to specify blocks inside the network. Args: string_list (list[str]): A list of strings, each string is a notation of block. Returns: blocks_args: A list of BlockArgs namedtuples of block args. )rVrWrrr) string_list blocks_argsrr&r&r'decodes zBlockDecoder.decodecCs"g}|D] }|t|q|S)aEncode a list of BlockArgs to a list of strings. Args: blocks_args (list[namedtuples]): A list of BlockArgs namedtuples of block args. Returns: block_strings: A list of strings, each string is a notation of block. )rrr)r block_stringsrr&r&r'encodeszBlockDecoder.encodeN) r*r+r,rr8rrrrr&r&r&r'rSs !   rc Cs"ddddddddd d d }||S) zMap EfficientNet model name to parameter coefficients. Args: model_name (str): Model name to be queried. Returns: params_dict[model_name]: A (width,depth,res,dropout) tuple. )?rp皙?)r皙?r)r333333?i333333?)rffffff?i,r)r?i|皙?)g?皙@ir)rg@i?)g@g@iXr)rg @ir)g333333@g333333@i r) zefficientnet-b0zefficientnet-b1zefficientnet-b2zefficientnet-b3zefficientnet-b4zefficientnet-b5zefficientnet-b6zefficientnet-b7zefficientnet-b8zefficientnet-l2r&) model_name params_dictr&r&r'efficientnet_paramss rrTc Cs8gd}t|}t|||||dd|dd|d }||fS)aUCreate BlockArgs and GlobalParams for efficientnet model. Args: width_coefficient (float) depth_coefficient (float) image_size (int) dropout_rate (float) drop_connect_rate (float) num_classes (int) Meaning as the name suggests. Returns: blocks_args, global_params. )zr1_k3_s11_e1_i32_o16_se0.25zr2_k3_s22_e6_i16_o24_se0.25zr2_k5_s22_e6_i24_o40_se0.25zr3_k3_s22_e6_i40_o80_se0.25zr3_k5_s11_e6_i80_o112_se0.25zr4_k5_s22_e6_i112_o192_se0.25zr1_k3_s11_e6_i192_o320_se0.25gGz?gMbP?Nr)rrr) rr r r rr rrr?r&r&r' efficientnets   rcCsZ|drt|\}}}}t||||d\}}ntd||r)|jdi|}||fS)zGet the block args and global params for a given model name. Args: model_name (str): Model's name. override_params (dict): A dict to modify global_params. Returns: blocks_args, global_params r)rr r r z!model name is not pre-defined: {}Nr&) startswithrrNotImplementedErrorformat_replace)roverride_paramswdrrNrr?r&r&r'get_model_paramss  rFc Cst|tr t|}n |rtnt}t||}|r-|j|dd}|j r,Jd |j n$| d| d|j|dd}t |j t ddgksQJd |j |j r\Jd |j |rgtd |dSdS)aZLoads pretrained weights from weights path or download using url. Args: model (Module): The whole model of efficientnet. model_name (str): Model name of efficientnet. weights_path (None or str): str: path to pretrained weights file on the local disk. None: use pretrained weights downloaded from the Internet. load_fc (bool): Whether to load pretrained weights for fc layer at the end of the model. advprop (bool): Whether to load pretrained weights trained with advprop (valid when weights_path is None). F)strictz0Missing keys when loading pretrained weights: {}z _fc.weightz_fc.biasz Loaded pretrained weights for {}N)rVrr!loadurl_map_advpropurl_maprload_urlload_state_dict missing_keysrpopsetunexpected_keysprint) modelr weights_pathload_fcadvpropverbose state_dicturl_map_retr&r&r'load_pretrained_weightss2         r)NNNrrrT)NTFT). collectionsrDr functoolsrr!rtorch.nnrrs torch.utilsr namedtuplerrrg_fields__new__ __defaults__hasattrrr ModuleautogradFunctionr.r9rCrGrUrZr^rcConv2drarbr MaxPool2drrobjectrrrrrr&r&r&r'sX      # )&&Y /