o ۷i_1@sddlZddlZddlmZddlmmZddlmZee Z zddl m Z Wne y8ddlmZ YnwdZGdddejZdd Zd d ZGd d d ejjjZGdddejjjZGdddejjjZGdddejjjZdS)N) init_logger)load_state_dict_from_url)load_urlzjhttps://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pthcsHeZdZdZdZdddddZefddddffd d Zd d ZZS) InceptionV3z5Pretrained InceptionV3 network returning feature mapsr)@TFc sPtt|||_||_t||_t||_|jdks Jdt |_ |r+t }nt dd}|j|j|jt jdddg}|j t j||jdkra|j|jt jdddg}|j t j||jdkr|j|j|j|j|j|j|j|jg} |j t j| |jdkr|j|j|j t j!dd g} |j t j| |"D]} || _#qd S) aMBuild pretrained InceptionV3 Parameters ---------- output_blocks : list of int Indices of blocks to return features of. Possible values are: - 0: corresponds to output of first max pooling - 1: corresponds to output of second max pooling - 2: corresponds to output which is fed to aux classifier - 3: corresponds to output of final average pooling resize_input : bool If true, bilinearly resizes input to width and height 299 before feeding input to model. As the network without fully connected layers is fully convolutional, it should be able to handle inputs of arbitrary size, so resizing might not be strictly needed normalize_input : bool If true, scales the input from range (0, 1) to the range the pretrained Inception network expects, namely (-1, 1) requires_grad : bool If true, parameters of the model require gradients. Possibly useful for finetuning the network use_fid_inception : bool If true, uses the pretrained Inception model used in Tensorflow's FID implementation. If false, uses the pretrained Inception model available in torchvision. The FID Inception model has different weights and a slightly different structure from torchvision's Inception model. If you want to compute FID scores, you are strongly advised to set this parameter to true to get comparable results. rz%Last possible output block index is 3DEFAULT)weightsr) kernel_sizestrider)rr) output_sizeN)$superr__init__ resize_inputnormalize_inputsorted output_blocksmaxlast_needed_blocknn ModuleListblocksfid_inception_v3 _inception_v3 Conv2d_1a_3x3 Conv2d_2a_3x3 Conv2d_2b_3x3 MaxPool2dappend Sequential Conv2d_3b_1x1 Conv2d_4a_3x3Mixed_5bMixed_5cMixed_5dMixed_6aMixed_6bMixed_6cMixed_6dMixed_6eMixed_7aMixed_7bMixed_7cAdaptiveAvgPool2d parameters requires_grad) selfrrrr4use_fid_inception inceptionblock0block1block2block3param __class__Q/home/ubuntu/vllm_env/lib/python3.10/site-packages/cache_dit/metrics/inception.pyr$sV&           zInceptionV3.__init__cCsvg}|}|jrtj|dddd}|jrd|d}t|jD]\}}||}||jvr0||||jkr8|Sq|S)aiGet Inception feature maps Parameters ---------- inp : torch.autograd.Variable Input tensor of shape Bx3xHxW. Values are expected to be in range (0, 1) Returns ------- List of torch.autograd.Variable, corresponding to the selected output block, sorted ascending by index )+rAbilinearF)sizemode align_cornersrr) rF interpolater enumeraterrr#r)r5inpoutpxidxblockr?r?r@forwards    zInceptionV3.forward) __name__ __module__ __qualname____doc__DEFAULT_BLOCK_INDEXBLOCK_INDEX_BY_DIMrrN __classcell__r?r?r=r@rs crcOszttttjddd}Wn tyd}Ynw|dkr%d|d<|dkrQd |vrQ|d d kr8d |d <n|d durCd|d <n td |d tj|d =tjj |i|S)z'Wraps `torchvision.models.inception_v3`.Nr)r)rF init_weights)r rr T pretrainedz+weights=={} not supported in torchvision {}) tuplemapint torchvision __version__split ValueErrorformatmodels inception_v3)argskwargsversionr?r?r@rs&"      rcCstdddd}tddd|_tdd d|_td d d|_td d d |_td dd |_td dd |_td dd |_ t d|_ t d|_ ttdd}|||S)aaBuild pretrained Inception model for FID computation The Inception model for FID computation uses a different set of weights and has a slightly different structure than torchvision's Inception. This method first constructs torchvision's Inception and then patches the necessary parts that are different in the FID Inception model. iFN) num_classes aux_logitsrr ) pool_featuresr i r ) channels_7x7ir )progress)r FIDInceptionAr'r(r) FIDInceptionCr+r,r-r.FIDInceptionE_1r0FIDInceptionE_2r1rFID_WEIGHTS_URLload_state_dict)r7 state_dictr?r?r@rs     rc(eZdZdZfddZddZZS)rqz,InceptionA block patched for FID computationctt|||dSN)rrqr)r5 in_channelsrkr=r?r@rzFIDInceptionA.__init__cCsr||}||}||}||}||}||}tj|ddddd}||}||||g}t |dSNrrFrrpaddingcount_include_pad) branch1x1 branch5x5_1 branch5x5_2branch3x3dbl_1branch3x3dbl_2branch3x3dbl_3rF avg_pool2d branch_pooltorchcat)r5rKr branch5x5 branch3x3dblroutputsr?r?r@rNs         zFIDInceptionA.forwardrOrPrQrRrrNrUr?r?r=r@rq rqcrx)rrz,InceptionC block patched for FID computationcryrz)rrrr)r5r{rnr=r?r@rr|zFIDInceptionC.__init__cCs||}||}||}||}||}||}||}||}||}t j |ddddd}| |}||||g}t |dSr})r branch7x7_1 branch7x7_2 branch7x7_3branch7x7dbl_1branch7x7dbl_2branch7x7dbl_3branch7x7dbl_4branch7x7dbl_5rFrrrr)r5rKr branch7x7 branch7x7dblrrr?r?r@rNs            zFIDInceptionC.forwardrr?r?r=r@rrrrrcrx)rsz2First InceptionE block patched for FID computationctt||dSrz)rrsrr5r{r=r?r@rzFIDInceptionE_1.__init__cCs||}||}||||g}t|d}||}||}||| |g}t|d}t j |ddddd}| |}||||g}t|dS)NrrFr~) r branch3x3_1 branch3x3_2a branch3x3_2brrrrbranch3x3dbl_3abranch3x3dbl_3brFrrr5rKr branch3x3rrrr?r?r@rNs         zFIDInceptionE_1.forwardrr?r?r=r@rsrrscrx)rtz3Second InceptionE block patched for FID computationcrrz)rrtrrr=r?r@r:rzFIDInceptionE_2.__init__cCs||}||}||||g}t|d}||}||}||| |g}t|d}t j |dddd}| |}||||g}t|dS)Nrr)rrr) rrrrrrrrrrrF max_pool2drrr?r?r@rN=s         zFIDInceptionE_2.forwardrr?r?r=r@rt7rrt)rr^torch.nnrtorch.nn.functional functionalrFcache_dit.loggerrrOloggertorchvision.models.utilsr ImportErrortorch.utils.model_zoorruModulerrrrcr7 InceptionArq InceptionCrr InceptionErsrtr?r?r?r@s(   !