o yih4 @sddlmZddlmZmZmZmZddlZddl Z ddl m Z ddl m Z ddl mZddlmZddlmZdd lmZmZerIdd lmZn Gd d d eZd dgZer[ddlZGdddeZGddde ZejZdde de de de dede f ddZGdd d eZ dS))deepcopy)AnyListOptionalUnionN)Tensor)Function)Module)Metric)rank_zero_info)_SCIPY_AVAILABLE_TORCH_FIDELITY_AVAILABLE)FeatureExtractorInceptionV3c@s eZdZdS)rN)__name__ __module__ __qualname__rrJ/home/ubuntu/.local/lib/python3.10/site-packages/torchmetrics/image/fid.pyrsrFrechetInceptionDistanceFIDc sfeZdZ ddedeedeeddffdd Zdeddffd d Zd e de ffd d Z Z S)NoTrainInceptionV3Nname features_listfeature_extractor_weights_pathreturncst||||dSN)super__init__eval)selfrrr __class__rrr*s zNoTrainInceptionV3.__init__modecs tdS)zRthe inception network should not be able to be switched away from evaluation mode.F)rtrain)rr"r rrr#4s zNoTrainInceptionV3.trainxcs"t|}|d|jddS)Nr)rforwardreshapeshape)rr$outr rrr&8s zNoTrainInceptionV3.forwardr) rrrstrrrrboolr#rr& __classcell__rrr rr)s rc@sDeZdZdZedededefddZedededefdd Zd S) MatrixSquareRootzoSquare root of a positive definite matrix. All credit to `Square Root of a Positive Definite Matrix`_ ctx input_datarcCsL|tj}tjj|dd\}}t |j  |}| ||S)NF)disp)detachcpunumpyastypenpfloat_scipylinalgsqrtmtorch from_numpyrealtosave_for_backward)r.r/m scipy_res_r9rrrr&Cs  zMatrixSquareRoot.forward grad_outputcCsfd}|jdr1|j\}|jtj}|jtj}tj |||}t | |}|S)Nr)needs_input_grad saved_tensorsdatar2r3r4r5r6r7r8solve_sylvesterr:r;r=)r.rB grad_inputr9gm grad_sqrtmrrrbackwardMs zMatrixSquareRoot.backwardN) rrr__doc__ staticmethodrrr&rJrrrrr-=s  r-ư>mu1sigma1mu2sigma2epsrc Cs||}t||}t|s4td|dtj|d|j|j d|}t||||}t |}| |t |t |d|S)aAdjusted version of `Fid Score`_ The Frechet Inception Distance between two multivariate Gaussians X_x ~ N(mu_1, sigm_1) and X_y ~ N(mu_2, sigm_2) is d^2 = ||mu_1 - mu_2||^2 + Tr(sigm_1 + sigm_2 - 2*sqrt(sigm_1*sigm_2)). Args: mu1: mean of activations calculated on predicted (x) samples sigma1: covariance matrix over activations calculated on predicted (x) samples mu2: mean of activations calculated on target (y) samples sigma2: covariance matrix over activations calculated on target (y) samples eps: offset constant - used if sigma_1 @ sigma_2 matrix is singular Returns: Scalar value of the distance between sets. z2FID calculation produces singular product; adding z$ to diagonal of covariance estimatesr)devicedtype) r9mmr:isfiniteallr eyesizerSrTtracedot) rNrOrPrQrRdiffcovmeanoffset tr_covmeanrrr _compute_fidbs &rac seZdZUdZdZeed<dZeed<dZeed<e ed<e ed<e ed<e ed <e ed <e ed < dde e e fdedede ddf fdd Zde deddfddZde fddZdfdd ZZS) ru[Calculates Fréchet inception distance (FID_) which is used to access the quality of generated images. Given by. .. math:: FID = |\mu - \mu_w| + tr(\Sigma + \Sigma_w - 2(\Sigma \Sigma_w)^{\frac{1}{2}}) where :math:`\mathcal{N}(\mu, \Sigma)` is the multivariate normal distribution estimated from Inception v3 (`fid ref1`_) features calculated on real life images and :math:`\mathcal{N}(\mu_w, \Sigma_w)` is the multivariate normal distribution estimated from Inception v3 features calculated on generated (fake) images. The metric was originally proposed in `fid ref1`_. Using the default feature extraction (Inception v3 using the original weights from `fid ref2`_), the input is expected to be mini-batches of 3-channel RGB images of shape ``(3 x H x W)``. If argument ``normalize`` is ``True`` images are expected to be dtype ``float`` and have values in the ``[0, 1]`` range, else if ``normalize`` is set to ``False`` images are expected to have dtype ``uint8`` and take values in the ``[0, 255]`` range. All images will be resized to 299 x 299 which is the size of the original training data. The boolian flag ``real`` determines if the images should update the statistics of the real distribution or the fake distribution. .. note:: using this metrics requires you to have ``scipy`` install. Either install as ``pip install torchmetrics[image]`` or ``pip install scipy`` .. note:: using this metric with the default feature extractor requires that ``torch-fidelity`` is installed. Either install as ``pip install torchmetrics[image]`` or ``pip install torch-fidelity`` As input to ``forward`` and ``update`` the metric accepts the following input - ``imgs`` (:class:`~torch.Tensor`): tensor with images feed to the feature extractor with - ``real`` (:class:`~bool`): bool indicating if ``imgs`` belong to the real or the fake distribution As output of `forward` and `compute` the metric returns the following output - ``fid`` (:class:`~torch.Tensor`): float scalar tensor with mean FID value over samples Args: feature: Either an integer or ``nn.Module``: - an integer will indicate the inceptionv3 feature layer to choose. Can be one of the following: 64, 192, 768, 2048 - an ``nn.Module`` for using a custom feature extractor. Expects that its forward method returns an ``(N,d)`` matrix where ``N`` is the batch size and ``d`` is the feature size. reset_real_features: Whether to also reset the real features. Since in many cases the real dataset does not change, the features can cached them to avoid recomputing them which is costly. Set this to ``False`` if your dataset does not change. kwargs: Additional keyword arguments, see :ref:`Metric kwargs` for more info. Raises: ValueError: If ``feature`` is set to an ``int`` (default settings) and ``torch-fidelity`` is not installed ValueError: If ``feature`` is set to an ``int`` not in [64, 192, 768, 2048] TypeError: If ``feature`` is not an ``str``, ``int`` or ``torch.nn.Module`` ValueError: If ``reset_real_features`` is not an ``bool`` Example: >>> import torch >>> _ = torch.manual_seed(123) >>> from torchmetrics.image.fid import FrechetInceptionDistance >>> fid = FrechetInceptionDistance(feature=64) >>> # generate two slightly overlapping image intensity distributions >>> imgs_dist1 = torch.randint(0, 200, (100, 3, 299, 299), dtype=torch.uint8) >>> imgs_dist2 = torch.randint(100, 255, (100, 3, 299, 299), dtype=torch.uint8) >>> fid.update(imgs_dist1, real=True) >>> fid.update(imgs_dist2, real=False) >>> fid.compute() tensor(12.7202) Fhigher_is_betteris_differentiablefull_state_updatereal_features_sumreal_features_cov_sumreal_features_num_samplesfake_features_sumfake_features_cov_sumfake_features_num_samplesTfeaturereset_real_features normalizekwargsrNc  s~tjdi|t|tr4|}tstdgd}||vr)td|d|dtdt|gd|_ nt|t rO||_ t j dd d t j d }| |jd }ntd t|ts\td||_t|tshtd||_||f}|jdt |dd|jdt |dd|jdt ddd|jdt |dd|jdt |dd|jdt ddddS)NzFrechetInceptionDistance metric requires that `Torch-fidelity` is installed. Either install as `pip install torchmetrics[image]` or `pip install torch-fidelity`.)@irkz3Integer input to argument `feature` must be one of z , but got .zinception-v3-compat)rrr)+rv)rTr%z'Got unknown input to argument `feature`z4Argument `reset_real_features` expected to be a boolz*Argument `normalize` expected to be a boolresum)dist_reduce_fxrfrgrhrirjr)rr isinstanceintr ModuleNotFoundError ValueErrorrr* inceptionr r:randintuint8r( TypeErrorr+rmrn add_statezerosdoubletensorlong) rrlrmrnro num_featuresvalid_int_input dummy_image mx_nb_feetsr rrrs>    z!FrechetInceptionDistance.__init__imgsr<cCs|jr |dn|}||}|j|_|}|dkr#|d}|rH|j|j dd7_|j | |7_ |j |jd7_ dS|j|j dd7_|j| |7_|j|jd7_dS)z)Update the state with extracted features.rsrtr)dimN)rnbyter}rT orig_dtyperr unsqueezererwrftrVrgr(rhrirj)rrr<featuresrrrupdates   zFrechetInceptionDistance.updatecCs|j|jd}|j|jd}|j|j||}||jd}|j|j||}||jd}t | d|| d| |j S)zWCalculate FID score based on accumulated extracted features from the two distributions.rrt) rergrrhrjrfrrVrirasqueezer=r)r mean_real mean_fake cov_real_numcov_real cov_fake_numcov_fakerrrcomputes"z FrechetInceptionDistance.computecsR|js"t|j}t|j}t|j}t||_||_||_dStdSr)rmrrerfrgrreset)rrerfrgr rrr"s     zFrechetInceptionDistance.reset)rkTF)rN)rrrrKrbr+__annotations__rcrdrrrzr rrrrrr,rrr rrs8  I   0 )rM)!copyrtypingrrrrr3r5r:rtorch.autogradrtorch.nnr torchmetrics.metricr torchmetrics.utilitiesr torchmetrics.utilities.importsr r ,torch_fidelity.feature_extractor_inceptionv3r__doctest_skip__r7rr-applyr9floatrarrrrrs*      "$