o .wiS @sddlmZddlmZddlmZmZmZddlZddlm Z ddl m Z ddl m Z ddlmZdd lmZmZdd lmZmZesGd gZer\dd lmZdd lmZddlmZnGddde ZdZdZdd gZGdddeZde de de de de f ddZGdddeZ dS))Sequence)deepcopy)AnyOptionalUnionN)Tensor)Module)adaptive_avg_pool2d)Metric)_MATPLOTLIB_AVAILABLE_TORCH_FIDELITY_AVAILABLE)_AX_TYPE_PLOT_OUT_TYPEFrechetInceptionDistance.plot)FeatureExtractorInceptionV3)vassert))interpolate_bilinear_2d_like_tensorflow1xc@s eZdZdS)_FeatureExtractorInceptionV3N)__name__ __module__ __qualname__rrS/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/torchmetrics/image/fid.pyr$srFrechetInceptionDistancec seZdZdZ ddedeedeeddffdd Zd eddffd d Z d e de e d ffddZ d e de fddZ ZS)NoTrainInceptionV3z)Module that never leaves evaluation mode.Nname features_listfeature_extractor_weights_pathreturncs(tstdt||||dS)NzNoTrainInceptionV3 module requires that `Torch-fidelity` is installed. Either install as `pip install torchmetrics[image]` or `pip install torch-fidelity`.)r ModuleNotFoundErrorsuper__init__eval)selfrrr __class__rrr!0s  zNoTrainInceptionV3.__init__modecs tdS)z.Force network to always be in evaluation mode.F)r train)r#r&r$rrr'@s zNoTrainInceptionV3.trainx.cstt|o |jtjkdi|j}t|dr ||j n|tj }t ||j |j fdd}|dd}| |}||}||}||}d|vrtt|ddd d d<|dt|d krttfd d |jDS||}||}||}d |vrt|ddd d d <|d t|d krtfdd |jDS||}||}||}||}||}||}||}| |}d|vrt|ddd d d<|dt|d krtfdd |jDS|!|}|"|}|#|}|$|}t%|d}d|vr8|d<|dt|d kr8tfdd |jDSd|vrk|&|j'j(j)}|d<|dt|d kratfdd |jDS||j'j*+d }n|'|}|d<tfdd |jDS)aForward method of inception net. Copy of the forward method from this file: https://github.com/toshas/torch-fidelity/blob/master/torch_fidelity/feature_extractor_inceptionv3.py with a single line change regarding the casting of `x` in the beginning. Corresponding license file (Apache License, Version 2.0): https://github.com/toshas/torch-fidelity/blob/master/LICENSE.md z6Expecting image as torch.Tensor with dtype=torch.uint8_dtypeF)size align_corners64)r.) output_sizerc3|]}|VqdSNr.0afeaturesrr dz=NoTrainInceptionV3._torch_fidelity_forward..192c3r1r2rr3r6rrr8nr9768c3r1r2rr3r6rrr8}r9r.2048c3r1r2rr3r6rrr8r9logits_unbiasedc3r1r2rr3r6rrr8r9logitsc3r1r2rr3r6rrr8r9),rtorch is_tensordtypeuint8rcopyhasattrtor)floatrINPUT_IMAGE_SIZE Conv2d_1a_3x3 Conv2d_2a_3x3 Conv2d_2b_3x3 MaxPool_1r squeezeremovelentuple Conv2d_3b_1x1 Conv2d_4a_3x3 MaxPool_2Mixed_5bMixed_5cMixed_5dMixed_6aMixed_6bMixed_6cMixed_6dMixed_6eMixed_7aMixed_7bMixed_7cAvgPoolflattenmmfcweightTbias unsqueeze)r#r(remaining_featuresrr6r_torch_fidelity_forwardDsx  "                                 z*NoTrainInceptionV3._torch_fidelity_forwardcCs ||}|d|jddS)z8Forward pass of neural network with reshaping of output.rr0)rgreshapeshape)r#r(outrrrforwards zNoTrainInceptionV3.forwardr2)rrr__doc__strlistrr!boolr'rrOrgrk __classcell__rrr$rr-sVrmu1sigma1mu2sigma2rcCsR||jdd}||}tj||jjdd}||d|S)arCompute adjusted 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 Returns: Scalar value of the distance between sets. r0dim)squaresumtracer?linalgeigvalssqrtreal)rqrrrsrtr5bcrrr _compute_fidsrcsVeZdZUdZdZeed<dZeed<dZeed<dZ e ed<e ed<e ed <e ed <e ed <e ed <e ed <e ed<dZ eed<     d-deee fdededeeeefdeededdffdd Zde deddfdd Zde fd!d"Zd.fd#d$ Zd%eeejfdd&ffd'd( Z d/d)eee ee fd*eedefd+d,ZZS)0ru<Calculate Fréchet inception distance (FID_) which is used to assess the quality of generated images. .. math:: FID = \|\mu - \mu_w\|^2 + 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 ``(3xHxW)``. 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. Using custom feature extractor is also possible. One can give a torch.nn.Module as `feature` argument. This custom feature extractor is expected to have output shape of ``(1, num_features)``. This would change the used feature extractor from default (Inception v3) to the given network. In case network doesn't have ``num_features`` attribute, a random tensor will be given to the network to infer feature dimensionality. Size of this tensor can be controlled by ``input_img_size`` argument and type of the tensor can be controlled with ``normalize`` argument (``True`` uses float32 tensors and ``False`` uses int8 tensors). In this case, update method expects to have the tensor given to `imgs` argument to be in the correct shape and type that is compatible to the custom feature extractor. This metric is known to be unstable in its calculatations, and we recommend for the best results using this metric that you calculate using `torch.float64` (default is `torch.float32`) which can be set using the `.set_dtype` method of the metric. .. hint:: 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 be cached them to avoid recomputing them which is costly. Set this to ``False`` if your dataset does not change. normalize: Argument for controlling the input image dtype normalization: - If default feature extractor is used, controls whether input imgs have values in range [0, 1] or not: - True: if input imgs have values ranged in [0, 1]. They are cast to int8/byte tensors. - False: if input imgs have values ranged in [0, 255]. No casting is done. - If custom feature extractor module is used, controls type of the input img tensors: - True: if input imgs are expected to be in the data type of torch.float32. - False: if input imgs are expected to be in the data type of torch.int8. input_img_size: tuple of integers. Indicates input img size to the custom feature extractor network if provided. kwargs: Additional keyword arguments, see :ref:`Metric kwargs` for more info. Raises: ValueError: If torch version is lower than 1.9 ModuleNotFoundError: 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: >>> from torch import rand >>> 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(12.6388) Fhigher_is_betteris_differentiablefull_state_updategplot_lower_boundreal_features_sumreal_features_cov_sumreal_features_num_samplesfake_features_sumfake_features_cov_sumfake_features_num_samples inceptionfeature_networkT+rNfeaturereset_real_features normalizeinput_img_sizerkwargsrc  stjdi|t|tstd||_d|_t|trB|}ts%t dd}||vr6td|d|dt dt |g|d |_ n]t|t r||_ d |_t|j d rst|j jtr_|j j}n@t|j jtrot|j j}n0td |jrtjd g|Rdtji} ntjddd g|Rtjd} | | jd}ntdt|tstd||_||f} |jdt|dd|jdt| dd|jdtddd|jdt|dd|jdt| dd|jdtddddS)Nz*Argument `normalize` expected to be a boolFzFrechetInceptionDistance metric requires that `Torch-fidelity` is installed. Either install as `pip install torchmetrics[image]` or `pip install torch-fidelity`.)@irz3Integer input to argument `feature` must be one of z , but got .zinception-v3-compat)rrrT num_featureszCExpected `self.inception.num_features` to be of type int or Tensor.r.rAr)rAr0z'Got unknown input to argument `feature`z4Argument `reset_real_features` expected to be a boolrry)dist_reduce_fxrrrrrr)r r! isinstancero ValueErrorrused_custom_modelintr rrrmrrrDrritem TypeErrorr?randfloat32randintrBrir add_statezerosdoubletensorlong) r#rrrrrrrvalid_int_input dummy_image mx_num_featsr$rrr!)sZ      z!FrechetInceptionDistance.__init__imgsr~cCs|jr |js |dn|}||}|j|_|}|dkr&|d}|rK|j |j dd7_ |j | |7_ |j|jd7_dS|j|j dd7_|j| |7_|j|jd7_dS)aUpdate the state with extracted features. Args: imgs: Input img tensors to evaluate. If used custom feature extractor please make sure dtype and size is correct for the model. real: Whether given image is real or fake. rr.rruN)rrbyterrA orig_dtyperrvrerryrtr`rrirrr)r#rr~r7rrrupdatels   zFrechetInceptionDistance.updatecCs|jdks |jdkrtd|j|jd}|j|jd}|j|j||}||jd}|j |j||}||jd}t | d|| d| |j S)zWCalculate FID score based on accumulated extracted features from the two distributions.rwzVMore than one sample is required for both the real and fake distributed to compute FIDrr.)rr RuntimeErrorrrerrrr`rrrLrEr)r# mean_real mean_fake cov_real_numcov_real cov_fake_numcov_fakerrrcomputes"z FrechetInceptionDistance.computecsR|js"t|j}t|j}t|j}t||_||_||_dStdS)zReset metric states.N)rrrrrr reset)r#rrrr$rrrs     zFrechetInceptionDistance.resetdst_typer cs$t|}t|jtr||j_|S)zTransfer all metric state to specific dtype. Special version of standard `type` method. Arguments: dst_type: the desired type as ``torch.dtype`` or string )r set_dtyperrrr))r#rrjr$rrrs  z"FrechetInceptionDistance.set_dtypevalaxcCs |||S)ajPlot a single or multiple values from the metric. Args: val: Either a single result from calling `metric.forward` or `metric.compute` or a list of these results. If no value is provided, will automatically call `metric.compute` and plot that result. ax: An matplotlib axis object. If provided will add plot to that axis Returns: Figure and Axes object Raises: ModuleNotFoundError: If `matplotlib` is not installed .. plot:: :scale: 75 >>> # Example plotting a single value >>> import torch >>> from torchmetrics.image.fid import FrechetInceptionDistance >>> 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) >>> metric = FrechetInceptionDistance(feature=64) >>> metric.update(imgs_dist1, real=True) >>> metric.update(imgs_dist2, real=False) >>> fig_, ax_ = metric.plot() .. plot:: :scale: 75 >>> # Example plotting multiple values >>> import torch >>> from torchmetrics.image.fid import FrechetInceptionDistance >>> imgs_dist1 = lambda: torch.randint(0, 200, (100, 3, 299, 299), dtype=torch.uint8) >>> imgs_dist2 = lambda: torch.randint(100, 255, (100, 3, 299, 299), dtype=torch.uint8) >>> metric = FrechetInceptionDistance(feature=64) >>> values = [ ] >>> for _ in range(3): ... metric.update(imgs_dist1(), real=True) ... metric.update(imgs_dist2(), real=False) ... values.append(metric.compute()) ... metric.reset() >>> fig_, ax_ = metric.plot(values) )_plot)r#rrrrrplots 0r)rTFrN)rN)NN) rrrrlrro__annotations__rrrrFrrrrmrrrOrrr!rrrr?rArrr rrrprrr$rrs\  a      C  )!collections.abcrrCrtypingrrrr?rtorch.nnrtorch.nn.functionalr torchmetrics.metricr torchmetrics.utilities.importsr r torchmetrics.utilities.plotr r__doctest_skip__,torch_fidelity.feature_extractor_inceptionv3rrtorch_fidelity.helpersr,torch_fidelity.interpolate_compat_tensorflowrrrrrrrrs.       s