# from https://github.com/mseitzer/pytorch-fid import numpy as np import torch import torch.nn as nn import torch.nn.functional as F import torchvision from scipy import linalg from torch.nn.functional import adaptive_avg_pool2d try: from torchvision.models.utils import load_state_dict_from_url except ImportError: from torch.utils.model_zoo import load_url as load_state_dict_from_url FID_WEIGHTS_URL = ( "https://github.com/mseitzer/pytorch-fid/releases/download/" + "fid_weights/pt_inception-2015-12-05-6726825d.pth" ) class InceptionV3(nn.Module): """Pretrained InceptionV3 network returning feature maps""" # Index of default block of inception to return, # corresponds to output of final average pooling DEFAULT_BLOCK_INDEX = 3 # Maps feature dimensionality to their output blocks indices BLOCK_INDEX_BY_DIM = { 64: 0, # First max pooling features 192: 1, # Second max pooling features 768: 2, # Pre-aux classifier features 2048: 3, # Final average pooling features } def __init__( self, output_blocks=[DEFAULT_BLOCK_INDEX], resize_input=True, normalize_input=True, requires_grad=False, use_fid_inception=True, ): """Build 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. """ super(InceptionV3, self).__init__() self.resize_input = resize_input self.normalize_input = normalize_input self.output_blocks = sorted(output_blocks) self.last_needed_block = max(output_blocks) assert self.last_needed_block <= 3, "Last possible output block index is 3" self.blocks = nn.ModuleList() if use_fid_inception: inception = fid_inception_v3() else: inception = _inception_v3(pretrained=True) # Block 0: input to maxpool1 block0 = [ inception.Conv2d_1a_3x3, inception.Conv2d_2a_3x3, inception.Conv2d_2b_3x3, nn.MaxPool2d(kernel_size=3, stride=2), ] self.blocks.append(nn.Sequential(*block0)) # Block 1: maxpool1 to maxpool2 if self.last_needed_block >= 1: block1 = [ inception.Conv2d_3b_1x1, inception.Conv2d_4a_3x3, nn.MaxPool2d(kernel_size=3, stride=2), ] self.blocks.append(nn.Sequential(*block1)) # Block 2: maxpool2 to aux classifier if self.last_needed_block >= 2: block2 = [ inception.Mixed_5b, inception.Mixed_5c, inception.Mixed_5d, inception.Mixed_6a, inception.Mixed_6b, inception.Mixed_6c, inception.Mixed_6d, inception.Mixed_6e, ] self.blocks.append(nn.Sequential(*block2)) # Block 3: aux classifier to final avgpool if self.last_needed_block >= 3: block3 = [ inception.Mixed_7a, inception.Mixed_7b, inception.Mixed_7c, nn.AdaptiveAvgPool2d(output_size=(1, 1)), ] self.blocks.append(nn.Sequential(*block3)) for param in self.parameters(): param.requires_grad = requires_grad def forward(self, inp): """Get 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 """ outp = [] x = inp if self.resize_input: x = F.interpolate(x, size=(299, 299), mode="bilinear", align_corners=False) if self.normalize_input: x = 2 * x - 1 # Scale from range (0, 1) to range (-1, 1) for idx, block in enumerate(self.blocks): x = block(x) if idx in self.output_blocks: outp.append(x) if idx == self.last_needed_block: break return outp def _inception_v3(*args, **kwargs): """Wraps `torchvision.models.inception_v3` Skips default weight initialization if supported by torchvision version. See https://github.com/mseitzer/pytorch-fid/issues/28. """ try: version = tuple(map(int, torchvision.__version__.split(".")[:2])) except ValueError: # Just a caution against weird version strings version = (0,) if version >= (0, 6): kwargs["init_weights"] = False return torchvision.models.inception_v3(*args, **kwargs) def fid_inception_v3(): """Build 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. """ inception = _inception_v3(num_classes=1008, aux_logits=False, pretrained=False) inception.Mixed_5b = FIDInceptionA(192, pool_features=32) inception.Mixed_5c = FIDInceptionA(256, pool_features=64) inception.Mixed_5d = FIDInceptionA(288, pool_features=64) inception.Mixed_6b = FIDInceptionC(768, channels_7x7=128) inception.Mixed_6c = FIDInceptionC(768, channels_7x7=160) inception.Mixed_6d = FIDInceptionC(768, channels_7x7=160) inception.Mixed_6e = FIDInceptionC(768, channels_7x7=192) inception.Mixed_7b = FIDInceptionE_1(1280) inception.Mixed_7c = FIDInceptionE_2(2048) state_dict = load_state_dict_from_url(FID_WEIGHTS_URL, progress=True) inception.load_state_dict(state_dict) return inception class FIDInceptionA(torchvision.models.inception.InceptionA): """InceptionA block patched for FID computation""" def __init__(self, in_channels, pool_features): super(FIDInceptionA, self).__init__(in_channels, pool_features) def forward(self, x): branch1x1 = self.branch1x1(x) branch5x5 = self.branch5x5_1(x) branch5x5 = self.branch5x5_2(branch5x5) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) # Patch: Tensorflow's average pool does not use the padded zero's in # its average calculation branch_pool = F.avg_pool2d( x, kernel_size=3, stride=1, padding=1, count_include_pad=False ) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch5x5, branch3x3dbl, branch_pool] return torch.cat(outputs, 1) class FIDInceptionC(torchvision.models.inception.InceptionC): """InceptionC block patched for FID computation""" def __init__(self, in_channels, channels_7x7): super(FIDInceptionC, self).__init__(in_channels, channels_7x7) def forward(self, x): branch1x1 = self.branch1x1(x) branch7x7 = self.branch7x7_1(x) branch7x7 = self.branch7x7_2(branch7x7) branch7x7 = self.branch7x7_3(branch7x7) branch7x7dbl = self.branch7x7dbl_1(x) branch7x7dbl = self.branch7x7dbl_2(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_3(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_4(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_5(branch7x7dbl) # Patch: Tensorflow's average pool does not use the padded zero's in # its average calculation branch_pool = F.avg_pool2d( x, kernel_size=3, stride=1, padding=1, count_include_pad=False ) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch7x7, branch7x7dbl, branch_pool] return torch.cat(outputs, 1) class FIDInceptionE_1(torchvision.models.inception.InceptionE): """First InceptionE block patched for FID computation""" def __init__(self, in_channels): super(FIDInceptionE_1, self).__init__(in_channels) def forward(self, x): branch1x1 = self.branch1x1(x) branch3x3 = self.branch3x3_1(x) branch3x3 = [ self.branch3x3_2a(branch3x3), self.branch3x3_2b(branch3x3), ] branch3x3 = torch.cat(branch3x3, 1) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = [ self.branch3x3dbl_3a(branch3x3dbl), self.branch3x3dbl_3b(branch3x3dbl), ] branch3x3dbl = torch.cat(branch3x3dbl, 1) # Patch: Tensorflow's average pool does not use the padded zero's in # its average calculation branch_pool = F.avg_pool2d( x, kernel_size=3, stride=1, padding=1, count_include_pad=False ) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch3x3, branch3x3dbl, branch_pool] return torch.cat(outputs, 1) class FIDInceptionE_2(torchvision.models.inception.InceptionE): """Second InceptionE block patched for FID computation""" def __init__(self, in_channels): super(FIDInceptionE_2, self).__init__(in_channels) def forward(self, x): branch1x1 = self.branch1x1(x) branch3x3 = self.branch3x3_1(x) branch3x3 = [ self.branch3x3_2a(branch3x3), self.branch3x3_2b(branch3x3), ] branch3x3 = torch.cat(branch3x3, 1) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = [ self.branch3x3dbl_3a(branch3x3dbl), self.branch3x3dbl_3b(branch3x3dbl), ] branch3x3dbl = torch.cat(branch3x3dbl, 1) # Patch: The FID Inception model uses max pooling instead of average # pooling. This is likely an error in this specific Inception # implementation, as other Inception models use average pooling here # (which matches the description in the paper). branch_pool = F.max_pool2d(x, kernel_size=3, stride=1, padding=1) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch3x3, branch3x3dbl, branch_pool] return torch.cat(outputs, 1) def compute_val_fid(trainer, verbose=0): """ Compute the fid score between the n=opts.train.fid.n_images real images from the validation set (domain is rf) and n fake images pained from those n validation images Args: trainer (climategan.Trainer): trainer to compute the val fid for Returns: float: FID score """ # get opts params batch_size = trainer.opts.train.fid.get("batch_size", 50) dims = trainer.opts.train.fid.get("dims", 2048) # set inception model block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims] model = InceptionV3([block_idx]).to(trainer.device) # first fid computation: compute the real stats, only once if trainer.real_val_fid_stats is None: if verbose > 0: print("Computing real_val_fid_stats for the first time") set_real_val_fid_stats(trainer, model, batch_size, dims) # get real stats real_m = trainer.real_val_fid_stats["m"] real_s = trainer.real_val_fid_stats["s"] # compute fake images fakes = compute_fakes(trainer) if verbose > 0: print("Computing fake activation statistics") # get fake stats fake_m, fake_s = calculate_activation_statistics( fakes, model, batch_size=batch_size, dims=dims, device=trainer.device ) # compute FD between the real and the fake inception stats return calculate_frechet_distance(real_m, real_s, fake_m, fake_s) def set_real_val_fid_stats(trainer, model, batch_size, dims): """ Sets the real_val_fid_stats attribute of the trainer with the m and s outputs of calculate_activation_statistics on the real data. This needs to be done only once since nothing changes during training here. Args: trainer (climategan.Trainer): trainer instance to compute the stats for model (InceptionV3): inception model to get the activations from batch_size (int): inception inference batch size dims (int): dimension selected in the model """ # in the rf domain display_size may be different from fid.n_images limit = trainer.opts.train.fid.n_images display_x = torch.stack( [sample["data"]["x"] for sample in trainer.display_images["val"]["rf"][:limit]] ).to(trainer.device) m, s = calculate_activation_statistics( display_x, model, batch_size=batch_size, dims=dims, device=trainer.device ) trainer.real_val_fid_stats = {"m": m, "s": s} def compute_fakes(trainer, verbose=0): """ Compute current fake inferences Args: trainer (climategan.Trainer): trainer instance verbose (int, optional): Print level. Defaults to 0. Returns: torch.Tensor: trainer.opts.train.fid.n_images painted images """ # in the rf domain display_size may be different from fid.n_images n = trainer.opts.train.fid.n_images bs = trainer.opts.data.loaders.batch_size display_batches = [ (sample["data"]["x"], sample["data"]["m"]) for sample in trainer.display_images["val"]["rf"][:n] ] display_x = torch.stack([b[0] for b in display_batches]).to(trainer.device) display_m = torch.stack([b[0] for b in display_batches]).to(trainer.device) nbs = len(display_x) // bs + 1 fakes = [] for b in range(nbs): if verbose > 0: print("computing fakes {}/{}".format(b + 1, nbs), end="\r", flush=True) with torch.no_grad(): x = display_x[b * bs : (b + 1) * bs] m = display_m[b * bs : (b + 1) * bs] fake = trainer.G.paint(m, x) fakes.append(fake) return torch.cat(fakes, dim=0) def calculate_activation_statistics( images, model, batch_size=50, dims=2048, device="cpu" ): """Calculation of the statistics used by the FID. Params: -- images : List of images -- model : Instance of inception model -- batch_size : The images numpy array is split into batches with batch size batch_size. A reasonable batch size depends on the hardware. -- dims : Dimensionality of features returned by Inception -- device : Device to run calculations Returns: -- mu : The mean over samples of the activations of the pool_3 layer of the inception model. -- sigma : The covariance matrix of the activations of the pool_3 layer of the inception model. """ act = get_activations(images, model, batch_size, dims, device) mu = np.mean(act, axis=0) sigma = np.cov(act, rowvar=False) return mu, sigma def get_activations(images, model, batch_size=50, dims=2048, device="cpu"): """Calculates the activations of the pool_3 layer for all images. Params: -- images : List of images -- model : Instance of inception model -- batch_size : Batch size of images for the model to process at once. Make sure that the number of samples is a multiple of the batch size, otherwise some samples are ignored. This behavior is retained to match the original FID score implementation. -- dims : Dimensionality of features returned by Inception -- device : Device to run calculations Returns: -- A numpy array of dimension (num images, dims) that contains the activations of the given tensor when feeding inception with the query tensor. """ model.eval() pred_arr = np.empty((len(images), dims)) start_idx = 0 nbs = len(images) // batch_size + 1 for b in range(nbs): batch = images[b * batch_size : (b + 1) * batch_size].to(device) if not batch.nelement(): continue with torch.no_grad(): pred = model(batch)[0] # If model output is not scalar, apply global spatial average pooling. # This happens if you choose a dimensionality not equal 2048. if pred.size(2) != 1 or pred.size(3) != 1: pred = adaptive_avg_pool2d(pred, output_size=(1, 1)) pred = pred.squeeze(3).squeeze(2).cpu().numpy() pred_arr[start_idx : start_idx + pred.shape[0]] = pred start_idx = start_idx + pred.shape[0] return pred_arr def calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-6): """Numpy implementation of the Frechet Distance. The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) and X_2 ~ N(mu_2, C_2) is d^2 = ||mu_1 - mu_2||^2 + Tr(C_1 + C_2 - 2*sqrt(C_1*C_2)). Stable version by Dougal J. Sutherland. Params: -- mu1 : Numpy array containing the activations of a layer of the inception net (like returned by the function 'get_predictions') for generated samples. -- mu2 : The sample mean over activations, precalculated on an representative data set. -- sigma1: The covariance matrix over activations for generated samples. -- sigma2: The covariance matrix over activations, precalculated on an representative data set. Returns: -- : The Frechet Distance. """ mu1 = np.atleast_1d(mu1) mu2 = np.atleast_1d(mu2) sigma1 = np.atleast_2d(sigma1) sigma2 = np.atleast_2d(sigma2) assert ( mu1.shape == mu2.shape ), "Training and test mean vectors have different lengths" assert ( sigma1.shape == sigma2.shape ), "Training and test covariances have different dimensions" diff = mu1 - mu2 # Product might be almost singular covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False) if not np.isfinite(covmean).all(): msg = ( "fid calculation produces singular product; " "adding %s to diagonal of cov estimates" ) % eps print(msg) offset = np.eye(sigma1.shape[0]) * eps covmean = linalg.sqrtm((sigma1 + offset).dot(sigma2 + offset)) # Numerical error might give slight imaginary component if np.iscomplexobj(covmean): if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3): m = np.max(np.abs(covmean.imag)) raise ValueError("Imaginary component {}".format(m)) covmean = covmean.real tr_covmean = np.trace(covmean) return diff.dot(diff) + np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean