''' Inception utilities This file contains methods for calculating IS and FID, using either the original numpy code or an accelerated fully-pytorch version that uses a fast newton-schulz approximation for the matrix sqrt. There are also methods for acquiring a desired number of samples from the Generator, and parallelizing the inbuilt PyTorch inception network. NOTE that Inception Scores and FIDs calculated using these methods will *not* be directly comparable to values calculated using the original TF IS/FID code. You *must* use the TF model if you wish to report and compare numbers. This code tends to produce IS values that are 5-10% lower than those obtained through TF. ''' import numpy as np from scipy import linalg # For numpy FID import time import torch import torch.nn as nn import torch.nn.functional as F from torch.nn import Parameter as P from torchvision.models.inception import inception_v3 # Module that wraps the inception network to enable use with dataparallel and # returning pool features and logits. class WrapInception(nn.Module): def __init__(self, net): super(WrapInception,self).__init__() self.net = net self.mean = P(torch.tensor([0.485, 0.456, 0.406]).view(1, -1, 1, 1), requires_grad=False) self.std = P(torch.tensor([0.229, 0.224, 0.225]).view(1, -1, 1, 1), requires_grad=False) def forward(self, x): # Normalize x x = (x + 1.) / 2.0 x = (x - self.mean) / self.std # Upsample if necessary if x.shape[2] != 299 or x.shape[3] != 299: x = F.interpolate(x, size=(299, 299), mode='bilinear', align_corners=True) # 299 x 299 x 3 x = self.net.Conv2d_1a_3x3(x) # 149 x 149 x 32 x = self.net.Conv2d_2a_3x3(x) # 147 x 147 x 32 x = self.net.Conv2d_2b_3x3(x) # 147 x 147 x 64 x = F.max_pool2d(x, kernel_size=3, stride=2) # 73 x 73 x 64 x = self.net.Conv2d_3b_1x1(x) # 73 x 73 x 80 x = self.net.Conv2d_4a_3x3(x) # 71 x 71 x 192 x = F.max_pool2d(x, kernel_size=3, stride=2) # 35 x 35 x 192 x = self.net.Mixed_5b(x) # 35 x 35 x 256 x = self.net.Mixed_5c(x) # 35 x 35 x 288 x = self.net.Mixed_5d(x) # 35 x 35 x 288 x = self.net.Mixed_6a(x) # 17 x 17 x 768 x = self.net.Mixed_6b(x) # 17 x 17 x 768 x = self.net.Mixed_6c(x) # 17 x 17 x 768 x = self.net.Mixed_6d(x) # 17 x 17 x 768 x = self.net.Mixed_6e(x) # 17 x 17 x 768 # 17 x 17 x 768 x = self.net.Mixed_7a(x) # 8 x 8 x 1280 x = self.net.Mixed_7b(x) # 8 x 8 x 2048 x = self.net.Mixed_7c(x) # 8 x 8 x 2048 pool = torch.mean(x.view(x.size(0), x.size(1), -1), 2) # 1 x 1 x 2048 logits = self.net.fc(F.dropout(pool, training=False).view(pool.size(0), -1)) # 1000 (num_classes) return pool, logits # A pytorch implementation of cov, from Modar M. Alfadly # https://discuss.pytorch.org/t/covariance-and-gradient-support/16217/2 def torch_cov(m, rowvar=False): '''Estimate a covariance matrix given data. Covariance indicates the level to which two variables vary together. If we examine N-dimensional samples, `X = [x_1, x_2, ... x_N]^T`, then the covariance matrix element `C_{ij}` is the covariance of `x_i` and `x_j`. The element `C_{ii}` is the variance of `x_i`. Args: m: A 1-D or 2-D array containing multiple variables and observations. Each row of `m` represents a variable, and each column a single observation of all those variables. rowvar: If `rowvar` is True, then each row represents a variable, with observations in the columns. Otherwise, the relationship is transposed: each column represents a variable, while the rows contain observations. Returns: The covariance matrix of the variables. ''' if m.dim() > 2: raise ValueError('m has more than 2 dimensions') if m.dim() < 2: m = m.view(1, -1) if not rowvar and m.size(0) != 1: m = m.t() # m = m.type(torch.double) # uncomment this line if desired fact = 1.0 / (m.size(1) - 1) m -= torch.mean(m, dim=1, keepdim=True) mt = m.t() # if complex: mt = m.t().conj() return fact * m.matmul(mt).squeeze() # Pytorch implementation of matrix sqrt, from Tsung-Yu Lin, and Subhransu Maji # https://github.com/msubhransu/matrix-sqrt def sqrt_newton_schulz(A, numIters, dtype=None): with torch.no_grad(): if dtype is None: dtype = A.type() batchSize = A.shape[0] dim = A.shape[1] normA = A.mul(A).sum(dim=1).sum(dim=1).sqrt() Y = A.div(normA.view(batchSize, 1, 1).expand_as(A)); I = torch.eye(dim,dim).view(1, dim, dim).repeat(batchSize,1,1).type(dtype) Z = torch.eye(dim,dim).view(1, dim, dim).repeat(batchSize,1,1).type(dtype) for i in range(numIters): T = 0.5*(3.0*I - Z.bmm(Y)) Y = Y.bmm(T) Z = T.bmm(Z) sA = Y*torch.sqrt(normA).view(batchSize, 1, 1).expand_as(A) return sA # FID calculator from TTUR--consider replacing this with GPU-accelerated cov # calculations using torch? def numpy_calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-6): """Numpy implementation of the Frechet Distance. Taken from https://github.com/bioinf-jku/TTUR 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 representive data set. -- sigma1: The covariance matrix over activations for generated samples. -- sigma2: The covariance matrix over activations, precalculated on an representive 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): print('wat') 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) out = diff.dot(diff) + np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean return out def torch_calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-6): """Pytorch implementation of the Frechet Distance. Taken from https://github.com/bioinf-jku/TTUR 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 representive data set. -- sigma1: The covariance matrix over activations for generated samples. -- sigma2: The covariance matrix over activations, precalculated on an representive data set. Returns: -- : The Frechet Distance. """ 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 # Run 50 itrs of newton-schulz to get the matrix sqrt of sigma1 dot sigma2 covmean = sqrt_newton_schulz(sigma1.mm(sigma2).unsqueeze(0), 50).squeeze() out = (diff.dot(diff) + torch.trace(sigma1) + torch.trace(sigma2) - 2 * torch.trace(covmean)) return out # Calculate Inception Score mean + std given softmax'd logits and number of splits def calculate_inception_score(pred, num_splits=10): scores = [] for index in range(num_splits): pred_chunk = pred[index * (pred.shape[0] // num_splits): (index + 1) * (pred.shape[0] // num_splits), :] kl_inception = pred_chunk * (np.log(pred_chunk) - np.log(np.expand_dims(np.mean(pred_chunk, 0), 0))) kl_inception = np.mean(np.sum(kl_inception, 1)) scores.append(np.exp(kl_inception)) return np.mean(scores), np.std(scores) # Loop and run the sampler and the net until it accumulates num_inception_images # activations. Return the pool, the logits, and the labels (if one wants # Inception Accuracy the labels of the generated class will be needed) def accumulate_inception_activations(sample, net, num_inception_images=50000): pool, logits, labels = [], [], [] while (torch.cat(logits, 0).shape[0] if len(logits) else 0) < num_inception_images: with torch.no_grad(): images, labels_val = sample() pool_val, logits_val = net(images.float()) pool += [pool_val] logits += [F.softmax(logits_val, 1)] labels += [labels_val] return torch.cat(pool, 0), torch.cat(logits, 0), torch.cat(labels, 0) # Load and wrap the Inception model def load_inception_net(parallel=False): inception_model = inception_v3(pretrained=True, transform_input=False) inception_model = WrapInception(inception_model.eval()).cuda() if parallel: print('Parallelizing Inception module...') inception_model = nn.DataParallel(inception_model) return inception_model # This produces a function which takes in an iterator which returns a set number of samples # and iterates until it accumulates config['num_inception_images'] images. # The iterator can return samples with a different batch size than used in # training, using the setting confg['inception_batchsize'] def prepare_inception_metrics(dataset, parallel, no_fid=False): # Load metrics; this is intentionally not in a try-except loop so that # the script will crash here if it cannot find the Inception moments. # By default, remove the "hdf5" from dataset dataset = dataset.strip('_hdf5') data_mu = np.load(dataset+'_inception_moments.npz')['mu'] data_sigma = np.load(dataset+'_inception_moments.npz')['sigma'] # Load network net = load_inception_net(parallel) def get_inception_metrics(sample, num_inception_images, num_splits=10, prints=True, use_torch=True): if prints: print('Gathering activations...') pool, logits, labels = accumulate_inception_activations(sample, net, num_inception_images) if prints: print('Calculating Inception Score...') IS_mean, IS_std = calculate_inception_score(logits.cpu().numpy(), num_splits) if no_fid: FID = 9999.0 else: if prints: print('Calculating means and covariances...') if use_torch: mu, sigma = torch.mean(pool, 0), torch_cov(pool, rowvar=False) else: mu, sigma = np.mean(pool.cpu().numpy(), axis=0), np.cov(pool.cpu().numpy(), rowvar=False) if prints: print('Covariances calculated, getting FID...') if use_torch: FID = torch_calculate_frechet_distance(mu, sigma, torch.tensor(data_mu).float().cuda(), torch.tensor(data_sigma).float().cuda()) FID = float(FID.cpu().numpy()) else: FID = numpy_calculate_frechet_distance(mu.cpu().numpy(), sigma.cpu().numpy(), data_mu, data_sigma) # Delete mu, sigma, pool, logits, and labels, just in case del mu, sigma, pool, logits, labels return IS_mean, IS_std, FID return get_inception_metrics