# Copyright 2017 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Model evaluation tools for TFGAN. These methods come from https://arxiv.org/abs/1606.03498, https://arxiv.org/abs/1706.08500, and https://arxiv.org/abs/1801.01401. NOTE: This implementation uses the same weights as in https://github.com/openai/improved-gan/blob/master/inception_score/model.py, but is more numerically stable and is an unbiased estimator of the true Inception score even when splitting the inputs into batches. """ from __future__ import absolute_import from __future__ import division from __future__ import print_function import functools import os import sys import tarfile from six.moves import urllib from tensorflow.contrib.layers.python.layers import layers from tensorflow.core.framework import graph_pb2 from tensorflow.python.framework import dtypes from tensorflow.python.framework import importer from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import functional_ops from tensorflow.python.ops import image_ops from tensorflow.python.ops import linalg_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops import nn_impl from tensorflow.python.ops import nn_ops from tensorflow.python.platform import gfile from tensorflow.python.platform import resource_loader __all__ = [ 'get_graph_def_from_disk', 'get_graph_def_from_resource', 'get_graph_def_from_url_tarball', 'preprocess_image', 'run_image_classifier', 'run_inception', 'inception_score', 'classifier_score', 'classifier_score_from_logits', 'frechet_inception_distance', 'frechet_classifier_distance', 'frechet_classifier_distance_from_activations', 'mean_only_frechet_classifier_distance_from_activations', 'diagonal_only_frechet_classifier_distance_from_activations', 'kernel_inception_distance', 'kernel_inception_distance_and_std', 'kernel_classifier_distance', 'kernel_classifier_distance_and_std', 'kernel_classifier_distance_from_activations', 'kernel_classifier_distance_and_std_from_activations', 'INCEPTION_DEFAULT_IMAGE_SIZE', ] INCEPTION_URL = 'http://download.tensorflow.org/models/frozen_inception_v1_2015_12_05.tar.gz' INCEPTION_FROZEN_GRAPH = 'inceptionv1_for_inception_score.pb' INCEPTION_INPUT = 'Mul:0' INCEPTION_OUTPUT = 'logits:0' INCEPTION_FINAL_POOL = 'pool_3:0' INCEPTION_DEFAULT_IMAGE_SIZE = 299 def _validate_images(images, image_size): images = ops.convert_to_tensor(images) images.shape.with_rank(4) images.shape.assert_is_compatible_with([None, image_size, image_size, None]) return images def _symmetric_matrix_square_root(mat, eps=1e-10): """Compute square root of a symmetric matrix. Note that this is different from an elementwise square root. We want to compute M' where M' = sqrt(mat) such that M' * M' = mat. Also note that this method **only** works for symmetric matrices. Args: mat: Matrix to take the square root of. eps: Small epsilon such that any element less than eps will not be square rooted to guard against numerical instability. Returns: Matrix square root of mat. """ # Unlike numpy, tensorflow's return order is (s, u, v) s, u, v = linalg_ops.svd(mat) # sqrt is unstable around 0, just use 0 in such case si = array_ops.where(math_ops.less(s, eps), s, math_ops.sqrt(s)) # Note that the v returned by Tensorflow is v = V # (when referencing the equation A = U S V^T) # This is unlike Numpy which returns v = V^T return math_ops.matmul( math_ops.matmul(u, array_ops.diag(si)), v, transpose_b=True) def preprocess_image(images, height=INCEPTION_DEFAULT_IMAGE_SIZE, width=INCEPTION_DEFAULT_IMAGE_SIZE, scope=None): """Prepare a batch of images for evaluation. This is the preprocessing portion of the graph from http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz. Note that it expects Tensors in [0, 255]. This function maps pixel values to [-1, 1] and resizes to match the InceptionV1 network. Args: images: 3-D or 4-D Tensor of images. Values are in [0, 255]. height: Integer. Height of resized output image. width: Integer. Width of resized output image. scope: Optional scope for name_scope. Returns: 3-D or 4-D float Tensor of prepared image(s). Values are in [-1, 1]. """ is_single = images.shape.ndims == 3 with ops.name_scope(scope, 'preprocess', [images, height, width]): if not images.dtype.is_floating: images = math_ops.to_float(images) if is_single: images = array_ops.expand_dims(images, axis=0) resized = image_ops.resize_bilinear(images, [height, width]) resized = (resized - 128.0) / 128.0 if is_single: resized = array_ops.squeeze(resized, axis=0) return resized def _kl_divergence(p, p_logits, q): """Computes the Kullback-Liebler divergence between p and q. This function uses p's logits in some places to improve numerical stability. Specifically: KL(p || q) = sum[ p * log(p / q) ] = sum[ p * ( log(p) - log(q) ) ] = sum[ p * ( log_softmax(p_logits) - log(q) ) ] Args: p: A 2-D floating-point Tensor p_ij, where `i` corresponds to the minibatch example and `j` corresponds to the probability of being in class `j`. p_logits: A 2-D floating-point Tensor corresponding to logits for `p`. q: A 1-D floating-point Tensor, where q_j corresponds to the probability of class `j`. Returns: KL divergence between two distributions. Output dimension is 1D, one entry per distribution in `p`. Raises: ValueError: If any of the inputs aren't floating-point. ValueError: If p or p_logits aren't 2D. ValueError: If q isn't 1D. """ for tensor in [p, p_logits, q]: if not tensor.dtype.is_floating: raise ValueError('Input %s must be floating type.', tensor.name) p.shape.assert_has_rank(2) p_logits.shape.assert_has_rank(2) q.shape.assert_has_rank(1) return math_ops.reduce_sum( p * (nn_ops.log_softmax(p_logits) - math_ops.log(q)), axis=1) def get_graph_def_from_disk(filename): """Get a GraphDef proto from a disk location.""" with gfile.FastGFile(filename, 'rb') as f: return graph_pb2.GraphDef.FromString(f.read()) def get_graph_def_from_resource(filename): """Get a GraphDef proto from within a .par file.""" return graph_pb2.GraphDef.FromString(resource_loader.load_resource(filename)) def get_graph_def_from_url_tarball(url, filename, tar_filename=None): """Get a GraphDef proto from a tarball on the web. Args: url: Web address of tarball filename: Filename of graph definition within tarball tar_filename: Temporary download filename (None = always download) Returns: A GraphDef loaded from a file in the downloaded tarball. """ if not (tar_filename and os.path.exists(tar_filename)): def _progress(count, block_size, total_size): sys.stdout.write('\r>> Downloading %s %.1f%%' % (url, float(count * block_size) / float(total_size) * 100.0)) sys.stdout.flush() tar_filename, _ = urllib.request.urlretrieve(url, tar_filename, _progress) with tarfile.open(tar_filename, 'r:gz') as tar: proto_str = tar.extractfile(filename).read() return graph_pb2.GraphDef.FromString(proto_str) def _default_graph_def_fn(): return get_graph_def_from_url_tarball(INCEPTION_URL, INCEPTION_FROZEN_GRAPH, os.path.basename(INCEPTION_URL)) def run_inception(images, graph_def=None, default_graph_def_fn=_default_graph_def_fn, image_size=INCEPTION_DEFAULT_IMAGE_SIZE, input_tensor=INCEPTION_INPUT, output_tensor=INCEPTION_OUTPUT): """Run images through a pretrained Inception classifier. Args: images: Input tensors. Must be [batch, height, width, channels]. Input shape and values must be in [-1, 1], which can be achieved using `preprocess_image`. graph_def: A GraphDef proto of a pretrained Inception graph. If `None`, call `default_graph_def_fn` to get GraphDef. default_graph_def_fn: A function that returns a GraphDef. Used if `graph_def` is `None. By default, returns a pretrained InceptionV3 graph. image_size: Required image width and height. See unit tests for the default values. input_tensor: Name of input Tensor. output_tensor: Name or list of output Tensors. This function will compute activations at the specified layer. Examples include INCEPTION_V3_OUTPUT and INCEPTION_V3_FINAL_POOL which would result in this function computing the final logits or the penultimate pooling layer. Returns: Tensor or Tensors corresponding to computed `output_tensor`. Raises: ValueError: If images are not the correct size. ValueError: If neither `graph_def` nor `default_graph_def_fn` are provided. """ images = _validate_images(images, image_size) if graph_def is None: if default_graph_def_fn is None: raise ValueError('If `graph_def` is `None`, must provide ' '`default_graph_def_fn`.') graph_def = default_graph_def_fn() activations = run_image_classifier(images, graph_def, input_tensor, output_tensor) if isinstance(activations, list): for i, activation in enumerate(activations): if array_ops.rank(activation) != 2: activations[i] = layers.flatten(activation) else: if array_ops.rank(activations) != 2: activations = layers.flatten(activations) return activations def run_image_classifier(tensor, graph_def, input_tensor, output_tensor, scope='RunClassifier'): """Runs a network from a frozen graph. Args: tensor: An Input tensor. graph_def: A GraphDef proto. input_tensor: Name of input tensor in graph def. output_tensor: A tensor name or list of tensor names in graph def. scope: Name scope for classifier. Returns: Classifier output if `output_tensor` is a string, or a list of outputs if `output_tensor` is a list. Raises: ValueError: If `input_tensor` or `output_tensor` aren't in the graph_def. """ input_map = {input_tensor: tensor} is_singleton = isinstance(output_tensor, str) if is_singleton: output_tensor = [output_tensor] classifier_outputs = importer.import_graph_def( graph_def, input_map, output_tensor, name=scope) if is_singleton: classifier_outputs = classifier_outputs[0] return classifier_outputs def classifier_score(images, classifier_fn, num_batches=1): """Classifier score for evaluating a conditional generative model. This is based on the Inception Score, but for an arbitrary classifier. This technique is described in detail in https://arxiv.org/abs/1606.03498. In summary, this function calculates exp( E[ KL(p(y|x) || p(y)) ] ) which captures how different the network's classification prediction is from the prior distribution over classes. NOTE: This function consumes images, computes their logits, and then computes the classifier score. If you would like to precompute many logits for large batches, use classifier_score_from_logits(), which this method also uses. Args: images: Images to calculate the classifier score for. classifier_fn: A function that takes images and produces logits based on a classifier. num_batches: Number of batches to split `generated_images` in to in order to efficiently run them through the classifier network. Returns: The classifier score. A floating-point scalar of the same type as the output of `classifier_fn`. """ generated_images_list = array_ops.split( images, num_or_size_splits=num_batches) # Compute the classifier splits using the memory-efficient `map_fn`. logits = functional_ops.map_fn( fn=classifier_fn, elems=array_ops.stack(generated_images_list), parallel_iterations=1, back_prop=False, swap_memory=True, name='RunClassifier') logits = array_ops.concat(array_ops.unstack(logits), 0) return classifier_score_from_logits(logits) def classifier_score_from_logits(logits): """Classifier score for evaluating a generative model from logits. This method computes the classifier score for a set of logits. This can be used independently of the classifier_score() method, especially in the case of using large batches during evaluation where we would like precompute all of the logits before computing the classifier score. This technique is described in detail in https://arxiv.org/abs/1606.03498. In summary, this function calculates: exp( E[ KL(p(y|x) || p(y)) ] ) which captures how different the network's classification prediction is from the prior distribution over classes. Args: logits: Precomputed 2D tensor of logits that will be used to compute the classifier score. Returns: The classifier score. A floating-point scalar of the same type as the output of `logits`. """ logits.shape.assert_has_rank(2) # Use maximum precision for best results. logits_dtype = logits.dtype if logits_dtype != dtypes.float64: logits = math_ops.to_double(logits) p = nn_ops.softmax(logits) q = math_ops.reduce_mean(p, axis=0) kl = _kl_divergence(p, logits, q) kl.shape.assert_has_rank(1) log_score = math_ops.reduce_mean(kl) final_score = math_ops.exp(log_score) if logits_dtype != dtypes.float64: final_score = math_ops.cast(final_score, logits_dtype) return final_score inception_score = functools.partial( classifier_score, classifier_fn=functools.partial( run_inception, output_tensor=INCEPTION_OUTPUT)) def trace_sqrt_product(sigma, sigma_v): """Find the trace of the positive sqrt of product of covariance matrices. '_symmetric_matrix_square_root' only works for symmetric matrices, so we cannot just take _symmetric_matrix_square_root(sigma * sigma_v). ('sigma' and 'sigma_v' are symmetric, but their product is not necessarily). Let sigma = A A so A = sqrt(sigma), and sigma_v = B B. We want to find trace(sqrt(sigma sigma_v)) = trace(sqrt(A A B B)) Note the following properties: (i) forall M1, M2: eigenvalues(M1 M2) = eigenvalues(M2 M1) => eigenvalues(A A B B) = eigenvalues (A B B A) (ii) if M1 = sqrt(M2), then eigenvalues(M1) = sqrt(eigenvalues(M2)) => eigenvalues(sqrt(sigma sigma_v)) = sqrt(eigenvalues(A B B A)) (iii) forall M: trace(M) = sum(eigenvalues(M)) => trace(sqrt(sigma sigma_v)) = sum(eigenvalues(sqrt(sigma sigma_v))) = sum(sqrt(eigenvalues(A B B A))) = sum(eigenvalues(sqrt(A B B A))) = trace(sqrt(A B B A)) = trace(sqrt(A sigma_v A)) A = sqrt(sigma). Both sigma and A sigma_v A are symmetric, so we **can** use the _symmetric_matrix_square_root function to find the roots of these matrices. Args: sigma: a square, symmetric, real, positive semi-definite covariance matrix sigma_v: same as sigma Returns: The trace of the positive square root of sigma*sigma_v """ # Note sqrt_sigma is called "A" in the proof above sqrt_sigma = _symmetric_matrix_square_root(sigma) # This is sqrt(A sigma_v A) above sqrt_a_sigmav_a = math_ops.matmul(sqrt_sigma, math_ops.matmul(sigma_v, sqrt_sigma)) return math_ops.trace(_symmetric_matrix_square_root(sqrt_a_sigmav_a)) def frechet_classifier_distance(real_images, generated_images, classifier_fn, num_batches=1): """Classifier distance for evaluating a generative model. This is based on the Frechet Inception distance, but for an arbitrary classifier. This technique is described in detail in https://arxiv.org/abs/1706.08500. Given two Gaussian distribution with means m and m_w and covariance matrices C and C_w, this function calculates |m - m_w|^2 + Tr(C + C_w - 2(C * C_w)^(1/2)) which captures how different the distributions of real images and generated images (or more accurately, their visual features) are. Note that unlike the Inception score, this is a true distance and utilizes information about real world images. Note that when computed using sample means and sample covariance matrices, Frechet distance is biased. It is more biased for small sample sizes. (e.g. even if the two distributions are the same, for a small sample size, the expected Frechet distance is large). It is important to use the same sample size to compute Frechet classifier distance when comparing two generative models. NOTE: This function consumes images, computes their activations, and then computes the classifier score. If you would like to precompute many activations for real and generated images for large batches, please use frechet_clasifier_distance_from_activations(), which this method also uses. Args: real_images: Real images to use to compute Frechet Inception distance. generated_images: Generated images to use to compute Frechet Inception distance. classifier_fn: A function that takes images and produces activations based on a classifier. num_batches: Number of batches to split images in to in order to efficiently run them through the classifier network. Returns: The Frechet Inception distance. A floating-point scalar of the same type as the output of `classifier_fn`. """ real_images_list = array_ops.split( real_images, num_or_size_splits=num_batches) generated_images_list = array_ops.split( generated_images, num_or_size_splits=num_batches) real_imgs = array_ops.stack(real_images_list) generated_imgs = array_ops.stack(generated_images_list) # Compute the activations using the memory-efficient `map_fn`. def compute_activations(elems): return functional_ops.map_fn(fn=classifier_fn, elems=elems, parallel_iterations=1, back_prop=False, swap_memory=True, name='RunClassifier') real_a = compute_activations(real_imgs) gen_a = compute_activations(generated_imgs) # Ensure the activations have the right shapes. real_a = array_ops.concat(array_ops.unstack(real_a), 0) gen_a = array_ops.concat(array_ops.unstack(gen_a), 0) return frechet_classifier_distance_from_activations(real_a, gen_a) def mean_only_frechet_classifier_distance_from_activations( real_activations, generated_activations): """Classifier distance for evaluating a generative model from activations. Given two Gaussian distribution with means m and m_w and covariance matrices C and C_w, this function calcuates |m - m_w|^2 which captures how different the distributions of real images and generated images (or more accurately, their visual features) are. Note that unlike the Inception score, this is a true distance and utilizes information about real world images. Note that when computed using sample means and sample covariance matrices, Frechet distance is biased. It is more biased for small sample sizes. (e.g. even if the two distributions are the same, for a small sample size, the expected Frechet distance is large). It is important to use the same sample size to compute frechet classifier distance when comparing two generative models. In this variant, we only compute the difference between the means of the fitted Gaussians. The computation leads to O(n) vs. O(n^2) memory usage, yet still retains much of the same information as FID. Args: real_activations: 2D array of activations of real images of size [num_images, num_dims] to use to compute Frechet Inception distance. generated_activations: 2D array of activations of generated images of size [num_images, num_dims] to use to compute Frechet Inception distance. Returns: The mean-only Frechet Inception distance. A floating-point scalar of the same type as the output of the activations. """ real_activations.shape.assert_has_rank(2) generated_activations.shape.assert_has_rank(2) activations_dtype = real_activations.dtype if activations_dtype != dtypes.float64: real_activations = math_ops.to_double(real_activations) generated_activations = math_ops.to_double(generated_activations) # Compute means of activations. m = math_ops.reduce_mean(real_activations, 0) m_w = math_ops.reduce_mean(generated_activations, 0) # Next the distance between means. mean = math_ops.reduce_sum( math_ops.squared_difference(m, m_w)) # Equivalent to L2 but more stable. mofid = mean if activations_dtype != dtypes.float64: mofid = math_ops.cast(mofid, activations_dtype) return mofid def diagonal_only_frechet_classifier_distance_from_activations( real_activations, generated_activations): """Classifier distance for evaluating a generative model. This is based on the Frechet Inception distance, but for an arbitrary classifier. This technique is described in detail in https://arxiv.org/abs/1706.08500. Given two Gaussian distribution with means m and m_w and covariance matrices C and C_w, this function calcuates |m - m_w|^2 + (sigma + sigma_w - 2(sigma x sigma_w)^(1/2)) which captures how different the distributions of real images and generated images (or more accurately, their visual features) are. Note that unlike the Inception score, this is a true distance and utilizes information about real world images. In this variant, we compute diagonal-only covariance matrices. As a result, instead of computing an expensive matrix square root, we can do something much simpler, and has O(n) vs O(n^2) space complexity. Note that when computed using sample means and sample covariance matrices, Frechet distance is biased. It is more biased for small sample sizes. (e.g. even if the two distributions are the same, for a small sample size, the expected Frechet distance is large). It is important to use the same sample size to compute frechet classifier distance when comparing two generative models. Args: real_activations: Real images to use to compute Frechet Inception distance. generated_activations: Generated images to use to compute Frechet Inception distance. Returns: The diagonal-only Frechet Inception distance. A floating-point scalar of the same type as the output of the activations. Raises: ValueError: If the shape of the variance and mean vectors are not equal. """ real_activations.shape.assert_has_rank(2) generated_activations.shape.assert_has_rank(2) activations_dtype = real_activations.dtype if activations_dtype != dtypes.float64: real_activations = math_ops.to_double(real_activations) generated_activations = math_ops.to_double(generated_activations) # Compute mean and covariance matrices of activations. m, var = nn_impl.moments(real_activations, axes=[0]) m_w, var_w = nn_impl.moments(generated_activations, axes=[0]) actual_shape = var.get_shape() expected_shape = m.get_shape() if actual_shape != expected_shape: raise ValueError('shape: {} must match expected shape: {}'.format( actual_shape, expected_shape)) # Compute the two components of FID. # First the covariance component. # Here, note that trace(A + B) = trace(A) + trace(B) trace = math_ops.reduce_sum( (var + var_w) - 2.0 * math_ops.sqrt(math_ops.multiply(var, var_w))) # Next the distance between means. mean = math_ops.reduce_sum( math_ops.squared_difference(m, m_w)) # Equivalent to L2 but more stable. dofid = trace + mean if activations_dtype != dtypes.float64: dofid = math_ops.cast(dofid, activations_dtype) return dofid def frechet_classifier_distance_from_activations(real_activations, generated_activations): """Classifier distance for evaluating a generative model. This methods computes the Frechet classifier distance from activations of real images and generated images. This can be used independently of the frechet_classifier_distance() method, especially in the case of using large batches during evaluation where we would like precompute all of the activations before computing the classifier distance. This technique is described in detail in https://arxiv.org/abs/1706.08500. Given two Gaussian distribution with means m and m_w and covariance matrices C and C_w, this function calculates |m - m_w|^2 + Tr(C + C_w - 2(C * C_w)^(1/2)) which captures how different the distributions of real images and generated images (or more accurately, their visual features) are. Note that unlike the Inception score, this is a true distance and utilizes information about real world images. Note that when computed using sample means and sample covariance matrices, Frechet distance is biased. It is more biased for small sample sizes. (e.g. even if the two distributions are the same, for a small sample size, the expected Frechet distance is large). It is important to use the same sample size to compute frechet classifier distance when comparing two generative models. Args: real_activations: 2D Tensor containing activations of real data. Shape is [batch_size, activation_size]. generated_activations: 2D Tensor containing activations of generated data. Shape is [batch_size, activation_size]. Returns: The Frechet Inception distance. A floating-point scalar of the same type as the output of the activations. """ real_activations.shape.assert_has_rank(2) generated_activations.shape.assert_has_rank(2) activations_dtype = real_activations.dtype if activations_dtype != dtypes.float64: real_activations = math_ops.to_double(real_activations) generated_activations = math_ops.to_double(generated_activations) # Compute mean and covariance matrices of activations. m = math_ops.reduce_mean(real_activations, 0) m_w = math_ops.reduce_mean(generated_activations, 0) num_examples_real = math_ops.to_double(array_ops.shape(real_activations)[0]) num_examples_generated = math_ops.to_double( array_ops.shape(generated_activations)[0]) # sigma = (1 / (n - 1)) * (X - mu) (X - mu)^T real_centered = real_activations - m sigma = math_ops.matmul( real_centered, real_centered, transpose_a=True) / ( num_examples_real - 1) gen_centered = generated_activations - m_w sigma_w = math_ops.matmul( gen_centered, gen_centered, transpose_a=True) / ( num_examples_generated - 1) # Find the Tr(sqrt(sigma sigma_w)) component of FID sqrt_trace_component = trace_sqrt_product(sigma, sigma_w) # Compute the two components of FID. # First the covariance component. # Here, note that trace(A + B) = trace(A) + trace(B) trace = math_ops.trace(sigma + sigma_w) - 2.0 * sqrt_trace_component # Next the distance between means. mean = math_ops.reduce_sum( math_ops.squared_difference(m, m_w)) # Equivalent to L2 but more stable. fid = trace + mean if activations_dtype != dtypes.float64: fid = math_ops.cast(fid, activations_dtype) return fid frechet_inception_distance = functools.partial( frechet_classifier_distance, classifier_fn=functools.partial( run_inception, output_tensor=INCEPTION_FINAL_POOL)) def kernel_classifier_distance(real_images, generated_images, classifier_fn, num_classifier_batches=1, max_block_size=1024, dtype=None): """Kernel "classifier" distance for evaluating a generative model. This is based on the Kernel Inception distance, but for an arbitrary embedding. This technique is described in detail in https://arxiv.org/abs/1801.01401. Given two distributions P and Q of activations, this function calculates E_{X, X' ~ P}[k(X, X')] + E_{Y, Y' ~ Q}[k(Y, Y')] - 2 E_{X ~ P, Y ~ Q}[k(X, Y)] where k is the polynomial kernel k(x, y) = ( x^T y / dimension + 1 )^3. This captures how different the distributions of real and generated images' visual features are. Like the Frechet distance (and unlike the Inception score), this is a true distance and incorporates information about the target images. Unlike the Frechet score, this function computes an *unbiased* and asymptotically normal estimator, which makes comparing estimates across models much more intuitive. The estimator used takes time quadratic in max_block_size. Larger values of max_block_size will decrease the variance of the estimator but increase the computational cost. This differs slightly from the estimator used by the original paper; it is the block estimator of https://arxiv.org/abs/1307.1954. NOTE: the blocking code assumes that real_activations and generated_activations are both in random order. If either is sorted in a meaningful order, the estimator will behave poorly. NOTE: This function consumes images, computes their activations, and then computes the classifier score. If you would like to precompute many activations for real and generated images for large batches, or to compute multiple scores based on the same images, please use kernel_clasifier_distance_from_activations(), which this method also uses. Args: real_images: Real images to use to compute Kernel Inception distance. generated_images: Generated images to use to compute Kernel Inception distance. classifier_fn: A function that takes images and produces activations based on a classifier. num_classifier_batches: Number of batches to split images in to in order to efficiently run them through the classifier network. max_estimator_block_size: integer, default 1024. The distance estimator splits samples into blocks for computational efficiency. Larger values are more computationally expensive but decrease the variance of the distance estimate. dtype: if not None, coerce activations to this dtype before computations. Returns: The Kernel Inception Distance. A floating-point scalar of the same type as the output of the activations. """ return kernel_classifier_distance_and_std( real_images, generated_images, classifier_fn, num_classifier_batches=num_classifier_batches, max_block_size=max_block_size, dtype=dtype)[0] kernel_inception_distance = functools.partial( kernel_classifier_distance, classifier_fn=functools.partial( run_inception, output_tensor=INCEPTION_FINAL_POOL)) def kernel_classifier_distance_and_std(real_images, generated_images, classifier_fn, num_classifier_batches=1, max_block_size=1024, dtype=None): """Kernel "classifier" distance for evaluating a generative model. This is based on the Kernel Inception distance, but for an arbitrary embedding. Also returns an estimate of the standard error of the distance estimator. This technique is described in detail in https://arxiv.org/abs/1801.01401. Given two distributions P and Q of activations, this function calculates E_{X, X' ~ P}[k(X, X')] + E_{Y, Y' ~ Q}[k(Y, Y')] - 2 E_{X ~ P, Y ~ Q}[k(X, Y)] where k is the polynomial kernel k(x, y) = ( x^T y / dimension + 1 )^3. This captures how different the distributions of real and generated images' visual features are. Like the Frechet distance (and unlike the Inception score), this is a true distance and incorporates information about the target images. Unlike the Frechet score, this function computes an *unbiased* and asymptotically normal estimator, which makes comparing estimates across models much more intuitive. The estimator used takes time quadratic in max_block_size. Larger values of max_block_size will decrease the variance of the estimator but increase the computational cost. This differs slightly from the estimator used by the original paper; it is the block estimator of https://arxiv.org/abs/1307.1954. NOTE: the blocking code assumes that real_activations and generated_activations are both in random order. If either is sorted in a meaningful order, the estimator will behave poorly. NOTE: This function consumes images, computes their activations, and then computes the classifier score. If you would like to precompute many activations for real and generated images for large batches, or to compute multiple scores based on the same images, please use kernel_clasifier_distance_from_activations(), which this method also uses. Args: real_images: Real images to use to compute Kernel Inception distance. generated_images: Generated images to use to compute Kernel Inception distance. classifier_fn: A function that takes images and produces activations based on a classifier. num_classifier_batches: Number of batches to split images in to in order to efficiently run them through the classifier network. max_estimator_block_size: integer, default 1024. The distance estimator splits samples into blocks for computational efficiency. Larger values are more computationally expensive but decrease the variance of the distance estimate. Having a smaller block size also gives a better estimate of the standard error. dtype: if not None, coerce activations to this dtype before computations. Returns: The Kernel Inception Distance. A floating-point scalar of the same type as the output of the activations. An estimate of the standard error of the distance estimator (a scalar of the same type). """ real_images_list = array_ops.split( real_images, num_or_size_splits=num_classifier_batches) generated_images_list = array_ops.split( generated_images, num_or_size_splits=num_classifier_batches) real_imgs = array_ops.stack(real_images_list) generated_imgs = array_ops.stack(generated_images_list) # Compute the activations using the memory-efficient `map_fn`. def compute_activations(elems): return functional_ops.map_fn( fn=classifier_fn, elems=elems, parallel_iterations=1, back_prop=False, swap_memory=True, name='RunClassifier') real_a = compute_activations(real_imgs) gen_a = compute_activations(generated_imgs) # Ensure the activations have the right shapes. real_a = array_ops.concat(array_ops.unstack(real_a), 0) gen_a = array_ops.concat(array_ops.unstack(gen_a), 0) return kernel_classifier_distance_and_std_from_activations( real_a, gen_a, max_block_size=max_block_size) kernel_inception_distance_and_std = functools.partial( kernel_classifier_distance_and_std, classifier_fn=functools.partial( run_inception, output_tensor=INCEPTION_FINAL_POOL)) def kernel_classifier_distance_from_activations(real_activations, generated_activations, max_block_size=1024, dtype=None): """Kernel "classifier" distance for evaluating a generative model. This methods computes the kernel classifier distance from activations of real images and generated images. This can be used independently of the kernel_classifier_distance() method, especially in the case of using large batches during evaluation where we would like to precompute all of the activations before computing the classifier distance, or if we want to compute multiple metrics based on the same images. This technique is described in detail in https://arxiv.org/abs/1801.01401. Given two distributions P and Q of activations, this function calculates E_{X, X' ~ P}[k(X, X')] + E_{Y, Y' ~ Q}[k(Y, Y')] - 2 E_{X ~ P, Y ~ Q}[k(X, Y)] where k is the polynomial kernel k(x, y) = ( x^T y / dimension + 1 )^3. This captures how different the distributions of real and generated images' visual features are. Like the Frechet distance (and unlike the Inception score), this is a true distance and incorporates information about the target images. Unlike the Frechet score, this function computes an *unbiased* and asymptotically normal estimator, which makes comparing estimates across models much more intuitive. The estimator used takes time quadratic in max_block_size. Larger values of max_block_size will decrease the variance of the estimator but increase the computational cost. This differs slightly from the estimator used by the original paper; it is the block estimator of https://arxiv.org/abs/1307.1954. NOTE: the blocking code assumes that real_activations and generated_activations are both in random order. If either is sorted in a meaningful order, the estimator will behave poorly. Args: real_activations: 2D Tensor containing activations of real data. Shape is [batch_size, activation_size]. generated_activations: 2D Tensor containing activations of generated data. Shape is [batch_size, activation_size]. max_block_size: integer, default 1024. The distance estimator splits samples into blocks for computational efficiency. Larger values are more computationally expensive but decrease the variance of the distance estimate. dtype: if not None, coerce activations to this dtype before computations. Returns: The Kernel Inception Distance. A floating-point scalar of the same type as the output of the activations. """ return kernel_classifier_distance_and_std_from_activations( real_activations, generated_activations, max_block_size=max_block_size)[0] def kernel_classifier_distance_and_std_from_activations(real_activations, generated_activations, max_block_size=1024, dtype=None): """Kernel "classifier" distance for evaluating a generative model. This methods computes the kernel classifier distance from activations of real images and generated images. This can be used independently of the kernel_classifier_distance() method, especially in the case of using large batches during evaluation where we would like to precompute all of the activations before computing the classifier distance, or if we want to compute multiple metrics based on the same images. It also returns a rough estimate of the standard error of the estimator. This technique is described in detail in https://arxiv.org/abs/1801.01401. Given two distributions P and Q of activations, this function calculates E_{X, X' ~ P}[k(X, X')] + E_{Y, Y' ~ Q}[k(Y, Y')] - 2 E_{X ~ P, Y ~ Q}[k(X, Y)] where k is the polynomial kernel k(x, y) = ( x^T y / dimension + 1 )^3. This captures how different the distributions of real and generated images' visual features are. Like the Frechet distance (and unlike the Inception score), this is a true distance and incorporates information about the target images. Unlike the Frechet score, this function computes an *unbiased* and asymptotically normal estimator, which makes comparing estimates across models much more intuitive. The estimator used takes time quadratic in max_block_size. Larger values of max_block_size will decrease the variance of the estimator but increase the computational cost. This differs slightly from the estimator used by the original paper; it is the block estimator of https://arxiv.org/abs/1307.1954. The estimate of the standard error will also be more reliable when there are more blocks, i.e. when max_block_size is smaller. NOTE: the blocking code assumes that real_activations and generated_activations are both in random order. If either is sorted in a meaningful order, the estimator will behave poorly. Args: real_activations: 2D Tensor containing activations of real data. Shape is [batch_size, activation_size]. generated_activations: 2D Tensor containing activations of generated data. Shape is [batch_size, activation_size]. max_block_size: integer, default 1024. The distance estimator splits samples into blocks for computational efficiency. Larger values are more computationally expensive but decrease the variance of the distance estimate. Having a smaller block size also gives a better estimate of the standard error. dtype: if not None, coerce activations to this dtype before computations. Returns: The Kernel Inception Distance. A floating-point scalar of the same type as the output of the activations. An estimate of the standard error of the distance estimator (a scalar of the same type). """ real_activations.shape.assert_has_rank(2) generated_activations.shape.assert_has_rank(2) real_activations.shape[1].assert_is_compatible_with( generated_activations.shape[1]) if dtype is None: dtype = real_activations.dtype assert generated_activations.dtype == dtype else: real_activations = math_ops.cast(real_activations, dtype) generated_activations = math_ops.cast(generated_activations, dtype) # Figure out how to split the activations into blocks of approximately # equal size, with none larger than max_block_size. n_r = array_ops.shape(real_activations)[0] n_g = array_ops.shape(generated_activations)[0] n_bigger = math_ops.maximum(n_r, n_g) n_blocks = math_ops.to_int32(math_ops.ceil(n_bigger / max_block_size)) v_r = n_r // n_blocks v_g = n_g // n_blocks n_plusone_r = n_r - v_r * n_blocks n_plusone_g = n_g - v_g * n_blocks sizes_r = array_ops.concat([ array_ops.fill([n_blocks - n_plusone_r], v_r), array_ops.fill([n_plusone_r], v_r + 1), ], 0) sizes_g = array_ops.concat([ array_ops.fill([n_blocks - n_plusone_g], v_g), array_ops.fill([n_plusone_g], v_g + 1), ], 0) zero = array_ops.zeros([1], dtype=dtypes.int32) inds_r = array_ops.concat([zero, math_ops.cumsum(sizes_r)], 0) inds_g = array_ops.concat([zero, math_ops.cumsum(sizes_g)], 0) dim = math_ops.cast(real_activations.shape[1], dtype) def compute_kid_block(i): 'Compute the ith block of the KID estimate.' r_s = inds_r[i] r_e = inds_r[i + 1] r = real_activations[r_s:r_e] m = math_ops.cast(r_e - r_s, dtype) g_s = inds_g[i] g_e = inds_g[i + 1] g = generated_activations[g_s:g_e] n = math_ops.cast(g_e - g_s, dtype) k_rr = (math_ops.matmul(r, r, transpose_b=True) / dim + 1)**3 k_rg = (math_ops.matmul(r, g, transpose_b=True) / dim + 1)**3 k_gg = (math_ops.matmul(g, g, transpose_b=True) / dim + 1)**3 return (-2 * math_ops.reduce_mean(k_rg) + (math_ops.reduce_sum(k_rr) - math_ops.trace(k_rr)) / (m * (m - 1)) + (math_ops.reduce_sum(k_gg) - math_ops.trace(k_gg)) / (n * (n - 1))) ests = functional_ops.map_fn( compute_kid_block, math_ops.range(n_blocks), dtype=dtype, back_prop=False) mn = math_ops.reduce_mean(ests) # nn_impl.moments doesn't use the Bessel correction, which we want here n_blocks_ = math_ops.cast(n_blocks, dtype) var = control_flow_ops.cond( math_ops.less_equal(n_blocks, 1), lambda: array_ops.constant(float('nan'), dtype=dtype), lambda: math_ops.reduce_sum(math_ops.square(ests - mn)) / (n_blocks_ - 1)) return mn, math_ops.sqrt(var / n_blocks_)