TestMu Conference is Back! Future of Testing + AIRegister for Free

How to use log_kernel method in autotest

Best Python code snippet using autotest_python

svgd.py

Source:svgd.py

1# Copyright (c) 2017-2019 Uber Technologies, Inc.2# SPDX-License-Identifier: Apache-2.03import math4from abc import ABCMeta, abstractmethod5import torch6from torch.distributions import biject_to7import pyro8from pyro import poutine9from pyro.distributions import Delta10from pyro.distributions.util import copy_docs_from11from pyro.infer.autoguide.guides import AutoContinuous12from pyro.infer.autoguide.initialization import init_to_sample13from pyro.infer.trace_elbo import Trace_ELBO14def vectorize(fn, num_particles, max_plate_nesting):15    def _fn(*args, **kwargs):16        with pyro.plate(17            "num_particles_vectorized", num_particles, dim=-max_plate_nesting - 118        ):19            return fn(*args, **kwargs)20    return _fn21class _SVGDGuide(AutoContinuous):22    """23    This modification of :class:`AutoContinuous` is used internally in the24    :class:`SVGD` inference algorithm.25    """26    def __init__(self, model):27        super().__init__(model, init_loc_fn=init_to_sample)28    def get_posterior(self, *args, **kwargs):29        svgd_particles = pyro.param("svgd_particles", self._init_loc)30        return Delta(svgd_particles, event_dim=1)31class SteinKernel(object, metaclass=ABCMeta):32    """33    Abstract class for kernels used in the :class:`SVGD` inference algorithm.34    """35    @abstractmethod36    def log_kernel_and_grad(self, particles):37        """38        Compute the component kernels and their gradients.39        :param particles: a tensor with shape (N, D)40        :returns: A pair (`log_kernel`, `kernel_grad`) where `log_kernel` is a (N, N, D)-shaped41            tensor equal to the logarithm of the kernel and `kernel_grad` is a (N, N, D)-shaped42            tensor where the entry (n, m, d) represents the derivative of `log_kernel` w.r.t.43            x_{m,d}, where x_{m,d} is the d^th dimension of particle m.44        """45        raise NotImplementedError46@copy_docs_from(SteinKernel)47class RBFSteinKernel(SteinKernel):48    """49    A RBF kernel for use in the SVGD inference algorithm. The bandwidth of the kernel is chosen from the50    particles using a simple heuristic as in reference [1].51    :param float bandwidth_factor: Optional factor by which to scale the bandwidth, defaults to 1.0.52    :ivar float ~.bandwidth_factor: Property that controls the factor by which to scale the bandwidth53        at each iteration.54    References55    [1] "Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm,"56        Qiang Liu, Dilin Wang57    """58    def __init__(self, bandwidth_factor=None):59        """60        :param float bandwidth_factor: Optional factor by which to scale the bandwidth61        """62        self.bandwidth_factor = bandwidth_factor63    def _bandwidth(self, norm_sq):64        """65        Compute the bandwidth along each dimension using the median pairwise squared distance between particles.66        """67        num_particles = norm_sq.size(0)68        index = torch.arange(num_particles)69        norm_sq = norm_sq[index > index.unsqueeze(-1), ...]70        median = norm_sq.median(dim=0)[0]71        if self.bandwidth_factor is not None:72            median = self.bandwidth_factor * median73        assert median.shape == norm_sq.shape[-1:]74        return median / math.log(num_particles + 1)75    @torch.no_grad()76    def log_kernel_and_grad(self, particles):77        delta_x = particles.unsqueeze(0) - particles.unsqueeze(1)  # N N D78        assert delta_x.dim() == 379        norm_sq = delta_x.pow(2.0)  # N N D80        h = self._bandwidth(norm_sq)  # D81        log_kernel = -(norm_sq / h)  # N N D82        grad_term = 2.0 * delta_x / h  # N N D83        assert log_kernel.shape == grad_term.shape84        return log_kernel, grad_term85    @property86    def bandwidth_factor(self):87        return self._bandwidth_factor88    @bandwidth_factor.setter89    def bandwidth_factor(self, bandwidth_factor):90        """91        :param float bandwidth_factor: Optional factor by which to scale the bandwidth92        """93        if bandwidth_factor is not None:94            assert bandwidth_factor > 0.0, "bandwidth_factor must be positive."95        self._bandwidth_factor = bandwidth_factor96@copy_docs_from(SteinKernel)97class IMQSteinKernel(SteinKernel):98    r"""99    An IMQ (inverse multi-quadratic) kernel for use in the SVGD inference algorithm [1]. The bandwidth of the kernel100    is chosen from the particles using a simple heuristic as in reference [2]. The kernel takes the form101    :math:`K(x, y) = (\alpha + ||x-y||^2/h)^{\beta}`102    where :math:`\alpha` and :math:`\beta` are user-specified parameters and :math:`h` is the bandwidth.103    :param float alpha: Kernel hyperparameter, defaults to 0.5.104    :param float beta: Kernel hyperparameter, defaults to -0.5.105    :param float bandwidth_factor: Optional factor by which to scale the bandwidth, defaults to 1.0.106    :ivar float ~.bandwidth_factor: Property that controls the factor by which to scale the bandwidth107        at each iteration.108    References109    [1] "Stein Points," Wilson Ye Chen, Lester Mackey, Jackson Gorham, Francois-Xavier Briol, Chris. J. Oates.110    [2] "Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm," Qiang Liu, Dilin Wang111    """112    def __init__(self, alpha=0.5, beta=-0.5, bandwidth_factor=None):113        """114        :param float alpha: Kernel hyperparameter, defaults to 0.5.115        :param float beta: Kernel hyperparameter, defaults to -0.5.116        :param float bandwidth_factor: Optional factor by which to scale the bandwidth117        """118        assert alpha > 0.0, "alpha must be positive."119        assert beta < 0.0, "beta must be negative."120        self.alpha = alpha121        self.beta = beta122        self.bandwidth_factor = bandwidth_factor123    def _bandwidth(self, norm_sq):124        """125        Compute the bandwidth along each dimension using the median pairwise squared distance between particles.126        """127        num_particles = norm_sq.size(0)128        index = torch.arange(num_particles)129        norm_sq = norm_sq[index > index.unsqueeze(-1), ...]130        median = norm_sq.median(dim=0)[0]131        if self.bandwidth_factor is not None:132            median = self.bandwidth_factor * median133        assert median.shape == norm_sq.shape[-1:]134        return median / math.log(num_particles + 1)135    @torch.no_grad()136    def log_kernel_and_grad(self, particles):137        delta_x = particles.unsqueeze(0) - particles.unsqueeze(1)  # N N D138        assert delta_x.dim() == 3139        norm_sq = delta_x.pow(2.0)  # N N D140        h = self._bandwidth(norm_sq)  # D141        base_term = self.alpha + norm_sq / h142        log_kernel = self.beta * torch.log(base_term)  # N N D143        grad_term = (-2.0 * self.beta) * delta_x / h  # N N D144        grad_term = grad_term / base_term145        assert log_kernel.shape == grad_term.shape146        return log_kernel, grad_term147    @property148    def bandwidth_factor(self):149        return self._bandwidth_factor150    @bandwidth_factor.setter151    def bandwidth_factor(self, bandwidth_factor):152        """153        :param float bandwidth_factor: Optional factor by which to scale the bandwidth154        """155        if bandwidth_factor is not None:156            assert bandwidth_factor > 0.0, "bandwidth_factor must be positive."157        self._bandwidth_factor = bandwidth_factor158class SVGD:159    """160    A basic implementation of Stein Variational Gradient Descent as described in reference [1].161    :param model: The model (callable containing Pyro primitives). Model must be fully vectorized162        and may only contain continuous latent variables.163    :param kernel: a SVGD compatible kernel like :class:`RBFSteinKernel`.164    :param optim: A wrapper for a PyTorch optimizer.165    :type optim: pyro.optim.PyroOptim166    :param int num_particles: The number of particles used in SVGD.167    :param int max_plate_nesting: The max number of nested :func:`pyro.plate` contexts in the model.168    :param str mode: Whether to use a Kernelized Stein Discrepancy that makes use of `multivariate`169        test functions (as in [1]) or `univariate` test functions (as in [2]). Defaults to `univariate`.170    Example usage:171    .. code-block:: python172        from pyro.infer import SVGD, RBFSteinKernel173        from pyro.optim import Adam174        kernel = RBFSteinKernel()175        adam = Adam({"lr": 0.1})176        svgd = SVGD(model, kernel, adam, num_particles=50, max_plate_nesting=0)177        for step in range(500):178            svgd.step(model_arg1, model_arg2)179        final_particles = svgd.get_named_particles()180    References181    [1] "Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm,"182        Qiang Liu, Dilin Wang183    [2] "Kernelized Complete Conditional Stein Discrepancy,"184        Raghav Singhal, Saad Lahlou, Rajesh Ranganath185    """186    def __init__(187        self, model, kernel, optim, num_particles, max_plate_nesting, mode="univariate"188    ):189        assert callable(model)190        assert isinstance(kernel, SteinKernel), "Must provide a valid SteinKernel"191        assert isinstance(192            optim, pyro.optim.PyroOptim193        ), "Must provide a valid Pyro optimizer"194        assert num_particles > 1, "Must use at least two particles"195        assert max_plate_nesting >= 0196        assert mode in [197            "univariate",198            "multivariate",199        ], "mode must be one of (univariate, multivariate)"200        self.model = vectorize(model, num_particles, max_plate_nesting)201        self.kernel = kernel202        self.optim = optim203        self.num_particles = num_particles204        self.max_plate_nesting = max_plate_nesting205        self.mode = mode206        self.loss = Trace_ELBO().differentiable_loss207        self.guide = _SVGDGuide(self.model)208    def get_named_particles(self):209        """210        Create a dictionary mapping name to vectorized value, of the form ``{name: tensor}``.211        The leading dimension of each tensor corresponds to particles, i.e. this creates a struct of arrays.212        """213        return {214            site["name"]: biject_to(site["fn"].support)(unconstrained_value)215            for site, unconstrained_value in self.guide._unpack_latent(216                pyro.param("svgd_particles")217            )218        }219    @torch.no_grad()220    def step(self, *args, **kwargs):221        """222        Computes the SVGD gradient, passing args and kwargs to the model,223        and takes a gradient step.224        :return dict: A dictionary of the form {name: float}, where each float225            is a mean squared gradient. This can be used to monitor the convergence of SVGD.226        """227        # compute gradients of log model joint228        with torch.enable_grad(), poutine.trace(param_only=True) as param_capture:229            loss = self.loss(self.model, self.guide, *args, **kwargs)230            loss.backward()231        # get particles used in the _SVGDGuide and reshape to have num_particles leading dimension232        particles = pyro.param("svgd_particles").unconstrained()233        reshaped_particles = particles.reshape(self.num_particles, -1)234        reshaped_particles_grad = particles.grad.reshape(self.num_particles, -1)235        # compute kernel ingredients236        log_kernel, kernel_grad = self.kernel.log_kernel_and_grad(reshaped_particles)237        if self.mode == "multivariate":238            kernel = log_kernel.sum(-1).exp()239            assert kernel.shape == (self.num_particles, self.num_particles)240            attractive_grad = torch.mm(kernel, reshaped_particles_grad)241            repulsive_grad = torch.einsum("nm,nm...->n...", kernel, kernel_grad)242        elif self.mode == "univariate":243            kernel = log_kernel.exp()244            assert kernel.shape == (245                self.num_particles,246                self.num_particles,247                reshaped_particles.size(-1),248            )249            attractive_grad = torch.einsum(250                "nmd,md->nd", kernel, reshaped_particles_grad251            )252            repulsive_grad = torch.einsum("nmd,nmd->nd", kernel, kernel_grad)253        # combine the attractive and repulsive terms in the SVGD gradient254        assert attractive_grad.shape == repulsive_grad.shape255        particles.grad = (attractive_grad + repulsive_grad).reshape(256            particles.shape257        ) / self.num_particles258        # compute per-parameter mean squared gradients259        squared_gradients = {260            site["name"]: value.mean().item()261            for site, value in self.guide._unpack_latent(particles.grad.pow(2.0))262        }263        # torch.optim objects gets instantiated for any params that haven't been seen yet264        params = set(265            site["value"].unconstrained() for site in param_capture.trace.nodes.values()266        )267        self.optim(params)268        # zero gradients269        pyro.infer.util.zero_grads(params)270        # return per-parameter mean squared gradients to user...

losses.py

Source:losses.py

1from keras.losses import binary_crossentropy2import keras.backend as K3import cv24import numpy as np5import tensorflow as tf6def dice_loss(y_true, y_pred):7    #smooth = 1.8    y_true_f = K.flatten(y_true)9    y_pred_f = K.flatten(y_pred)10    intersection = K.sum(y_true_f * y_pred_f)11    return (2. * intersection) / (K.sum(y_true_f) + K.sum(y_pred_f))12def bce_dice_loss(y_true, y_pred):13    return binary_crossentropy(y_true, y_pred) + (1 - dice_loss(y_true, y_pred))14def gauss2D(shape=(3,3),sigma=0.5):15    m, n = [(ss-1.)/2. for ss in shape]16    y, x = np.ogrid[-m:m+1,-n:n+1]17    h = np.exp(-(x*x + y*y) / (2.*sigma*sigma))18    h[h < np.finfo(h.dtype).eps*h.max()] = 019    sumh = h.sum()20    if sumh != 0:21        h /= sumh22    return h23log = np.array([24    [0.0448, 0.0468, 0.0564, 0.0468, 0.0448],25    [0.0468, 0.3167, 0.7146, 0.3167, 0.0468],26    [0.0564, 0.7146, -4.9048, 0.7146, 0.0564],27    [0.0468, 0.3167, 0.7146, 0.3167, 0.0468],28    [0.0448, 0.0468, 0.0564, 0.0468, 0.0448]]).astype(np.float32)29def weights_mask(mask):30    log_kernel = tf.convert_to_tensor(log)31    log_kernel = tf.reshape(log_kernel, [5, 5, 1, 1])32    log_kernel = tf.to_float(log_kernel)33    mask = tf.to_float(mask)34    edges = tf.nn.conv2d(mask, log_kernel, padding='SAME', strides=[1, 1, 1, 1])35    edges = edges > 0.9536    edges = tf.to_float(edges)37    gauss_kernel = tf.convert_to_tensor(gauss2D((5, 5), 2))38    gauss_kernel = tf.reshape(gauss_kernel, [5, 5, 1, 1])39    gauss_kernel = tf.to_float(gauss_kernel)40    return tf.nn.conv2d(edges, gauss_kernel, padding='SAME', strides=[1, 1, 1, 1])41def edges_dice_loss(y_true, y_pred):42    y_true_f = K.flatten(y_true)43    y_pred_f = K.flatten(y_pred)44    weights = K.flatten(weights_mask(y_true))45    intersection = K.sum(y_true_f * y_pred_f)46    eq = tf.equal(y_true_f, y_pred_f)47    eq = tf.to_float(eq)48    weightedEdges = K.sum(eq * weights)49    return (2. * intersection + weightedEdges) / (K.sum(y_true_f) + K.sum(y_pred_f) + K.sum(weights))50def bce_edges(y_true, y_pred):51    return binary_crossentropy(y_true, y_pred) + (1 - edges_dice_loss(y_true, y_pred))52# weight: weighted tensor(same shape with mask image)53def weighted_bce_loss(y_true, y_pred, weight):54    # avoiding overflow55    epsilon = 1e-756    y_pred = K.clip(y_pred, epsilon, 1. - epsilon)57    logit_y_pred = K.log(y_pred / (1. - y_pred))58    # https://www.tensorflow.org/api_docs/python/tf/nn/weighted_cross_entropy_with_logits59    loss = (1. - y_true) * logit_y_pred + (1. + (weight - 1.) * y_true) * \60                                          (K.log(1. + K.exp(-K.abs(logit_y_pred))) + K.maximum(-logit_y_pred, 0.))61    return K.sum(loss) / K.sum(weight)62def weighted_dice_loss(y_true, y_pred, weight):63    smooth = 1.64    w, m1, m2 = weight * weight, y_true, y_pred65    intersection = (m1 * m2)66    score = (2. * K.sum(w * intersection) + smooth) / (K.sum(w * m1) + K.sum(w * m2) + smooth)67    loss = 1. - K.sum(score)68    return loss69def weighted_bce_dice_loss(y_true, y_pred):70    y_true = K.cast(y_true, 'float32')71    y_pred = K.cast(y_pred, 'float32')72    # if we want to get same size of output, kernel size must be odd number73    averaged_mask = K.pool2d(74        y_true, pool_size=(11, 11), strides=(1, 1), padding='same', pool_mode='avg')75    border = K.cast(K.greater(averaged_mask, 0.005), 'float32') * K.cast(K.less(averaged_mask, 0.995), 'float32')76    weight = K.ones_like(averaged_mask)77    w0 = K.sum(weight)78    weight += border * 279    w1 = K.sum(weight)80    weight *= (w0 / w1)81    loss = weighted_bce_loss(y_true, y_pred, weight) + \82           weighted_dice_loss(y_true, y_pred, weight)...

generate_log.py

Source:generate_log.py

1import math2import numpy as np3# Function for calculating the laplacian of the gaussian at a given point and with a given variance4def log(x, y, sigma):5    numerator = ( (y**2)+(x**2)-2*(sigma**2) )6    denominator = ( (2*math.pi*(sigma**6) ))7    exponential = math.exp(-((x**2)+(y**2))/(2*(sigma**2)))8    return numerator*exponential/denominator9def generate_log(sigma):10    size = max(1,2*round(sigma*3)+1)11    w = math.ceil(float(size)*float(sigma))12    # If the dimension is an even number, make it odd13    if(w%2 == 0):14        w = w + 115    log_kernel = []16    w_range = int(math.floor(w/2))17 18    for i in range(-w_range, w_range):19        for j in range(-w_range, w_range):20            log_kernel.append(log(i,j,sigma))21    log_kernel = np.array(log_kernel)22    log_kernel = np.reshape(log_kernel, (w-1,w-1))23    ...

Automation Testing Tutorials

Learn to execute automation testing from scratch with LambdaTest Learning Hub. Right from setting up the prerequisites to run your first automation test, to following best practices and diving deeper into advanced test scenarios. LambdaTest Learning Hubs compile a list of step-by-step guides to help you be proficient with different test automation frameworks i.e. Selenium, Cypress, TestNG etc.