# Copyright 2018 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. # ============================================================================== """Gradients for operators defined in random_ops.py.""" from __future__ import absolute_import from __future__ import division from __future__ import print_function import numpy as np from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import clip_ops from tensorflow.python.ops import gen_array_ops from tensorflow.python.ops import gen_random_ops from tensorflow.python.ops import math_ops def add_leading_unit_dimensions(x, num_dimensions): # pylint: disable=invalid-name new_shape = array_ops.concat( [array_ops.ones([num_dimensions], dtype=dtypes.int32), array_ops.shape(x)], axis=0) return array_ops.reshape(x, new_shape) @ops.RegisterGradient("RandomGamma") def _RandomGammaGrad(op, grad): # pylint: disable=invalid-name """Returns the gradient of a Gamma sample w.r.t. alpha. The gradient is computed using implicit differentiation (Figurnov et al., 2018). Args: op: A `RandomGamma` operation. We assume that the inputs to the operation are `shape` and `alpha` tensors, and the output is the `sample` tensor. grad: The incoming gradient `dloss / dsample` of the same shape as `op.outputs[0]`. Returns: A `Tensor` with derivatives `dloss / dalpha`. References: Implicit Reparameterization Gradients: [Figurnov et al., 2018] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients) ([pdf] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients.pdf)) """ shape = op.inputs[0] alpha = op.inputs[1] sample = op.outputs[0] with ops.control_dependencies([grad]): # Make the parameters alpha broadcastable with samples by appending # unit dimensions. num_sample_dimensions = array_ops.shape(shape)[0] alpha_broadcastable = add_leading_unit_dimensions( alpha, num_sample_dimensions) partial_a = gen_random_ops.random_gamma_grad(alpha_broadcastable, sample) # The first input is shape; the second input is alpha. return (None, math_ops.reduce_sum( grad * partial_a, axis=math_ops.range(num_sample_dimensions))) @ops.RegisterGradient("StatelessRandomGammaV2") def _StatelessRandomGammaV2Grad(op, grad): # pylint: disable=invalid-name """Returns the gradient of a Gamma sample w.r.t. alpha. The gradient is computed using implicit differentiation (Figurnov et al., 2018). Args: op: A `StatelessRandomGamma` operation. We assume that the inputs to the operation are `shape`, `seed` and `alpha` tensors, and the output is the `sample` tensor. grad: The incoming gradient `dloss / dsample` of the same shape as `op.outputs[0]`. Returns: A `Tensor` with derivatives `dloss / dalpha`. References: Implicit Reparameterization Gradients: [Figurnov et al., 2018] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients) ([pdf] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients.pdf)) """ shape = op.inputs[0] alpha = op.inputs[2] sample = op.outputs[0] with ops.control_dependencies([grad]): # Note that the shape handling is slightly different for stateless_gamma, # in particular num_sample_dimensions is different. num_sample_dimensions = array_ops.shape(shape)[0] - array_ops.rank(alpha) # Make the parameters alpha broadcastable with samples by appending # unit dimensions. alpha_broadcastable = add_leading_unit_dimensions(alpha, num_sample_dimensions) partial_a = gen_random_ops.random_gamma_grad(alpha_broadcastable, sample) # The first two inputs are shape, seed, third input is alpha. return (None, None, math_ops.reduce_sum( grad * partial_a, axis=math_ops.range(num_sample_dimensions))) def _Ndtr(x): """Normal distribution function.""" half_sqrt_2 = constant_op.constant( 0.5 * np.sqrt(2.), dtype=x.dtype, name="half_sqrt_2") w = x * half_sqrt_2 z = math_ops.abs(w) y = array_ops.where( z < half_sqrt_2, 1. + math_ops.erf(w), array_ops.where( w > 0., 2. - math_ops.erfc(z), math_ops.erfc(z))) return 0.5 * y @ops.RegisterGradient("StatelessParameterizedTruncatedNormal") def _StatelessParameterizedTruncatedNormalGrad(op, grad): # pylint: disable=invalid-name """Returns the gradient of a TruncatedNormal sample w.r.t. parameters. The gradient is computed using implicit differentiation (Figurnov et al., 2018). Args: op: A `StatelessParameterizedTruncatedNormal` operation. We assume that the inputs to the operation are `shape`, `seed`, `mean`, `stddev`, `minval`, and `maxval` tensors, and the output is the `sample` tensor. grad: The incoming gradient `dloss / dsample` of the same shape as `op.outputs[0]`. Returns: A list of `Tensor` with derivates with respect to each parameter. References: Implicit Reparameterization Gradients: [Figurnov et al., 2018] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients) ([pdf] (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients.pdf)) """ shape = op.inputs[0] mean = op.inputs[2] stddev = op.inputs[3] minval = op.inputs[4] maxval = op.inputs[5] sample = op.outputs[0] with ops.control_dependencies([grad]): minval_std = (minval - mean) / stddev maxval_std = (maxval - mean) / stddev sample_std = (sample - mean) / stddev cdf_sample = (_Ndtr(sample_std) - _Ndtr(minval_std)) / ( _Ndtr(maxval_std) - _Ndtr(minval_std)) # Clip to avoid zero argument for log_cdf expression tiny = np.finfo(mean.dtype.as_numpy_dtype).tiny eps = np.finfo(mean.dtype.as_numpy_dtype).eps cdf_sample = clip_ops.clip_by_value(cdf_sample, tiny, 1 - eps) dmaxval = math_ops.exp(0.5 * (sample_std ** 2 - maxval_std ** 2) + math_ops.log(cdf_sample)) dminval = math_ops.exp(0.5 * (sample_std ** 2 - minval_std ** 2) + math_ops.log1p(-cdf_sample)) dmean = array_ops.ones_like(sample_std) dstddev = sample_std # Reduce over extra dimensions caused by `shape`. We need to get the # difference in rank from shape vs. the broadcasted rank. mean_shape = array_ops.shape(mean) stddev_shape = array_ops.shape(stddev) minval_shape = array_ops.shape(minval) maxval_shape = array_ops.shape(maxval) broadcast_shape = array_ops.broadcast_dynamic_shape( mean_shape, stddev_shape) broadcast_shape = array_ops.broadcast_dynamic_shape( minval_shape, broadcast_shape) broadcast_shape = array_ops.broadcast_dynamic_shape( maxval_shape, broadcast_shape) extra_dims = math_ops.range( array_ops.size(shape) - array_ops.size(broadcast_shape)) grad_mean = math_ops.reduce_sum(grad * dmean, axis=extra_dims) grad_stddev = math_ops.reduce_sum(grad * dstddev, axis=extra_dims) grad_minval = math_ops.reduce_sum(grad * dminval, axis=extra_dims) grad_maxval = math_ops.reduce_sum(grad * dmaxval, axis=extra_dims) _, rmean = gen_array_ops.broadcast_gradient_args( broadcast_shape, mean_shape) _, rstddev = gen_array_ops.broadcast_gradient_args( broadcast_shape, stddev_shape) _, rminval = gen_array_ops.broadcast_gradient_args( broadcast_shape, minval_shape) _, rmaxval = gen_array_ops.broadcast_gradient_args( broadcast_shape, maxval_shape) grad_mean = array_ops.reshape( math_ops.reduce_sum(grad_mean, axis=rmean, keepdims=True), mean_shape) grad_stddev = array_ops.reshape( math_ops.reduce_sum(grad_stddev, axis=rstddev, keepdims=True), stddev_shape) grad_minval = array_ops.reshape( math_ops.reduce_sum(grad_minval, axis=rminval, keepdims=True), minval_shape) grad_maxval = array_ops.reshape( math_ops.reduce_sum(grad_maxval, axis=rmaxval, keepdims=True), maxval_shape) # The first two inputs are shape. return (None, None, grad_mean, grad_stddev, grad_minval, grad_maxval)