1# Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================== 15"""Weight broadcasting operations. 16 17In `tf.losses` and `tf.metrics`, we support limited weight broadcasting. This 18file includes operations for those broadcasting rules. 19""" 20 21from __future__ import absolute_import 22from __future__ import division 23from __future__ import print_function 24 25from tensorflow.python.framework import ops 26from tensorflow.python.framework import tensor_util 27from tensorflow.python.ops import array_ops 28from tensorflow.python.ops import control_flow_ops 29from tensorflow.python.ops import math_ops 30from tensorflow.python.ops import sets 31 32 33def _has_valid_dims(weights_shape, values_shape): 34 with ops.name_scope( 35 None, "has_invalid_dims", (weights_shape, values_shape)) as scope: 36 values_shape_2d = array_ops.expand_dims(values_shape, -1) 37 valid_dims = array_ops.concat( 38 (values_shape_2d, array_ops.ones_like(values_shape_2d)), axis=1) 39 weights_shape_2d = array_ops.expand_dims(weights_shape, -1) 40 invalid_dims = sets.set_difference(weights_shape_2d, valid_dims) 41 num_invalid_dims = array_ops.size( 42 invalid_dims.values, name="num_invalid_dims") 43 return math_ops.equal(0, num_invalid_dims, name=scope) 44 45 46def _has_valid_nonscalar_shape( 47 weights_rank, weights_shape, values_rank, values_shape): 48 with ops.name_scope( 49 None, "has_valid_nonscalar_shape", 50 (weights_rank, weights_shape, values_rank, values_shape)) as scope: 51 is_same_rank = math_ops.equal( 52 values_rank, weights_rank, name="is_same_rank") 53 return control_flow_ops.cond( 54 is_same_rank, 55 lambda: _has_valid_dims(weights_shape, values_shape), 56 lambda: is_same_rank, 57 name=scope) 58 59 60_ASSERT_BROADCASTABLE_ERROR_PREFIX = "weights can not be broadcast to values." 61 62 63def assert_broadcastable(weights, values): 64 """Asserts `weights` can be broadcast to `values`. 65 66 In `tf.losses` and `tf.metrics`, we support limited weight broadcasting. We 67 let weights be either scalar, or the same rank as the target values, with each 68 dimension either 1, or the same as the corresponding values dimension. 69 70 Args: 71 weights: `Tensor` of weights. 72 values: `Tensor` of values to which weights are applied. 73 74 Returns: 75 `Operation` raising `InvalidArgumentError` if `weights` has incorrect shape. 76 `no_op` if static checks determine `weights` has correct shape. 77 78 Raises: 79 ValueError: If static checks determine `weights` has incorrect shape. 80 """ 81 with ops.name_scope(None, "assert_broadcastable", (weights, values)) as scope: 82 with ops.name_scope(None, "weights", (weights,)) as weights_scope: 83 weights = ops.convert_to_tensor(weights, name=weights_scope) 84 weights_shape = array_ops.shape(weights, name="shape") 85 weights_rank = array_ops.rank(weights, name="rank") 86 weights_rank_static = tensor_util.constant_value(weights_rank) 87 88 with ops.name_scope(None, "values", (values,)) as values_scope: 89 values = ops.convert_to_tensor(values, name=values_scope) 90 values_shape = array_ops.shape(values, name="shape") 91 values_rank = array_ops.rank(values, name="rank") 92 values_rank_static = tensor_util.constant_value(values_rank) 93 94 # Try static checks. 95 if weights_rank_static is not None and values_rank_static is not None: 96 if weights_rank_static == 0: 97 return control_flow_ops.no_op(name="static_scalar_check_success") 98 if weights_rank_static != values_rank_static: 99 raise ValueError( 100 "%s values.rank=%s. weights.rank=%s." 101 " values.shape=%s. weights.shape=%s." % ( 102 _ASSERT_BROADCASTABLE_ERROR_PREFIX, values_rank_static, 103 weights_rank_static, values.shape, weights.shape)) 104 weights_shape_static = tensor_util.constant_value(weights_shape) 105 values_shape_static = tensor_util.constant_value(values_shape) 106 if weights_shape_static is not None and values_shape_static is not None: 107 # Sanity check, this should always be true since we checked rank above. 108 ndims = len(values_shape_static) 109 assert ndims == len(weights_shape_static) 110 111 for i in range(ndims): 112 if weights_shape_static[i] not in (1, values_shape_static[i]): 113 raise ValueError( 114 "%s Mismatch at dim %s. values.shape=%s weights.shape=%s." % ( 115 _ASSERT_BROADCASTABLE_ERROR_PREFIX, i, values_shape_static, 116 weights_shape_static)) 117 return control_flow_ops.no_op(name="static_dims_check_success") 118 119 # Dynamic checks. 120 is_scalar = math_ops.equal(0, weights_rank, name="is_scalar") 121 data = ( 122 _ASSERT_BROADCASTABLE_ERROR_PREFIX, 123 "weights.shape=", weights.name, weights_shape, 124 "values.shape=", values.name, values_shape, 125 "is_scalar=", is_scalar, 126 ) 127 is_valid_shape = control_flow_ops.cond( 128 is_scalar, 129 lambda: is_scalar, 130 lambda: _has_valid_nonscalar_shape( # pylint: disable=g-long-lambda 131 weights_rank, weights_shape, values_rank, values_shape), 132 name="is_valid_shape") 133 return control_flow_ops.Assert(is_valid_shape, data, name=scope) 134 135 136def broadcast_weights(weights, values): 137 """Broadcast `weights` to the same shape as `values`. 138 139 This returns a version of `weights` following the same broadcast rules as 140 `mul(weights, values)`, but limited to the weights shapes allowed by 141 `assert_broadcastable`. When computing a weighted average, use this function 142 to broadcast `weights` before summing them; e.g., 143 `reduce_sum(w * v) / reduce_sum(_broadcast_weights(w, v))`. 144 145 Args: 146 weights: `Tensor` whose shape is broadcastable to `values` according to the 147 rules of `assert_broadcastable`. 148 values: `Tensor` of any shape. 149 150 Returns: 151 `weights` broadcast to `values` shape according to the rules of 152 `assert_broadcastable`. 153 """ 154 with ops.name_scope(None, "broadcast_weights", (weights, values)) as scope: 155 values = ops.convert_to_tensor(values, name="values") 156 weights = ops.convert_to_tensor( 157 weights, dtype=values.dtype.base_dtype, name="weights") 158 159 # Try static check for exact match. 160 weights_shape = weights.get_shape() 161 values_shape = values.get_shape() 162 if (weights_shape.is_fully_defined() and 163 values_shape.is_fully_defined() and 164 weights_shape.is_compatible_with(values_shape)): 165 return weights 166 167 with ops.control_dependencies((assert_broadcastable(weights, values),)): 168 return math_ops.multiply( 169 weights, array_ops.ones_like(values), name=scope) 170