1# Copyright 2018 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# pylint: disable=protected-access 16# pylint: disable=g-classes-have-attributes 17"""Contains the InputSpec class.""" 18 19from tensorflow.python.framework import dtypes 20from tensorflow.python.framework import tensor_shape 21from tensorflow.python.framework import tensor_spec 22from tensorflow.python.keras import backend 23from tensorflow.python.util import nest 24from tensorflow.python.util.tf_export import keras_export 25from tensorflow.python.util.tf_export import tf_export 26 27 28@keras_export('keras.layers.InputSpec') 29@tf_export(v1=['layers.InputSpec']) 30class InputSpec(object): 31 """Specifies the rank, dtype and shape of every input to a layer. 32 33 Layers can expose (if appropriate) an `input_spec` attribute: 34 an instance of `InputSpec`, or a nested structure of `InputSpec` instances 35 (one per input tensor). These objects enable the layer to run input 36 compatibility checks for input structure, input rank, input shape, and 37 input dtype. 38 39 A None entry in a shape is compatible with any dimension, 40 a None shape is compatible with any shape. 41 42 Args: 43 dtype: Expected DataType of the input. 44 shape: Shape tuple, expected shape of the input 45 (may include None for unchecked axes). Includes the batch size. 46 ndim: Integer, expected rank of the input. 47 max_ndim: Integer, maximum rank of the input. 48 min_ndim: Integer, minimum rank of the input. 49 axes: Dictionary mapping integer axes to 50 a specific dimension value. 51 allow_last_axis_squeeze: If True, then allow inputs of rank N+1 as long 52 as the last axis of the input is 1, as well as inputs of rank N-1 53 as long as the last axis of the spec is 1. 54 name: Expected key corresponding to this input when passing data as 55 a dictionary. 56 57 Example: 58 59 ```python 60 class MyLayer(Layer): 61 def __init__(self): 62 super(MyLayer, self).__init__() 63 # The layer will accept inputs with shape (?, 28, 28) & (?, 28, 28, 1) 64 # and raise an appropriate error message otherwise. 65 self.input_spec = InputSpec( 66 shape=(None, 28, 28, 1), 67 allow_last_axis_squeeze=True) 68 ``` 69 """ 70 71 def __init__(self, 72 dtype=None, 73 shape=None, 74 ndim=None, 75 max_ndim=None, 76 min_ndim=None, 77 axes=None, 78 allow_last_axis_squeeze=False, 79 name=None): 80 self.dtype = dtypes.as_dtype(dtype).name if dtype is not None else None 81 shape = tensor_shape.TensorShape(shape) 82 if shape.rank is None: 83 shape = None 84 else: 85 shape = tuple(shape.as_list()) 86 if shape is not None: 87 self.ndim = len(shape) 88 self.shape = shape 89 else: 90 self.ndim = ndim 91 self.shape = None 92 self.max_ndim = max_ndim 93 self.min_ndim = min_ndim 94 self.name = name 95 self.allow_last_axis_squeeze = allow_last_axis_squeeze 96 try: 97 axes = axes or {} 98 self.axes = {int(k): axes[k] for k in axes} 99 except (ValueError, TypeError): 100 raise TypeError('The keys in axes must be integers.') 101 102 if self.axes and (self.ndim is not None or self.max_ndim is not None): 103 max_dim = (self.ndim if self.ndim else self.max_ndim) - 1 104 max_axis = max(self.axes) 105 if max_axis > max_dim: 106 raise ValueError('Axis {} is greater than the maximum allowed value: {}' 107 .format(max_axis, max_dim)) 108 109 def __repr__(self): 110 spec = [('dtype=' + str(self.dtype)) if self.dtype else '', 111 ('shape=' + str(self.shape)) if self.shape else '', 112 ('ndim=' + str(self.ndim)) if self.ndim else '', 113 ('max_ndim=' + str(self.max_ndim)) if self.max_ndim else '', 114 ('min_ndim=' + str(self.min_ndim)) if self.min_ndim else '', 115 ('axes=' + str(self.axes)) if self.axes else ''] 116 return 'InputSpec(%s)' % ', '.join(x for x in spec if x) 117 118 def get_config(self): 119 return { 120 'dtype': self.dtype, 121 'shape': self.shape, 122 'ndim': self.ndim, 123 'max_ndim': self.max_ndim, 124 'min_ndim': self.min_ndim, 125 'axes': self.axes} 126 127 @classmethod 128 def from_config(cls, config): 129 return cls(**config) 130 131 132def to_tensor_shape(spec): 133 """Returns a tf.TensorShape object that matches the shape specifications. 134 135 If the InputSpec's shape or ndim is defined, this method will return a fully 136 or partially-known shape. Otherwise, the returned TensorShape is None. 137 138 Args: 139 spec: an InputSpec object. 140 141 Returns: 142 a tf.TensorShape object 143 """ 144 if spec.ndim is None and spec.shape is None: 145 return tensor_shape.TensorShape(None) 146 elif spec.shape is not None: 147 return tensor_shape.TensorShape(spec.shape) 148 else: 149 shape = [None] * spec.ndim 150 for a in spec.axes: 151 shape[a] = spec.axes[a] # Assume that axes is defined 152 return tensor_shape.TensorShape(shape) 153 154 155def assert_input_compatibility(input_spec, inputs, layer_name): 156 """Checks compatibility between the layer and provided inputs. 157 158 This checks that the tensor(s) `inputs` verify the input assumptions 159 of a layer (if any). If not, a clear and actional exception gets raised. 160 161 Args: 162 input_spec: An InputSpec instance, list of InputSpec instances, a nested 163 structure of InputSpec instances, or None. 164 inputs: Input tensor, list of input tensors, or a nested structure of 165 input tensors. 166 layer_name: String, name of the layer (for error message formatting). 167 168 Raises: 169 ValueError: in case of mismatch between 170 the provided inputs and the expectations of the layer. 171 """ 172 if not input_spec: 173 return 174 175 input_spec = nest.flatten(input_spec) 176 if isinstance(inputs, dict): 177 # Flatten `inputs` by reference order if input spec names are provided 178 names = [spec.name for spec in input_spec] 179 if all(names): 180 list_inputs = [] 181 for name in names: 182 if name not in inputs: 183 raise ValueError('Missing data for input "%s". ' 184 'You passed a data dictionary with keys %s. ' 185 'Expected the following keys: %s' % 186 (name, list(inputs.keys()), names)) 187 list_inputs.append(inputs[name]) 188 inputs = list_inputs 189 190 inputs = nest.flatten(inputs) 191 for x in inputs: 192 # Having a shape/dtype is the only commonality of the various tensor-like 193 # objects that may be passed. The most common kind of invalid type we are 194 # guarding for is a Layer instance (Functional API), which does not 195 # have a `shape` attribute. 196 if not hasattr(x, 'shape'): 197 raise TypeError('Inputs to a layer should be tensors. Got: %s' % (x,)) 198 199 if len(inputs) != len(input_spec): 200 raise ValueError('Layer ' + layer_name + ' expects ' + 201 str(len(input_spec)) + ' input(s), ' 202 'but it received ' + str(len(inputs)) + 203 ' input tensors. Inputs received: ' + str(inputs)) 204 for input_index, (x, spec) in enumerate(zip(inputs, input_spec)): 205 if spec is None: 206 continue 207 208 shape = tensor_shape.TensorShape(x.shape) 209 if shape.rank is None: 210 return 211 # Check ndim. 212 if spec.ndim is not None and not spec.allow_last_axis_squeeze: 213 ndim = shape.rank 214 if ndim != spec.ndim: 215 raise ValueError('Input ' + str(input_index) + ' of layer ' + 216 layer_name + ' is incompatible with the layer: ' 217 'expected ndim=' + str(spec.ndim) + ', found ndim=' + 218 str(ndim) + '. Full shape received: ' + 219 str(tuple(shape))) 220 if spec.max_ndim is not None: 221 ndim = x.shape.rank 222 if ndim is not None and ndim > spec.max_ndim: 223 raise ValueError('Input ' + str(input_index) + ' of layer ' + 224 layer_name + ' is incompatible with the layer: ' 225 'expected max_ndim=' + str(spec.max_ndim) + 226 ', found ndim=' + str(ndim)) 227 if spec.min_ndim is not None: 228 ndim = x.shape.rank 229 if ndim is not None and ndim < spec.min_ndim: 230 raise ValueError('Input ' + str(input_index) + ' of layer ' + 231 layer_name + ' is incompatible with the layer: ' 232 ': expected min_ndim=' + str(spec.min_ndim) + 233 ', found ndim=' + str(ndim) + 234 '. Full shape received: ' + 235 str(tuple(shape))) 236 # Check dtype. 237 if spec.dtype is not None: 238 if x.dtype.name != spec.dtype: 239 raise ValueError('Input ' + str(input_index) + ' of layer ' + 240 layer_name + ' is incompatible with the layer: ' 241 'expected dtype=' + str(spec.dtype) + 242 ', found dtype=' + str(x.dtype)) 243 244 # Check specific shape axes. 245 shape_as_list = shape.as_list() 246 if spec.axes: 247 for axis, value in spec.axes.items(): 248 if hasattr(value, 'value'): 249 value = value.value 250 if value is not None and shape_as_list[int(axis)] not in {value, None}: 251 raise ValueError( 252 'Input ' + str(input_index) + ' of layer ' + layer_name + ' is' 253 ' incompatible with the layer: expected axis ' + str(axis) + 254 ' of input shape to have value ' + str(value) + 255 ' but received input with shape ' + display_shape(x.shape)) 256 # Check shape. 257 if spec.shape is not None and shape.rank is not None: 258 spec_shape = spec.shape 259 if spec.allow_last_axis_squeeze: 260 if shape_as_list and shape_as_list[-1] == 1: 261 shape_as_list = shape_as_list[:-1] 262 if spec_shape and spec_shape[-1] == 1: 263 spec_shape = spec_shape[:-1] 264 for spec_dim, dim in zip(spec_shape, shape_as_list): 265 if spec_dim is not None and dim is not None: 266 if spec_dim != dim: 267 raise ValueError('Input ' + str(input_index) + 268 ' is incompatible with layer ' + layer_name + 269 ': expected shape=' + str(spec.shape) + 270 ', found shape=' + display_shape(x.shape)) 271 272 273def display_shape(shape): 274 return str(tuple(shape.as_list())) 275 276 277def to_tensor_spec(input_spec, default_dtype=None): 278 """Converts a Keras InputSpec object to a TensorSpec.""" 279 default_dtype = default_dtype or backend.floatx() 280 if isinstance(input_spec, InputSpec): 281 dtype = input_spec.dtype or default_dtype 282 return tensor_spec.TensorSpec(to_tensor_shape(input_spec), dtype) 283 return tensor_spec.TensorSpec(None, default_dtype) 284