Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/keras/src/premade_models/wide_deep.py: 22%

95 statements  

« prev     ^ index     » next       coverage.py v7.4.0, created at 2024-01-03 07:57 +0000

1# Copyright 2019 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"""Built-in WideNDeep model classes.""" 

16 

17import tensorflow.compat.v2 as tf 

18 

19from keras.src import activations 

20from keras.src import backend 

21from keras.src import layers as layer_module 

22from keras.src.engine import base_layer 

23from keras.src.engine import data_adapter 

24from keras.src.engine import training as keras_training 

25from keras.src.saving import serialization_lib 

26 

27# isort: off 

28from tensorflow.python.util import deprecation 

29from tensorflow.python.util.tf_export import keras_export 

30 

31 

32@keras_export( 

33 "keras.experimental.WideDeepModel", 

34 v1=["keras.experimental.WideDeepModel", "keras.models.WideDeepModel"], 

35) 

36@deprecation.deprecated_endpoints("keras.experimental.WideDeepModel") 

37class WideDeepModel(keras_training.Model): 

38 r"""Wide & Deep Model for regression and classification problems. 

39 

40 This model jointly train a linear and a dnn model. 

41 

42 Example: 

43 

44 ```python 

45 linear_model = LinearModel() 

46 dnn_model = keras.Sequential([keras.layers.Dense(units=64), 

47 keras.layers.Dense(units=1)]) 

48 combined_model = WideDeepModel(linear_model, dnn_model) 

49 combined_model.compile(optimizer=['sgd', 'adam'], 

50 loss='mse', metrics=['mse']) 

51 # define dnn_inputs and linear_inputs as separate numpy arrays or 

52 # a single numpy array if dnn_inputs is same as linear_inputs. 

53 combined_model.fit([linear_inputs, dnn_inputs], y, epochs) 

54 # or define a single `tf.data.Dataset` that contains a single tensor or 

55 # separate tensors for dnn_inputs and linear_inputs. 

56 dataset = tf.data.Dataset.from_tensors(([linear_inputs, dnn_inputs], y)) 

57 combined_model.fit(dataset, epochs) 

58 ``` 

59 

60 Both linear and dnn model can be pre-compiled and trained separately 

61 before jointly training: 

62 

63 Example: 

64 ```python 

65 linear_model = LinearModel() 

66 linear_model.compile('adagrad', 'mse') 

67 linear_model.fit(linear_inputs, y, epochs) 

68 dnn_model = keras.Sequential([keras.layers.Dense(units=1)]) 

69 dnn_model.compile('rmsprop', 'mse') 

70 dnn_model.fit(dnn_inputs, y, epochs) 

71 combined_model = WideDeepModel(linear_model, dnn_model) 

72 combined_model.compile(optimizer=['sgd', 'adam'], 

73 loss='mse', metrics=['mse']) 

74 combined_model.fit([linear_inputs, dnn_inputs], y, epochs) 

75 ``` 

76 

77 """ 

78 

79 def __init__(self, linear_model, dnn_model, activation=None, **kwargs): 

80 """Create a Wide & Deep Model. 

81 

82 Args: 

83 linear_model: a premade LinearModel, its output must match the output 

84 of the dnn model. 

85 dnn_model: a `tf.keras.Model`, its output must match the output of the 

86 linear model. 

87 activation: Activation function. Set it to None to maintain a linear 

88 activation. 

89 **kwargs: The keyword arguments that are passed on to 

90 BaseLayer.__init__. Allowed keyword arguments include `name`. 

91 """ 

92 super().__init__(**kwargs) 

93 base_layer.keras_premade_model_gauge.get_cell("WideDeep").set(True) 

94 self.linear_model = linear_model 

95 self.dnn_model = dnn_model 

96 self.activation = activations.get(activation) 

97 

98 def call(self, inputs, training=None): 

99 if not isinstance(inputs, (tuple, list)) or len(inputs) != 2: 

100 linear_inputs = dnn_inputs = inputs 

101 else: 

102 linear_inputs, dnn_inputs = inputs 

103 linear_output = self.linear_model(linear_inputs) 

104 

105 if self.dnn_model._expects_training_arg: 

106 if training is None: 

107 training = backend.learning_phase() 

108 dnn_output = self.dnn_model(dnn_inputs, training=training) 

109 else: 

110 dnn_output = self.dnn_model(dnn_inputs) 

111 output = tf.nest.map_structure( 

112 lambda x, y: (x + y), linear_output, dnn_output 

113 ) 

114 if self.activation: 

115 return tf.nest.map_structure(self.activation, output) 

116 return output 

117 

118 # This does not support gradient scaling and LossScaleOptimizer. 

119 def train_step(self, data): 

120 x, y, sample_weight = data_adapter.unpack_x_y_sample_weight(data) 

121 with tf.GradientTape() as tape: 

122 y_pred = self(x, training=True) 

123 loss = self.compiled_loss( 

124 y, y_pred, sample_weight, regularization_losses=self.losses 

125 ) 

126 self.compiled_metrics.update_state(y, y_pred, sample_weight) 

127 

128 if isinstance(self.optimizer, (list, tuple)): 

129 linear_vars = self.linear_model.trainable_variables 

130 dnn_vars = self.dnn_model.trainable_variables 

131 linear_grads, dnn_grads = tape.gradient( 

132 loss, (linear_vars, dnn_vars) 

133 ) 

134 

135 linear_optimizer = self.optimizer[0] 

136 dnn_optimizer = self.optimizer[1] 

137 linear_optimizer.apply_gradients(zip(linear_grads, linear_vars)) 

138 dnn_optimizer.apply_gradients(zip(dnn_grads, dnn_vars)) 

139 else: 

140 trainable_variables = self.trainable_variables 

141 grads = tape.gradient(loss, trainable_variables) 

142 self.optimizer.apply_gradients(zip(grads, trainable_variables)) 

143 

144 return {m.name: m.result() for m in self.metrics} 

145 

146 def _make_train_function(self): 

147 # Only needed for graph mode and model_to_estimator. 

148 has_recompiled = self._recompile_weights_loss_and_weighted_metrics() 

149 self._check_trainable_weights_consistency() 

150 # If we have re-compiled the loss/weighted metric sub-graphs then create 

151 # train function even if one exists already. This is because 

152 # `_feed_sample_weights` list has been updated on re-compile. 

153 if getattr(self, "train_function", None) is None or has_recompiled: 

154 # Restore the compiled trainable state. 

155 current_trainable_state = self._get_trainable_state() 

156 self._set_trainable_state(self._compiled_trainable_state) 

157 

158 inputs = ( 

159 self._feed_inputs 

160 + self._feed_targets 

161 + self._feed_sample_weights 

162 ) 

163 if not isinstance(backend.symbolic_learning_phase(), int): 

164 inputs += [backend.symbolic_learning_phase()] 

165 

166 if isinstance(self.optimizer, (list, tuple)): 

167 linear_optimizer = self.optimizer[0] 

168 dnn_optimizer = self.optimizer[1] 

169 else: 

170 linear_optimizer = self.optimizer 

171 dnn_optimizer = self.optimizer 

172 

173 with backend.get_graph().as_default(): 

174 with backend.name_scope("training"): 

175 # Training updates 

176 updates = [] 

177 linear_updates = linear_optimizer.get_updates( 

178 params=self.linear_model.trainable_weights, 

179 loss=self.total_loss, 

180 ) 

181 updates += linear_updates 

182 dnn_updates = dnn_optimizer.get_updates( 

183 params=self.dnn_model.trainable_weights, 

184 loss=self.total_loss, 

185 ) 

186 updates += dnn_updates 

187 # Unconditional updates 

188 updates += self.get_updates_for(None) 

189 # Conditional updates relevant to this model 

190 updates += self.get_updates_for(self.inputs) 

191 

192 metrics = self._get_training_eval_metrics() 

193 metrics_tensors = [ 

194 m._call_result 

195 for m in metrics 

196 if hasattr(m, "_call_result") 

197 ] 

198 

199 with backend.name_scope("training"): 

200 # Gets loss and metrics. Updates weights at each call. 

201 fn = backend.function( 

202 inputs, 

203 [self.total_loss] + metrics_tensors, 

204 updates=updates, 

205 name="train_function", 

206 **self._function_kwargs 

207 ) 

208 setattr(self, "train_function", fn) 

209 

210 # Restore the current trainable state 

211 self._set_trainable_state(current_trainable_state) 

212 

213 def get_config(self): 

214 linear_config = serialization_lib.serialize_keras_object( 

215 self.linear_model 

216 ) 

217 dnn_config = serialization_lib.serialize_keras_object(self.dnn_model) 

218 config = { 

219 "linear_model": linear_config, 

220 "dnn_model": dnn_config, 

221 "activation": activations.serialize(self.activation), 

222 } 

223 base_config = base_layer.Layer.get_config(self) 

224 return dict(list(base_config.items()) + list(config.items())) 

225 

226 @classmethod 

227 def from_config(cls, config, custom_objects=None): 

228 linear_config = config.pop("linear_model") 

229 linear_model = layer_module.deserialize(linear_config, custom_objects) 

230 dnn_config = config.pop("dnn_model") 

231 dnn_model = layer_module.deserialize(dnn_config, custom_objects) 

232 activation = activations.deserialize( 

233 config.pop("activation", None), custom_objects=custom_objects 

234 ) 

235 return cls( 

236 linear_model=linear_model, 

237 dnn_model=dnn_model, 

238 activation=activation, 

239 **config 

240 ) 

241