| commit | author | age | ||
| 425bf7 | 1 | import abc |
| JK | 2 | import json |
| 3 | ||
| 4 | import numpy | |
| 5 | import scipy.linalg | |
| 6 | from keras import Sequential | |
| 7 | from keras.layers import Dense | |
| 8 | from keras.models import load_model | |
| 9 | ||
| 10 | model_types = {} | |
| 11 | ||
| 12 | ||
| 13 | class Model: | |
| 14 | @abc.abstractmethod | |
| 15 | def get_temperature(self, flat_history: numpy.array) -> float: | |
| 16 | pass | |
| 17 | ||
| 18 | @abc.abstractmethod | |
| 19 | def save(self, filename): | |
| 20 | pass | |
| 21 | ||
| 22 | @classmethod | |
| 23 | def load(cls, config, filename): | |
| 24 | return model_types[config['model_type']]._load(filename) | |
| 25 | ||
| 26 | @abc.abstractclassmethod | |
| 27 | def _load(cls, filename): | |
| 28 | pass | |
| 29 | ||
| 30 | @classmethod | |
| 31 | def generate(cls, config: dict): | |
| 32 | return model_types[config['model_type']].generate(config) | |
| 33 | ||
| 34 | ||
| 35 | class LinearModel(Model): | |
| 36 | def __init__(self, model): | |
| 37 | self.model = model | |
| 38 | ||
| 39 | def get_temperature(self, flat_history): | |
| 40 | return sum(numpy.multiply(self.model, flat_history)) | |
| 41 | ||
| 42 | def save(self, filename): | |
| 43 | with open(filename, 'w') as fp: | |
| 44 | json.dump(self.model.tolist(), fp) | |
| 45 | ||
| 46 | @classmethod | |
| 47 | def _load(cls, filename): | |
| 48 | with open(filename, 'r') as fp: | |
| 49 | model = json.load(fp) | |
| 50 | model = numpy.array(model) | |
| 51 | return cls(model) | |
| 52 | ||
| 53 | @classmethod | |
| 54 | def generate(cls, config: dict): | |
| 55 | from .ArgParser import get_sliding_window_from_config | |
| 56 | sliding_window = get_sliding_window_from_config(config) | |
| 57 | xs = [] | |
| 58 | ys = [] | |
| 59 | for x in sliding_window: | |
| 60 | xs.append(x.get_model_values()) | |
| 61 | ys.append(x.get_model_target()) | |
| 62 | return cls(scipy.linalg.lstsq(xs, ys)[0]) | |
| 63 | ||
| 64 | ||
| 65 | model_types['linear'] = LinearModel | |
| 66 | ||
| 67 | ||
| 68 | class NeuralModel(Model): | |
| 69 | def __init__(self, model): | |
| 70 | self.model = model | |
| 71 | ||
| 72 | def get_temperature(self, flat_history): | |
| 73 | state = numpy.reshape(flat_history, [1, len(flat_history)]) | |
| 74 | result = self.model.predict(state) | |
| 75 | return result[0][0] | |
| 76 | ||
| 77 | @staticmethod | |
| 78 | def _build_model(state_size): | |
| 79 | model = Sequential() | |
| 80 | model.add(Dense(int(state_size / 2) or 1, activation='linear', | |
| 81 | input_dim=state_size, | |
| 82 | kernel_initializer='random_uniform')) | |
| 83 | model.add(Dense(int(state_size / 4) or 1, activation='linear', | |
| 84 | kernel_initializer='random_uniform')) | |
| 85 | model.add(Dense(int(state_size / 8) or 1, activation='linear', | |
| 86 | kernel_initializer='random_uniform')) | |
| 87 | model.add(Dense(1, activation='linear', | |
| 88 | kernel_initializer='random_uniform')) | |
| 89 | model.compile(optimizer='adagrad', | |
| 90 | loss='mean_squared_error', | |
| 91 | metrics=['accuracy']) | |
| 92 | return model | |
| 93 | ||
| 94 | def save(self, filename): | |
| 95 | self.model.save(filename) | |
| 96 | ||
| 97 | @classmethod | |
| 98 | def _load(cls, filename): | |
| 99 | model = load_model(filename) | |
| 100 | return cls(model) | |
| 101 | ||
| 102 | @classmethod | |
| 103 | def generate(cls, config: dict): | |
| 104 | from .ArgParser import get_sliding_window_from_config | |
| 105 | sliding_window = get_sliding_window_from_config(config) | |
| 106 | window = next(sliding_window) | |
| 107 | model = cls._build_model(window.get_model_size()) | |
| 108 | ||
| 109 | xs = [] | |
| 110 | ys = [] | |
| 111 | for window in sliding_window: | |
| 112 | xs.append(window.get_model_values()) | |
| 113 | ys.append([window.get_next_value('temp_in')]) | |
| 114 | ||
| 115 | model.fit(numpy.array(xs), numpy.array(ys), epochs=config['model_epochs'], batch_size=32, shuffle=True) | |
| 116 | return cls(model) | |
| 117 | ||
| 118 | ||
| 119 | model_types['neural'] = NeuralModel | |
| 120 | ||
| 121 | ||
| 122 | class NeuralModelLess(NeuralModel): | |
| 123 | @staticmethod | |
| 124 | def _build_model(state_size): | |
| 125 | model = Sequential() | |
| 126 | model.add(Dense(int(state_size / 2) or 1, activation='linear', | |
| 127 | input_dim=state_size, | |
| 128 | kernel_initializer='random_uniform')) | |
| 129 | model.add(Dense(int(state_size / 4) or 1, activation='linear', | |
| 130 | kernel_initializer='random_uniform')) | |
| 131 | model.add(Dense(int(state_size / 8) or 1, activation='linear', | |
| 132 | kernel_initializer='random_uniform')) | |
| 133 | model.add(Dense(1, activation='linear', | |
| 134 | kernel_initializer='random_uniform')) | |
| 135 | model.compile(optimizer='adagrad', | |
| 136 | loss='mean_squared_error', | |
| 137 | metrics=['accuracy']) | |
| 138 | return model | |
| 139 | ||
| 140 | model_types['neural_less'] = NeuralModelLess | |
| 141 | ||
| 142 | ||
| 143 | class NeuralModelMore(NeuralModel): | |
| 144 | @staticmethod | |
| 145 | def _build_model(state_size): | |
| 146 | model = Sequential() | |
| 147 | model.add(Dense(int(state_size / 2) or 1, activation='linear', | |
| 148 | input_dim=state_size, | |
| 149 | kernel_initializer='random_uniform')) | |
| 150 | model.add(Dense(int(state_size / 4) or 1, activation='linear', | |
| 151 | kernel_initializer='random_uniform')) | |
| 152 | model.add(Dense(int(state_size / 8) or 1, activation='linear', | |
| 153 | kernel_initializer='random_uniform')) | |
| 154 | model.add(Dense(1, activation='linear', | |
| 155 | kernel_initializer='random_uniform')) | |
| 156 | model.compile(optimizer='adagrad', | |
| 157 | loss='mean_squared_error', | |
| 158 | metrics=['accuracy']) | |
| 159 | return model | |
| 160 | ||
| 161 | model_types['neural_more'] = NeuralModelMore | |
| 162 | ||
| 163 | ||
| 164 | class NeuralModelRelu(NeuralModel): | |
| 165 | @staticmethod | |
| 166 | def _build_model(state_size): | |
| 167 | model = Sequential() | |
| 168 | model.add(Dense(int(state_size / 2) or 1, activation='linear', | |
| 169 | input_dim=state_size, | |
| 170 | kernel_initializer='random_uniform')) | |
| 171 | model.add(Dense(int(state_size / 4) or 1, activation='relu', | |
| 172 | kernel_initializer='random_uniform')) | |
| 173 | model.add(Dense(int(state_size / 8) or 1, activation='relu', | |
| 174 | kernel_initializer='random_uniform')) | |
| 175 | model.add(Dense(1, activation='linear', | |
| 176 | kernel_initializer='random_uniform')) | |
| 177 | model.compile(optimizer='adagrad', | |
| 178 | loss='mean_squared_error', | |
| 179 | metrics=['accuracy']) | |
| 180 | return model | |
| 181 | ||
| 182 | model_types['neural_relu'] = NeuralModelRelu | |
| 183 | ||
| 184 | ||
| 185 | class NeuralModelReduced(NeuralModel): | |
| 186 | @staticmethod | |
| 187 | def _build_model(state_size): | |
| 188 | model = Sequential() | |
| 189 | model.add(Dense(int(state_size / 2) or 1, activation='linear', | |
| 190 | input_dim=state_size, | |
| 191 | kernel_initializer='random_uniform')) | |
| 192 | model.add(Dense(1, activation='linear', | |
| 193 | kernel_initializer='random_uniform')) | |
| 194 | model.compile(optimizer='adagrad', | |
| 195 | loss='mean_squared_error', | |
| 196 | metrics=['accuracy']) | |
| 197 | return model | |
| 198 | ||
| 199 | model_types['neural_half'] = NeuralModelReduced | |
| 200 | ||
| 201 | ||
| 202 | class NeuralModelSqueezeExpand(NeuralModel): | |
| 203 | @staticmethod | |
| 204 | def _build_model(state_size): | |
| 205 | model = Sequential() | |
| 206 | model.add(Dense(int(state_size / 2) or 1, activation='linear', | |
| 207 | input_dim=state_size, | |
| 208 | kernel_initializer='random_uniform')) | |
| 209 | model.add(Dense(int(state_size / 4) or 1, activation='linear', | |
| 210 | kernel_initializer='random_uniform')) | |
| 211 | model.add(Dense(int(state_size / 2) or 1, activation='linear', | |
| 212 | kernel_initializer='random_uniform')) | |
| 213 | model.add(Dense(1, activation='linear', | |
| 214 | kernel_initializer='random_uniform')) | |
| 215 | model.compile(optimizer='adagrad', | |
| 216 | loss='mean_squared_error', | |
| 217 | metrics=['accuracy']) | |
| 218 | return model | |
| 219 | ||
| 220 | model_types['neural_hqh'] = NeuralModelSqueezeExpand | |
