-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrouter.py
172 lines (155 loc) · 8.79 KB
/
router.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
import numpy as np
from .base_algorithm import Algorithm
from .lambda_strategy import ConstantStrategy
from collections import Counter
class Router(Algorithm):
def __init__(self, quality_computer, cost_computer, models, max_expected_cost,
rounding_digits=8, strategies=[ConstantStrategy(10000, 50)]):
"""
Initialize the Router object.
Args:
quality_computer: The quality computer object used for computing model qualities.
cost_computer: The cost computer object used for computing model costs.
models: A list of model objects.
max_expected_cost: The maximum expected cost.
rounding_digits: The number of digits to round the results to (default: 8).
strategies: A list of strategy objects for hyperparameter optimization
(default: [ConstantStrategy(10000, 50)]).
"""
super().__init__(quality_computer, cost_computer, models,
max_expected_cost, strategies, rounding_digits)
self.lambda_ = None
self.qualities = None
self.costs = None
def compute_budget(self, qualities, costs, ground_truth_qualities=None, ground_truth_costs=None):
"""
Computes the optimal lambdas for model selection based on the given qualities and costs.
Args:
qualities (list): List of qualities for each model.
costs (list): List of costs for each model.
ground_truth_qualities (list, optional): List of ground truth qualities for each model.
Defaults to None.
ground_truth_costs (list, optional): List of ground truth costs for each model.
Defaults to None.
Returns:
None
"""
best_lambda = None
best_quality = None
for strategy in self.strategies:
lambdas, cost, quality = strategy.compute_lambdas([0], self._execute, self.max_expected_cost,
(qualities, costs, ground_truth_qualities, ground_truth_costs))
if best_lambda is None or cost < self.max_expected_cost and (best_lambda is None or quality > best_quality):
best_lambda = lambdas[0]
best_quality = quality
self.lambda_ = best_lambda
output_cheap = self._execute(self.lambda_, qualities, costs, ground_truth_qualities,
ground_truth_costs, cheapest=True, most_expensive=False)
output_expensive = self._execute(self.lambda_, qualities, costs, ground_truth_qualities,
ground_truth_costs, cheapest=False, most_expensive=True)
cost_cheap = output_cheap['cost']
cost_expensive = output_expensive['cost']
if cost_cheap == cost_expensive:
self.gamma = 1
else:
self.gamma = (cost_expensive - self.max_expected_cost) / (cost_expensive - cost_cheap)
self.gamma = min(1, max(0, self.gamma))
def get_lambdas(self):
return [self.lambda_]
def fit(self, questions, model_answers, ground_truth_qualities=None, ground_truth_costs=None):
self.quality_computer.trigger_training(True)
self.cost_computer.trigger_training(True)
qualities, _ = self.quality_computer.predict(questions, [[None] * len(self.models)] * len(questions))
costs = self.cost_computer.predict(questions, [[None] * len(self.models)] * len(questions))
self.compute_budget(qualities, costs, ground_truth_qualities, ground_truth_costs)
self.quality_computer.trigger_training(False)
self.cost_computer.trigger_training(False)
def predict(self, questions, model_answers=None):
qualities_question, _ = self.quality_computer.predict(questions, model_answers)
costs_question = self.cost_computer.predict(questions, model_answers)
return [self._predict_model(qualities_question[i], costs_question[i], cheapest=False, most_expensive=False) for i in range(len(questions))]
def select_answer(self, questions, model_answers):
models_selected = []
for i, model_answer in enumerate(model_answers):
indices_with_answer = [j for j in range(len(model_answer)) if model_answer[j] is not None]
if len(indices_with_answer) == 0:
models_selected.append(None)
else:
models_selected.append(self.models[indices_with_answer[0]])
return models_selected
def _predict_model(self, qualities_question, costs_model,
lambda_=None, cheapest=True, most_expensive=False):
"""
Predicts the best model based on the qualities of the question and the costs of the models.
Args:
qualities_question (numpy.ndarray): An array representing the qualities of the question.
costs_model (numpy.ndarray): An array representing the costs of the models.
lambda_ (float, optional): A parameter used to adjust the importance of qualities and costs. If not provided, the default value from the class instance will be used.
cheapest (bool, optional): If True, the cheapest model among the best models will be selected. Default is True.
most_expensive (bool, optional): If True, the most expensive model among the best models will be selected. Default is False.
Returns:
best_model: The best model based on the given qualities and costs.
"""
if lambda_ is None:
lambda_ = self.lambda_
highest = np.round(qualities_question - lambda_ * costs_model, self.rounding_digits)
# check which model has the highest value. If multiple, return the cheapest
max_value = np.max(highest)
# collect all models with the highest value
best_models = np.where(highest == max_value)[0]
# select the cheapest model
if len(best_models) > 1:
best_model_cheapest = best_models[np.argmin(costs_model[best_models])]
best_model_expensive = best_models[np.argmax(costs_model[best_models])]
if cheapest:
best_model = best_model_cheapest
elif most_expensive:
best_model = best_model_expensive
elif np.random.rand() < self.gamma:
best_model = best_model_expensive
else:
best_model = best_model_cheapest
else:
best_model = best_models[0]
best_model = self.models[best_model]
return best_model
def _execute(self, lambda_, qualities, costs, ground_truth_qualities=None, ground_truth_costs=None,
cheapest=True, most_expensive=False):
"""
Executes the model selection process based on the given parameters.
Args:
lambda_ (float or list): The regularization parameter(s) for the model selection.
qualities (list): The list of quality values for each model.
costs (list): The list of cost values for each model.
ground_truth_qualities (list, optional): The list of ground truth quality values for each model. Default is None.
ground_truth_costs (list, optional): The list of ground truth cost values for each model. Default is None.
cheapest (bool, optional): Flag indicating whether to select the cheapest model. Default is True.
most_expensive (bool, optional): Flag indicating whether to select the most expensive model. Default is False.
Returns:
- dict: A dictionary containing the following keys:
- 'cost': The mean cost value of the selected models.
- 'quality': The mean quality value of the selected models.
- 'models_run': A Counter object containing the count of each model selected.
"""
if isinstance(lambda_, list):
lambda_ = lambda_[0]
all_costs = []
all_qualities = []
models_run = []
for i in range(len(qualities)):
best_model = self._predict_model(qualities[i], costs[i], lambda_, cheapest, most_expensive)
best_model_index = self.models.index(best_model)
models_run.append(best_model_index)
if ground_truth_costs is not None:
all_costs.append(ground_truth_costs[i][best_model_index])
else:
all_costs.append(costs[i][best_model_index])
if ground_truth_qualities is not None:
all_qualities.append(ground_truth_qualities[i][best_model_index])
else:
all_qualities.append(qualities[i][best_model_index])
return {
'cost': np.mean(all_costs),
'quality': np.mean(all_qualities),
'models_run': Counter(models_run)
}