|
1
|
|
|
import torch |
|
2
|
|
|
import torch.nn as nn |
|
3
|
|
|
import numpy as np |
|
4
|
|
|
from torch.utils.data import DataLoader, TensorDataset |
|
5
|
|
|
from gradient_free_optimizers import ( |
|
6
|
|
|
HillClimbingOptimizer, |
|
7
|
|
|
RandomSearchOptimizer, |
|
8
|
|
|
RepulsingHillClimbingOptimizer, |
|
9
|
|
|
PowellsMethod, |
|
10
|
|
|
) |
|
11
|
|
|
|
|
12
|
|
|
# Define a synthetic dataset |
|
13
|
|
|
# np.random.seed(42) |
|
14
|
|
|
X = np.random.rand(1000, 20) |
|
15
|
|
|
true_weights = np.random.rand(20, 1) |
|
16
|
|
|
y = X @ true_weights + 0.1 * np.random.randn(1000, 1) |
|
17
|
|
|
|
|
18
|
|
|
X = torch.Tensor(X) |
|
19
|
|
|
y = torch.Tensor(y) |
|
20
|
|
|
|
|
21
|
|
|
# Create a DataLoader |
|
22
|
|
|
dataset = TensorDataset(X, y) |
|
23
|
|
|
dataloader = DataLoader(dataset, batch_size=64, shuffle=True) |
|
24
|
|
|
|
|
25
|
|
|
|
|
26
|
|
|
num_epochs = 10 |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
# Define a more complex neural network |
|
30
|
|
|
class ComplexModel(nn.Module): |
|
31
|
|
|
def __init__(self): |
|
32
|
|
|
super(ComplexModel, self).__init__() |
|
33
|
|
|
self.network = nn.Sequential( |
|
34
|
|
|
nn.Linear(20, 64), |
|
35
|
|
|
nn.ReLU(), |
|
36
|
|
|
nn.Linear(64, 64), |
|
37
|
|
|
nn.ReLU(), |
|
38
|
|
|
nn.Linear(64, 1), |
|
39
|
|
|
) |
|
40
|
|
|
|
|
41
|
|
|
def forward(self, x): |
|
42
|
|
|
return self.network(x) |
|
43
|
|
|
|
|
44
|
|
|
|
|
45
|
|
|
# Initialize the model |
|
46
|
|
|
model = ComplexModel() |
|
47
|
|
|
|
|
48
|
|
|
# Define a loss function |
|
49
|
|
|
criterion = nn.MSELoss() |
|
50
|
|
|
|
|
51
|
|
|
|
|
52
|
|
|
# Define the custom optimizer with GFO |
|
53
|
|
|
class GFOOptimizer(torch.optim.Optimizer): |
|
54
|
|
|
def __init__(self, params, model, dataloader, criterion, lr=1e-3): |
|
55
|
|
|
self.model = model |
|
56
|
|
|
self.dataloader = dataloader |
|
57
|
|
|
self.criterion = criterion |
|
58
|
|
|
self.lr = lr |
|
59
|
|
|
|
|
60
|
|
|
# Flatten the initial model parameters |
|
61
|
|
|
self.initial_weights = {} |
|
62
|
|
|
self.params = [] |
|
63
|
|
|
counter = 0 |
|
64
|
|
|
for param in self.model.parameters(): |
|
65
|
|
|
self.params.extend(param.data.cpu().numpy().flatten()) |
|
66
|
|
|
|
|
67
|
|
|
for value in param.data.flatten(): |
|
68
|
|
|
self.initial_weights[f"x{counter}"] = ( |
|
69
|
|
|
value.item() |
|
70
|
|
|
) # Convert tensor value to Python scalar |
|
71
|
|
|
counter += 1 |
|
72
|
|
|
|
|
73
|
|
|
# Define the search space |
|
74
|
|
|
self.search_space = { |
|
75
|
|
|
f"x{i}": np.arange(-1.0, 1.0, 0.1, dtype=np.float32) |
|
76
|
|
|
for i in range(len(self.params)) |
|
77
|
|
|
} |
|
78
|
|
|
|
|
79
|
|
|
# Initialize the GFO optimizer |
|
80
|
|
|
self.optimizer = HillClimbingOptimizer( |
|
81
|
|
|
self.search_space, initialize={"warm_start": [self.initial_weights]} |
|
82
|
|
|
) |
|
83
|
|
|
|
|
84
|
|
|
self.optimizer.init_search( |
|
85
|
|
|
objective_function=self.objective_function, |
|
86
|
|
|
n_iter=num_epochs * len(dataloader), |
|
87
|
|
|
max_time=None, |
|
88
|
|
|
max_score=None, |
|
89
|
|
|
early_stopping=None, |
|
90
|
|
|
memory=True, |
|
91
|
|
|
memory_warm_start=None, |
|
92
|
|
|
verbosity=[], |
|
93
|
|
|
) |
|
94
|
|
|
|
|
95
|
|
|
defaults = dict(lr=lr) |
|
96
|
|
|
super().__init__(params, defaults) |
|
97
|
|
|
|
|
98
|
|
|
def objective_function(self, opt_params): |
|
99
|
|
|
opt_params_l = list(opt_params.values()) |
|
100
|
|
|
|
|
101
|
|
|
# Set model parameters |
|
102
|
|
|
start = 0 |
|
103
|
|
|
for param in self.model.parameters(): |
|
104
|
|
|
param_length = param.numel() |
|
105
|
|
|
param.data = torch.tensor( |
|
106
|
|
|
opt_params_l[start : start + param_length] |
|
107
|
|
|
).view(param.shape) |
|
108
|
|
|
start += param_length |
|
109
|
|
|
|
|
110
|
|
|
# Compute the loss |
|
111
|
|
|
total_loss = 0.0 |
|
112
|
|
|
with torch.no_grad(): |
|
113
|
|
|
for batch_X, batch_y in self.dataloader: |
|
114
|
|
|
outputs = self.model(batch_X) |
|
115
|
|
|
loss = self.criterion(outputs, batch_y) |
|
116
|
|
|
total_loss += loss.item() |
|
117
|
|
|
return total_loss / len(self.dataloader) |
|
118
|
|
|
|
|
119
|
|
|
def step(self, closure=None): |
|
120
|
|
|
if closure is not None: |
|
121
|
|
|
closure() |
|
122
|
|
|
|
|
123
|
|
|
# Use GFO to find the best parameters |
|
124
|
|
|
self.optimizer.search_step() |
|
125
|
|
|
best_params = self.optimizer.pos_new |
|
126
|
|
|
|
|
127
|
|
|
print("self.optimizer.score_new", self.optimizer.score_new) |
|
128
|
|
|
|
|
129
|
|
|
# Set the best parameters to the model |
|
130
|
|
|
start = 0 |
|
131
|
|
|
for param in self.model.parameters(): |
|
132
|
|
|
param_length = param.numel() |
|
133
|
|
|
""" |
|
134
|
|
|
param.data.copy_( |
|
135
|
|
|
torch.tensor( |
|
136
|
|
|
best_params[start : start + param_length], |
|
137
|
|
|
dtype=torch.float32, |
|
138
|
|
|
).view(param.shape) |
|
139
|
|
|
) |
|
140
|
|
|
""" |
|
141
|
|
|
start += param_length |
|
142
|
|
|
|
|
143
|
|
|
self.params = best_params |
|
144
|
|
|
|
|
145
|
|
|
|
|
146
|
|
|
# Initialize the custom optimizer |
|
147
|
|
|
optimizer = GFOOptimizer( |
|
148
|
|
|
model.parameters(), model, dataloader, criterion, lr=0.01 |
|
149
|
|
|
) |
|
150
|
|
|
# optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9) |
|
151
|
|
|
|
|
152
|
|
|
|
|
153
|
|
|
# Training loop |
|
154
|
|
|
for epoch in range(num_epochs): |
|
155
|
|
|
for batch_X, batch_y in dataloader: |
|
156
|
|
|
# Zero the gradients |
|
157
|
|
|
optimizer.zero_grad() |
|
158
|
|
|
|
|
159
|
|
|
# Forward pass |
|
160
|
|
|
outputs = model(batch_X) |
|
161
|
|
|
loss = criterion(outputs, batch_y) |
|
162
|
|
|
|
|
163
|
|
|
# Backward pass |
|
164
|
|
|
loss.backward() |
|
165
|
|
|
|
|
166
|
|
|
# Update the weights |
|
167
|
|
|
optimizer.step() |
|
168
|
|
|
|
|
169
|
|
|
# Print the loss for every epoch |
|
170
|
|
|
print(f"Epoch [{epoch+1}/{num_epochs}], Loss: {loss.item():.4f}") |
|
171
|
|
|
|
|
172
|
|
|
print("Training completed!") |
|
173
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
