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
|
"""Script to generate baseline values from PyTorch optimization algorithms"""
import argparse
import math
import sys
import torch
import torch.optim
HEADER = """
#include <torch/types.h>
#include <vector>
namespace expected_parameters {
"""
FOOTER = "} // namespace expected_parameters"
PARAMETERS = "inline std::vector<std::vector<torch::Tensor>> {}() {{"
OPTIMIZERS = {
"LBFGS": lambda p: torch.optim.LBFGS(p, 1.0),
"LBFGS_with_line_search": lambda p: torch.optim.LBFGS(
p, 1.0, line_search_fn="strong_wolfe"
),
"Adam": lambda p: torch.optim.Adam(p, 1.0),
"Adam_with_weight_decay": lambda p: torch.optim.Adam(p, 1.0, weight_decay=1e-2),
"Adam_with_weight_decay_and_amsgrad": lambda p: torch.optim.Adam(
p, 1.0, weight_decay=1e-6, amsgrad=True
),
"AdamW": lambda p: torch.optim.AdamW(p, 1.0),
"AdamW_without_weight_decay": lambda p: torch.optim.AdamW(p, 1.0, weight_decay=0),
"AdamW_with_amsgrad": lambda p: torch.optim.AdamW(p, 1.0, amsgrad=True),
"Adagrad": lambda p: torch.optim.Adagrad(p, 1.0),
"Adagrad_with_weight_decay": lambda p: torch.optim.Adagrad(
p, 1.0, weight_decay=1e-2
),
"Adagrad_with_weight_decay_and_lr_decay": lambda p: torch.optim.Adagrad(
p, 1.0, weight_decay=1e-6, lr_decay=1e-3
),
"RMSprop": lambda p: torch.optim.RMSprop(p, 0.1),
"RMSprop_with_weight_decay": lambda p: torch.optim.RMSprop(
p, 0.1, weight_decay=1e-2
),
"RMSprop_with_weight_decay_and_centered": lambda p: torch.optim.RMSprop(
p, 0.1, weight_decay=1e-6, centered=True
),
"RMSprop_with_weight_decay_and_centered_and_momentum": lambda p: torch.optim.RMSprop(
p, 0.1, weight_decay=1e-6, centered=True, momentum=0.9
),
"SGD": lambda p: torch.optim.SGD(p, 0.1),
"SGD_with_weight_decay": lambda p: torch.optim.SGD(p, 0.1, weight_decay=1e-2),
"SGD_with_weight_decay_and_momentum": lambda p: torch.optim.SGD(
p, 0.1, momentum=0.9, weight_decay=1e-2
),
"SGD_with_weight_decay_and_nesterov_momentum": lambda p: torch.optim.SGD(
p, 0.1, momentum=0.9, weight_decay=1e-6, nesterov=True
),
}
def weight_init(module):
if isinstance(module, torch.nn.Linear):
stdev = 1.0 / math.sqrt(module.weight.size(1))
for p in module.parameters():
p.data.uniform_(-stdev, stdev)
def run(optimizer_name, iterations, sample_every):
torch.manual_seed(0)
model = torch.nn.Sequential(
torch.nn.Linear(2, 3),
torch.nn.Sigmoid(),
torch.nn.Linear(3, 1),
torch.nn.Sigmoid(),
)
model = model.to(torch.float64).apply(weight_init)
optimizer = OPTIMIZERS[optimizer_name](model.parameters())
input = torch.tensor([[0.1, 0.2], [0.3, 0.4], [0.5, 0.6]], dtype=torch.float64)
values = []
for i in range(iterations):
optimizer.zero_grad()
output = model.forward(input)
loss = output.sum()
loss.backward()
def closure():
return torch.tensor([10.0])
optimizer.step(closure)
if i % sample_every == 0:
values.append(
[p.clone().flatten().data.numpy() for p in model.parameters()]
)
return values
def emit(optimizer_parameter_map):
# Don't write generated with an @ in front, else this file is recognized as generated.
print("// @{} from {}".format("generated", __file__))
print(HEADER)
for optimizer_name, parameters in optimizer_parameter_map.items():
print(PARAMETERS.format(optimizer_name))
print(" return {")
for sample in parameters:
print(" {")
for parameter in sample:
parameter_values = "{{{}}}".format(", ".join(map(str, parameter)))
print(f" torch::tensor({parameter_values}),")
print(" },")
print(" };")
print("}\n")
print(FOOTER)
def main():
parser = argparse.ArgumentParser(
"Produce optimization output baseline from PyTorch"
)
parser.add_argument("-i", "--iterations", default=1001, type=int)
parser.add_argument("-s", "--sample-every", default=100, type=int)
options = parser.parse_args()
optimizer_parameter_map = {}
for optimizer in OPTIMIZERS.keys():
sys.stderr.write(f"Evaluating {optimizer} ...\n")
optimizer_parameter_map[optimizer] = run(
optimizer, options.iterations, options.sample_every
)
emit(optimizer_parameter_map)
if __name__ == "__main__":
main()
|
