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
|
"""
# -*- coding: utf-8 -*-
#
# Copyright 2021 Michael Büsch <m@bues.ch>
#
# Licensed under the Apache License version 2.0
# or the MIT license, at your option.
# SPDX-License-Identifier: Apache-2.0 OR MIT
#
"""
from mlplib.activation import *
from mlplib.backward import *
from mlplib.forward import *
from mlplib.init import *
from mlplib.loss import *
from mlplib.optimize import *
from mlplib.parameters import *
from mlplib.util import *
import numpy as np
def make_net():
seed(42)
nr_inputs = 4
layout = (6, 9, 2)
params = Parameters(
weights=init_layers_weights(nr_inputs, layout),
biases=init_layers_biases(layout),
actvns=[
ReLU(),
ReLU(),
Sigmoid(),
],
)
return params
def run_net(params, optimizer, decay=False):
x = standard_normal((20, params.nr_inputs))
y = standard_normal((20, params.layout[-1]))
lossfn = MSE()
prev_loss = 9999.0
prev_alpha = optimizer.alpha + 9999.0
for i in range(100):
gradients, yh = backward_prop(x, y, params, lossfn)
new_loss = lossfn.fn(yh, y)
assert new_loss < prev_loss
prev_loss = new_loss
optimizer.apply(i, gradients)
if decay:
assert optimizer.alpha < prev_alpha
prev_alpha = optimizer.alpha
def test_gradient_descent():
params = make_net()
optimizer = GradDescent(params=params, alpha=0.1)
run_net(params, optimizer)
def test_momentum():
params = make_net()
optimizer = Momentum(params=params, alpha=0.1, beta=0.9)
run_net(params, optimizer)
def test_rms_prop():
params = make_net()
optimizer = RMSProp(params=params, alpha=0.01, beta=0.9)
run_net(params, optimizer)
def test_adam():
params = make_net()
optimizer = Adam(params=params, alpha=0.01, beta1=0.9, beta2=0.9)
run_net(params, optimizer)
def test_decay_simple():
params = make_net()
optimizer = AlphaDecaySimple(
Adam(params=params, alpha=0.01, beta1=0.9, beta2=0.9),
decay_rate=0.15)
run_net(params, optimizer, decay=True)
def test_decay_exp():
params = make_net()
optimizer = AlphaDecayExp(
Adam(params=params, alpha=0.01, beta1=0.9, beta2=0.9),
decay_rate=0.15)
run_net(params, optimizer, decay=True)
def test_decay_sqrt():
params = make_net()
optimizer = AlphaDecaySqrt(
Adam(params=params, alpha=0.01, beta1=0.9, beta2=0.9),
decay_rate=0.15)
run_net(params, optimizer, decay=True)
# vim: ts=4 sw=4 expandtab
|
