def test_get(self):
nn1 = core.FeedForward(momentum=0.1, learn_rate=0.1)
nn1 += layers.Tanh(2, 2)
nn1 += layers.Linear(2, 2)
nn2 = core.FeedForward(momentum=0.1, learn_rate=0.1)
nn2 += layers.Tanh(2, 2)
nn2 += layers.Linear(2, 2)
ensemble = core.Ensemble(nn1, nn2)
ensemble.fit([0, 1], [2, 1])
stack = numpy.vstack((nn1.get([0, 0]), nn2.get([0, 0])))
self.assertEqual(round(ensemble.get([0, 0])[0] * 1000),
round((stack.sum(axis=0) / len(stack))[0] * 1000))
def test_dropout_after_training(self):
n = core.FeedForward(momentum=0.1, learn_rate=0.1)
drop = layers.Dropout(layers.Tanh(2, 2), percentage=0.5)
n += layers.Linear(2, 2)
n += drop
n += layers.Linear(2, 1)
s = [
([0, 0], [0]),
([0, 1], [1]),
([1, 0], [1]),
([1, 1], [0]),
]
n.fit(*s[1])
n.fit(*s[0])
n.fit(*s[2])
n.fit(*s[0])
n.fit(*s[1])
zeros = 0
for row in drop.y:
if row[0] == 0:
zeros += 1
self.assertEqual(zeros, len(drop.w) // 2)
def test_layers_addition(self):
v = core.FeedForward()
v += layers.Linear(2, 3)
v += layers.Tanh(3, 2)
v += layers.Linear(2, 1)
self.assertEqual(len(v.layers), 3)
self.assertEqual(len(v.layers[1].v), 2)
def test_cv(self):
import core.estimators
n = core.FeedForward(momentum=0.1, learn_rate=0.1)
n += layers.Linear(2, 2)
n += layers.Tanh(2, 1)
n += layers.Linear(1, 0)
s = [
([0, 0], [0]),
([0, 1], [1]),
([1, 0], [1]),
([1, 1], [0]),
]
error = core.estimators.cv(n, s)
self.assertTrue(type(error) is float)
def test_by_xor(self):
error = 0.1
n = core.FeedForward(
momentum=0.1,
learn_rate=0.1) # .create([2, 2, 1], default=layers.Tanh)
n += layers.Tanh(2, 2)
n += layers.Linear(2, 1)
s = [
([0, 0], [0]),
([0, 1], [1]),
([1, 0], [1]),
([1, 1], [0]),
]
for i in range(10_000):
r = random.randint(0, len(s) - 1)
n.fit(*s[r])
def test_xor_by_train(self):
error = 0.1
n = core.FeedForward(momentum=0.1, learn_rate=0.1)
n += layers.Tanh(2, 2)
n += layers.Linear(2, 1)
s = [
([0, 0], [0]),
([0, 1], [1]),
([1, 0], [1]),
([1, 1], [0]),
]
n.train(s, 10_000)
for v in s:
res = n.get(v[0])
self.assertTrue(abs(v[1][0] - res[0]) < error)
for v in s:
print(n.get(v[0]), end='\n\n')
def test_tanh(self):
l = layers.Tanh(3, 3)
self.assertEqual(l.a(0), 0)
self.assertEqual(l.der(0), 1)
import matplotlib.pyplot as plot
import random
import layers
from core import FeedForward
nn = FeedForward(momentum=0.1, learn_rate=0.1,
weight_decay=0.2) # .create([2, 2, 1], default=layers.Tanh)
nn += layers.Tanh(2, 2)
nn += layers.Linear(2, 1)
s = [
([0, 0], [0]),
([0, 1], [1]),
([1, 0], [1]),
([1, 1], [0]),
]
E = []
for i in range(10_000):
r = random.randint(0, len(s) - 1)
nn.fit(*s[r])
E.append(nn.error)
print(nn)
for v in s:
print(nn.get(v[0]), end='\n\n')
def __init__(self, *args):
self.layer = layers.Tanh()
self.inputs = args
import numpy
import layers
from core import FeedForward, separate_data
from core.estimators import cv
def f(x):
return 0.5 * numpy.sin(numpy.exp(x)) - numpy.cos(numpy.exp(-1 * x))
nn = FeedForward(momentum=0.2, learn_rate=0.05,
weight_decay=0.2) # .create([2, 2, 1], default=layers.Tanh)
nn += layers.Tanh(1, 10)
nn += layers.Tanh(10, 10)
nn += layers.Tanh(10, 10)
nn += layers.Tanh(10, 10)
nn += layers.Linear(10, 1)
data = [([x], [f(x)]) for x in numpy.linspace(-2.2, 2.5, 150)]
ts, vs = separate_data(data, 0.15)
# duplicate x and y for easy plotting
x = numpy.linspace(-2.2, 2.5, 150)
y = f(x)
error = []
v_error = []
from core.estimators import cv
def normal(arr):
s = numpy.sum(numpy.abs(arr))
return numpy.round(numpy.abs(arr) / s, decimals=2)
training, validation = separate_data(
from_csv("D:\\DELETE\\Дипломмо\\output.csv"), 0.15)
# noise(training, from_range=(0, 2), axis=0)
# noise(training, from_range=(-0.05, 0.05), axis=1)
ff1 = FeedForward(learn_rate=0.05, momentum=0.2, weight_decay=0.5)
ff1 += layers.Tanh(6, 23)
ff1 += layers.Dropout(layers.Tanh(23, 28), percentage=0.3)
ff1 += layers.Dropout(layers.Tanh(28, 28), percentage=0.3)
ff1 += layers.Dropout(layers.Tanh(28, 28), percentage=0.3)
ff1 += layers.Dropout(layers.Tanh(28, 28), percentage=0.3)
ff1 += layers.Linear(28, 8)
ff2 = FeedForward(learn_rate=0.07, momentum=0.2, weight_decay=0.23)
ff2 += layers.Tanh(6, 23)
ff2 += layers.Dropout(layers.Tanh(23, 28), percentage=0.3)
ff2 += layers.Dropout(layers.Tanh(28, 28), percentage=0.3)
ff2 += layers.Dropout(layers.Tanh(28, 28), percentage=0.3)
ff2 += layers.Dropout(layers.Tanh(28, 28), percentage=0.3)
ff2 += layers.Linear(28, 8)
ff3 = FeedForward(learn_rate=0.04, momentum=0.6, weight_decay=0.4)