Не работает нейросеть на tensorflow

@mikinol · Регистрация: 16.12.2021

Студворк — интернет-сервис помощи студентам

Нейросеть постоянно выводит 1

Нейросеть:

Python
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
model = keras.Sequential([
    Input(shape=(10,)),
    Dense(10,activation='relu'),
    Dense(9,activation='relu'),
    Dense(9,activation='relu'),
    Dense(8,activation='relu'),
    Dense(8,activation='relu'),
    Dense(7,activation='relu'),
    Dense(7,activation='relu'),
    Dense(6,activation='relu'),
    Dense(6,activation='relu'),
    Dense(5,activation='relu'),
    Dense(5,activation='relu'),
    Dense(4,activation='relu'),
    Dense(4,activation='relu'),
    Dense(3,activation='relu'),
    Dense(3,activation='relu'),
    Dense(2,activation='relu'),
    Dense(2,activation='relu'),
    Dense(1,activation='relu'),
    Dense(1,activation='softmax')
])
model.compile(optimizer='adam', loss=tf.keras.losses.BinaryCrossentropy(from_logits=True), metrics=['accuracy','TruePositives'])

Датасет:

Python
1
2
3
4
5
6
data_in = tf.constant([
[0.0753442,0.0755078,0.0755801,0.0755446,0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489],[0.0755078,0.0755801,0.0755446,0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864],[0.0755801,0.0755446,0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456],[0.0755446,0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452],[0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724],[0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746],[0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397],[0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398],[0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408],[0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969],[0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338],[0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355],[0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047],[0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539],[0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889],[0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353],[0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548],[0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724],[0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221],[0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072],[0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807],[0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354],[0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792],[0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196],[0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856],[0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451],[0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861],[0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253],[0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539],[0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539,0.0753813],[0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539,0.0753813,0.075529],[0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539,0.0753813,0.075529,0.0754916],[0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539,0.0753813,0.075529,0.0754916,0.0757147]
])
data_out = tf.constant([
1,0,1,0,0,1,1,0,0,0,1,1,0,0,0,1,1,0,0,0,1,1,0,1,0,1,1,1,0,1,0,1,0
])

@u235 · 06.03.2023, 08:19

mikinol, попробуйте нормировать/шкалировать входные данные. У вас очень низкая вариация входных данных. И модель, думаю, переусложнена для такого малого количества входных данных.

@VTsaregorodtsev · 06.03.2023, 10:03

Сообщение от mikinol

Нейросеть:

Выкинь все промежуточные слои между первым и последним dense.
Вот если не будет точности обучения - только тогда и напихаешь ещё нейронов в первый слой (либо вставишь 1 промежуточный слой).

Сообщение от mikinol

Нейросеть постоянно выводит 1

Софтмакс не работает для сетки с одним выходом. Точнее, он в этом случае и будет всегда выдавать 1.
Если бы ты посмотрел на формулу - то сам бы это понял. Там каждый выход нормируется (делится) на сумму выходов, т.е. при одном выходе будет деление величины самОй на себя, что всегда будет давать 1.

@mikinol · 06.03.2023, 11:37 **[ТС]**

Я испавил вот так:

Python
1
2
3
4
5
6
model = keras.Sequential([
    Input(shape=(10,)),
    Dense(6,activation='relu'),
    Dense(2,activation='softmax')
])
model.compile(optimizer='adam', loss=tf.keras.losses.BinaryCrossentropy(from_logits=True), metrics=['accuracy','TruePositives'])

Python
1
2
3
4
5
6
data_in = tf.constant([
[0.0753442,0.0755078,0.0755801,0.0755446,0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489],[0.0755078,0.0755801,0.0755446,0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864],[0.0755801,0.0755446,0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456],[0.0755446,0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452],[0.0755074,0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724],[0.0755408,0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746],[0.0754946,0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397],[0.0755715,0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398],[0.075464,0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408],[0.0754489,0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969],[0.0754864,0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338],[0.075456,0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355],[0.0755452,0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047],[0.0754724,0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539],[0.0753746,0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889],[0.0754397,0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353],[0.0754398,0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548],[0.075408,0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724],[0.0753969,0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221],[0.075338,0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072],[0.075355,0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807],[0.0754047,0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354],[0.07539,0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792],[0.0753889,0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196],[0.0753353,0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856],[0.0753548,0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451],[0.0753724,0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861],[0.0752221,0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253],[0.0752072,0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539],[0.0751807,0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539,0.0753813],[0.0752354,0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539,0.0753813,0.075529],[0.0752792,0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539,0.0753813,0.075529,0.0754916],[0.075196,0.0752856,0.0751451,0.0751861,0.075253,0.07539,0.0753813,0.075529,0.0754916,0.0757147]
])
data_out = tf.constant([
[1,0],[0,1],[1,0],[0,1],[0,1],[1,0],[1,0],[0,1],[0,1],[0,1],[1,0],[1,0],[0,1],[0,1],[0,1],[1,0],[1,0],[0,1],[0,1],[0,1],[1,0],[1,0],[0,1],[1,0],[0,1],[1,0],[1,0],[1,0],[0,1],[1,0],[0,1],[1,0],[0,1]
])

Но всегда получается [0.51,0.49] или [0.49,0.51]

Добавлено через 16 минут
Всегда получается через 5 эпох обучения все ответы [0.51,0.49], а при более большом количестве эпох все ответы [0.49,0.51]

@u235 · 06.03.2023, 14:57

mikinol, обычно не делают два класса с софтмаксом, а делают один класс и сигмоиду с бинарной кроссэнтропией, насколько я знаю..

@Михалыч · 06.03.2023, 15:13

Не по теме:

Сообщение от VTsaregorodtsev

Там каждый выход нормируется (делится) на сумму выходов, т.е. при одном выходе будет деление величины самОй на себя, что всегда будет давать 1.

Сообщение от u235

делают один класс и сигмоиду с бинарной кроссэнтропией

как жаль что в этом во всем не разбираюсь, но так хочется, пока на практике куда применить еще не знаю. Здорово что на форуме есть люди к кому потом можно будет обратиться.

@u235 · 06.03.2023, 19:29

Не по теме:

Михалыч, я сам, можно сказать, начинающий в машинном обучении и нейросетях, тема обширная и интересная. Тут на форуме есть люди гораздо более компетентные (VTsaregorodtsev, passant) в этих вопросах, в отличие от меня. Но чем смогу - помогу.

Новые блоги и статьи Все статьи Все блоги /
Учёным и волонтёрам проекта «Einstein@home» удалось обнаружить четыре гамма-лучевых пульсара в джете Млечного Пути Programma_Boinc 01.01.2026 Учёным и волонтёрам проекта «Einstein@home» удалось обнаружить четыре гамма-лучевых пульсара в джете Млечного Пути Сочетание глобально распределённой вычислительной мощности и инновационных. . .	Советы по крайней бережливости. Внимание, это ОЧЕНЬ длинный пост. Programma_Boinc 28.12.2025 Советы по крайней бережливости. Внимание, это ОЧЕНЬ длинный пост. Налог на собак: https:/ / ********/ gallery/ V06K53e Финансовый отчет в Excel: https:/ / ********/ gallery/ bKBkQFf Пост отсюда. . .	Кто-нибудь знает, где можно бесплатно получить настольный компьютер или ноутбук? США. Programma_Boinc 26.12.2025 Нашел на реддите интересную статью под названием Anyone know where to get a free Desktop or Laptop? Ниже её машинный перевод. После долгих разбирательств я наконец-то вернула себе. . .	Thinkpad X220 Tablet — это лучший бюджетный ноутбук для учёбы, точка. Programma_Boinc 23.12.2025 Рецензия / Мнение/ Перевод Нашел на реддите интересную статью под названием The Thinkpad X220 Tablet is the best budget school laptop period . Ниже её машинный перевод. Thinkpad X220 Tablet —. . .	PhpStorm 2025.3: WSL Terminal всегда стартует в ~ and_y87 14.12.2025 PhpStorm 2025. 3: WSL Terminal всегда стартует в ~ (home), игнорируя директорию проекта Симптом: После обновления до PhpStorm 2025. 3 встроенный терминал WSL открывается в домашней директории. . .
Как объединить две одинаковые БД Access с разными данными VikBal 11.12.2025 Помогите пожалуйста !! Как объединить 2 одинаковые БД Access с разными данными.	Новый ноутбук volvo 07.12.2025 Всем привет. По скидке в "черную пятницу" взял себе новый ноутбук Lenovo ThinkBook 16 G7 на Амазоне: Ryzen 5 7533HS 64 Gb DDR5 1Tb NVMe 16" Full HD Display Win11 Pro	Музыка, написанная Искусственным Интеллектом volvo 04.12.2025 Всем привет. Некоторое время назад меня заинтересовало, что уже умеет ИИ в плане написания музыки для песен, и, собственно, исполнения этих самых песен. Стихов у нас много, уже вышли 4 книги, еще 3. . .	От async/await к виртуальным потокам в Python IndentationError 23.11.2025 Армин Ронахер поставил под сомнение async/ await. Создатель Flask заявляет: цветные функции - провал, виртуальные потоки - решение. Не threading-динозавры, а новое поколение лёгких потоков. Откат?. . .	Поиск "дружественных имён" СОМ портов Argus19 22.11.2025 Поиск "дружественных имён" СОМ портов На странице: https:/ / norseev. ru/ 2018/ 01/ 04/ comportlist_windows/ нашёл схожую тему. Там приведён код на С++, который показывает только имена СОМ портов, типа,. . .

@u235 5514 / 2867 / 571 Регистрация: 07.11.2019 Сообщений: 4,757
	06.03.2023, 08:19
	mikinol, попробуйте нормировать/шкалировать входные данные. У вас очень низкая вариация входных данных. И модель, думаю, переусложнена для такого малого количества входных данных. 1

@u235 5514 / 2867 / 571 Регистрация: 07.11.2019 Сообщений: 4,757
	06.03.2023, 14:57
	mikinol, обычно не делают два класса с софтмаксом, а делают один класс и сигмоиду с бинарной кроссэнтропией, насколько я знаю.. 1

@u235
	06.03.2023, 19:29
	Не по теме: Михалыч, я сам, можно сказать, начинающий в машинном обучении и нейросетях, тема обширная и интересная. Тут на форуме есть люди гораздо более компетентные (VTsaregorodtsev, passant) в этих вопросах, в отличие от меня. Но чем смогу - помогу. 0