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 /
Отправка уведомления на почту при изменении наименования справочника Maks 24.03.2026 Программная отправка письма электронной почты на примере изменения наименования типового справочника "Склады" в конфигурации БП3. Перед реализацией необходимо выполнить настройку системной учетной. . .	модель ЗдравоСохранения 5. Меньше увольнений- больше дохода! anaschu 24.03.2026 Теперь система здравосохранения уменьшает количество увольнений. 9TO2GP2bpX4 a42b81fb172ffc12ca589c7898261ccb/ https:/ / rutube. ru/ video/ a42b81fb172ffc12ca589c7898261ccb/ Слева синяя линия -. . .	Midnight Chicago Blues kumehtar 24.03.2026 Такой Midnight Chicago Blues, знаешь?. . Когда вечерние улицы становятся ночными, а ты не можешь уснуть. Ты идёшь в любимый старый бар, и бармен наливает тебе виски. Ты смотришь на пролетающие. . .	SDL3 для Desktop (MinGW): Вывод текста со шрифтом TTF с помощью библиотеки SDL3_ttf на Си и C++ 8Observer8 24.03.2026 Содержание блога Финальные проекты на Си и на C++: finish-text-sdl3-c. zip finish-text-sdl3-cpp. zip
Жизнь в неопределённости kumehtar 23.03.2026 Жизнь — это постоянное существование в неопределённости. Например, даже если у тебя есть список дел, невозможно дойти до точки, где всё окончательно завершено и больше ничего не осталось. В принципе,. . .	Модель здравоСохранения: работники работают быстрее после её введения. anaschu 23.03.2026 geJalZw1fLo Корпорация до введения программа здравоохранения имела много невыполненных работниками заданий, после введения программы количество заданий выросло. Но на выплатах по больничным это. . .	Контроль уникальности заводского номера Maks 23.03.2026 Алгоритм контроля уникальности заводского (или серийного) номера на примере нетипового документа выдачи шин для спецтехники с табличной частью, разработанного в конфигурации КА2. Данные берутся из. . .	Хочу заставить корпорации вкладываться в здоровье сотрудников: делаю мат модель здравосохранения anaschu 22.03.2026 e7EYtONaj8Y Z4Tv2zpXVVo https:/ / github. com/ shumilovas/ med2. git

@u235 5516 / 2869 / 571 Регистрация: 07.11.2019 Сообщений: 4,760
	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 5516 / 2869 / 571 Регистрация: 07.11.2019 Сообщений: 4,760
	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