Why does tanh activation work better with Pytorch than with Keras?

I'm doing a neural network to recognize written Cyrillic letters, and I found out that, when I use tanh activation function, it works WAY better with PyTorch than with Keras.

Keras code:

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten, Lambda, BatchNormalization
from tensorflow.keras.callbacks import ReduceLROnPlateau, EarlyStopping

a_param = 0.5
activation_type = 'tanh'

custom_activation = create_activation_function(a_param, activation_type)

input_shape = (100, 100, 4)

model = Sequential()
model.add(tf.keras.layers.InputLayer(input_shape=input_shape))
model.add(Flatten())

model.add(Dense(784))
model.add(BatchNormalization())
model.add(Lambda(custom_activation))

model.add(Dense(30))
model.add(BatchNormalization())
model.add(Lambda(custom_activation))

model.add(Dense(10))
model.add(Lambda(custom_activation))

model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3),
              loss='categorical_crossentropy',
              metrics=['accuracy'])

model.summary()

epochs = 30
batch_size = 128

lr_schedule = ReduceLROnPlateau(
    monitor='val_loss', factor=0.5, patience=3, min_lr=1e-6
)

early_stopping = EarlyStopping(
    monitor='val_loss', patience=3, restore_best_weights=True
)

history = model.fit(
    X_train, y_train,
    epochs=epochs,
    batch_size=batch_size,
    validation_data=(X_test, y_test),
    shuffle=True,
    callbacks=[lr_schedule, early_stopping]
)

Pytorch code:

import torch
import torch.nn as nn

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        # 100*100 изображения с 4 каналами (RGBA)
        self.fc1 = nn.Linear(100 * 100 * 4, 128)  # Входной слой для изображений 100x100 с 4 каналами
        self.fc2 = nn.Linear(128, 64)
        self.fc3 = nn.Linear(64, 10)

    def forward(self, x):
        # Изменение формы тензора в вектор длины 100 * 100 * 4
        x = x.view(-1, 100 * 100 * 4)
        x = torch.tanh(self.fc1(x))  # Использование tanh
        x = torch.tanh(self.fc2(x))
        x = self.fc3(x)
        return torch.tanh(x)  # Применение tanh на выходе

While Keras the accuracy is around +/- 10%, while Pytorch is WAY above - up to 90%. What can be cause of it? Did I do something wrong or is it supposed to work this way?