由于計算機內(nèi)存的限制，如果使用梯度下降訓練神經(jīng)網(wǎng)絡(luò)的時候，通常只能一次輸入一個小批次數(shù)據(jù)，而不能一次性地輸入所有數(shù)據(jù)。minibatch 可以通過分批次處理訓練數(shù)據(jù)，解決內(nèi)存不足的問題，同時提高了訓練的效率。

2、有效提高計算速度。

當數(shù)據(jù)增大時，迭代次數(shù)越多，訓練越耗時。minibatch的應(yīng)用可以提高計算速度，提升訓練效率。

三、minibatch的實現(xiàn)方式

1、手動生成minibatch。

import numpy as np

def gen_minibatch(inputs, targets, batch_size):
    '''
    inputs和targets為輸入的數(shù)據(jù)和對應(yīng)的數(shù)據(jù)標簽
    batch_size為每個batch的大小
    '''
    input_batches = inputs.reshape(-1, batch_size, inputs.shape[1])
    target_batches = targets.reshape(-1, batch_size, targets.shape[1])
        
    for i in range(len(input_batches)):
        yield input_batches[i], target_batches[i]

# 輸入數(shù)據(jù)和標記
X = np.random.rand(40, 4)
y = np.random.randint(0, 2, 40)

# 批大小
batch_size = 10

for minibatch in gen_minibatch(X, y, batch_size):
    input_data, target_data = minibatch
    # do something

import torch
from torch.utils.data import DataLoader, TensorDataset

# 定義數(shù)據(jù)集并轉(zhuǎn)化為 DataLoader
train_dataset = TensorDataset(torch.Tensor(X_train), torch.Tensor(y_train))
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)

# 訓練模型
for i, (X_batch, y_batch) in enumerate(train_dataloader):
    # do something with X_batch and y_batch

shuffled_X, shuffled_y = shuffle(X_train, y_train)
train_dataset = TensorDataset(torch.Tensor(shuffled_X), torch.Tensor(shuffled_y))
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)