使用 Python 的 numpy 实现随机梯度下降

Stochastic gradient descent implementation with Python's numpy

我必须使用 python numpy 库实现随机梯度下降。为此,我给出了以下函数定义:

def compute_stoch_gradient(y, tx, w):
    """Compute a stochastic gradient for batch data."""

def stochastic_gradient_descent(
        y, tx, initial_w, batch_size, max_epochs, gamma):
    """Stochastic gradient descent algorithm."""

我还获得了以下帮助功能:

def batch_iter(y, tx, batch_size, num_batches=1, shuffle=True):
    """
    Generate a minibatch iterator for a dataset.
    Takes as input two iterables (here the output desired values 'y' and the input data 'tx')
    Outputs an iterator which gives mini-batches of `batch_size` matching elements from `y` and `tx`.
    Data can be randomly shuffled to avoid ordering in the original data messing with the randomness of the minibatches.
    Example of use :
    for minibatch_y, minibatch_tx in batch_iter(y, tx, 32):
        <DO-SOMETHING>
    """
    data_size = len(y)

    if shuffle:
        shuffle_indices = np.random.permutation(np.arange(data_size))
        shuffled_y = y[shuffle_indices]
        shuffled_tx = tx[shuffle_indices]
    else:
        shuffled_y = y
        shuffled_tx = tx
    for batch_num in range(num_batches):
        start_index = batch_num * batch_size
        end_index = min((batch_num + 1) * batch_size, data_size)
        if start_index != end_index:
            yield shuffled_y[start_index:end_index], shuffled_tx[start_index:end_index]

我实现了以下两个功能:

def compute_stoch_gradient(y, tx, w):
    """Compute a stochastic gradient for batch data."""
    e = y - tx.dot(w)
    return (-1/y.shape[0])*tx.transpose().dot(e)


def stochastic_gradient_descent(y, tx, initial_w, batch_size, max_epochs, gamma):
    """Stochastic gradient descent algorithm."""
    ws = [initial_w]
    losses = []
    w = initial_w
    for n_iter in range(max_epochs):
        for minibatch_y,minibatch_x in batch_iter(y,tx,batch_size):
            w = ws[n_iter] - gamma * compute_stoch_gradient(minibatch_y,minibatch_x,ws[n_iter])
            ws.append(np.copy(w))
            loss = y - tx.dot(w)
            losses.append(loss)

    return losses, ws

我不确定应该在范围 (max_epochs) 或更大的范围内进行迭代。我这样说是因为我读到一个纪元是 "each time we run through the entire data set"。所以我认为一个时代包含更多的一次迭代...

在典型的实现中,批量大小为 B 的小批量梯度下降应从数据集中随机选取 B 个数据点,并根据该子集上计算出的梯度更新权重。这个过程本身会持续很多次,直到收敛或某个阈值最大迭代。 B=1 的 mini-batch 是 SGD,有时会有噪音。

除了上述评论外,您可能还想尝试一下批量大小和学习率(步长),因为它们对随机和小批量梯度下降的收敛速度有重要影响。

下图显示了这两个参数对 SGDlogistic regression 的收敛速度的影响,同时对亚马逊产品评论数据集进行情感分析,该作业出现在 coursera 课程中机器学习 - 华盛顿大学的分类:

更详细的信息可以参考https://sandipanweb.wordpress.com/2017/03/31/online-learning-sentiment-analysis-with-logistic-regression-via-stochastic-gradient-ascent/?frame-nonce=987e584e16