区分梯度答案

【问题标题】：Differentiate gradients区分梯度
【发布时间】：2018-08-15 10:11:13
【问题描述】：

有没有办法在 PyTorch 中区分渐变？

例如，我可以在 TensorFlow 中做到这一点：

from pylab import *
import tensorflow as tf
tf.reset_default_graph()
sess = tf.InteractiveSession()

def gradient_descent( loss_fnc, w, max_its, lr):
    '''a gradient descent "RNN" '''    
    for k in range(max_its):
        w = w - lr * tf.gradients( loss_fnc(w), w )[0]
    return w

lr = tf.Variable( 0.0, dtype=tf.float32)
w = tf.Variable( tf.zeros(10), dtype=tf.float32)
reg = tf.Variable( 1.0, dtype=tf.float32 )

def loss_fnc(w):
    return tf.reduce_sum((tf.ones(10) - w)**2) + reg * tf.reduce_sum( w**2 )

w_n = gradient_descent( loss_fnc, w, 10, lr )

sess.run( tf.initialize_all_variables())

# differentiate through the gradient_descent RNN with respnect to the initial weight 
print(tf.gradients( w_n, w))

# differentiate through the gradient_descent RNN with respnect to the learning rate
print(tf.gradients( w_n, lr))

输出是

[<tf.Tensor 'gradients_10/AddN_9:0' shape=(10,) dtype=float32>]

[<tf.Tensor 'gradients_11/AddN_9:0' shape=() dtype=float32>]

我如何在 PyTorch 中做类似的事情？

【问题讨论】：

标签： python neural-network deep-learning gradient-descent pytorch

【解决方案1】：

您只需要使用函数torch.autograd.grad，它的作用与tf.gradients完全相同。

所以在 pytorch 中这将是：

from torch.autograd import Variable, grad
import torch



def gradient_descent( loss_fnc, w, max_its, lr):
    '''a gradient descent "RNN" '''    
    for k in range(max_its):
        w = w - lr * grad( loss_fnc(w), w )
    return w

lr = Variable(torch.zeros(1), , requires_grad=True)
w = Variable( torch.zeros(10), requires_grad=True)
reg = Variable( torch.ones(1) , requires_grad=True)

def loss_fnc(w):
    return torch.sum((Variable(torch.ones(10)) - w)**2) + reg * torch.sum( w**2 )

w_n = gradient_descent( loss_fnc, w, 10, lr )


# differentiate through the gradient_descent RNN with respnect to the initial weight 
print(grad( w_n, w))

# differentiate through the gradient_descent RNN with respnect to the learning rate
print(grad( w_n, lr))

【讨论】：