pytorch模型中的参数如何不出现在计算图中?


10

我正在尝试更新/更改神经网络模型的参数,然后使更新的神经网络的正向传递在计算图中(无论我们进行了多少更改/更新)。

我尝试了这个想法,但是每当我这样做时,pytorch都会将更新的张量(在模型内部)设置为叶子,这会终止渐变流到我要接收渐变的网络。它杀死了梯度流,因为叶子节点不是我希望它们成为计算图形的一部分(因为它们不是真正的叶子)。

我已经尝试了多种方法,但似乎没有任何效果。我创建了一个自包含的虚拟代码,该代码打印了我希望具有渐变的网络的渐变:

import torch
import torch.nn as nn

import copy

from collections import OrderedDict

# img = torch.randn([8,3,32,32])
# targets = torch.LongTensor([1, 2, 0, 6, 2, 9, 4, 9])
# img = torch.randn([1,3,32,32])
# targets = torch.LongTensor([1])
x = torch.randn(1)
target = 12.0*x**2

criterion = nn.CrossEntropyLoss()

#loss_net = nn.Sequential(OrderedDict([('conv0',nn.Conv2d(in_channels=3,out_channels=10,kernel_size=32))]))
loss_net = nn.Sequential(OrderedDict([('fc0', nn.Linear(in_features=1,out_features=1))]))

hidden = torch.randn(size=(1,1),requires_grad=True)
updater_net = nn.Sequential(OrderedDict([('fc0',nn.Linear(in_features=1,out_features=1))]))
print(f'updater_net.fc0.weight.is_leaf = {updater_net.fc0.weight.is_leaf}')
#
nb_updates = 2
for i in range(nb_updates):
    print(f'i = {i}')
    new_params = copy.deepcopy( loss_net.state_dict() )
    ## w^<t> := f(w^<t-1>,delta^<t-1>)
    for (name, w) in loss_net.named_parameters():
        print(f'name = {name}')
        print(w.size())
        hidden = updater_net(hidden).view(1)
        print(hidden.size())
        #delta = ((hidden**2)*w/2)
        delta = w + hidden
        wt = w + delta
        print(wt.size())
        new_params[name] = wt
        #del loss_net.fc0.weight
        #setattr(loss_net.fc0, 'weight', nn.Parameter( wt ))
        #setattr(loss_net.fc0, 'weight', wt)
        #loss_net.fc0.weight = wt
        #loss_net.fc0.weight = nn.Parameter( wt )
    ##
    loss_net.load_state_dict(new_params)
#
print()
print(f'updater_net.fc0.weight.is_leaf = {updater_net.fc0.weight.is_leaf}')
outputs = loss_net(x)
loss_val = 0.5*(target - outputs)**2
loss_val.backward()
print()
print(f'-- params that dont matter if they have gradients --')
print(f'loss_net.grad = {loss_net.fc0.weight.grad}')
print('-- params we want to have gradients --')
print(f'hidden.grad = {hidden.grad}')
print(f'updater_net.fc0.weight.grad = {updater_net.fc0.weight.grad}')
print(f'updater_net.fc0.bias.grad = {updater_net.fc0.bias.grad}')

如果有人知道如何执行此操作,请给我ping命令...我将更新次数设置为2,因为更新操作应在计算图中进行任意次数...因此它必须适用于2。


密切相关的帖子:

交叉发布:


您是否尝试了争论backward?即retain_graph=True和/或create_graph=True
Szymon Maszke

Answers:


3

无法正常工作,因为命名参数模块被删除。


似乎可行:

import torch
import torch.nn as nn

from torchviz import make_dot

import copy

from collections import OrderedDict

# img = torch.randn([8,3,32,32])
# targets = torch.LongTensor([1, 2, 0, 6, 2, 9, 4, 9])
# img = torch.randn([1,3,32,32])
# targets = torch.LongTensor([1])
x = torch.randn(1)
target = 12.0*x**2

criterion = nn.CrossEntropyLoss()

#loss_net = nn.Sequential(OrderedDict([('conv0',nn.Conv2d(in_channels=3,out_channels=10,kernel_size=32))]))
loss_net = nn.Sequential(OrderedDict([('fc0', nn.Linear(in_features=1,out_features=1))]))

hidden = torch.randn(size=(1,1),requires_grad=True)
updater_net = nn.Sequential(OrderedDict([('fc0',nn.Linear(in_features=1,out_features=1))]))
print(f'updater_net.fc0.weight.is_leaf = {updater_net.fc0.weight.is_leaf}')
#
def del_attr(obj, names):
    if len(names) == 1:
        delattr(obj, names[0])
    else:
        del_attr(getattr(obj, names[0]), names[1:])
def set_attr(obj, names, val):
    if len(names) == 1:
        setattr(obj, names[0], val)
    else:
        set_attr(getattr(obj, names[0]), names[1:], val)

nb_updates = 2
for i in range(nb_updates):
    print(f'i = {i}')
    new_params = copy.deepcopy( loss_net.state_dict() )
    ## w^<t> := f(w^<t-1>,delta^<t-1>)
    for (name, w) in list(loss_net.named_parameters()):
        hidden = updater_net(hidden).view(1)
        #delta = ((hidden**2)*w/2)
        delta = w + hidden
        wt = w + delta
        del_attr(loss_net, name.split("."))
        set_attr(loss_net, name.split("."), wt)
    ##
#
print()
print(f'updater_net.fc0.weight.is_leaf = {updater_net.fc0.weight.is_leaf}')
print(f'loss_net.fc0.weight.is_leaf = {loss_net.fc0.weight.is_leaf}')
outputs = loss_net(x)
loss_val = 0.5*(target - outputs)**2
loss_val.backward()
print()
print(f'-- params that dont matter if they have gradients --')
print(f'loss_net.grad = {loss_net.fc0.weight.grad}')
print('-- params we want to have gradients --')
print(f'hidden.grad = {hidden.grad}') # None because this is not a leaf, it is overriden in the for loop above.
print(f'updater_net.fc0.weight.grad = {updater_net.fc0.weight.grad}')
print(f'updater_net.fc0.bias.grad = {updater_net.fc0.bias.grad}')
make_dot(loss_val)

输出:

updater_net.fc0.weight.is_leaf = True
i = 0
i = 1

updater_net.fc0.weight.is_leaf = True
loss_net.fc0.weight.is_leaf = False

-- params that dont matter if they have gradients --
loss_net.grad = None
-- params we want to have gradients --
hidden.grad = None
updater_net.fc0.weight.grad = tensor([[0.7152]])
updater_net.fc0.bias.grad = tensor([-7.4249])

致谢:来自pytorch团队的强大分子:https ://discuss.pytorch.org/t/how-does-one-have-the-parameters-of-a-model-not-be-leafs/70076/9?u =匹诺曹


伙计们,这是错误的,请不要使用此代码,因为它不允许传播渐变超过1个步骤。改用它:github.com/facebookresearch/higher
Pinocchio

这是行不通的!
Pinocchio

高等图书馆对我也不起作用。
Pinocchio

By using our site, you acknowledge that you have read and understand our Cookie Policy and Privacy Policy.
Licensed under cc by-sa 3.0 with attribution required.