How does one have parameters in a pytorch model not be leafs and be in the computation graph?

问题

I am trying to update/change the parameters of a neural net model and then having the forward pass of the updated neural net be in the computation graph (no matter how many changes/updates we do).

I tried this idea but whenever I do it pytorch sets my updated tensors (inside the model) to be leafs, which kills the flow of gradients to the networks I want to receive gradients. It kills the flow of gradients because leaf nodes are not part of the computation graph the way I want them to be (since they aren't truly leafs).

I've tried multiple things but nothing seems to work. I created a dummy code that is self contained that prints the gradients of the networks I desire to have gradients:

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}')

if anyone knows how to do this please give me a ping...I set the the number of times to update to be 2 because the update operation should be in the computation graph an arbitrary number of times...so it MUST work for 2.

Strongly related post:

SO: How does one have parameters in a pytorch model not be leafs and be in the computation graph?
pytorch forum: https://discuss.pytorch.org/t/how-does-one-have-the-parameters-of-a-model-not-be-leafs/70076

Cross-posted:

Quora: https://www.quora.com/unanswered/How-does-one-have-parameters-in-a-PyTorch-model-not-be-leaves-and-be-in-the-computation-graph
reddit: https://www.reddit.com/r/pytorch/comments/f5gu3g/how_does_one_have_parameters_in_a_pytorch_model/

回答1:

DOESNT WORK PROPERLY cuz the named parameter modules get deleted.

Seems this works:

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)

output:

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])

Acknowledgement: mighty albanD from pytorch team: https://discuss.pytorch.org/t/how-does-one-have-the-parameters-of-a-model-not-be-leafs/70076/9?u=pinocchio

回答2:

You should try to keep the same tensors, not create new ones.

Go for their data attribute and set a new value.

for (name, w) in loss_net.named_parameters():
    ....
    w.data = wt.data

This worked for me in this question: How to assign a new value to a pytorch Variable without breaking backpropagation?

来源：https://stackoverflow.com/questions/60271131/how-does-one-have-parameters-in-a-pytorch-model-not-be-leafs-and-be-in-the-compu

标签

python

machine-learning

neural-network

pytorch