Unable to use first-order MAML approaches

[chnsh]

Hi,

Thank you for developing pytorch-meta - I have been trying to use the following method in the inner loop:

params = gradient_update_parameters(
                self.model, inner_losses, params=params, step_size=self.inner_step_size, first_order=True
            )

When I use this code and call .backward, step on the outer loss, I get an error as follows: RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.

Have you encountered something like this?

Pytorch version: 1.5.1

[tristandeleu]

Sorry for the late reply. I haven't been able to reproduce this issue unfortunately (PyTorch 1.6, Torchmeta 1.5.0). Here is an example

import torch
import torch.nn as nn

from torchmeta.modules import MetaSequential, MetaLinear
from torchmeta.utils import gradient_update_parameters

model = MetaSequential(
    MetaLinear(2, 3),
    nn.ReLU(),
    MetaLinear(3, 5)
)

train_inputs = torch.randn(7, 2)

train_outputs_1 = model(train_inputs)
inner_loss_1 = train_outputs_1.sum()  # Dummy loss
params = gradient_update_parameters(model, inner_loss_1, first_order=True)

train_outputs_2 = model(train_inputs, params=params)
inner_loss_2 = train_outputs_2.sum()  # Dummy loss
params = gradient_update_parameters(model, inner_loss_2, params=params, first_order=True)

test_inputs = torch.randn(7, 2)

test_outputs = model(test_inputs, params=params)
outer_loss = test_outputs.sum()

outer_loss.backward()

Can you give me a minimal example where you get this issue?