Learner layer_groups parameter

I’m having trouble with the layer_groups parameter on the Learner class. When I use the default, my model converges just fine, but when I set the parameter to a list of the model’s modules, it stubbornly refuses to make progress. Here’s a distilled example:

import numpy as np

import torch
import torch.nn.functional as F
import torch.nn as nn

from torch.utils.data import DataLoader
from torch.utils.data.dataset import TensorDataset

from fastai import *

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()

        self.linear1 = nn.Linear(1, 5)
        self.linear2 = nn.Linear(5, 1)

    def forward(self, x):
        x = self.linear1(x)
        x = self.linear2(x)

        return x
    
def generate_data(size):
    x = np.random.uniform(size=(size, 1))
    y = x * 2.0
    return torch.FloatTensor(x), torch.FloatTensor(y)

train_x, train_y = generate_data(1000)
val_x, val_y = generate_data(100)

train_ds = TensorDataset(train_x, train_y)
val_ds = TensorDataset(val_x, val_y)

train_dl = DataLoader(train_ds, batch_size=8)
val_dl = DataLoader(val_ds, batch_size=8)

data_bunch = DataBunch(train_dl, val_dl)

model = SimpleModel()
learn = Learner(data_bunch, model, loss_func=F.mse_loss)
learn.fit_one_cycle(1)

# ... converges nicely

model = SimpleModel()
learn = Learner(data_bunch, model, loss_func=F.mse_loss, layer_groups=[model.linear1, model.linear2])
learn.fit_one_cycle(1)

# ... makes no discernible progress

Any idea what I might be doing wrong?

2 Likes

I found a way to make it go. The groups need to be wrapped in nn.Sequential for some reason:

model = SimpleModel()
layer_groups=[nn.Sequential(model.linear1), nn.Sequential(model.linear2)]
learn = Learner(data_bunch, model, loss_func=F.mse_loss, layer_groups=layer_groups)
learn.fit_one_cycle(1)
3 Likes

Thanks for sharing this. When I run either of these, and I get to learn.fit_one_cycle(1), it spits out a bunch of errors and ends with

TypeError: on_epoch_end() missing 1 required positional argument: ‘smooth_loss’

Interesting… this works in fastai 1.0.19 but not in any of the subsequent versions. I wonder what changed.

The problem is the batch size. DataLoader has a default batch size of 1, and it seems that fastai stopped supporting that. I updated the code snippet above with a new batch size and it now works fine with fastai 1.0.22.

1 Like

Oh cool. Thanks.
Last night I merged your work with Jeremy’s Lesson 5 code and produced the following gist which also works: https://gist.github.com/drscotthawley/2288e92f23b02e7fd5352708fc6cd125

1 Like

Hello @sgugger, I was using fastai2 for one of my projects and wanted to make sure whether the layer_groups have been applied successfully. With fastai-v1, I believe Learner had a layer_groups attribute which we could use for debugging purpose. I found no equivalent in fastai2, let me know if I’m missing something and how could I do the same with fastai2.

Thanks!

In v2, you need to provide a splitter to split your model in several parameter groups. Look at the notebook vision.learner or text.learner to see some examples.

1 Like

splitter is the way we apply splits but layer_groups used to be kind of sanity check whether splits have been applied correctly. I was talking about this parameter from v1.

Like the one utility function by @PPPW allows us to check layer groups of any custom architecture

The parameter groups are directly inside the optimizer now, so if you want to check them you need to create one (learn.create_opt()) then check learn.opt.pgs.

1 Like

Okay. Thanks!