using LBFGS optimizer in pytorch lightening the model is not converging as compared to native pytoch + LBFGS

[ghost]

Common bugs:

Comparing the results of LBFGS + Pytorch lightening to native pytorch + LBFGS, Pytorch lightening is not able to update wights and model is not converging. there are some issues to point out:

1. Adam + Pytorch lightening on MNIST works fine, however LBFGS + Pytorch lightening is not working as expected.
2. LBFGS + Native pytorch works very well, however when we try LBFGS + Pytorch lightening it does not work as expected.

🐛 Bug

LBFGS + Pytorch Lightening has problem converging and weights are updating as compared to Adam + Pytorch lightening.

Code sample

import os
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torchvision.datasets import MNIST
from torchvision import transforms,datasets
from torch.utils.data import DataLoader,random_split
import pytorch_lightning as pl 
from IPython.display import clear_output

class LightningMNISTClassifier(pl.LightningModule):
  def __init__(self):
    super(LightningMNISTClassifier,self).__init__()
    self.layer_1 = nn.Linear(28 * 28, 128)
    self.layer_2 = nn.Linear(128, 256)
    self.layer_3 = nn.Linear(256, 10)

  def forward(self, x):
    batch_size, channels, width, height = x.size()
    x=x.view(batch_size,-1)
    # layer 1
    x = self.layer_1(x)
    x = torch.relu(x)
    # layer 2
    x = self.layer_2(x)
    x = torch.relu(x) 
    # layer 3
    x = self.layer_3(x)
    # probability distribution over labels
    x = torch.log_softmax(x, dim=1)  
    return x 
  def prepare_data(self):
    transform=transforms.Compose([transforms.ToTensor(), 
                                  transforms.Normalize((0.1307,), (0.3081,))])
    # prepare transforms standard to MNIST
    mnist_train = MNIST(os.getcwd(), train=True, download=True, transform=transform)
    mnist_test = MNIST(os.getcwd(), train=False, download=True, transform=transform)  
    self.mnist_train, self.mnist_val = random_split(mnist_train, [55000, 5000])

  def train_dataloader(self):
    return DataLoader(self.mnist_train,batch_size=1024)

  # def val_dataloader(self):
  #   return DataLoader(self.mnist_val,batch_size=1024)
  # def test_dataloader(self):
  #   return DataLoader(self.mnist_test,batch_size=1024)

  def configure_optimizers(self):
    # optimizer=optim.Adam(self.parameters(),lr=1e-3)
    optimizer = optim.LBFGS(self.parameters(), lr=1e-2)
    return optimizer

  # def backward(self, trainer, loss, optimizer):
  #   loss.backward(retain_graph=True)

  def optimizer_step(self, current_epoch, batch_nb, optimizer, optimizer_idx,
                     second_order_closure, on_tpu=False, using_native_amp=False,
                     using_lbfgs=False):
        # update params
      optimizer.step(second_order_closure) 

  def cross_entropy_loss(self,logits,labels):
    return F.nll_loss(logits,labels)

  def training_step(self,train_batch,batch_idx):
    x,y=train_batch
    logits=self.forward(x)
    loss=self.cross_entropy_loss(logits,y)
    return  {'loss':loss}

  def training_epoch_end(self,outputs):
    avg_loss=torch.stack([x['loss'] for x in outputs]).mean()
    print('epoch={}, avg_Train_loss={:.2f}'.format(self.current_epoch,avg_loss.item()))
    # return {'avg_train_loss':avg_loss}

  # def validation_step(self,val_batch,batch_idx):
  #   x,y=val_batch
  #   logits=self.forward(x)
  #   loss=self.cross_entropy_loss(logits,y)
  #   return {'val_loss':loss}
  # def validation_epoch_end(self,outputs):
  #   avg_loss=torch.stack([x['val_loss'] for x in outputs]).mean()
  #   print('epoch={}, avg_Test_loss={:.2f}'.format(self.current_epoch,avg_loss.item()))
  #   return {'avg_val_loss':avg_loss}

model=LightningMNISTClassifier()
#from pytorch_lightning.callbacks import EarlyStopping
trainer=pl.Trainer(max_epochs=400,gpus=1,
                  #  check_val_every_n_epoch=2,
                  #  accumulate_grad_batches=5,
#                   early_stop_callback=early_stop,
                  #  limit_train_batches=50,
#                   val_check_interval=0.25,
                   progress_bar_refresh_rate=0,
#                   num_sanity_val_steps=0,
                   weights_summary=None)
clear_output(wait=True)
trainer.fit(model)Preformatted text.

Expected behavior

Environment

Please copy and paste the output from our environment collection script (or fill out the checklist below manually).

You can get the script and run it with:

wget https://raw.githubusercontent.com/PyTorchLightning/pytorch-lightning/master/tests/collect_env_details.py
# For security purposes, please check the contents of collect_env_details.py before running it.
python collect_env_details.py

Environment: -Colab and pycharm -PyTorch version: 1.6.0+CPU and GPU -pytorch-lightning==1.0.0rc3

[rohitgr7]

do you have the code for native PyTorch + LBFGS for the same?

[ghost]

this is the code including MNIST and LBFGS that works fine with native pytorch:

import os
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torchvision.datasets import MNIST
from torchvision import transforms,datasets
from torch.utils.data import DataLoader,random_split

class PytorchMNISTClassifier(nn.Module):
  def __init__(self):
    super(PytorchMNISTClassifier,self).__init__()
    self.layer_1 = nn.Linear(28 * 28, 128)
    self.layer_2 = nn.Linear(128, 256)
    self.layer_3 = nn.Linear(256, 10)
  def forward(self, x):
    batch_size, channels, width, height = x.size()
    x=x.view(batch_size,-1)
    # layer 1
    x = self.layer_1(x)
    x = torch.relu(x)
    # layer 2
    x = self.layer_2(x)
    x = torch.relu(x) 
    # layer 3
    x = self.layer_3(x)
    # probability distribution over labels
    x = torch.log_softmax(x, dim=1)  
    return x 

def cross_entropy_loss(logits,labels):
  return F.nll_loss(logits,labels)

if __name__ == '__main__':

  if torch.cuda.is_available():
    device=torch.device('cuda:0')
  else:
    device=torch.device('cpu')

  model=PytorchMNISTClassifier()
  model=model.to(device)
  # optimizer=optim.Adam(model.parameters(),lr=1e-3)
  optimizer = optim.LBFGS(model.parameters(),lr=0.01)

  transform=transforms.Compose([transforms.ToTensor(), 
                                  transforms.Normalize((0.1307,), (0.3081,))])
  # prepare transforms standard to MNIST
  mnist_train = MNIST(os.getcwd(), train=True, download=True, transform=transform)
  mnist_test = MNIST(os.getcwd(), train=False, download=True, transform=transform)  
  mnist_train, mnist_val = random_split(mnist_train, [55000, 5000])

  data=DataLoader(mnist_train,batch_size=1024)

  for Epoch in range(10):
    loss_total=0.
    for i,(x,y) in enumerate(data):
      x=x.to(device)
      y=y.to(device)
      def closure():
        logits=model(x)
        optimizer.zero_grad()
        loss=cross_entropy_loss(logits,y)
        loss.backward(retain_graph=True)
        return loss
    loss_out = optimizer.step(closure)
    loss_total+=loss_out.item()
    print('total_loss--->', loss_total)

[williamFalcon]

You don't need to override optimizer_step... you're only doing it to pass in the second_order closure, but that's exactly what the default implementation does

        if on_tpu:
            xm.optimizer_step(optimizer)
        elif using_native_amp:
            self.trainer.scaler.step(optimizer)
        elif using_lbfgs:
            optimizer.step(second_order_closure)
        else:
            optimizer.step()

[rohitgr7]

@williamFalcon still should converge right?? even if the overridden method is doing the same update. Maybe a bug here if it's not converging in pl. will check this.

[ghost]

@williamFalcon we modified the code by removing optimizer_step, however it dose not help solving the issue.

[rohitgr7]

ok found something. not sure if it's correct or not since I haven't used LBFGS before.

I checked that optim.LBFGS calls closure 20 times for each step and in this example it doesn't call any step and .backward() explicitly but relies on optimizer.step(closure) to do that. Also in every 20 steps the underlying loss is different.

But pl calls an explicit training_step with the closure obviously that means it will be called 21 times + an explicit loss.backward() is called always.

These are my observations. Anyone with prior experience with LBFGS optimizer can confirm the right way to do this.

[williamFalcon]

how many times does it get called with pytorch?

LBFGS is a quasi knewton method which means it does not compute the hessian directly but instead it approximates it.

I assume pytorch calls step multiple times to do this approximation?

[rohitgr7]

how many times does it get called with pytorch?

the given example calls it 20 times. I think it always calls it 20 times, checked a few examples.

[ghost]

the default value for the number of iterations is 20 times , based on the pytorch help:

torch.optim.LBFGS(params, lr=1, max_iter=20, max_eval=None, tolerance_grad=1e-07, tolerance_change=1e-09, history_size=100, line_search_fn=None)

[ghost]

@williamFalcon We are in the process of developing a code that requires me to use LBFGS optimizer. I'd like to use pytorch-lightening platform for this code. do you think that the LBFGS issue can be resolved any time soon in the later versions?

[justusschock]

@rohitgr7 unfortunately it does not seem to be fixed by #4190 even though the number of backward calls are now correct (there is a test for that). The loss is still not decreasing though (haven't investigated further)

[rohitgr7]

ok will check this if I get some time :)

[ghost]

@williamFalcon it seems that the LBFGS optimizer in the latest version of pytorch-lightening carries the same issue as the previous versions. Is there a way to fix this issue temporarily up to the time that bug gets fixed.

[ghost]

@Borda , @edenlightning ,LBFGS issue dose not seem be fixed in the latest version of pytorch Lightening. should we hope that this issue could be fixed in the near future? we started a project using pytorch lightening and got stuck because of not being able to use LBFGS optimizer. if it is not fixed yet, would be possible to expedite resolving this issue?

[stale[bot]]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!

[Bajo1994]

@williamFalcon @Borda @edenlightning Since this thread will be closed automatically within the next 48 hours, I decided to mention you guys with the hope that the bug gets fixed in a meaningful period. I also appreciate @justusschock for his efforts to fix the issue. Ignoring a bug will not fix it, and it dramatically stops the research activities of people who trusted lightning. Please help us with fixing the bug.

[Bajo1994]

@carmocca I am very thankful if you take a look at the discussion made here to see whether you can help us fix the issue. The LBFGS bug in lightning has dramatically impacted an important project that I am working on.

[edenlightning]

Apologize for the delay! We try our best to take a look at every issue with the resources that we have. We bumped the priority for this one and will try to prioritize in the next sprints!

[Bajo1994]

@edenlightning I greatly appreciate your help on this subject.

[akihironitta]

As @justusschock added the tests in https://github.com/PyTorchLightning/pytorch-lightning/pull/4190 and I confirmed locally with cProfile, the number of backward passes (the number of times closure was called) in PL is 20 which is the same as native PyTorch, so this should be no problem.

PL code example (originally from @peymanpoozesh)

```python import os import warnings import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.utils.data import DataLoader, random_split from torchvision import transforms from torchvision.datasets import MNIST import pytorch_lightning as pl warnings.filterwarnings("ignore") pl.seed_everything(42) class LightningMNISTClassifier(pl.LightningModule): def __init__(self): super(LightningMNISTClassifier, self).__init__() self.layer_1 = nn.Linear(28 * 28, 128) self.layer_2 = nn.Linear(128, 256) self.layer_3 = nn.Linear(256, 10) def forward(self, x): batch_size, channels, width, height = x.size() x = x.view(batch_size, -1) x = self.layer_1(x) x = torch.relu(x) x = self.layer_2(x) x = torch.relu(x) x = self.layer_3(x) x = torch.log_softmax(x, dim=1) return x def prepare_data(self): transform = transforms.Compose( [transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))] ) mnist_train = MNIST(os.getcwd(), train=True, download=True, transform=transform) self.mnist_train, self.mnist_val = random_split( mnist_train, [55000, 5000], generator=torch.Generator().manual_seed(42) ) def train_dataloader(self): dl = DataLoader(self.mnist_train, batch_size=1024, num_workers=0) return dl def configure_optimizers(self): # optimizer = optim.Adam(self.parameters(), lr=1e-3) optimizer = optim.LBFGS(self.parameters(), lr=0.01, max_iter=20) return optimizer def training_step(self, train_batch, batch_idx): x, y = train_batch logits = self.forward(x) loss = F.nll_loss(logits, y) return {"loss": loss} def training_step_end(self, outputs): print("closure_loss:", outputs["loss"].item()) return outputs def main(): model = LightningMNISTClassifier() trainer = pl.Trainer( max_epochs=30, progress_bar_refresh_rate=0, weights_summary=None, # fast_dev_run=20, ) trainer.fit(model) if __name__ == "__main__": main() ```

native PyTorch code example (originally from @peymanpoozesh)

```python import os import warnings import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.utils.data import DataLoader, random_split from torchvision import transforms from torchvision.datasets import MNIST from pytorch_lightning import seed_everything warnings.filterwarnings("ignore") seed_everything(42) class PytorchMNISTClassifier(nn.Module): def __init__(self): super(PytorchMNISTClassifier, self).__init__() self.layer_1 = nn.Linear(28 * 28, 128) self.layer_2 = nn.Linear(128, 256) self.layer_3 = nn.Linear(256, 10) def forward(self, x): batch_size, channels, width, height = x.size() x = x.view(batch_size, -1) x = self.layer_1(x) x = torch.relu(x) x = self.layer_2(x) x = torch.relu(x) x = self.layer_3(x) x = torch.log_softmax(x, dim=1) return x def main(): device = torch.device("cpu") model = PytorchMNISTClassifier().to(device) # optimizer=optim.Adam(model.parameters(),lr=1e-3) optimizer = optim.LBFGS(model.parameters(), lr=0.01, max_iter=20) transform = transforms.Compose( [transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))] ) mnist_train = MNIST(os.getcwd(), train=True, download=True, transform=transform) mnist_train, mnist_val = random_split( mnist_train, [55000, 5000], generator=torch.Generator().manual_seed(42) ) dl = DataLoader(mnist_train, batch_size=1024, num_workers=0) for epoch in range(30): for i, (x, y) in enumerate(dl): x = x.to(device) y = y.to(device) def closure(): logits = model(x) optimizer.zero_grad() loss = F.nll_loss(logits, y) loss.backward(retain_graph=True) print("closure_loss:", loss.item()) return loss loss_out = optimizer.step(closure=closure) if __name__ == "__main__": main() ```

my env

```console $ wget https://raw.githubusercontent.com/PyTorchLightning/pytorch-lightning/master/tests/collect_env_details.py $ python collect_env_details.py * CUDA: - GPU: - available: False - version: None * Packages: - numpy: 1.19.5 - pyTorch_debug: False - pyTorch_version: 1.7.1+cpu - pytorch-lightning: 1.1.4 - tqdm: 4.56.0 * System: - OS: Linux - architecture: - 64bit - ELF - processor: - python: 3.8.5 - version: #1 SMP Debian 4.19.160-2 (2020-11-28) ```

I have no idea how I could investigate this further. @carmocca @rohitgr7 Could you help here if you have time...?

---

EDIT (Jan 28, 2021): Not sure how this helps us debug, but I realised that if we change the value of torch.optim.LBFGS(..., max_iter=20) from 20 (by default) to 1 or 2, both PL and native PyTorch behave exactly the same which I confirmed with my example code above. (Both don't converge though.)

[akihironitta]

@peymanpoozesh @Bajo1994 Sorry for the delay. I haven't figured out why LBFGS behaves differently between Lightning and native PyTorch, but I found an easy workaround, so let me share it here.

The workaround is to use the manual optimization instead of the default automatic optimization (See my notebook linked below for the complete code using BoringModel):

class Model(pl.LightningModule):
    def __init__(self, ...):
        self.automatic_optimization = False
        ...

    def training_step(self, batch, batch_idx):
        optimizer = self.optimizers()
        def closure():
            output = self.layer(batch)
            loss = self.loss(batch, output)
            optimizer.zero_grad()
            self.manual_backward(loss)
            return loss
        optimizer.step(closure=closure)

See also: Full code of the workaround - Google Colab Manual optimization - PyTorch Lightning Docs

[akihironitta]

Here are the minimal code examples using BoringModel.

Lightning code

```python import pytorch_lightning as pl import torch from torch.utils.data import DataLoader, Dataset pl.seed_everything(42) class RandomDataset(Dataset): def __init__(self, size, num_samples): self.len = num_samples self.data = torch.randn(num_samples, size) def __getitem__(self, index): return self.data[index] def __len__(self): return self.len class BoringModel(pl.LightningModule): def __init__(self): super().__init__() self.layer = torch.nn.Linear(32, 2) def forward(self, x): return self.layer(x) def loss(self, batch, prediction): return torch.nn.functional.mse_loss(prediction, torch.ones_like(prediction)) def training_step(self, batch, batch_idx): output = self.layer(batch) loss = self.loss(batch, output) return {"loss": loss} def training_step_end(self, training_step_outputs): loss = training_step_outputs["loss"] print("loss:", loss.item()) return training_step_outputs def configure_optimizers(self): # optimizer = torch.optim.Adam(self.parameters(), lr=1e-3) optimizer = torch.optim.LBFGS(self.parameters(), lr=0.01, max_iter=20) return optimizer def main(): ds = RandomDataset(32, 100000) dl = DataLoader(ds, batch_size=1024) model = BoringModel() trainer = pl.Trainer( progress_bar_refresh_rate=0, fast_dev_run=1, ) trainer.fit(model, dl) if __name__ == "__main__": main() ```

Pure PyTorch code

```python import torch import torch.nn as nn from pytorch_lightning import seed_everything from torch.utils.data import DataLoader, Dataset seed_everything(42) class RandomDataset(Dataset): def __init__(self, size, num_samples): self.len = num_samples self.data = torch.randn(num_samples, size) def __getitem__(self, index): return self.data[index] def __len__(self): return self.len class Model(nn.Module): def __init__(self): super(Model, self).__init__() self.layer = torch.nn.Linear(32, 2) def forward(self, x): return self.layer(x) def main(): ds = RandomDataset(32, 100000) dl = DataLoader(ds, batch_size=1024) model = Model() # optimizer = torch.optim.Adam(model.parameters(), lr=1e-3) optimizer = torch.optim.LBFGS(model.parameters(), lr=0.01, max_iter=20) for epoch in range(3): for i, x in enumerate(dl): def closure(): prediction = model(x) loss = torch.nn.functional.mse_loss(prediction, torch.ones_like(prediction)) optimizer.zero_grad() # removing this line causes the same bug as in Lightning script loss.backward() print("loss:", loss.item()) return loss loss_out = optimizer.step(closure=closure) if __name__ == '__main__': main() ```

cc: @carmocca @tchaton

[justusschock]

@akihironitta Why doesn't optimizer.step(closure=closure) work? Why do you have to unwrap it? Because without unwrapping you also get all the precision support from lightning :)

[akihironitta]

@justusschock Fixed! (It was just for print debugging from another script because LightningOptimizer doesn't return the output of closure())

[Bajo1994]

@akihironitta @carmocca I am very thankful for your great effort on this bug. I am looking forward to resuming my project as soon as you update the pl package. In my code, I like to switch between LBFGS and Adam optimizers. I like to use the LBFGS when the loss is large and then switch to Adam. I hope switching between these two optimizers would be smooth in pl (I had difficulties in switching between these two optimizers in native PyTorch). I will keep you posted if there is any problem.