Effective Training Techniques

Lightning implements various techniques to help during training that can help make the training smoother.

---

Accumulate Gradients

Accumulated gradients run K small batches of size N before doing a backward pass. The effect is a large effective batch size of size KxN, where N is the batch size. Internally it doesn’t stack up the batches and do a forward pass rather it accumulates the gradients for K batches and then do an optimizer.step to make sure the effective batch size is increased but there is no memory overhead.

Warning

When using distributed training for eg. DDP, with let’s say with P devices, each device accumulates independently i.e. it stores the gradients after each loss.backward() and doesn’t sync the gradients across the devices until we call optimizer.step(). So for each accumulation step, the effective batch size on each device will remain N*K but right before the optimizer.step(), the gradient sync will make the effective batch size as P*N*K. For DP, since the batch is split across devices, the final effective batch size will be N*K.

See also

Trainer

# DEFAULT (ie: no accumulated grads)
trainer = Trainer(accumulate_grad_batches=1)

# Accumulate gradients for 7 batches
trainer = Trainer(accumulate_grad_batches=7)

You can set different values for it at different epochs by passing a dictionary, where the key represents the epoch at which the value for gradient accumulation should be updated.

# till 5th epoch, it will accumulate every 8 batches. From 5th epoch
# till 9th epoch it will accumulate every 4 batches and after that no accumulation
# will happen. Note that you need to use zero-indexed epoch keys here
trainer = Trainer(accumulate_grad_batches={0: 8, 4: 4, 8: 1})

Or, you can create custom GradientAccumulationScheduler

from pytorch_lightning.callbacks import GradientAccumulationScheduler


# till 5th epoch, it will accumulate every 8 batches. From 5th epoch
# till 9th epoch it will accumulate every 4 batches and after that no accumulation
# will happen. Note that you need to use zero-indexed epoch keys here
accumulator = GradientAccumulationScheduler(scheduling={0: 8, 4: 4, 8: 1})
trainer = Trainer(callbacks=accumulator)

---

Gradient Clipping

Gradient clipping can be enabled to avoid exploding gradients. By default, this will clip the gradient norm by calling torch.nn.utils.clip_grad_norm_() computed over all model parameters together. If the Trainer’s gradient_clip_algorithm is set to 'value' ('norm' by default), this will use instead torch.nn.utils.clip_grad_value_() for each parameter instead.

Note

If using mixed precision, the gradient_clip_val does not need to be changed as the gradients are unscaled before applying the clipping function.

See also

Trainer

# DEFAULT (ie: don't clip)
trainer = Trainer(gradient_clip_val=0)

# clip gradients' global norm to <=0.5 using gradient_clip_algorithm='norm' by default
trainer = Trainer(gradient_clip_val=0.5)

# clip gradients' maximum magnitude to <=0.5
trainer = Trainer(gradient_clip_val=0.5, gradient_clip_algorithm="value")

Read more about Configuring Gradient Clipping for advanced use-cases.

---

Stochastic Weight Averaging

Stochastic Weight Averaging (SWA) can make your models generalize better at virtually no additional cost. This can be used with both non-trained and trained models. The SWA procedure smooths the loss landscape thus making it harder to end up in a local minimum during optimization.

For a more detailed explanation of SWA and how it works, read this post by the PyTorch team.

See also

The StochasticWeightAveraging callback

# Enable Stochastic Weight Averaging using the callback
trainer = Trainer(callbacks=[StochasticWeightAveraging(swa_lrs=1e-2)])

---

Batch Size Finder

Auto-scaling of batch size can be enabled to find the largest batch size that fits into memory. Large batch size often yields a better estimation of the gradients, but may also result in longer training time. Inspired by https://github.com/BlackHC/toma.

See also

Trainer

# DEFAULT (ie: don't scale batch size automatically)
trainer = Trainer(auto_scale_batch_size=None)

# Autoscale batch size
trainer = Trainer(auto_scale_batch_size=None | "power" | "binsearch")

# Find the batch size
trainer.tune(model)

Currently, this feature supports two modes 'power' scaling and 'binsearch' scaling. In 'power' scaling, starting from a batch size of 1 keeps doubling the batch size until an out-of-memory (OOM) error is encountered. Setting the argument to 'binsearch' will initially also try doubling the batch size until it encounters an OOM, after which it will do a binary search that will finetune the batch size. Additionally, it should be noted that the batch size scaler cannot search for batch sizes larger than the size of the training dataset.

Note

This feature expects that a batch_size field is either located as a model attribute i.e. model.batch_size or as a field in your hparams i.e. model.hparams.batch_size. Similarly it can work with datamodules too. The field should exist and will be updated by the results of this algorithm. Additionally, your train_dataloader() method should depend on this field for this feature to work i.e.

# using LightningModule
class LitModel(LightningModule):
    def __init__(self, batch_size):
        super().__init__()
        self.save_hyperparameters()
        # or
        self.batch_size = batch_size

    def train_dataloader(self):
        return DataLoader(train_dataset, batch_size=self.batch_size | self.hparams.batch_size)


trainer = Trainer(...)
model = LitModel(batch_size=32)
trainer.tune(model)

# using LightningDataModule
class LitDataModule(LightningDataModule):
    def __init__(self, batch_size):
        super().__init__()
        self.save_hyperparameters()
        # or
        self.batch_size = batch_size

    def train_dataloader(self):
        return DataLoader(train_dataset, batch_size=self.batch_size | self.hparams.batch_size)


trainer = Trainer(...)
model = MyModel()
datamodule = LitDataModule(batch_size=32)
trainer.tune(model, datamodule=datamodule)

Note that the train_dataloader can be either part of the LightningModule or LightningDataModule as shown above. If both the LightningModule and the LightningDataModule contain a train_dataloader, the LightningDataModule takes precedence.

Warning

Due to the constraints listed above, this features does NOT work when passing dataloaders directly to .fit().

The scaling algorithm has a number of parameters that the user can control by invoking the scale_batch_size() method:

# Use default in trainer construction
trainer = Trainer()
tuner = Tuner(trainer)

# Invoke method
new_batch_size = tuner.scale_batch_size(model, *extra_parameters_here)

# Override old batch size (this is done automatically)
model.hparams.batch_size = new_batch_size

# Fit as normal
trainer.fit(model)

The algorithm in short works by:

1. Dumping the current state of the model and trainer

2. Iteratively until convergence or maximum number of tries max_trials (default 25) has been reached:

   • Call fit() method of trainer. This evaluates steps_per_trial (default 3) number of optimization steps. Each training step can trigger an OOM error if the tensors (training batch, weights, gradients, etc.) allocated during the steps have a too large memory footprint.

   • If an OOM error is encountered, decrease batch size else increase it. How much the batch size is increased/decreased is determined by the chosen strategy.

3. The found batch size is saved to either model.batch_size or model.hparams.batch_size

4. Restore the initial state of model and trainer

Warning

Batch size finder is not yet supported for DDP or any of its variations, it is coming soon.

---

Learning Rate Finder

For training deep neural networks, selecting a good learning rate is essential for both better performance and faster convergence. Even optimizers such as Adam that are self-adjusting the learning rate can benefit from more optimal choices.

To reduce the amount of guesswork concerning choosing a good initial learning rate, a learning rate finder can be used. As described in this paper a learning rate finder does a small run where the learning rate is increased after each processed batch and the corresponding loss is logged. The result of this is a lr vs. loss plot that can be used as guidance for choosing an optimal initial learning rate.

Warning

For the moment, this feature only works with models having a single optimizer. LR Finder support for DDP and any of its variations is not implemented yet. It is coming soon.

Using Lightning’s built-in LR finder

To enable the learning rate finder, your lightning module needs to have a learning_rate or lr attribute (or as a field in your hparams i.e. hparams.learning_rate or hparams.lr). Then, set Trainer(auto_lr_find=True) during trainer construction, and then call trainer.tune(model) to run the LR finder. The suggested learning_rate will be written to the console and will be automatically set to your lightning module, which can be accessed via self.learning_rate or self.lr.

See also

trainer.tune.

class LitModel(LightningModule):
    def __init__(self, learning_rate):
        super().__init__()
        self.learning_rate = learning_rate
        self.model = Model(...)

    def configure_optimizers(self):
        return Adam(self.parameters(), lr=(self.lr or self.learning_rate))


model = LitModel()

# finds learning rate automatically
# sets hparams.lr or hparams.learning_rate to that learning rate
trainer = Trainer(auto_lr_find=True)

trainer.tune(model)

If your model is using an arbitrary value instead of self.lr or self.learning_rate, set that value as auto_lr_find:

model = LitModel()

# to set to your own hparams.my_value
trainer = Trainer(auto_lr_find="my_value")

trainer.tune(model)

You can also inspect the results of the learning rate finder or just play around with the parameters of the algorithm. This can be done by invoking the lr_find() method. A typical example of this would look like:

model = MyModelClass(hparams)
trainer = Trainer()

# Run learning rate finder
lr_finder = trainer.tuner.lr_find(model)

# Results can be found in
print(lr_finder.results)

# Plot with
fig = lr_finder.plot(suggest=True)
fig.show()

# Pick point based on plot, or get suggestion
new_lr = lr_finder.suggestion()

# update hparams of the model
model.hparams.lr = new_lr

# Fit model
trainer.fit(model)

The figure produced by lr_finder.plot() should look something like the figure below. It is recommended to not pick the learning rate that achieves the lowest loss, but instead something in the middle of the sharpest downward slope (red point). This is the point returned py lr_finder.suggestion().

---

Advanced GPU Optimizations

When training on single or multiple GPU machines, Lightning offers a host of advanced optimizations to improve throughput, memory efficiency, and model scaling. Refer to Advanced GPU Optimized Training for more details.

---

Sharing Datasets Across Process Boundaries

The LightningDataModule class provides an organized way to decouple data loading from training logic, with prepare_data() being used for downloading and pre-processing the dataset on a single process, and setup() loading the pre-processed data for each process individually:

class MNISTDataModule(pl.LightningDataModule):
    def prepare_data(self):
        MNIST(self.data_dir, download=True)

    def setup(self, stage: Optional[str] = None):
        self.mnist = MNIST(self.data_dir)

    def train_loader(self):
        return DataLoader(self.mnist, batch_size=128)

However, for in-memory datasets, that means that each process will hold a (redundant) replica of the dataset in memory, which may be impractical when using many processes while utilizing datasets that nearly fit into CPU memory, as the memory consumption will scale up linearly with the number of processes. For example, when training Graph Neural Networks, a common strategy is to load the entire graph into CPU memory for fast access to the entire graph structure and its features, and to then perform neighbor sampling to obtain mini-batches that fit onto the GPU.

A simple way to prevent redundant dataset replicas is to rely on torch.multiprocessing to share the data automatically between spawned processes via shared memory. For this, all data pre-loading should be done on the main process inside DataModule.__init__(). As a result, all tensor-data will get automatically shared when using the DDPSpawnStrategy strategy.

Warning

torch.multiprocessing will send a handle of each individual tensor to other processes. In order to prevent any errors due to too many open file handles, try to reduce the number of tensors to share, e.g., by stacking your data into a single tensor.

class MNISTDataModule(pl.LightningDataModule):
    def __init__(self, data_dir: str):
        self.mnist = MNIST(data_dir, download=True, transform=T.ToTensor())

    def train_loader(self):
        return DataLoader(self.mnist, batch_size=128)


model = Model(...)
datamodule = MNISTDataModule("data/MNIST")

trainer = Trainer(accelerator="gpu", devices=2, strategy="ddp_spawn")
trainer.fit(model, datamodule)

See the graph-level and node-level prediction examples in PyTorch Geometric for practical use-cases.