Trainer

The Trainer and TFTrainer classes provide an API for feature-complete training in most standard use cases. It’s used in most of the example scripts.

Before instantiating your Trainer/TFTrainer, create a TrainingArguments/TFTrainingArguments to access all the points of customization during training.

The API supports distributed training on multiple GPUs/TPUs, mixed precision through NVIDIA Apex and Native AMP for PyTorch and tf.keras.mixed_precision for TensorFlow.

Both Trainer and TFTrainer contain the basic training loop which supports the above features. To inject custom behavior you can subclass them and override the following methods:

  • get_train_dataloader/get_train_tfdataset – Creates the training DataLoader (PyTorch) or TF Dataset.

  • get_eval_dataloader/get_eval_tfdataset – Creates the evaluation DataLoader (PyTorch) or TF Dataset.

  • get_test_dataloader/get_test_tfdataset – Creates the test DataLoader (PyTorch) or TF Dataset.

  • log – Logs information on the various objects watching training.

  • create_optimizer_and_scheduler – Sets up the optimizer and learning rate scheduler if they were not passed at init. Note, that you can also subclass or override the create_optimizer and create_scheduler methods separately.

  • create_optimizer – Sets up the optimizer if it wasn’t passed at init.

  • create_scheduler – Sets up the learning rate scheduler if it wasn’t passed at init.

  • compute_loss - Computes the loss on a batch of training inputs.

  • training_step – Performs a training step.

  • prediction_step – Performs an evaluation/test step.

  • run_model (TensorFlow only) – Basic pass through the model.

  • evaluate – Runs an evaluation loop and returns metrics.

  • predict – Returns predictions (with metrics if labels are available) on a test set.

Warning

The Trainer class is optimized for 🤗 Transformers models and can have surprising behaviors when you use it on other models. When using it on your own model, make sure:

  • your model always return tuples or subclasses of ModelOutput.

  • your model can compute the loss if a labels argument is provided and that loss is returned as the first element of the tuple (if your model returns tuples)

  • your model can accept multiple label arguments (use the label_names in your TrainingArguments to indicate their name to the Trainer) but none of them should be named "label".

Here is an example of how to customize Trainer using a custom loss function for multi-label classification:

import torch
from transformers import Trainer

class MultilabelTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False):
        labels = inputs.pop("labels")
        outputs = model(**inputs)
        logits = outputs.logits
        loss_fct = torch.nn.BCEWithLogitsLoss()
        loss = loss_fct(logits.view(-1, self.model.config.num_labels),
                        labels.float().view(-1, self.model.config.num_labels))
        return (loss, outputs) if return_outputs else loss

Another way to customize the training loop behavior for the PyTorch Trainer is to use callbacks that can inspect the training loop state (for progress reporting, logging on TensorBoard or other ML platforms…) and take decisions (like early stopping).

Trainer

class transformers.Trainer(model: torch.nn.modules.module.Module = None, args: transformers.training_args.TrainingArguments = None, data_collator: Optional[NewType.<locals>.new_type] = None, train_dataset: Optional[torch.utils.data.dataset.Dataset] = None, eval_dataset: Optional[torch.utils.data.dataset.Dataset] = None, tokenizer: Optional[transformers.tokenization_utils_base.PreTrainedTokenizerBase] = None, model_init: Callable[transformers.modeling_utils.PreTrainedModel] = None, compute_metrics: Optional[Callable[transformers.trainer_utils.EvalPrediction, Dict]] = None, callbacks: Optional[List[transformers.trainer_callback.TrainerCallback]] = None, optimizers: Tuple[torch.optim.optimizer.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None))[source]

Trainer is a simple but feature-complete training and eval loop for PyTorch, optimized for 🤗 Transformers.

Parameters
  • model (PreTrainedModel or torch.nn.Module, optional) –

    The model to train, evaluate or use for predictions. If not provided, a model_init must be passed.

    Note

    Trainer is optimized to work with the PreTrainedModel provided by the library. You can still use your own models defined as torch.nn.Module as long as they work the same way as the 🤗 Transformers models.

  • args (TrainingArguments, optional) – The arguments to tweak for training. Will default to a basic instance of TrainingArguments with the output_dir set to a directory named tmp_trainer in the current directory if not provided.

  • data_collator (DataCollator, optional) – The function to use to form a batch from a list of elements of train_dataset or eval_dataset. Will default to default_data_collator() if no tokenizer is provided, an instance of DataCollatorWithPadding() otherwise.

  • train_dataset (torch.utils.data.dataset.Dataset, optional) – The dataset to use for training. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed.

  • eval_dataset (torch.utils.data.dataset.Dataset, optional) – The dataset to use for evaluation. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed.

  • tokenizer (PreTrainedTokenizerBase, optional) – The tokenizer used to preprocess the data. If provided, will be used to automatically pad the inputs the maximum length when batching inputs, and it will be saved along the model to make it easier to rerun an interrupted training or reuse the fine-tuned model.

  • model_init (Callable[[], PreTrainedModel], optional) –

    A function that instantiates the model to be used. If provided, each call to train() will start from a new instance of the model as given by this function.

    The function may have zero argument, or a single one containing the optuna/Ray Tune trial object, to be able to choose different architectures according to hyper parameters (such as layer count, sizes of inner layers, dropout probabilities etc).

  • compute_metrics (Callable[[EvalPrediction], Dict], optional) – The function that will be used to compute metrics at evaluation. Must take a EvalPrediction and return a dictionary string to metric values.

  • callbacks (List of TrainerCallback, optional) –

    A list of callbacks to customize the training loop. Will add those to the list of default callbacks detailed in here.

    If you want to remove one of the default callbacks used, use the Trainer.remove_callback() method.

  • optimizers (Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR, optional) – A tuple containing the optimizer and the scheduler to use. Will default to an instance of AdamW on your model and a scheduler given by get_linear_schedule_with_warmup() controlled by args.

Important attributes:

  • model – Always points to the core model. If using a transformers model, it will be a PreTrainedModel subclass.

  • model_wrapped – Always points to the most external model in case one or more other modules wrap the original model. This is the model that should be used for the forward pass. For example, under DeepSpeed, the inner model is wrapped in DeepSpeed and then again in torch.nn.DistributedDataParallel. If the inner model hasn’t been wrapped, then self.model_wrapped is the same as self.model.

  • is_model_parallel – Whether or not a model has been switched to a model parallel mode (different from data parallelism, this means some of the model layers are split on different GPUs).

  • place_model_on_device – Whether or not to automatically place the model on the device - it will be set to False if model parallel or deepspeed is used, or if the default TrainingArguments.place_model_on_device is overridden to return False .

  • is_in_train – Whether or not a model is currently running train (e.g. when evaluate is called while in train)

add_callback(callback)[source]

Add a callback to the current list of TrainerCallback.

Parameters

callback (type or TrainerCallback) – A TrainerCallback class or an instance of a TrainerCallback. In the first case, will instantiate a member of that class.

compute_loss(model, inputs, return_outputs=False)[source]

How the loss is computed by Trainer. By default, all models return the loss in the first element.

Subclass and override for custom behavior.

create_optimizer()[source]

Setup the optimizer.

We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the Trainer’s init through optimizers, or subclass and override this method in a subclass.

create_optimizer_and_scheduler(num_training_steps: int)[source]

Setup the optimizer and the learning rate scheduler.

We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the Trainer’s init through optimizers, or subclass and override this method (or create_optimizer and/or create_scheduler) in a subclass.

create_scheduler(num_training_steps: int)[source]

Setup the scheduler. The optimizer of the trainer must have been set up before this method is called.

Parameters

num_training_steps (int) – The number of training steps to do.

evaluate(eval_dataset: Optional[torch.utils.data.dataset.Dataset] = None, ignore_keys: Optional[List[str]] = None, metric_key_prefix: str = 'eval') → Dict[str, float][source]

Run evaluation and returns metrics.

The calling script will be responsible for providing a method to compute metrics, as they are task-dependent (pass it to the init compute_metrics argument).

You can also subclass and override this method to inject custom behavior.

Parameters
  • eval_dataset (Dataset, optional) – Pass a dataset if you wish to override self.eval_dataset. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement the __len__ method.

  • ignore_keys (Lst[str], optional) – A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.

  • metric_key_prefix (str, optional, defaults to "eval") – An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “eval_bleu” if the prefix is “eval” (default)

Returns

A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The dictionary also contains the epoch number which comes from the training state.

floating_point_ops(inputs: Dict[str, Union[torch.Tensor, Any]])[source]

For models that inherit from PreTrainedModel, uses that method to compute the number of floating point operations for every backward + forward pass. If using another model, either implement such a method in the model or subclass and override this method.

Parameters

inputs (Dict[str, Union[torch.Tensor, Any]]) – The inputs and targets of the model.

Returns

The number of floating-point operations.

Return type

int

get_eval_dataloader(eval_dataset: Optional[torch.utils.data.dataset.Dataset] = None) → torch.utils.data.dataloader.DataLoader[source]

Returns the evaluation DataLoader.

Subclass and override this method if you want to inject some custom behavior.

Parameters

eval_dataset (torch.utils.data.dataset.Dataset, optional) – If provided, will override self.eval_dataset. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement __len__.

get_test_dataloader(test_dataset: torch.utils.data.dataset.Dataset) → torch.utils.data.dataloader.DataLoader[source]

Returns the test DataLoader.

Subclass and override this method if you want to inject some custom behavior.

Parameters

test_dataset (torch.utils.data.dataset.Dataset, optional) – The test dataset to use. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement __len__.

get_train_dataloader() → torch.utils.data.dataloader.DataLoader[source]

Returns the training DataLoader.

Will use no sampler if self.train_dataset does not implement __len__, a random sampler (adapted to distributed training if necessary) otherwise.

Subclass and override this method if you want to inject some custom behavior.

Launch an hyperparameter search using optuna or Ray Tune. The optimized quantity is determined by compute_objective, which defaults to a function returning the evaluation loss when no metric is provided, the sum of all metrics otherwise.

Warning

To use this method, you need to have provided a model_init when initializing your Trainer: we need to reinitialize the model at each new run. This is incompatible with the optimizers argument, so you need to subclass Trainer and override the method create_optimizer_and_scheduler() for custom optimizer/scheduler.

Parameters
  • hp_space (Callable[["optuna.Trial"], Dict[str, float]], optional) – A function that defines the hyperparameter search space. Will default to default_hp_space_optuna() or default_hp_space_ray() depending on your backend.

  • compute_objective (Callable[[Dict[str, float]], float], optional) – A function computing the objective to minimize or maximize from the metrics returned by the evaluate method. Will default to default_compute_objective().

  • n_trials (int, optional, defaults to 100) – The number of trial runs to test.

  • direction (str, optional, defaults to "minimize") – Whether to optimize greater or lower objects. Can be "minimize" or "maximize", you should pick "minimize" when optimizing the validation loss, "maximize" when optimizing one or several metrics.

  • backend (str or HPSearchBackend, optional) – The backend to use for hyperparameter search. Will default to optuna or Ray Tune, depending on which one is installed. If both are installed, will default to optuna.

  • kwargs

    Additional keyword arguments passed along to optuna.create_study or ray.tune.run. For more information see:

Returns

All the information about the best run.

Return type

transformers.trainer_utils.BestRun

is_local_process_zero() → bool[source]

Whether or not this process is the local (e.g., on one machine if training in a distributed fashion on several machines) main process.

is_world_process_zero() → bool[source]

Whether or not this process is the global main process (when training in a distributed fashion on several machines, this is only going to be True for one process).

log(logs: Dict[str, float]) → None[source]

Log logs on the various objects watching training.

Subclass and override this method to inject custom behavior.

Parameters

logs (Dict[str, float]) – The values to log.

log_metrics(split, metrics)

Log metrics in a specially formatted way

Under distributed environment this is done only for a process with rank 0.

Parameters
  • split (str) – Mode/split name: one of train, eval, test

  • metrics (Dict[str, float]) – The metrics returned from train/evaluate/predictmetrics: metrics dict

Notes on memory reports:

In order to get memory usage report you need to install psutil. You can do that with pip install psutil.

Now when this method is run, you will see a report that will include:

init_mem_cpu_alloc_delta   =     1301MB
init_mem_cpu_peaked_delta  =      154MB
init_mem_gpu_alloc_delta   =      230MB
init_mem_gpu_peaked_delta  =        0MB
train_mem_cpu_alloc_delta  =     1345MB
train_mem_cpu_peaked_delta =        0MB
train_mem_gpu_alloc_delta  =      693MB
train_mem_gpu_peaked_delta =        7MB

Understanding the reports:

  • the first segment, e.g., train__, tells you which stage the metrics are for. Reports starting with init_ will be added to the first stage that gets run. So that if only evaluation is run, the memory usage for the __init__ will be reported along with the eval_ metrics.

  • the third segment, is either cpu or gpu, tells you whether it’s the general RAM or the gpu0 memory metric.

  • *_alloc_delta - is the difference in the used/allocated memory counter between the end and the start of the stage - it can be negative if a function released more memory than it allocated.

  • *_peaked_delta - is any extra memory that was consumed and then freed - relative to the current allocated memory counter - it is never negative. When you look at the metrics of any stage you add up alloc_delta + peaked_delta and you know how much memory was needed to complete that stage.

The reporting happens only for process of rank 0 and gpu 0 (if there is a gpu). Typically this is enough since the main process does the bulk of work, but it could be not quite so if model parallel is used and then other GPUs may use a different amount of gpu memory. This is also not the same under DataParallel where gpu0 may require much more memory than the rest since it stores the gradient and optimizer states for all participating GPUS. Perhaps in the future these reports will evolve to measure those too.

The CPU RAM metric measures RSS (Resident Set Size) includes both the memory which is unique to the process and the memory shared with other processes. It is important to note that it does not include swapped out memory, so the reports could be imprecise.

The CPU peak memory is measured using a sampling thread. Due to python’s GIL it may miss some of the peak memory if that thread didn’t get a chance to run when the highest memory was used. Therefore this report can be less than reality. Using tracemalloc would have reported the exact peak memory, but it doesn’t report memory allocations outside of python. So if some C++ CUDA extension allocated its own memory it won’t be reported. And therefore it was dropped in favor of the memory sampling approach, which reads the current process memory usage.

The GPU allocated and peak memory reporting is done with torch.cuda.memory_allocated() and torch.cuda.max_memory_allocated(). This metric reports only “deltas” for pytorch-specific allocations, as torch.cuda memory management system doesn’t track any memory allocated outside of pytorch. For example, the very first cuda call typically loads CUDA kernels, which may take from 0.5 to 2GB of GPU memory.

Note that this tracker doesn’t account for memory allocations outside of Trainer’s __init__, train, evaluate and predict calls.

Because evaluation calls may happen during train, we can’t handle nested invocations because torch.cuda.max_memory_allocated is a single counter, so if it gets reset by a nested eval call, train’s tracker will report incorrect info. If this pytorch issue gets resolved it will be possible to change this class to be re-entrant. Until then we will only track the outer level of train, evaluate and predict methods. Which means that if eval is called during train, it’s the latter that will account for its memory usage and that of the former.

This also means that if any other tool that is used along the Trainer calls torch.cuda.reset_peak_memory_stats, the gpu peak memory stats could be invalid. And the Trainer will disrupt the normal behavior of any such tools that rely on calling torch.cuda.reset_peak_memory_stats themselves.

For best performance you may want to consider turning the memory profiling off for production runs.

metrics_format(metrics: Dict[str, float]) → Dict[str, float]

Reformat Trainer metrics values to a human-readable format

Parameters

metrics (Dict[str, float]) – The metrics returned from train/evaluate/predict

Returns

The reformatted metrics

Return type

metrics (Dict[str, float])

num_examples(dataloader: torch.utils.data.dataloader.DataLoader) → int[source]

Helper to get number of samples in a DataLoader by accessing its dataset.

Will raise an exception if the underlying dataset does not implement method __len__

pop_callback(callback)[source]

Remove a callback from the current list of TrainerCallback and returns it.

If the callback is not found, returns None (and no error is raised).

Parameters

callback (type or TrainerCallback) – A TrainerCallback class or an instance of a TrainerCallback. In the first case, will pop the first member of that class found in the list of callbacks.

Returns

The callback removed, if found.

Return type

TrainerCallback

predict(test_dataset: torch.utils.data.dataset.Dataset, ignore_keys: Optional[List[str]] = None, metric_key_prefix: str = 'test') → transformers.trainer_utils.PredictionOutput[source]

Run prediction and returns predictions and potential metrics.

Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method will also return metrics, like in evaluate().

Parameters
  • test_dataset (Dataset) – Dataset to run the predictions on. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed. Has to implement the method __len__

  • ignore_keys (Lst[str], optional) – A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.

  • metric_key_prefix (str, optional, defaults to "test") – An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “test_bleu” if the prefix is “test” (default)

Note

If your predictions or labels have different sequence length (for instance because you’re doing dynamic padding in a token classification task) the predictions will be padded (on the right) to allow for concatenation into one array. The padding index is -100.

Returns: NamedTuple A namedtuple with the following keys:

  • predictions (np.ndarray): The predictions on test_dataset.

  • label_ids (np.ndarray, optional): The labels (if the dataset contained some).

  • metrics (Dict[str, float], optional): The potential dictionary of metrics (if the dataset contained labels).

prediction_loop(dataloader: torch.utils.data.dataloader.DataLoader, description: str, prediction_loss_only: Optional[bool] = None, ignore_keys: Optional[List[str]] = None, metric_key_prefix: str = 'eval') → transformers.trainer_utils.PredictionOutput[source]

Prediction/evaluation loop, shared by Trainer.evaluate() and Trainer.predict().

Works both with or without labels.

prediction_step(model: torch.nn.modules.module.Module, inputs: Dict[str, Union[torch.Tensor, Any]], prediction_loss_only: bool, ignore_keys: Optional[List[str]] = None) → Tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]][source]

Perform an evaluation step on model using obj:inputs.

Subclass and override to inject custom behavior.

Parameters
  • model (nn.Module) – The model to evaluate.

  • inputs (Dict[str, Union[torch.Tensor, Any]]) –

    The inputs and targets of the model.

    The dictionary will be unpacked before being fed to the model. Most models expect the targets under the argument labels. Check your model’s documentation for all accepted arguments.

  • prediction_loss_only (bool) – Whether or not to return the loss only.

  • ignore_keys (Lst[str], optional) – A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.

Returns

A tuple with the loss, logits and labels (each being optional).

Return type

Tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]

remove_callback(callback)[source]

Remove a callback from the current list of TrainerCallback.

Parameters

callback (type or TrainerCallback) – A TrainerCallback class or an instance of a TrainerCallback. In the first case, will remove the first member of that class found in the list of callbacks.

save_metrics(split, metrics, combined=True)

Save metrics into a json file for that split, e.g. train_results.json.

Under distributed environment this is done only for a process with rank 0.

Parameters
  • split (str) – Mode/split name: one of train, eval, test, all

  • metrics (Dict[str, float]) – The metrics returned from train/evaluate/predict

  • combined (bool, optional, defaults to True) – Creates combined metrics by updating all_results.json with metrics of this call

To understand the metrics please read the docstring of log_metrics(). The only difference is that raw unformatted numbers are saved in the current method.

save_model(output_dir: Optional[str] = None)[source]

Will save the model, so you can reload it using from_pretrained().

Will only save from the main process.

save_state()

Saves the Trainer state, since Trainer.save_model saves only the tokenizer with the model

Under distributed environment this is done only for a process with rank 0.

train(resume_from_checkpoint: Optional[Union[str, bool]] = None, trial: Union[optuna.Trial, Dict[str, Any]] = None, **kwargs)[source]

Main training entry point.

Parameters
  • resume_from_checkpoint (str or bool, optional) – If a str, local path to a saved checkpoint as saved by a previous instance of Trainer. If a bool and equals True, load the last checkpoint in args.output_dir as saved by a previous instance of Trainer. If present, training will resume from the model/optimizer/scheduler states loaded here.

  • trial (optuna.Trial or Dict[str, Any], optional) – The trial run or the hyperparameter dictionary for hyperparameter search.

  • kwargs – Additional keyword arguments used to hide deprecated arguments

training_step(model: torch.nn.modules.module.Module, inputs: Dict[str, Union[torch.Tensor, Any]]) → torch.Tensor[source]

Perform a training step on a batch of inputs.

Subclass and override to inject custom behavior.

Parameters
  • model (nn.Module) – The model to train.

  • inputs (Dict[str, Union[torch.Tensor, Any]]) –

    The inputs and targets of the model.

    The dictionary will be unpacked before being fed to the model. Most models expect the targets under the argument labels. Check your model’s documentation for all accepted arguments.

Returns

The tensor with training loss on this batch.

Return type

torch.Tensor

Seq2SeqTrainer

class transformers.Seq2SeqTrainer(model: torch.nn.modules.module.Module = None, args: transformers.training_args.TrainingArguments = None, data_collator: Optional[NewType.<locals>.new_type] = None, train_dataset: Optional[torch.utils.data.dataset.Dataset] = None, eval_dataset: Optional[torch.utils.data.dataset.Dataset] = None, tokenizer: Optional[transformers.tokenization_utils_base.PreTrainedTokenizerBase] = None, model_init: Callable[transformers.modeling_utils.PreTrainedModel] = None, compute_metrics: Optional[Callable[transformers.trainer_utils.EvalPrediction, Dict]] = None, callbacks: Optional[List[transformers.trainer_callback.TrainerCallback]] = None, optimizers: Tuple[torch.optim.optimizer.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None))[source]
evaluate(eval_dataset: Optional[torch.utils.data.dataset.Dataset] = None, ignore_keys: Optional[List[str]] = None, metric_key_prefix: str = 'eval', max_length: Optional[int] = None, num_beams: Optional[int] = None) → Dict[str, float][source]

Run evaluation and returns metrics.

The calling script will be responsible for providing a method to compute metrics, as they are task-dependent (pass it to the init compute_metrics argument).

You can also subclass and override this method to inject custom behavior.

Parameters
  • eval_dataset (Dataset, optional) – Pass a dataset if you wish to override self.eval_dataset. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed. It must implement the __len__ method.

  • ignore_keys (List[str], optional) – A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.

  • metric_key_prefix (str, optional, defaults to "eval") – An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “eval_bleu” if the prefix is "eval" (default)

  • max_length (int, optional) – The maximum target length to use when predicting with the generate method.

  • num_beams (int, optional) – Number of beams for beam search that will be used when predicting with the generate method. 1 means no beam search.

Returns

A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The dictionary also contains the epoch number which comes from the training state.

predict(test_dataset: torch.utils.data.dataset.Dataset, ignore_keys: Optional[List[str]] = None, metric_key_prefix: str = 'eval', max_length: Optional[int] = None, num_beams: Optional[int] = None) → transformers.trainer_utils.PredictionOutput[source]

Run prediction and returns predictions and potential metrics.

Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method will also return metrics, like in evaluate().

Parameters
  • test_dataset (Dataset) – Dataset to run the predictions on. If it is an datasets.Dataset, columns not accepted by the model.forward() method are automatically removed. Has to implement the method __len__

  • ignore_keys (List[str], optional) – A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions.

  • metric_key_prefix (str, optional, defaults to "eval") – An optional prefix to be used as the metrics key prefix. For example the metrics “bleu” will be named “eval_bleu” if the prefix is "eval" (default)

  • max_length (int, optional) – The maximum target length to use when predicting with the generate method.

  • num_beams (int, optional) – Number of beams for beam search that will be used when predicting with the generate method. 1 means no beam search.

Note

If your predictions or labels have different sequence lengths (for instance because you’re doing dynamic padding in a token classification task) the predictions will be padded (on the right) to allow for concatenation into one array. The padding index is -100.

Returns: NamedTuple A namedtuple with the following keys:

  • predictions (np.ndarray): The predictions on test_dataset.

  • label_ids (np.ndarray, optional): The labels (if the dataset contained some).

  • metrics (Dict[str, float], optional): The potential dictionary of metrics (if the dataset contained labels).

TFTrainer

class transformers.TFTrainer(model: transformers.modeling_tf_utils.TFPreTrainedModel, args: transformers.training_args_tf.TFTrainingArguments, train_dataset: Optional[tensorflow.python.data.ops.dataset_ops.DatasetV2] = None, eval_dataset: Optional[tensorflow.python.data.ops.dataset_ops.DatasetV2] = None, compute_metrics: Optional[Callable[transformers.trainer_utils.EvalPrediction, Dict]] = None, tb_writer: Optional[tensorflow.python.ops.summary_ops_v2.SummaryWriter] = None, optimizers: Tuple[tensorflow.python.keras.optimizer_v2.optimizer_v2.OptimizerV2, tensorflow.python.keras.optimizer_v2.learning_rate_schedule.LearningRateSchedule] = None, None)[source]

TFTrainer is a simple but feature-complete training and eval loop for TensorFlow, optimized for 🤗 Transformers.

Parameters
  • model (TFPreTrainedModel) – The model to train, evaluate or use for predictions.

  • args (TFTrainingArguments) – The arguments to tweak training.

  • train_dataset (Dataset, optional) – The dataset to use for training. The dataset should yield tuples of (features, labels) where features is a dict of input features and labels is the labels. If labels is a tensor, the loss is calculated by the model by calling model(features, labels=labels). If labels is a dict, such as when using a QuestionAnswering head model with multiple targets, the loss is instead calculated by calling model(features, **labels).

  • eval_dataset (Dataset, optional) – The dataset to use for evaluation. The dataset should yield tuples of (features, labels) where features is a dict of input features and labels is the labels. If labels is a tensor, the loss is calculated by the model by calling model(features, labels=labels). If labels is a dict, such as when using a QuestionAnswering head model with multiple targets, the loss is instead calculated by calling model(features, **labels).

  • compute_metrics (Callable[[EvalPrediction], Dict], optional) – The function that will be used to compute metrics at evaluation. Must take a EvalPrediction and return a dictionary string to metric values.

  • tb_writer (tf.summary.SummaryWriter, optional) – Object to write to TensorBoard.

  • optimizers (Tuple[tf.keras.optimizers.Optimizer, tf.keras.optimizers.schedules.LearningRateSchedule], optional) – A tuple containing the optimizer and the scheduler to use. The optimizer default to an instance of tf.keras.optimizers.Adam if args.weight_decay_rate is 0 else an instance of AdamWeightDecay. The scheduler will default to an instance of tf.keras.optimizers.schedules.PolynomialDecay if args.num_warmup_steps is 0 else an instance of WarmUp.

create_optimizer_and_scheduler(num_training_steps: int)[source]

Setup the optimizer and the learning rate scheduler.

We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the TFTrainer’s init through optimizers, or subclass and override this method.

evaluate(eval_dataset: Optional[tensorflow.python.data.ops.dataset_ops.DatasetV2] = None) → Dict[str, float][source]

Run evaluation and returns metrics.

The calling script will be responsible for providing a method to compute metrics, as they are task-dependent (pass it to the init compute_metrics argument).

Parameters

eval_dataset (Dataset, optional) – Pass a dataset if you wish to override self.eval_dataset. The dataset should yield tuples of (features, labels) where features is a dict of input features and labels is the labels. If labels is a tensor, the loss is calculated by the model by calling model(features, labels=labels). If labels is a dict, such as when using a QuestionAnswering head model with multiple targets, the loss is instead calculated by calling model(features, **labels).

Returns

A dictionary containing the evaluation loss and the potential metrics computed from the predictions.

get_eval_tfdataset(eval_dataset: Optional[tensorflow.python.data.ops.dataset_ops.DatasetV2] = None) → tensorflow.python.data.ops.dataset_ops.DatasetV2[source]

Returns the evaluation Dataset.

Parameters

eval_dataset (Dataset, optional) – If provided, will override self.eval_dataset. The dataset should yield tuples of (features, labels) where features is a dict of input features and labels is the labels. If labels is a tensor, the loss is calculated by the model by calling model(features, labels=labels). If labels is a dict, such as when using a QuestionAnswering head model with multiple targets, the loss is instead calculated by calling model(features, **labels).

Subclass and override this method if you want to inject some custom behavior.

get_test_tfdataset(test_dataset: tensorflow.python.data.ops.dataset_ops.DatasetV2) → tensorflow.python.data.ops.dataset_ops.DatasetV2[source]

Returns a test Dataset.

Parameters

test_dataset (Dataset) – The dataset to use. The dataset should yield tuples of (features, labels) where features is a dict of input features and labels is the labels. If labels is a tensor, the loss is calculated by the model by calling model(features, labels=labels). If labels is a dict, such as when using a QuestionAnswering head model with multiple targets, the loss is instead calculated by calling model(features, **labels).

Subclass and override this method if you want to inject some custom behavior.

get_train_tfdataset() → tensorflow.python.data.ops.dataset_ops.DatasetV2[source]

Returns the training Dataset.

Subclass and override this method if you want to inject some custom behavior.

log(logs: Dict[str, float]) → None[source]

Log logs on the various objects watching training.

Subclass and override this method to inject custom behavior.

Parameters

logs (Dict[str, float]) – The values to log.

predict(test_dataset: tensorflow.python.data.ops.dataset_ops.DatasetV2) → transformers.trainer_utils.PredictionOutput[source]

Run prediction and returns predictions and potential metrics.

Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method will also return metrics, like in evaluate().

Parameters

test_dataset (Dataset) – Dataset to run the predictions on. The dataset should yield tuples of (features, labels) where features is a dict of input features and labels is the labels. If labels is a tensor, the loss is calculated by the model by calling model(features, labels=labels). If labels is a dict, such as when using a QuestionAnswering head model with multiple targets, the loss is instead calculated by calling model(features, **labels)

Returns: NamedTuple A namedtuple with the following keys:

  • predictions (np.ndarray): The predictions on test_dataset.

  • label_ids (np.ndarray, optional): The labels (if the dataset contained some).

  • metrics (Dict[str, float], optional): The potential dictionary of metrics (if the dataset contained labels).

prediction_loop(dataset: tensorflow.python.data.ops.dataset_ops.DatasetV2, steps: int, num_examples: int, description: str, prediction_loss_only: Optional[bool] = None) → transformers.trainer_utils.PredictionOutput[source]

Prediction/evaluation loop, shared by evaluate() and predict().

Works both with or without labels.

prediction_step(features: tensorflow.python.framework.ops.Tensor, labels: tensorflow.python.framework.ops.Tensor, nb_instances_in_global_batch: tensorflow.python.framework.ops.Tensor) → tensorflow.python.framework.ops.Tensor[source]

Compute the prediction on features and update the loss with labels.

Subclass and override to inject some custom behavior.

run_model(features, labels, training)[source]

Computes the loss of the given features and labels pair.

Subclass and override this method if you want to inject some custom behavior.

Parameters
  • features (tf.Tensor) – A batch of input features.

  • labels (tf.Tensor) – A batch of labels.

  • training (bool) – Whether or not to run the model in training mode.

Returns

The loss and logits.

Return type

A tuple of two tf.Tensor

save_model(output_dir: Optional[str] = None)[source]

Will save the model, so you can reload it using from_pretrained().

setup_comet()[source]

Setup the optional Comet.ml integration.

Environment:
COMET_MODE:

(Optional): str - “OFFLINE”, “ONLINE”, or “DISABLED”

COMET_PROJECT_NAME:

(Optional): str - Comet.ml project name for experiments

COMET_OFFLINE_DIRECTORY:

(Optional): str - folder to use for saving offline experiments when COMET_MODE is “OFFLINE”

For a number of configurable items in the environment, see here

setup_wandb()[source]

Setup the optional Weights & Biases (wandb) integration.

One can subclass and override this method to customize the setup if needed. Find more information here. You can also override the following environment variables:

Environment:
WANDB_PROJECT:

(Optional): str - “huggingface” by default, set this to a custom string to store results in a different project.

WANDB_DISABLED:

(Optional): boolean - defaults to false, set to “true” to disable wandb entirely.

train() → None[source]

Train method to train the model.

training_step(features, labels, nb_instances_in_global_batch)[source]

Perform a training step on features and labels.

Subclass and override to inject some custom behavior.

TrainingArguments

class transformers.TrainingArguments(output_dir: str, overwrite_output_dir: bool = False, do_train: bool = False, do_eval: bool = None, do_predict: bool = False, evaluation_strategy: transformers.trainer_utils.IntervalStrategy = 'no', prediction_loss_only: bool = False, per_device_train_batch_size: int = 8, per_device_eval_batch_size: int = 8, per_gpu_train_batch_size: Optional[int] = None, per_gpu_eval_batch_size: Optional[int] = None, gradient_accumulation_steps: int = 1, eval_accumulation_steps: Optional[int] = None, learning_rate: float = 5e-05, weight_decay: float = 0.0, adam_beta1: float = 0.9, adam_beta2: float = 0.999, adam_epsilon: float = 1e-08, max_grad_norm: float = 1.0, num_train_epochs: float = 3.0, max_steps: int = -1, lr_scheduler_type: transformers.trainer_utils.SchedulerType = 'linear', warmup_ratio: float = 0.0, warmup_steps: int = 0, logging_dir: Optional[str] = <factory>, logging_strategy: transformers.trainer_utils.IntervalStrategy = 'steps', logging_first_step: bool = False, logging_steps: int = 500, save_strategy: transformers.trainer_utils.IntervalStrategy = 'steps', save_steps: int = 500, save_total_limit: Optional[int] = None, no_cuda: bool = False, seed: int = 42, fp16: bool = False, fp16_opt_level: str = 'O1', fp16_backend: str = 'auto', fp16_full_eval: bool = False, local_rank: int = -1, tpu_num_cores: Optional[int] = None, tpu_metrics_debug: bool = False, debug: bool = False, dataloader_drop_last: bool = False, eval_steps: int = None, dataloader_num_workers: int = 0, past_index: int = -1, run_name: Optional[str] = None, disable_tqdm: Optional[bool] = None, remove_unused_columns: Optional[bool] = True, label_names: Optional[List[str]] = None, load_best_model_at_end: Optional[bool] = False, metric_for_best_model: Optional[str] = None, greater_is_better: Optional[bool] = None, ignore_data_skip: bool = False, sharded_ddp: str = '', deepspeed: Optional[str] = None, label_smoothing_factor: float = 0.0, adafactor: bool = False, group_by_length: bool = False, length_column_name: Optional[str] = 'length', report_to: Optional[List[str]] = None, ddp_find_unused_parameters: Optional[bool] = None, dataloader_pin_memory: bool = True, skip_memory_metrics: bool = False, mp_parameters: str = '')[source]

TrainingArguments is the subset of the arguments we use in our example scripts which relate to the training loop itself.

Using HfArgumentParser we can turn this class into argparse arguments that can be specified on the command line.

Parameters
  • output_dir (str) – The output directory where the model predictions and checkpoints will be written.

  • overwrite_output_dir (bool, optional, defaults to False) – If True, overwrite the content of the output directory. Use this to continue training if output_dir points to a checkpoint directory.

  • do_train (bool, optional, defaults to False) – Whether to run training or not. This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • do_eval (bool, optional) – Whether to run evaluation on the validation set or not. Will be set to True if evaluation_strategy is different from "no". This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • do_predict (bool, optional, defaults to False) – Whether to run predictions on the test set or not. This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • evaluation_strategy (str or IntervalStrategy, optional, defaults to "no") –

    The evaluation strategy to adopt during training. Possible values are:

    • "no": No evaluation is done during training.

    • "steps": Evaluation is done (and logged) every eval_steps.

    • "epoch": Evaluation is done at the end of each epoch.

  • prediction_loss_only (bool, optional, defaults to False) – When performing evaluation and generating predictions, only returns the loss.

  • per_device_train_batch_size (int, optional, defaults to 8) – The batch size per GPU/TPU core/CPU for training.

  • per_device_eval_batch_size (int, optional, defaults to 8) – The batch size per GPU/TPU core/CPU for evaluation.

  • gradient_accumulation_steps (int, optional, defaults to 1) –

    Number of updates steps to accumulate the gradients for, before performing a backward/update pass.

    Warning

    When using gradient accumulation, one step is counted as one step with backward pass. Therefore, logging, evaluation, save will be conducted every gradient_accumulation_steps * xxx_step training examples.

  • eval_accumulation_steps (int, optional) – Number of predictions steps to accumulate the output tensors for, before moving the results to the CPU. If left unset, the whole predictions are accumulated on GPU/TPU before being moved to the CPU (faster but requires more memory).

  • learning_rate (float, optional, defaults to 5e-5) – The initial learning rate for AdamW optimizer.

  • weight_decay (float, optional, defaults to 0) – The weight decay to apply (if not zero) to all layers except all bias and LayerNorm weights in AdamW optimizer.

  • adam_beta1 (float, optional, defaults to 0.9) – The beta1 hyperparameter for the AdamW optimizer.

  • adam_beta2 (float, optional, defaults to 0.999) – The beta2 hyperparameter for the AdamW optimizer.

  • adam_epsilon (float, optional, defaults to 1e-8) – The epsilon hyperparameter for the AdamW optimizer.

  • max_grad_norm (float, optional, defaults to 1.0) – Maximum gradient norm (for gradient clipping).

  • num_train_epochs (float, optional, defaults to 3.0) – Total number of training epochs to perform (if not an integer, will perform the decimal part percents of the last epoch before stopping training).

  • max_steps (int, optional, defaults to -1) – If set to a positive number, the total number of training steps to perform. Overrides num_train_epochs.

  • lr_scheduler_type (str or SchedulerType, optional, defaults to "linear") – The scheduler type to use. See the documentation of SchedulerType for all possible values.

  • warmup_ratio (float, optional, defaults to 0.0) – Ratio of total training steps used for a linear warmup from 0 to learning_rate.

  • warmup_steps (int, optional, defaults to 0) – Number of steps used for a linear warmup from 0 to learning_rate. Overrides any effect of warmup_ratio.

  • logging_dir (str, optional) – TensorBoard log directory. Will default to runs/**CURRENT_DATETIME_HOSTNAME**.

  • logging_strategy (str or IntervalStrategy, optional, defaults to "steps") –

    The logging strategy to adopt during training. Possible values are:

    • "no": No logging is done during training.

    • "epoch": Logging is done at the end of each epoch.

    • "steps": Logging is done every logging_steps.

  • logging_first_step (bool, optional, defaults to False) – Whether to log and evaluate the first global_step or not.

  • logging_steps (int, optional, defaults to 500) – Number of update steps between two logs if logging_strategy="steps".

  • save_strategy (str or IntervalStrategy, optional, defaults to "steps") –

    The checkpoint save strategy to adopt during training. Possible values are:

    • "no": No save is done during training.

    • "epoch": Save is done at the end of each epoch.

    • "steps": Save is done every save_steps.

  • save_steps (int, optional, defaults to 500) – Number of updates steps before two checkpoint saves if save_strategy="steps".

  • save_total_limit (int, optional) – If a value is passed, will limit the total amount of checkpoints. Deletes the older checkpoints in output_dir.

  • no_cuda (bool, optional, defaults to False) – Whether to not use CUDA even when it is available or not.

  • seed (int, optional, defaults to 42) – Random seed that will be set at the beginning of training. To ensure reproducibility across runs, use the model_init() function to instantiate the model if it has some randomly initialized parameters.

  • fp16 (bool, optional, defaults to False) – Whether to use 16-bit (mixed) precision training instead of 32-bit training.

  • fp16_opt_level (str, optional, defaults to ‘O1’) – For fp16 training, Apex AMP optimization level selected in [‘O0’, ‘O1’, ‘O2’, and ‘O3’]. See details on the Apex documentation.

  • fp16_backend (str, optional, defaults to "auto") – The backend to use for mixed precision training. Must be one of "auto", "amp" or "apex". "auto" will use AMP or APEX depending on the PyTorch version detected, while the other choices will force the requested backend.

  • fp16_full_eval (bool, optional, defaults to False) – Whether to use full 16-bit precision evaluation instead of 32-bit. This will be faster and save memory but can harm metric values.

  • local_rank (int, optional, defaults to -1) – Rank of the process during distributed training.

  • tpu_num_cores (int, optional) – When training on TPU, the number of TPU cores (automatically passed by launcher script).

  • debug (bool, optional, defaults to False) – When training on TPU, whether to print debug metrics or not.

  • dataloader_drop_last (bool, optional, defaults to False) – Whether to drop the last incomplete batch (if the length of the dataset is not divisible by the batch size) or not.

  • eval_steps (int, optional) – Number of update steps between two evaluations if evaluation_strategy="steps". Will default to the same value as logging_steps if not set.

  • dataloader_num_workers (int, optional, defaults to 0) – Number of subprocesses to use for data loading (PyTorch only). 0 means that the data will be loaded in the main process.

  • past_index (int, optional, defaults to -1) – Some models like TransformerXL or :doc`XLNet <../model_doc/xlnet>` can make use of the past hidden states for their predictions. If this argument is set to a positive int, the Trainer will use the corresponding output (usually index 2) as the past state and feed it to the model at the next training step under the keyword argument mems.

  • run_name (str, optional) – A descriptor for the run. Typically used for wandb logging.

  • disable_tqdm (bool, optional) – Whether or not to disable the tqdm progress bars and table of metrics produced by NotebookTrainingTracker in Jupyter Notebooks. Will default to True if the logging level is set to warn or lower (default), False otherwise.

  • remove_unused_columns (bool, optional, defaults to True) –

    If using datasets.Dataset datasets, whether or not to automatically remove the columns unused by the model forward method.

    (Note that this behavior is not implemented for TFTrainer yet.)

  • label_names (List[str], optional) –

    The list of keys in your dictionary of inputs that correspond to the labels.

    Will eventually default to ["labels"] except if the model used is one of the XxxForQuestionAnswering in which case it will default to ["start_positions", "end_positions"].

  • load_best_model_at_end (bool, optional, defaults to False) –

    Whether or not to load the best model found during training at the end of training.

    Note

    When set to True, the parameters save_strategy and save_steps will be ignored and the model will be saved after each evaluation.

  • metric_for_best_model (str, optional) –

    Use in conjunction with load_best_model_at_end to specify the metric to use to compare two different models. Must be the name of a metric returned by the evaluation with or without the prefix "eval_". Will default to "loss" if unspecified and load_best_model_at_end=True (to use the evaluation loss).

    If you set this value, greater_is_better will default to True. Don’t forget to set it to False if your metric is better when lower.

  • greater_is_better (bool, optional) –

    Use in conjunction with load_best_model_at_end and metric_for_best_model to specify if better models should have a greater metric or not. Will default to:

    • True if metric_for_best_model is set to a value that isn’t "loss" or "eval_loss".

    • False if metric_for_best_model is not set, or set to "loss" or "eval_loss".

  • ignore_data_skip (bool, optional, defaults to False) – When resuming training, whether or not to skip the epochs and batches to get the data loading at the same stage as in the previous training. If set to True, the training will begin faster (as that skipping step can take a long time) but will not yield the same results as the interrupted training would have.

  • sharded_ddp (bool, str or list of ShardedDDPOption, optional, defaults to False) –

    Use Sharded DDP training from FairScale (in distributed training only). This is an experimental feature.

    A list of options along the following:

    • "simple": to use first instance of sharded DDP released by fairscale (ShardedDDP) similar to ZeRO-2.

    • "zero_dp_2": to use the second instance of sharded DPP released by fairscale (FullyShardedDDP) in Zero-2 mode (with reshard_after_forward=False).

    • "zero_dp_3": to use the second instance of sharded DPP released by fairscale (FullyShardedDDP) in Zero-3 mode (with reshard_after_forward=True).

    • "offload": to add ZeRO-offload (only compatible with "zero_dp_2" and "zero_dp_3").

    If a string is passed, it will be split on space. If a bool is passed, it will be converted to an empty list for False and ["simple"] for True.

  • deepspeed (str or dict, optional) – Use Deepspeed. This is an experimental feature and its API may evolve in the future. The value is either the location of DeepSpeed json config file (e.g., ds_config.json) or an already loaded json file as a dict

  • label_smoothing_factor (float, optional, defaults to 0.0) – The label smoothing factor to use. Zero means no label smoothing, otherwise the underlying onehot-encoded labels are changed from 0s and 1s to label_smoothing_factor/num_labels and 1 - label_smoothing_factor + label_smoothing_factor/num_labels respectively.

  • adafactor (bool, optional, defaults to False) – Whether or not to use the Adafactor optimizer instead of AdamW.

  • group_by_length (bool, optional, defaults to False) – Whether or not to group together samples of roughly the same length in the training dataset (to minimize padding applied and be more efficient). Only useful if applying dynamic padding.

  • length_column_name (str, optional, defaults to "length") – Column name for precomputed lengths. If the column exists, grouping by length will use these values rather than computing them on train startup. Ignored unless group_by_length is True and the dataset is an instance of Dataset.

  • report_to (str or List[str], optional, defaults to "all") – The list of integrations to report the results and logs to. Supported platforms are "azure_ml", "comet_ml", "mlflow", "tensorboard" and "wandb". Use "all" to report to all integrations installed, "none" for no integrations.

  • ddp_find_unused_parameters (bool, optional) – When using distributed training, the value of the flag find_unused_parameters passed to DistributedDataParallel. Will default to False if gradient checkpointing is used, True otherwise.

  • dataloader_pin_memory (bool, optional, defaults to True)) – Whether you want to pin memory in data loaders or not. Will default to True.

  • skip_memory_metrics (bool, optional, defaults to False)) – Whether to skip adding of memory profiler reports to metrics. Defaults to False.

property device

The device used by this process.

property eval_batch_size

The actual batch size for evaluation (may differ from per_gpu_eval_batch_size in distributed training).

property n_gpu

The number of GPUs used by this process.

Note

This will only be greater than one when you have multiple GPUs available but are not using distributed training. For distributed training, it will always be 1.

property parallel_mode

The current mode used for parallelism if multiple GPUs/TPU cores are available. One of:

  • ParallelMode.NOT_PARALLEL: no parallelism (CPU or one GPU).

  • ParallelMode.NOT_DISTRIBUTED: several GPUs in one single process (uses torch.nn.DataParallel).

  • ParallelMode.DISTRIBUTED: several GPUs, each having its own process (uses torch.nn.DistributedDataParallel).

  • ParallelMode.TPU: several TPU cores.

property place_model_on_device

Can be subclassed and overridden for some specific integrations.

property process_index

The number of processes used in parallel.

to_dict()[source]

Serializes this instance while replace Enum by their values (for JSON serialization support).

to_json_string()[source]

Serializes this instance to a JSON string.

to_sanitized_dict() → Dict[str, Any][source]

Sanitized serialization to use with TensorBoard’s hparams

property train_batch_size

The actual batch size for training (may differ from per_gpu_train_batch_size in distributed training).

property world_size

The number of processes used in parallel.

Seq2SeqTrainingArguments

class transformers.Seq2SeqTrainingArguments(output_dir: str, overwrite_output_dir: bool = False, do_train: bool = False, do_eval: bool = None, do_predict: bool = False, evaluation_strategy: transformers.trainer_utils.IntervalStrategy = 'no', prediction_loss_only: bool = False, per_device_train_batch_size: int = 8, per_device_eval_batch_size: int = 8, per_gpu_train_batch_size: Optional[int] = None, per_gpu_eval_batch_size: Optional[int] = None, gradient_accumulation_steps: int = 1, eval_accumulation_steps: Optional[int] = None, learning_rate: float = 5e-05, weight_decay: float = 0.0, adam_beta1: float = 0.9, adam_beta2: float = 0.999, adam_epsilon: float = 1e-08, max_grad_norm: float = 1.0, num_train_epochs: float = 3.0, max_steps: int = -1, lr_scheduler_type: transformers.trainer_utils.SchedulerType = 'linear', warmup_ratio: float = 0.0, warmup_steps: int = 0, logging_dir: Optional[str] = <factory>, logging_strategy: transformers.trainer_utils.IntervalStrategy = 'steps', logging_first_step: bool = False, logging_steps: int = 500, save_strategy: transformers.trainer_utils.IntervalStrategy = 'steps', save_steps: int = 500, save_total_limit: Optional[int] = None, no_cuda: bool = False, seed: int = 42, fp16: bool = False, fp16_opt_level: str = 'O1', fp16_backend: str = 'auto', fp16_full_eval: bool = False, local_rank: int = -1, tpu_num_cores: Optional[int] = None, tpu_metrics_debug: bool = False, debug: bool = False, dataloader_drop_last: bool = False, eval_steps: int = None, dataloader_num_workers: int = 0, past_index: int = -1, run_name: Optional[str] = None, disable_tqdm: Optional[bool] = None, remove_unused_columns: Optional[bool] = True, label_names: Optional[List[str]] = None, load_best_model_at_end: Optional[bool] = False, metric_for_best_model: Optional[str] = None, greater_is_better: Optional[bool] = None, ignore_data_skip: bool = False, sharded_ddp: str = '', deepspeed: Optional[str] = None, label_smoothing_factor: float = 0.0, adafactor: bool = False, group_by_length: bool = False, length_column_name: Optional[str] = 'length', report_to: Optional[List[str]] = None, ddp_find_unused_parameters: Optional[bool] = None, dataloader_pin_memory: bool = True, skip_memory_metrics: bool = False, mp_parameters: str = '', sortish_sampler: bool = False, predict_with_generate: bool = False)[source]

TrainingArguments is the subset of the arguments we use in our example scripts which relate to the training loop itself.

Using HfArgumentParser we can turn this class into argparse arguments that can be specified on the command line.

Parameters
  • output_dir (str) – The output directory where the model predictions and checkpoints will be written.

  • overwrite_output_dir (bool, optional, defaults to False) – If True, overwrite the content of the output directory. Use this to continue training if output_dir points to a checkpoint directory.

  • do_train (bool, optional, defaults to False) – Whether to run training or not. This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • do_eval (bool, optional) – Whether to run evaluation on the validation set or not. Will be set to True if evaluation_strategy is different from "no". This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • do_predict (bool, optional, defaults to False) – Whether to run predictions on the test set or not. This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • evaluation_strategy (str or IntervalStrategy, optional, defaults to "no") –

    The evaluation strategy to adopt during training. Possible values are:

    • "no": No evaluation is done during training.

    • "steps": Evaluation is done (and logged) every eval_steps.

    • "epoch": Evaluation is done at the end of each epoch.

  • prediction_loss_only (bool, optional, defaults to False) – When performing evaluation and generating predictions, only returns the loss.

  • per_device_train_batch_size (int, optional, defaults to 8) – The batch size per GPU/TPU core/CPU for training.

  • per_device_eval_batch_size (int, optional, defaults to 8) – The batch size per GPU/TPU core/CPU for evaluation.

  • gradient_accumulation_steps (int, optional, defaults to 1) –

    Number of updates steps to accumulate the gradients for, before performing a backward/update pass.

    Warning

    When using gradient accumulation, one step is counted as one step with backward pass. Therefore, logging, evaluation, save will be conducted every gradient_accumulation_steps * xxx_step training examples.

  • eval_accumulation_steps (int, optional) – Number of predictions steps to accumulate the output tensors for, before moving the results to the CPU. If left unset, the whole predictions are accumulated on GPU/TPU before being moved to the CPU (faster but requires more memory).

  • learning_rate (float, optional, defaults to 5e-5) – The initial learning rate for AdamW optimizer.

  • weight_decay (float, optional, defaults to 0) – The weight decay to apply (if not zero) to all layers except all bias and LayerNorm weights in AdamW optimizer.

  • adam_beta1 (float, optional, defaults to 0.9) – The beta1 hyperparameter for the AdamW optimizer.

  • adam_beta2 (float, optional, defaults to 0.999) – The beta2 hyperparameter for the AdamW optimizer.

  • adam_epsilon (float, optional, defaults to 1e-8) – The epsilon hyperparameter for the AdamW optimizer.

  • max_grad_norm (float, optional, defaults to 1.0) – Maximum gradient norm (for gradient clipping).

  • num_train_epochs (float, optional, defaults to 3.0) – Total number of training epochs to perform (if not an integer, will perform the decimal part percents of the last epoch before stopping training).

  • max_steps (int, optional, defaults to -1) – If set to a positive number, the total number of training steps to perform. Overrides num_train_epochs.

  • lr_scheduler_type (str or SchedulerType, optional, defaults to "linear") – The scheduler type to use. See the documentation of SchedulerType for all possible values.

  • warmup_ratio (float, optional, defaults to 0.0) – Ratio of total training steps used for a linear warmup from 0 to learning_rate.

  • warmup_steps (int, optional, defaults to 0) – Number of steps used for a linear warmup from 0 to learning_rate. Overrides any effect of warmup_ratio.

  • logging_dir (str, optional) – TensorBoard log directory. Will default to runs/**CURRENT_DATETIME_HOSTNAME**.

  • logging_strategy (str or IntervalStrategy, optional, defaults to "steps") –

    The logging strategy to adopt during training. Possible values are:

    • "no": No logging is done during training.

    • "epoch": Logging is done at the end of each epoch.

    • "steps": Logging is done every logging_steps.

  • logging_first_step (bool, optional, defaults to False) – Whether to log and evaluate the first global_step or not.

  • logging_steps (int, optional, defaults to 500) – Number of update steps between two logs if logging_strategy="steps".

  • save_strategy (str or IntervalStrategy, optional, defaults to "steps") –

    The checkpoint save strategy to adopt during training. Possible values are:

    • "no": No save is done during training.

    • "epoch": Save is done at the end of each epoch.

    • "steps": Save is done every save_steps.

  • save_steps (int, optional, defaults to 500) – Number of updates steps before two checkpoint saves if save_strategy="steps".

  • save_total_limit (int, optional) – If a value is passed, will limit the total amount of checkpoints. Deletes the older checkpoints in output_dir.

  • no_cuda (bool, optional, defaults to False) – Whether to not use CUDA even when it is available or not.

  • seed (int, optional, defaults to 42) – Random seed that will be set at the beginning of training. To ensure reproducibility across runs, use the model_init() function to instantiate the model if it has some randomly initialized parameters.

  • fp16 (bool, optional, defaults to False) – Whether to use 16-bit (mixed) precision training instead of 32-bit training.

  • fp16_opt_level (str, optional, defaults to ‘O1’) – For fp16 training, Apex AMP optimization level selected in [‘O0’, ‘O1’, ‘O2’, and ‘O3’]. See details on the Apex documentation.

  • fp16_backend (str, optional, defaults to "auto") – The backend to use for mixed precision training. Must be one of "auto", "amp" or "apex". "auto" will use AMP or APEX depending on the PyTorch version detected, while the other choices will force the requested backend.

  • fp16_full_eval (bool, optional, defaults to False) – Whether to use full 16-bit precision evaluation instead of 32-bit. This will be faster and save memory but can harm metric values.

  • local_rank (int, optional, defaults to -1) – Rank of the process during distributed training.

  • tpu_num_cores (int, optional) – When training on TPU, the number of TPU cores (automatically passed by launcher script).

  • debug (bool, optional, defaults to False) – When training on TPU, whether to print debug metrics or not.

  • dataloader_drop_last (bool, optional, defaults to False) – Whether to drop the last incomplete batch (if the length of the dataset is not divisible by the batch size) or not.

  • eval_steps (int, optional) – Number of update steps between two evaluations if evaluation_strategy="steps". Will default to the same value as logging_steps if not set.

  • dataloader_num_workers (int, optional, defaults to 0) – Number of subprocesses to use for data loading (PyTorch only). 0 means that the data will be loaded in the main process.

  • past_index (int, optional, defaults to -1) – Some models like TransformerXL or :doc`XLNet <../model_doc/xlnet>` can make use of the past hidden states for their predictions. If this argument is set to a positive int, the Trainer will use the corresponding output (usually index 2) as the past state and feed it to the model at the next training step under the keyword argument mems.

  • run_name (str, optional) –

    A descriptor for the run. Typically used for wandb logging.

  • disable_tqdm (bool, optional) – Whether or not to disable the tqdm progress bars and table of metrics produced by NotebookTrainingTracker in Jupyter Notebooks. Will default to True if the logging level is set to warn or lower (default), False otherwise.

  • remove_unused_columns (bool, optional, defaults to True) –

    If using datasets.Dataset datasets, whether or not to automatically remove the columns unused by the model forward method.

    (Note that this behavior is not implemented for TFTrainer yet.)

  • label_names (List[str], optional) –

    The list of keys in your dictionary of inputs that correspond to the labels.

    Will eventually default to ["labels"] except if the model used is one of the XxxForQuestionAnswering in which case it will default to ["start_positions", "end_positions"].

  • load_best_model_at_end (bool, optional, defaults to False) –

    Whether or not to load the best model found during training at the end of training.

    Note

    When set to True, the parameters save_strategy and save_steps will be ignored and the model will be saved after each evaluation.

  • metric_for_best_model (str, optional) –

    Use in conjunction with load_best_model_at_end to specify the metric to use to compare two different models. Must be the name of a metric returned by the evaluation with or without the prefix "eval_". Will default to "loss" if unspecified and load_best_model_at_end=True (to use the evaluation loss).

    If you set this value, greater_is_better will default to True. Don’t forget to set it to False if your metric is better when lower.

  • greater_is_better (bool, optional) –

    Use in conjunction with load_best_model_at_end and metric_for_best_model to specify if better models should have a greater metric or not. Will default to:

    • True if metric_for_best_model is set to a value that isn’t "loss" or "eval_loss".

    • False if metric_for_best_model is not set, or set to "loss" or "eval_loss".

  • ignore_data_skip (bool, optional, defaults to False) – When resuming training, whether or not to skip the epochs and batches to get the data loading at the same stage as in the previous training. If set to True, the training will begin faster (as that skipping step can take a long time) but will not yield the same results as the interrupted training would have.

  • sharded_ddp (bool, str or list of ShardedDDPOption, optional, defaults to False) –

    Use Sharded DDP training from FairScale (in distributed training only). This is an experimental feature.

    A list of options along the following:

    • "simple": to use first instance of sharded DDP released by fairscale (ShardedDDP) similar to ZeRO-2.

    • "zero_dp_2": to use the second instance of sharded DPP released by fairscale (FullyShardedDDP) in Zero-2 mode (with reshard_after_forward=False).

    • "zero_dp_3": to use the second instance of sharded DPP released by fairscale (FullyShardedDDP) in Zero-3 mode (with reshard_after_forward=True).

    • "offload": to add ZeRO-offload (only compatible with "zero_dp_2" and "zero_dp_3").

    If a string is passed, it will be split on space. If a bool is passed, it will be converted to an empty list for False and ["simple"] for True.

  • deepspeed (str or dict, optional) – Use Deepspeed. This is an experimental feature and its API may evolve in the future. The value is either the location of DeepSpeed json config file (e.g., ds_config.json) or an already loaded json file as a dict

  • label_smoothing_factor (float, optional, defaults to 0.0) – The label smoothing factor to use. Zero means no label smoothing, otherwise the underlying onehot-encoded labels are changed from 0s and 1s to label_smoothing_factor/num_labels and 1 - label_smoothing_factor + label_smoothing_factor/num_labels respectively.

  • adafactor (bool, optional, defaults to False) – Whether or not to use the Adafactor optimizer instead of AdamW.

  • group_by_length (bool, optional, defaults to False) – Whether or not to group together samples of roughly the same length in the training dataset (to minimize padding applied and be more efficient). Only useful if applying dynamic padding.

  • length_column_name (str, optional, defaults to "length") – Column name for precomputed lengths. If the column exists, grouping by length will use these values rather than computing them on train startup. Ignored unless group_by_length is True and the dataset is an instance of Dataset.

  • report_to (str or List[str], optional, defaults to "all") – The list of integrations to report the results and logs to. Supported platforms are "azure_ml", "comet_ml", "mlflow", "tensorboard" and "wandb". Use "all" to report to all integrations installed, "none" for no integrations.

  • ddp_find_unused_parameters (bool, optional) – When using distributed training, the value of the flag find_unused_parameters passed to DistributedDataParallel. Will default to False if gradient checkpointing is used, True otherwise.

  • dataloader_pin_memory (bool, optional, defaults to True)) – Whether you want to pin memory in data loaders or not. Will default to True.

  • skip_memory_metrics (bool, optional, defaults to False)) – Whether to skip adding of memory profiler reports to metrics. Defaults to False.

sortish_sampler (bool, optional, defaults to False):

Whether to use a sortish sampler or not. Only possible if the underlying datasets are Seq2SeqDataset for now but will become generally available in the near future.

It sorts the inputs according to lengths in order to minimize the padding size, with a bit of randomness for the training set.

predict_with_generate (bool, optional, defaults to False):

Whether to use generate to calculate generative metrics (ROUGE, BLEU).

TFTrainingArguments

class transformers.TFTrainingArguments(output_dir: str, overwrite_output_dir: bool = False, do_train: bool = False, do_eval: bool = None, do_predict: bool = False, evaluation_strategy: transformers.trainer_utils.IntervalStrategy = 'no', prediction_loss_only: bool = False, per_device_train_batch_size: int = 8, per_device_eval_batch_size: int = 8, per_gpu_train_batch_size: Optional[int] = None, per_gpu_eval_batch_size: Optional[int] = None, gradient_accumulation_steps: int = 1, eval_accumulation_steps: Optional[int] = None, learning_rate: float = 5e-05, weight_decay: float = 0.0, adam_beta1: float = 0.9, adam_beta2: float = 0.999, adam_epsilon: float = 1e-08, max_grad_norm: float = 1.0, num_train_epochs: float = 3.0, max_steps: int = -1, lr_scheduler_type: transformers.trainer_utils.SchedulerType = 'linear', warmup_ratio: float = 0.0, warmup_steps: int = 0, logging_dir: Optional[str] = <factory>, logging_strategy: transformers.trainer_utils.IntervalStrategy = 'steps', logging_first_step: bool = False, logging_steps: int = 500, save_strategy: transformers.trainer_utils.IntervalStrategy = 'steps', save_steps: int = 500, save_total_limit: Optional[int] = None, no_cuda: bool = False, seed: int = 42, fp16: bool = False, fp16_opt_level: str = 'O1', fp16_backend: str = 'auto', fp16_full_eval: bool = False, local_rank: int = -1, tpu_num_cores: Optional[int] = None, tpu_metrics_debug: bool = False, debug: bool = False, dataloader_drop_last: bool = False, eval_steps: int = None, dataloader_num_workers: int = 0, past_index: int = -1, run_name: Optional[str] = None, disable_tqdm: Optional[bool] = None, remove_unused_columns: Optional[bool] = True, label_names: Optional[List[str]] = None, load_best_model_at_end: Optional[bool] = False, metric_for_best_model: Optional[str] = None, greater_is_better: Optional[bool] = None, ignore_data_skip: bool = False, sharded_ddp: str = '', deepspeed: Optional[str] = None, label_smoothing_factor: float = 0.0, adafactor: bool = False, group_by_length: bool = False, length_column_name: Optional[str] = 'length', report_to: Optional[List[str]] = None, ddp_find_unused_parameters: Optional[bool] = None, dataloader_pin_memory: bool = True, skip_memory_metrics: bool = False, mp_parameters: str = '', tpu_name: str = None, tpu_zone: str = None, gcp_project: str = None, poly_power: float = 1.0, xla: bool = False)[source]

TrainingArguments is the subset of the arguments we use in our example scripts which relate to the training loop itself.

Using HfArgumentParser we can turn this class into argparse arguments that can be specified on the command line.

Parameters
  • output_dir (str) – The output directory where the model predictions and checkpoints will be written.

  • overwrite_output_dir (bool, optional, defaults to False) – If True, overwrite the content of the output directory. Use this to continue training if output_dir points to a checkpoint directory.

  • do_train (bool, optional, defaults to False) – Whether to run training or not. This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • do_eval (bool, optional) – Whether to run evaluation on the validation set or not. Will be set to True if evaluation_strategy is different from "no". This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • do_predict (bool, optional, defaults to False) – Whether to run predictions on the test set or not. This argument is not directly used by Trainer, it’s intended to be used by your training/evaluation scripts instead. See the example scripts for more details.

  • evaluation_strategy (str or IntervalStrategy, optional, defaults to "no") –

    The evaluation strategy to adopt during training. Possible values are:

    • "no": No evaluation is done during training.

    • "steps": Evaluation is done (and logged) every eval_steps.

    • "epoch": Evaluation is done at the end of each epoch.

  • per_device_train_batch_size (int, optional, defaults to 8) – The batch size per GPU/TPU core/CPU for training.

  • per_device_eval_batch_size (int, optional, defaults to 8) – The batch size per GPU/TPU core/CPU for evaluation.

  • gradient_accumulation_steps

    (int, optional, defaults to 1): Number of updates steps to accumulate the gradients for, before performing a backward/update pass.

    Warning

    When using gradient accumulation, one step is counted as one step with backward pass. Therefore, logging, evaluation, save will be conducted every gradient_accumulation_steps * xxx_step training examples.

  • learning_rate (float, optional, defaults to 5e-5) – The initial learning rate for Adam.

  • weight_decay (float, optional, defaults to 0) – The weight decay to apply (if not zero).

  • adam_beta1 (float, optional, defaults to 0.9) – The beta1 hyperparameter for the Adam optimizer.

  • adam_beta2 (float, optional, defaults to 0.999) – The beta2 hyperparameter for the Adam optimizer.

  • adam_epsilon (float, optional, defaults to 1e-8) – The epsilon hyperparameter for the Adam optimizer.

  • max_grad_norm (float, optional, defaults to 1.0) – Maximum gradient norm (for gradient clipping).

  • num_train_epochs (float, optional, defaults to 3.0) – Total number of training epochs to perform.

  • max_steps (int, optional, defaults to -1) – If set to a positive number, the total number of training steps to perform. Overrides num_train_epochs.

  • warmup_ratio (float, optional, defaults to 0.0) – Ratio of total training steps used for a linear warmup from 0 to learning_rate.

  • warmup_steps (int, optional, defaults to 0) – Number of steps used for a linear warmup from 0 to learning_rate. Overrides any effect of warmup_ratio.

  • logging_dir (str, optional) – TensorBoard log directory. Will default to runs/**CURRENT_DATETIME_HOSTNAME**.

  • logging_strategy (str or IntervalStrategy, optional, defaults to "steps") –

    The logging strategy to adopt during training. Possible values are:

    • "no": No logging is done during training.

    • "epoch": Logging is done at the end of each epoch.

    • "steps": Logging is done every logging_steps.

  • logging_first_step (bool, optional, defaults to False) – Whether to log and evaluate the first global_step or not.

  • logging_steps (int, optional, defaults to 500) – Number of update steps between two logs if logging_strategy="steps".

  • save_strategy (str or IntervalStrategy, optional, defaults to "steps") –

    The checkpoint save strategy to adopt during training. Possible values are:

    • "no": No save is done during training.

    • "epoch": Save is done at the end of each epoch.

    • "steps": Save is done every save_steps.

  • save_steps (int, optional, defaults to 500) – Number of updates steps before two checkpoint saves if save_strategy="steps".

  • save_total_limit (int, optional) – If a value is passed, will limit the total amount of checkpoints. Deletes the older checkpoints in output_dir.

  • no_cuda (bool, optional, defaults to False) – Whether to not use CUDA even when it is available or not.

  • seed (int, optional, defaults to 42) – Random seed that will be set at the beginning of training.

  • fp16 (bool, optional, defaults to False) – Whether to use 16-bit (mixed) precision training (through NVIDIA Apex) instead of 32-bit training.

  • fp16_opt_level (str, optional, defaults to ‘O1’) – For fp16 training, Apex AMP optimization level selected in [‘O0’, ‘O1’, ‘O2’, and ‘O3’]. See details on the Apex documentation.

  • local_rank (int, optional, defaults to -1) – During distributed training, the rank of the process.

  • tpu_num_cores (int, optional) – When training on TPU, the number of TPU cores (automatically passed by launcher script).

  • debug (bool, optional, defaults to False) – Whether to activate the trace to record computation graphs and profiling information or not.

  • dataloader_drop_last (bool, optional, defaults to False) – Whether to drop the last incomplete batch (if the length of the dataset is not divisible by the batch size) or not.

  • eval_steps (int, optional, defaults to 1000) – Number of update steps before two evaluations.

  • past_index (int, optional, defaults to -1) – Some models like TransformerXL or :doc`XLNet <../model_doc/xlnet>` can make use of the past hidden states for their predictions. If this argument is set to a positive int, the Trainer will use the corresponding output (usually index 2) as the past state and feed it to the model at the next training step under the keyword argument mems.

  • tpu_name (str, optional) – The name of the TPU the process is running on.

  • tpu_zone (str, optional) – The zone of the TPU the process is running on. If not specified, we will attempt to automatically detect from metadata.

  • gcp_project (str, optional) – Google Cloud Project name for the Cloud TPU-enabled project. If not specified, we will attempt to automatically detect from metadata.

  • run_name (str, optional) – A descriptor for the run. Notably used for wandb logging.

  • xla (bool, optional) – Whether to activate the XLA compilation or not.

property eval_batch_size

The actual batch size for evaluation (may differ from per_gpu_eval_batch_size in distributed training).

property n_gpu

The number of replicas (CPUs, GPUs or TPU cores) used in this training.

property n_replicas

The number of replicas (CPUs, GPUs or TPU cores) used in this training.

property strategy

The strategy used for distributed training.

property train_batch_size

The actual batch size for training (may differ from per_gpu_train_batch_size in distributed training).

Trainer Integrations

The Trainer has been extended to support libraries that may dramatically improve your training time and fit much bigger models.

Currently it supports third party solutions, DeepSpeed and FairScale, which implement parts of the paper ZeRO: Memory Optimizations Toward Training Trillion Parameter Models, by Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, Yuxiong He.

This provided support is new and experimental as of this writing.

Installation Notes

As of this writing, both FairScale and Deepspeed require compilation of CUDA C++ code, before they can be used.

While all installation issues should be dealt with through the corresponding GitHub Issues of FairScale and Deepspeed, there are a few common issues that one may encounter while building any PyTorch extension that needs to build CUDA extensions.

Therefore, if you encounter a CUDA-related build issue while doing one of the following or both:

pip install fairscale
pip install deepspeed

please, read the following notes first.

In these notes we give examples for what to do when pytorch has been built with CUDA 10.2. If your situation is different remember to adjust the version number to the one you are after.

Possible problem #1

While, Pytorch comes with its own CUDA toolkit, to build these two projects you must have an identical version of CUDA installed system-wide.

For example, if you installed pytorch with cudatoolkit==10.2 in the Python environment, you also need to have CUDA 10.2 installed system-wide.

The exact location may vary from system to system, but /usr/local/cuda-10.2 is the most common location on many Unix systems. When CUDA is correctly set up and added to the PATH environment variable, one can find the installation location by doing:

which nvcc

If you don’t have CUDA installed system-wide, install it first. You will find the instructions by using your favorite search engine. For example, if you’re on Ubuntu you may want to search for: ubuntu cuda 10.2 install.

Possible problem #2

Another possible common problem is that you may have more than one CUDA toolkit installed system-wide. For example you may have:

/usr/local/cuda-10.2
/usr/local/cuda-11.0

Now, in this situation you need to make sure that your PATH and LD_LIBRARY_PATH environment variables contain the correct paths to the desired CUDA version. Typically, package installers will set these to contain whatever the last version was installed. If you encounter the problem, where the package build fails because it can’t find the right CUDA version despite you having it installed system-wide, it means that you need to adjust the 2 aforementioned environment variables.

First, you may look at their contents:

echo $PATH
echo $LD_LIBRARY_PATH

so you get an idea of what is inside.

It’s possible that LD_LIBRARY_PATH is empty.

PATH lists the locations of where executables can be found and LD_LIBRARY_PATH is for where shared libraries are to looked for. In both cases, earlier entries have priority over the later ones. : is used to separate multiple entries.

Now, to tell the build program where to find the specific CUDA toolkit, insert the desired paths to be listed first by doing:

export PATH=/usr/local/cuda-10.2/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-10.2/lib64:$LD_LIBRARY_PATH

Note that we aren’t overwriting the existing values, but prepending instead.

Of course, adjust the version number, the full path if need be. Check that the directories you assign actually do exist. lib64 sub-directory is where the various CUDA .so objects, like libcudart.so reside, it’s unlikely that your system will have it named differently, but if it is adjust it to reflect your reality.

Possible problem #3

Some older CUDA versions may refuse to build with newer compilers. For example, you my have gcc-9 but it wants gcc-7.

There are various ways to go about it.

If you can install the latest CUDA toolkit it typically should support the newer compiler.

Alternatively, you could install the lower version of the compiler in addition to the one you already have, or you may already have it but it’s not the default one, so the build system can’t see it. If you have gcc-7 installed but the build system complains it can’t find it, the following might do the trick:

sudo ln -s /usr/bin/gcc-7  /usr/local/cuda-10.2/bin/gcc
sudo ln -s /usr/bin/g++-7  /usr/local/cuda-10.2/bin/g++

Here, we are making a symlink to gcc-7 from /usr/local/cuda-10.2/bin/gcc and since /usr/local/cuda-10.2/bin/ should be in the PATH environment variable (see the previous problem’s solution), it should find gcc-7 (and g++7) and then the build will succeed.

As always make sure to edit the paths in the example to match your situation.

FairScale

By integrating FairScale the Trainer provides support for the following features from the ZeRO paper:

  1. Optimizer State Sharding

  2. Gradient Sharding

  3. Model Parameters Sharding (new and very experimental)

  4. CPU offload (new and very experimental)

You will need at least two GPUs to use this feature.

Installation:

Install the library via pypi:

pip install fairscale

or via transformersextras:

pip install transformers[fairscale]

(will become available starting from transformers==4.6.0)

or find more details on the FairScale’s GitHub page.

If you’re still struggling with the build, first make sure to read Installation Notes.

If it’s still not resolved the build issue, here are a few more ideas.

fairscale seems to have an issue with the recently introduced by pip build isolation feature. If you have a problem with it, you may want to try one of:

pip install fairscale --no-build-isolation .

or:

git clone https://github.com/facebookresearch/fairscale/
cd fairscale
rm -r dist build
python setup.py bdist_wheel
pip uninstall -y fairscale
pip install dist/fairscale-*.whl

fairscale also has issues with building against pytorch-nightly, so if you use it you may have to try one of:

pip uninstall -y fairscale; pip install fairscale --pre \
-f https://download.pytorch.org/whl/nightly/cu110/torch_nightly.html \
--no-cache --no-build-isolation

or:

pip install -v --disable-pip-version-check . \
-f https://download.pytorch.org/whl/nightly/cu110/torch_nightly.html --pre

Of course, adjust the urls to match the cuda version you use.

If after trying everything suggested you still encounter build issues, please, proceed with the GitHub Issue of FairScale.

Usage:

To use the first version of Sharded data-parallelism, add --sharded_ddp simple to the command line arguments, and make sure you have added the distributed launcher -m torch.distributed.launch --nproc_per_node=NUMBER_OF_GPUS_YOU_HAVE if you haven’t been using it already.

For example here is how you could use it for run_translation.py with 2 GPUs:

python -m torch.distributed.launch --nproc_per_node=2 examples/seq2seq/run_translation.py \
--model_name_or_path t5-small --per_device_train_batch_size 1   \
--output_dir output_dir --overwrite_output_dir \
--do_train --max_train_samples 500 --num_train_epochs 1 \
--dataset_name wmt16 --dataset_config "ro-en" \
--source_lang en --target_lang ro \
--fp16 --sharded_ddp simple

Notes:

  • This feature requires distributed training (so multiple GPUs).

  • It is not implemented for TPUs.

  • It works with --fp16 too, to make things even faster.

  • One of the main benefits of enabling --sharded_ddp simple is that it uses a lot less GPU memory, so you should be able to use significantly larger batch sizes using the same hardware (e.g. 3x and even bigger) which should lead to significantly shorter training time.

  1. To use the second version of Sharded data-parallelism, add --sharded_ddp zero_dp_2 or --sharded_ddp zero_dp_3 to the command line arguments, and make sure you have added the distributed launcher -m torch.distributed.launch --nproc_per_node=NUMBER_OF_GPUS_YOU_HAVE if you haven’t been using it already.

For example here is how you could use it for run_translation.py with 2 GPUs:

python -m torch.distributed.launch --nproc_per_node=2 examples/seq2seq/run_translation.py \
--model_name_or_path t5-small --per_device_train_batch_size 1   \
--output_dir output_dir --overwrite_output_dir \
--do_train --max_train_samples 500 --num_train_epochs 1 \
--dataset_name wmt16 --dataset_config "ro-en" \
--source_lang en --target_lang ro \
--fp16 --sharded_ddp zero_dp_2

zero_dp_2 is an optimized version of the simple wrapper, while zero_dp_3 fully shards model weights, gradients and optimizer states.

Both are compatible with adding cpu_offload to enable ZeRO-offload (activate it like this: --sharded_ddp "zero_dp_2 cpu_offload").

Notes:

  • This feature requires distributed training (so multiple GPUs).

  • It is not implemented for TPUs.

  • It works with --fp16 too, to make things even faster.

  • The cpu_offload additional option requires --fp16.

  • This is an area of active development, so make sure you have a source install of fairscale to use this feature as some bugs you encounter may have been fixed there already.

Known caveats:

  • This feature is incompatible with --predict_with_generate in the run_translation.py script.

  • Using --sharded_ddp zero_dp_3 requires wrapping each layer of the model in the special container FullyShardedDataParallelism of fairscale. It should be used with the option auto_wrap if you are not doing this yourself: --sharded_ddp "zero_dp_3 auto_wrap".

DeepSpeed

DeepSpeed implements everything described in the ZeRO paper. Currently it provides full support for:

  1. Optimizer State Partitioning (ZeRO stage 1)

  2. Gradient Partitioning (ZeRO stage 2)

  3. Param Partitioning (ZeRO stage 3)

  4. Custom mixed precision training handling

  5. A range of fast CUDA-extension-based Optimizers

  6. ZeRO-Offload

ZeRO-Offload has its own dedicated paper: ZeRO-Offload: Democratizing Billion-Scale Model Training.

DeepSpeed ZeRO-2 is currently used only for training, as all the currently available features are of no use to inference.

DeepSpeed ZeRO-3 can be used for inference as well, since it allows huge models to be loaded on multiple GPUs, which won’t be possible on a single GPU.

Installation

Install the library via pypi:

pip install deepspeed

or via transformersextras:

pip install transformers[deepspeed]

(will become available starting from transformers==4.6.0)

or find more details on the DeepSpeed’s GitHub page and advanced install.

If you’re still struggling with the build, first make sure to read Installation Notes.

If you don’t prebuild the extensions and rely on them to be built at run time and you tried all of the above solutions to no avail, the next thing to try is to pre-build the modules before installing them.

To make a local build for DeepSpeed:

git clone https://github.com/microsoft/DeepSpeed/
cd DeepSpeed
rm -rf build
TORCH_CUDA_ARCH_LIST="6.1;8.6" DS_BUILD_OPS=1 pip install . \
--global-option="build_ext" --global-option="-j8" --no-cache -v \
--disable-pip-version-check 2>&1 | tee build.log

Edit TORCH_CUDA_ARCH_LIST to insert the code for the architectures of the GPU cards you intend to use.

Or if you need to use the same setup on multiple machines, make a binary wheel:

git clone https://github.com/microsoft/DeepSpeed/
cd DeepSpeed
rm -rf build
TORCH_CUDA_ARCH_LIST="6.1;8.6" DS_BUILD_OPS=1 \
python setup.py build_ext -j8 bdist_wheel

it will generate something like dist/deepspeed-0.3.13+8cd046f-cp38-cp38-linux_x86_64.whl which now you can install as pip install deepspeed-0.3.13+8cd046f-cp38-cp38-linux_x86_64.whl locally or on any other machine.

Again, remember to ensure to adjust TORCH_CUDA_ARCH_LIST to the target architectures.

You can find the complete list of NVIDIA GPUs and their corresponding Compute Capabilities (same as arch in this context) here.

You can check the archs pytorch was built with using:

python -c "import torch; print(torch.cuda.get_arch_list())"

Here is how to find out the arch for one of the installed GPU. For example, for GPU 0:

CUDA_VISIBLE_DEVICES=0 python -c "import torch; \
print(torch.cuda.get_device_properties(torch.device('cuda')))"

If the output is:

_CudaDeviceProperties(name='GeForce RTX 3090', major=8, minor=6, total_memory=24268MB, multi_processor_count=82)

then you know that this card’s arch is 8.6.

You can also leave TORCH_CUDA_ARCH_LIST out completely and then the build program will automatically query the architecture of the GPUs the build is made on. This may or may not match the GPUs on the target machines, that’s why it’s best to specify the desired archs explicitly.

If after trying everything suggested you still encounter build issues, please, proceed with the GitHub Issue of Deepspeed,

Deployment with multiple GPUs

To deploy this feature with multiple GPUs adjust the Trainer command line arguments as following:

  1. replace python -m torch.distributed.launch with deepspeed.

  2. add a new argument --deepspeed ds_config.json, where ds_config.json is the DeepSpeed configuration file as documented here. The file naming is up to you.

Therefore, if your original command line looked as following:

python -m torch.distributed.launch --nproc_per_node=2 your_program.py <normal cl args>

Now it should be:

deepspeed --num_gpus=2 your_program.py <normal cl args> --deepspeed ds_config.json

Unlike, torch.distributed.launch where you have to specify how many GPUs to use with --nproc_per_node, with the deepspeed launcher you don’t have to use the corresponding --num_gpus if you want all of your GPUs used. The full details on how to configure various nodes and GPUs can be found here.

In fact, you can continue using -m torch.distributed.launch with DeepSpeed as long as you don’t need to use deepspeed launcher-specific arguments. Typically if you don’t need a multi-node setup you’re not required to use the deepspeed launcher. But since in the DeepSpeed documentation it’ll be used everywhere, for consistency we will use it here as well.

Here is an example of running run_translation.py under DeepSpeed deploying all available GPUs:

deepspeed examples/seq2seq/run_translation.py \
--deepspeed tests/deepspeed/ds_config.json \
--model_name_or_path t5-small --per_device_train_batch_size 1   \
--output_dir output_dir --overwrite_output_dir --fp16 \
--do_train --max_train_samples 500 --num_train_epochs 1 \
--dataset_name wmt16 --dataset_config "ro-en" \
--source_lang en --target_lang ro

Note that in the DeepSpeed documentation you are likely to see --deepspeed --deepspeed_config ds_config.json - i.e. two DeepSpeed-related arguments, but for the sake of simplicity, and since there are already so many arguments to deal with, we combined the two into a single argument.

For some practical usage examples, please, see this post.

Deployment with one GPU

To deploy DeepSpeed with one GPU adjust the Trainer command line arguments as following:

deepspeed --num_gpus=1 examples/seq2seq/run_translation.py \
--deepspeed tests/deepspeed/ds_config.json \
--model_name_or_path t5-small --per_device_train_batch_size 1   \
--output_dir output_dir --overwrite_output_dir --fp16 \
--do_train --max_train_samples 500 --num_train_epochs 1 \
--dataset_name wmt16 --dataset_config "ro-en" \
--source_lang en --target_lang ro

This is almost the same as with multiple-GPUs, but here we tell DeepSpeed explicitly to use just one GPU. By default, DeepSpeed deploys all GPUs it can see. If you have only 1 GPU to start with, then you don’t need this argument. The following documentation discusses the launcher options.

Why would you want to use DeepSpeed with just one GPU?

  1. It has a ZeRO-offload feature which can delegate some computations and memory to the host’s CPU and RAM, and thus leave more GPU resources for model’s needs - e.g. larger batch size, or enabling a fitting of a very big model which normally won’t fit.

  2. It provides a smart GPU memory management system, that minimizes memory fragmentation, which again allows you to fit bigger models and data batches.

While we are going to discuss the configuration in details next, the key to getting a huge improvement on a single GPU with DeepSpeed is to have at least the following configuration in the configuration file:

{
  "zero_optimization": {
     "stage": 2,
     "allgather_partitions": true,
     "allgather_bucket_size": 2e8,
     "reduce_scatter": true,
     "reduce_bucket_size": 2e8,
     "overlap_comm": true,
     "contiguous_gradients": true,
     "cpu_offload": true
  },
}

which enables cpu_offload and some other important features. You may experiment with the buffer sizes, you will find more details in the discussion below.

For a practical usage example of this type of deployment, please, see this post.

Notes:

  • if you need to run on a specific GPU, which is different from GPU 0, you can’t use CUDA_VISIBLE_DEVICES to limit the visible scope of available GPUs. Instead, you have to use the following syntax:

    deepspeed --include localhost:1 examples/seq2seq/run_translation.py ...
    

    In this example, we tell DeepSpeed to use GPU 1 (second gpu).

Deployment in Notebooks

The problem with running notebook cells as a script is that there is no normal deepspeed launcher to rely on, so under certain setups we have to emulate it.

If you’re using only 1 GPU, here is how you’d have to adjust your training code in the notebook to use DeepSpeed.

# DeepSpeed requires a distributed environment even when only one process is used.
# This emulates a launcher in the notebook
import os
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '9994' # modify if RuntimeError: Address already in use
os.environ['RANK'] = "0"
os.environ['LOCAL_RANK'] = "0"
os.environ['WORLD_SIZE'] = "1"

# Now proceed as normal, plus pass the deepspeed config file
training_args = TrainingArguments(..., deepspeed="ds_config.json")
trainer = Trainer(...)
trainer.train()

Note: ... stands for the normal arguments that you’d pass to the functions.

If you want to use more than 1 GPU, you must use a multi-process environment for DeepSpeed to work. That is, you have to use the launcher for that purpose and this cannot be accomplished by emulating the distributed environment presented at the beginning of this section.

If you want to create the config file on the fly in the notebook in the current directory, you could have a dedicated cell with:

%%bash
cat <<'EOT' > ds_config.json
{
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

    "zero_optimization": {
        "stage": 2,
        "allgather_partitions": true,
        "allgather_bucket_size": 2e8,
        "overlap_comm": true,
        "reduce_scatter": true,
        "reduce_bucket_size": 2e8,
        "contiguous_gradients": true,
        "cpu_offload": true
    },

    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": 3e-5,
            "betas": [0.8, 0.999],
            "eps": 1e-8,
            "weight_decay": 3e-7
        }
    },

    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": 0,
            "warmup_max_lr": 3e-5,
            "warmup_num_steps": 500
        }
    },

    "steps_per_print": 2000,
    "wall_clock_breakdown": false
}
EOT

If the training script is in a normal file and not in the notebook cells, you can launch deepspeed normally via shell from a cell. For example, to use run_translation.py you would launch it with:

!git clone https://github.com/huggingface/transformers
!cd transformers; deepspeed examples/seq2seq/run_translation.py ...

or with %%bash magic, where you can write a multi-line code for the shell program to run:

%%bash

git clone https://github.com/huggingface/transformers
cd transformers
deepspeed examples/seq2seq/run_translation.py ...

In such case you don’t need any of the code presented at the beginning of this section.

Note: %%bash magic is neat, but currently it buffers the output so you won’t see the logs until the process completes.

Configuration

For the complete guide to the DeepSpeed configuration options that can be used in its configuration file please refer to the following documentation.

You can find dozens of DeepSpeed configuration examples that address various practical needs in the DeepSpeedExamples repo:

git clone https://github.com/microsoft/DeepSpeedExamples
cd DeepSpeedExamples
find . -name '*json'

Continuing the code from above, let’s say you’re looking to configure the Lamb optimizer. So you can search through the example .json files with:

grep -i Lamb $(find . -name '*json')

Some more examples are to be found in the main repo as well.

When using DeepSpeed you always need to supply a DeepSpeed configuration file, yet some configuration parameters have to be configured via the command line. You will find the nuances in the rest of this guide.

To get an idea of what DeepSpeed configuration file looks like, here is one that activates ZeRO stage 2 features, enables FP16, uses AdamW optimizer and WarmupLR scheduler:

{
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

   "zero_optimization": {
       "stage": 2,
       "allgather_partitions": true,
       "allgather_bucket_size": 5e8,
       "overlap_comm": true,
       "reduce_scatter": true,
       "reduce_bucket_size": 5e8,
       "contiguous_gradients": true,
       "cpu_offload": true
   },

   "optimizer": {
     "type": "AdamW",
     "params": {
       "lr": 3e-5,
       "betas": [ 0.8, 0.999 ],
       "eps": 1e-8,
       "weight_decay": 3e-7
     }
   },

   "scheduler": {
     "type": "WarmupLR",
     "params": {
       "warmup_min_lr": 0,
       "warmup_max_lr": 3e-5,
       "warmup_num_steps": 500
     }
   }
}

When you execute the program, DeepSpeed will log the configuration it received from the Trainer to the console, so you can see exactly what was the final configuration passed to it.

Passing Configuration

As discussed in this document normally the DeepSpeed configuration is passed as a path to a json file, but if you’re not using the command line interface to configure the training, and instead instantiate the Trainer via TrainingArguments then for the deepspeed argument you can pass a nested dict. This allows you to create the configuration on the fly and doesn’t require you to write it to the file system before passing it to TrainingArguments.

To summarize you can do:

TrainingArguments(..., deespeed="/path/to/ds_config.json")

or:

ds_config_dict=dict(scheduler=scheduler_params, optimizer=optimizer_params)
TrainingArguments(..., deespeed=ds_config_dict)

Shared Configuration

Some configuration information is required by both the Trainer and DeepSpeed to function correctly, therefore, to prevent conflicting definitions, which could lead to hard to detect errors, we chose to configure those via the Trainer command line arguments.

Therefore, the following DeepSpeed configuration params shouldn’t be used with the Trainer:

  • train_batch_size

  • train_micro_batch_size_per_gpu

  • gradient_accumulation_steps

as these will be automatically derived from the run time environment and the following 2 command line arguments:

--per_device_train_batch_size 8 --gradient_accumulation_steps 2

which are always required to be supplied.

Of course, you will need to adjust the values in this example to your situation.

ZeRO

Zero Redundancy Optimizer (ZeRO) is the work horse of DeepSpeed. It support 3 different levels (stages) of optimization. The first one is not quite interesting for scalability purposes, therefore this document focuses on stages 2 and 3. You will find more indepth information in the DeepSpeed documentation.

The zero_optimization section of the configuration file is the most important part (docs), since that is where you define which ZeRO stages you want to enable and how to configure them. You will find the explanation for each parameter in the DeepSpeed docs.

This section has to be configured exclusively via DeepSpeed configuration - the Trainer provides no equivalent command line arguments.

Note: currently DeepSpeed doesn’t validate parameter names, so if you misspell any, it’ll use the default setting for the parameter that got misspelled. You can watch the DeepSpeed engine start up log messages to see what values it is going to use.

ZeRO-2 Config

The following is an example configuration for ZeRO stage 2:

{
    "zero_optimization": {
        "stage": 2,
        "allgather_partitions": true,
        "allgather_bucket_size": 5e8,
        "overlap_comm": true,
        "reduce_scatter": true,
        "reduce_bucket_size": 5e8,
        "contiguous_gradients": true,
        "cpu_offload": true
    }
}

Performance tuning:

  • enabling cpu_offload should reduce GPU RAM usage (it requires "stage": 2)

  • "overlap_comm": true trades off increased GPU RAM usage to lower all-reduce latency. overlap_comm uses 4.5x the allgather_bucket_size and reduce_bucket_size values. So if they are set to 5e8, this requires a 9GB footprint (5e8 x 2Bytes x 2 x 4.5). Therefore, if you have a GPU with 8GB or less RAM, to avoid getting OOM-errors you will need to reduce those parameters to about 2e8, which would require 3.6GB. You will want to do the same on larger capacity GPU as well, if you’re starting to hit OOM.

  • when reducing these buffers you’re trading communication speed to avail more GPU RAM. The smaller the buffer size, the slower the communication, and the more GPU RAM will be available to other tasks. So if a bigger batch size is important, getting a slightly slower training time could be a good trade.

ZeRO-3 Config

The following is an example configuration for ZeRO stage 3:

{
    "zero_optimization": {
        "stage": 3,
        "cpu_offload": true,
        "cpu_offload_params": true,
        "cpu_offload_use_pin_memory" : true,
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e14,
        "reduce_bucket_size": 1e6,
        "stage3_prefetch_bucket_size": 0.94e6,
        "stage3_param_persistence_threshold": 1e4,
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_fp16_weights_on_model_save": true
    }
}

Note: if you’re migrating from ZeRO-2 configuration that: allgather_partitions, allgather_bucket_size and reduce_scatter configuration parameters are not used in ZeRO-3. If you keep these they will just be ignored.

Performance tuning:

  • sub_group_size: 1e14

  • reduce_bucket_size: hidden_size*hidden_size

  • stage3_prefetch_bucket_size: 0.9 * hidden_size * hidden_size

  • stage3_param_persistence_threshold: 10 * hidden_size

  • stage3_max_live_parameters: 1e9

  • stage3_max_reuse_distance: 1e9

If hitting OOM reduce stage3_max_live_parameters and stage3_max_reuse_distance. They should have minimal impact on performance unless you are doing activation checkpointing. 1e9 would consume ~2GB. The memory is shared by stage3_max_live_parameters and stage3_max_reuse_distance, so its not additive, its just 2GB total.

stage3_max_live_parameters is the upper limit on how many full parameters you want to keep on the GPU at any given time. “reuse distance” is a metric we are using to figure out when will a parameter be used again in the future, and we use the stage3_max_reuse_distance to decide whether to throw away the parameter or to keep it. If a parameter is going to be used again in near future (less than stage3_max_reuse_distance) then we keep it to reduce communication overhead. This is super helpful when you have activation checkpointing enabled, where we do a forward recompute and backward passes a a single layer granularity and want to keep the parameter in the forward recompute till the backward

If you set reduce_bucket_size, stage3_prefetch_bucket_size and stage3_param_persistence_threshold as recommended above, they will already be fairly small so you won’t have to tune those much.

Since hidden_size varies from model to model, the Trainer will automatically set the needed value for the 3 config parameters that contain that variable (using model.config.hidden_size). Just set these values to 0 as shown below and the right configuration will be passed to DeepSpeed:

{
    "zero_optimization": {
        "stage": 3,
        "cpu_offload": true,
        "cpu_offload_params": true,
        "cpu_offload_use_pin_memory" : true,
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e14,
        "reduce_bucket_size": 0,
        "stage3_prefetch_bucket_size": 0,
        "stage3_param_persistence_threshold": 0,
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_fp16_weights_on_model_save": true
    }
}

stage3_gather_fp16_weights_on_model_save enables model fp16 weights consolidation when model gets saved. With large models and multiple GPUs this is an expensive operation both in terms of memory and speed. It’s currently required if you plan to resume the training. Watch out for future updates that will remove this limitation and make things more flexible.

ZeRO-2 vs ZeRO-3 Performance

ZeRO-3 is likely to be slower than ZeRO-2 if everything else is configured the same because the former has to gather model weights in addition to what ZeRO-2 does. If ZeRO-2 meets your needs and you don’t need to scale beyond a few GPUs then you may choose to stick to it. It’s important to understand that ZeRO-3 enables a much higher scalability capacity at a cost of speed.

It’s possible to adjust ZeRO-3 configuration to make it perform closer to ZeRO-2:

  • set stage3_param_persistence_threshold to a very large number - larger than the largest parameter, e.g., 6 * hidden_size * hidden_size. This will keep the parameters on the GPUs.

  • turn off cpu_offload_params since ZeRO-2 doesn’t have that option.

The performance will likely improve significantly with just cpu_offload_params turned off, even if you don’t change stage3_param_persistence_threshold. Of course, these changes will impact the size of the model you can train. So these help you to trade scalability for speed depending on your needs.

ZeRO-2 Example

Here is a full ZeRO-2 all-enabled configuration file ds_config_zero2.json:

{
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

    "zero_optimization": {
        "stage": 2,
        "allgather_partitions": true,
        "allgather_bucket_size": 2e8,
        "overlap_comm": true,
        "reduce_scatter": true,
        "reduce_bucket_size": 2e8,
        "contiguous_gradients": true,
        "cpu_offload": true
    },

    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": 3e-5,
            "betas": [0.8, 0.999],
            "eps": 1e-8,
            "weight_decay": 3e-7
        }
    },

    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": 0,
            "warmup_max_lr": 3e-5,
            "warmup_num_steps": 500
        }
    },

    "steps_per_print": 2000,
    "wall_clock_breakdown": false
}
ZeRO-3 Example

Here is a full ZeRO-3 all-enabled configuration file ds_config_zero3.json:

{
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

    "zero_optimization": {
        "stage": 3,
        "cpu_offload": true,
        "cpu_offload_params": true,
        "cpu_offload_use_pin_memory" : true,
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e14,
        "reduce_bucket_size": 1e6,
        "stage3_prefetch_bucket_size": 0.94e6,
        "stage3_param_persistence_threshold": 1e4,
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_fp16_weights_on_model_save": true
    },

    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": 3e-5,
            "betas": [0.8, 0.999],
            "eps": 1e-8,
            "weight_decay": 3e-7
        }
    },

    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": 0,
            "warmup_max_lr": 3e-5,
            "warmup_num_steps": 500
        }
    },

    "steps_per_print": 2000,
    "wall_clock_breakdown": false
}

Optimizer and Scheduler

As long as you don’t enable cpu_offload you can mix and match DeepSpeed and HuggingFace schedulers and optimizers, with the exception of using the combination of HuggingFace scheduler and DeepSpeed optimizer:

Combos

HF Scheduler

DS Scheduler

HF Optimizer

Yes

Yes

DS Optimizer

No

Yes

If cpu_offload is enabled you must use both DeepSpeed scheduler and DeepSpeed optimizer.

Optimizer

DeepSpeed’s main optimizers are Adam, AdamW, OneBitAdam, and Lamb. These have been thoroughly tested with ZeRO and are thus recommended to be used. It, however, can import other optimizers from torch. The full documentation is here.

If you don’t configure the optimizer entry in the configuration file, the Trainer will automatically set it to AdamW and will use the supplied values or the defaults for the following command line arguments: --learning_rate, --adam_beta1, --adam_beta2, --adam_epsilon and --weight_decay.

Here is an example of the pre-configured optimizer entry for AdamW:

{
   "optimizer": {
       "type": "AdamW",
       "params": {
         "lr": 0.001,
         "betas": [0.8, 0.999],
         "eps": 1e-8,
         "weight_decay": 3e-7
       }
     }
}

Note that the command line arguments will override the values in the configuration file. This is so that there is one definitive source of the values and to avoid hard to find errors when for example, the learning rate is set to different values in different places. Command line rules. The values that get overridden are:

  • lr with the value of --learning_rate

  • betas with the value of --adam_beta1 --adam_beta2

  • eps with the value of --adam_epsilon

  • weight_decay with the value of --weight_decay

Therefore please remember to tune the shared hyperparameters on the command line.

If you want to use another optimizer which is not listed above, you will have to add "zero_allow_untested_optimizer": true to the top level configuration.

If you want to use one of the officially supported optimizers, configure them explicitly in the configuration file, and make sure to adjust the values. e.g. if use Adam you will want weight_decay around 0.01.

Scheduler

DeepSpeed supports LRRangeTest, OneCycle, WarmupLR and WarmupDecayLR LR schedulers. The full documentation is here.

Here is where the schedulers overlap between 🤗 Transformers and DeepSpeed:

  • WarmupLR via --lr_scheduler_type constant_with_warmup

  • WarmupDecayLR via --lr_scheduler_type linear. This is also the default value for --lr_scheduler_type, therefore, if you don’t configure the scheduler this is scheduler that will get configured by default.

If you don’t configure the scheduler entry in the configuration file, the Trainer will use the values of --lr_scheduler_type, --learning_rate and --warmup_steps to configure a 🤗 Transformers version of it.

Here is an example of the pre-configured scheduler entry for WarmupLR:

{
   "scheduler": {
         "type": "WarmupLR",
         "params": {
             "warmup_min_lr": 0,
             "warmup_max_lr": 0.001,
             "warmup_num_steps": 1000
         }
     }
}

Note that the command line arguments will override the values in the configuration file. This is so that there is one definitive source of the values and to avoid hard to find errors when for example, the learning rate is set to different values in different places. Command line rules. The values that get overridden are:

  • warmup_max_lr with the value of --learning_rate

  • warmup_num_steps with the value of --warmup_steps

  • total_num_steps with either the value of --max_steps or if it is not provided, derived automatically at run time based on the environment and the size of the dataset and other command line arguments (needed for WarmupDecayLR).

Therefore please remember to tune the shared hyperparameters on the command line.

For example, for WarmupDecayLR, you can use the following entry:

{
   "scheduler": {
         "type": "WarmupDecayLR",
         "params": {
             "total_num_steps": 10,
             "last_batch_iteration": -1,
             "warmup_min_lr": 0,
             "warmup_max_lr": 0.001,
             "warmup_num_steps": 1000
         }
     }
}

and warmup_max_lr, warmup_num_steps and total_num_steps will be corrected at loading time.

Automatic Mixed Precision

You can use automatic mixed precision with either a pytorch-like AMP way or the apex-like way:

If you want to use an equivalent of the Pytorch native amp, you can either configure the fp16 entry in the configuration file, or use the following command line arguments: --fp16 --fp16_backend amp.

Here is an example of the fp16 configuration:

{
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
}

Here is the documentation.

If you want to use NVIDIA’s apex instead, you can can either configure the amp entry in the configuration file, or use the following command line arguments: --fp16 --fp16_backend apex --fp16_opt_level 01.

Here is an example of the amp configuration:

{
    "amp": {
        "enabled": true,
        "opt_level": "O1"
    }
}

Here is the documentation.

Gradient Accumulation

While normally DeepSpeed gets gradient accumulation configured with:

{
    "gradient_accumulation_steps": 3,
}

in this case, to enable gradient accumulation, pass the command line --gradient_accumulation_steps 3 argument as normal and it will get injected into the DeepSpeed configuration.

If you try to add it directly to the configuration file, you will receive an error from the Trainer - this is because this setting is needed by the Trainer too, and so this approach ensures that there is a single way of setting this value and thus avoid potential subtle errors.

Gradient Clipping

If you don’t configure the gradient_clipping entry in the configuration file, the Trainer will use the value of the --max_grad_norm command line argument to set it.

Here is an example of the gradient_clipping configuration:

{
    "gradient_clipping": 1.0,
}

Getting the model weights out

As long as you continue training and resuming using DeepSpeed you don’t need to worry about anything. DeepSpeed stores fp32 master weights in its custom checkpoint optimizer files, which are global_step*/*optim_states.pt (this is glob pattern), and are saved under the normal checkpoint.

FP16 Weights:

When a model is saved under ZeRO-2, you end up having the normal pytorch_model.bin file with the model weights, but they are only the fp16 version of the weights.

Under ZeRO-3, things are much more complicated, since the model weights are partitioned out over multiple GPUs, therefore "stage3_gather_fp16_weights_on_model_save": true is required to get the Trainer to save the fp16 version of the weights. If this setting is False pytorch_model.bin won’t be created. This is because by default DeepSpeed’s state_dict contains a placeholder and not the real weights. If we were to save this state_dict it won’t be possible to load it back.

FP32 Weights:

While the fp16 weights are fine for resuming training, if you finished finetuning your model and want to upload it to the models hub or pass it to someone else you most likely will want to get the fp32 weights. This cannot be done during training since this is a process that requires a lot of memory, and therefore this is performed offline.

DeepSpeed creates a special conversion script zero_to_fp32.py which it places in the top-level of the checkpoint folder. Using this script you can extract the weights at any point. The script is standalone and you no longer need to have the configuration file or a Trainer to do the extraction.

Let’s say your checkpoint folder looks like this:

$ ls -l output_dir/checkpoint-1/
-rw-rw-r-- 1 stas stas 1.4K Mar 27 20:42 config.json
drwxrwxr-x 2 stas stas 4.0K Mar 25 19:52 global_step1/
-rw-rw-r-- 1 stas stas   12 Mar 27 13:16 latest
-rw-rw-r-- 1 stas stas 827K Mar 27 20:42 optimizer.pt
-rw-rw-r-- 1 stas stas 231M Mar 27 20:42 pytorch_model.bin
-rw-rw-r-- 1 stas stas  623 Mar 27 20:42 scheduler.pt
-rw-rw-r-- 1 stas stas 1.8K Mar 27 20:42 special_tokens_map.json
-rw-rw-r-- 1 stas stas 774K Mar 27 20:42 spiece.model
-rw-rw-r-- 1 stas stas 1.9K Mar 27 20:42 tokenizer_config.json
-rw-rw-r-- 1 stas stas  339 Mar 27 20:42 trainer_state.json
-rw-rw-r-- 1 stas stas 2.3K Mar 27 20:42 training_args.bin
-rwxrw-r-- 1 stas stas 5.5K Mar 27 13:16 zero_to_fp32.py*

In this example there is just one DeepSpeed checkpoint sub-folder global_step1. Therefore to reconstruct the fp32 weights just run:

python zero_to_fp32.py global_step1 pytorch_model.bin

The script will automatically handle either ZeRO-2 or ZeRO-3 checkpoint.

python zero_to_fp32.py -h will give you usage details.

If you have multiple DeepSpeed checkpoint sub-folders, pick the one you know to have the desired weights.

This is it. pytorch_model.bin will now contain the full fp32 model weights consolidated from multiple GPUs.

Note: currently the script requires 2x general RAM of the final fp32 model weights.

ZeRO 3 Nuances

ZeRO 3 is quite different from ZeRO 2 because of its param sharding feature.

While all the efforts were made for things to just work without needing any special changes to your models, in certain circumstances you may find the following information to be needed.

Registering External Parameters

If layer A needs to access weights belonging to layer B, currently layer A needs to tell DeepSpeed about it. This is done with the help of deepspeed.zero.register_external_parameter that needs to be called in A.__init__ and can be seen in the following example:

class ModuleZ3(torch.nn.Module):
    def __init__(self, *args):
        super().__init__(self, *args)
        self.layer1 = SomeLayer()
        self.layer2 = OtherLayer()
        deepspeed.zero.register_external_parameter(self, self.layer1.weight)

    def forward(self, input):
        x = self.layer1(input)
        # self.layer1.weight is needed in ModuleZ3.forward
        y = self.layer2(x, self.layer1.weight)
        return y

In general transformers models don’t use this style of referring to other layer’s weights so most likely you won’t need to use it.

For full details on this method please refer to Registering External Parameters.

Constructing Massive Models

DeepSpeed/ZeRO-3 can handle models with Trillions of parameters which may not fit onto the existing RAM. In such cases, but also if you want the initialization to happen much faster, initialize the model using deepspeed.zero.Init() context manager (which is also a function decorator), like so:

from transformers import T5ForConditionalGeneration, T5Config
import deepspeed
with deepspeed.zero.Init():
   config = T5Config.from_pretrained("t5-small")
   model = T5ForConditionalGeneration(config)

As you can see this gives you a randomly initialized model.

If you want to use a pretrained model, model_class.from_pretrained will activate this feature as long as is_deepspeed_zero3_enabled() returns True, which can be set manually via deepspeed_zero3_enable(True). Therefore to enable this feature here is the required sequence:

from transformers.integrations import deepspeed_zero3_enable
deepspeed_zero3_enable(True)
model = T5ForConditionalGeneration.from_pretrained("t5-small")

If you’re using Trainer command line arguments which include --deepspeed ds_config.json with ZeRO-3 config enabled, then you can skip deepspeed_zero3_enable(True) as it will try to discover whether it’ll be run under ZeRO-3 and from_pretrained will automatically activate this feature.

Note: If the fp16 weights of the model can’t fit onto the memory of a single GPU this feature must be used.

For full details on this method and other related features please refer to Constructing Massive Models.

Gathering Parameters

Under ZeRO-3 on multiple GPUs no single GPU has all the parameters unless it’s the parameters for the currently executing layer. So if you need to access all parameters from all layers at once there is a specific method to do it. Most likely you won’t need it, but if you do please refer to Gathering Parameters

We do however use it internally in several places, one such example is when loading pretrained model weights in from_pretrained. We load one layer at a time and immediately partition it to all participating GPUs, as for very large models it won’t be possible to load it on one GPU and then spread it out to multiple GPUs, due to memory limitations.

Also under ZeRO-3, if you write your own code and run into a model parameter weight that looks like:

tensor([1.], device='cuda:0', dtype=torch.float16, requires_grad=True)

stress on tensor([1.]), or if you get an error where it says the parameter is of size 1, instead of some much larger multi-dimensional shape, this means that the parameter is partitioned and what you see is a ZeRO-3 placeholder.

Notes

  • DeepSpeed works with the PyTorch Trainer but not TF TFTrainer.

  • While DeepSpeed has a pip installable PyPI package, it is highly recommended that it gets installed from source to best match your hardware and also if you need to enable certain features, like 1-bit Adam, which aren’t available in the pypi distribution.

  • You don’t have to use the Trainer to use DeepSpeed with 🤗 Transformers - you can use any model with your own trainer, and you will have to adapt the latter according to the DeepSpeed integration instructions.

Main DeepSpeed Resources

Papers:

Finally, please, remember that, HuggingFace Trainer only integrates DeepSpeed, therefore if you have any problems or questions with regards to DeepSpeed usage, please, file an issue with DeepSpeed GitHub.