Trainer

The Trainer is a complete training and evaluation loop for PyTorch models implemented in the Transformers library. You only need to pass it the necessary pieces for training (model, tokenizer, dataset, evaluation function, training hyperparameters, etc.), and the Trainer class takes care of the rest. This makes it easier to start training faster without manually writing your own training loop. But at the same time, Trainer is very customizable and offers a ton of training options so you can tailor it to your exact training needs.

Before you start, make sure Accelerate - a library for enabling and running PyTorch training across distributed environments - is installed.

pip install accelerate

# upgrade
pip install accelerate --upgrade

This guide provides an overview of the Trainer class.

Basic usage

Trainer includes all the code you’ll find in a basic training loop:

1. perform a training step to calculate the loss
2. calculate the gradients with the backward method
3. update the weights based on the gradients
4. repeat this process until you’ve reached a predetermined number of epochs

The Trainer class abstracts all of this code away so you don’t have to worry about manually writing a training loop every time or if you’re just getting started with PyTorch and training. You only need to provide the essential components required for training, such as a model and a dataset, and the Trainer class handles everything else.

If you want to specify any training options or hyperparameters, you can find them in the TrainingArguments class. For example, let’s define where to save the model in output_dir and push the model to the Hub after training with push_to_hub=True.

from transformers import TrainingArguments

training_args = TrainingArguments(
    output_dir="your-model",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=2,
    weight_decay=0.01,
    eval_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
    push_to_hub=True,
)

Pass training_args to the Trainer along with a model, dataset, something to preprocess the dataset with (depending on your data type it could be a tokenizer, feature extractor or image processor), a data collator, and a function to compute the metrics you want to track during training.

Finally, call train() to start training!

from transformers import Trainer

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],
    tokenizer=tokenizer,
    data_collator=data_collator,
    compute_metrics=compute_metrics,
)

trainer.train()

Checkpoints

The Trainer class saves your model checkpoints to the directory specified in the output_dir parameter of TrainingArguments. You’ll find the checkpoints saved in a checkpoint-000 subfolder where the numbers at the end correspond to the training step. Saving checkpoints are useful for resuming training later.

# resume from latest checkpoint
trainer.train(resume_from_checkpoint=True)

# resume from specific checkpoint saved in output directory
trainer.train(resume_from_checkpoint="your-model/checkpoint-1000")

You can save your checkpoints (the optimizer state is not saved by default) to the Hub by setting push_to_hub=True in TrainingArguments to commit and push them. Other options for deciding how your checkpoints are saved are set up in the hub_strategy parameter:

• hub_strategy="checkpoint" pushes the latest checkpoint to a subfolder named “last-checkpoint” from which you can resume training
• hub_strategy="all_checkpoints" pushes all checkpoints to the directory defined in output_dir (you’ll see one checkpoint per folder in your model repository)

When you resume training from a checkpoint, the Trainer tries to keep the Python, NumPy, and PyTorch RNG states the same as they were when the checkpoint was saved. But because PyTorch has various non-deterministic default settings, the RNG states aren’t guaranteed to be the same. If you want to enable full determinism, take a look at the Controlling sources of randomness guide to learn what you can enable to make your training fully deterministic. Keep in mind though that by making certain settings deterministic, training may be slower.

Customize the Trainer

While the Trainer class is designed to be accessible and easy-to-use, it also offers a lot of customizability for more adventurous users. Many of the Trainer’s method can be subclassed and overridden to support the functionality you want, without having to rewrite the entire training loop from scratch to accommodate it. These methods include:

• get_train_dataloader() creates a training DataLoader
• get_eval_dataloader() creates an evaluation DataLoader
• get_test_dataloader() creates a test DataLoader
• log() logs information on the various objects that watch training
• create_optimizer_and_scheduler() creates an optimizer and learning rate scheduler if they weren’t passed in the __init__; these can also be separately customized with create_optimizer() and create_scheduler() respectively
• compute_loss() computes the loss on a batch of training inputs
• training_step() performs the training step
• prediction_step() performs the prediction and test step
• evaluate() evaluates the model and returns the evaluation metrics
• predict() makes predictions (with metrics if labels are available) on the test set

For example, if you want to customize the compute_loss() method to use a weighted loss instead.

from torch import nn
from transformers import Trainer

class CustomTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False):
        labels = inputs.pop("labels")
        # forward pass
        outputs = model(**inputs)
        logits = outputs.get("logits")
        # compute custom loss for 3 labels with different weights
        loss_fct = nn.CrossEntropyLoss(weight=torch.tensor([1.0, 2.0, 3.0], device=model.device))
        loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
        return (loss, outputs) if return_outputs else loss

Callbacks

Another option for customizing the Trainer is to use callbacks. Callbacks don’t change anything in the training loop. They inspect the training loop state and then execute some action (early stopping, logging results, etc.) depending on the state. In other words, a callback can’t be used to implement something like a custom loss function and you’ll need to subclass and override the compute_loss() method for that.

For example, if you want to add an early stopping callback to the training loop after 10 steps.

from transformers import TrainerCallback

class EarlyStoppingCallback(TrainerCallback):
    def __init__(self, num_steps=10):
        self.num_steps = num_steps
    
    def on_step_end(self, args, state, control, **kwargs):
        if state.global_step >= self.num_steps:
            return {"should_training_stop": True}
        else:
            return {}

Then pass it to the Trainer’s callback parameter.

from transformers import Trainer

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],
    tokenizer=tokenizer,
    data_collator=data_collator,
    compute_metrics=compute_metrics,
    callback=[EarlyStoppingCallback()],
)

Logging

The Trainer is set to logging.INFO by default which reports errors, warnings, and other basic information. A Trainer replica - in distributed environments - is set to logging.WARNING which only reports errors and warnings. You can change the logging level with the log_level and log_level_replica parameters in TrainingArguments.

To configure the log level setting for each node, use the log_on_each_node parameter to determine whether to use the log level on each node or only on the main node.

For example, to set your main code and modules to use the same log level according to each node:

logger = logging.getLogger(__name__)

logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
    datefmt="%m/%d/%Y %H:%M:%S",
    handlers=[logging.StreamHandler(sys.stdout)],
)

log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)

trainer = Trainer(...)

Use different combinations of log_level and log_level_replica to configure what gets logged on each of the nodes.

my_app.py ... --log_level warning --log_level_replica error

NEFTune

NEFTune is a technique that can improve performance by adding noise to the embedding vectors during training. To enable it in Trainer, set the neftune_noise_alpha parameter in TrainingArguments to control how much noise is added.

from transformers import TrainingArguments, Trainer

training_args = TrainingArguments(..., neftune_noise_alpha=0.1)
trainer = Trainer(..., args=training_args)

NEFTune is disabled after training to restore the original embedding layer to avoid any unexpected behavior.

GaLore

Gradient Low-Rank Projection (GaLore) is a memory-efficient low-rank training strategy that allows full-parameter learning but is more memory-efficient than common low-rank adaptation methods, such as LoRA.

First make sure to install GaLore official repository:

pip install galore-torch

Then simply add one of ["galore_adamw", "galore_adafactor", "galore_adamw_8bit"] in optim together with optim_target_modules, which can be a list of strings, regex or full path corresponding to the target module names you want to adapt. Below is an end-to-end example script (make sure to pip install trl datasets):

import torch
import datasets
import trl

from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM

train_dataset = datasets.load_dataset('imdb', split='train')

args = TrainingArguments(
    output_dir="./test-galore",
    max_steps=100,
    per_device_train_batch_size=2,
    optim="galore_adamw",
    optim_target_modules=["attn", "mlp"]
)

model_id = "google/gemma-2b"

config = AutoConfig.from_pretrained(model_id)

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)

trainer = trl.SFTTrainer(
    model=model, 
    args=args,
    train_dataset=train_dataset,
    dataset_text_field='text',
    max_seq_length=512,
)

trainer.train()

To pass extra arguments supports by GaLore, you should pass correctly optim_args, for example:

import torch
import datasets
import trl

from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM

train_dataset = datasets.load_dataset('imdb', split='train')

args = TrainingArguments(
    output_dir="./test-galore",
    max_steps=100,
    per_device_train_batch_size=2,
    optim="galore_adamw",
    optim_target_modules=["attn", "mlp"],
    optim_args="rank=64, update_proj_gap=100, scale=0.10",
)

model_id = "google/gemma-2b"

config = AutoConfig.from_pretrained(model_id)

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)

trainer = trl.SFTTrainer(
    model=model, 
    args=args,
    train_dataset=train_dataset,
    dataset_text_field='text',
    max_seq_length=512,
)

trainer.train()

You can read more about the method in the original repository or the paper.

Currently you can only train Linear layers that are considered as GaLore layers and will use low-rank decomposition to be trained while remaining layers will be optimized in the conventional manner.

Note it will take a bit of time before starting the training (~3 minutes for a 2B model on a NVIDIA A100), but training should go smoothly afterwards.

You can also perform layer-wise optimization by post-pending the optimizer name with layerwise like below:

import torch
import datasets
import trl

from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM

train_dataset = datasets.load_dataset('imdb', split='train')

args = TrainingArguments(
    output_dir="./test-galore",
    max_steps=100,
    per_device_train_batch_size=2,
    optim="galore_adamw_layerwise",
    optim_target_modules=["attn", "mlp"]
)

model_id = "google/gemma-2b"

config = AutoConfig.from_pretrained(model_id)

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)

trainer = trl.SFTTrainer(
    model=model, 
    args=args,
    train_dataset=train_dataset,
    dataset_text_field='text',
    max_seq_length=512,
)

trainer.train()

Note layerwise optimization is a bit experimental and does not support DDP (Distributed Data Parallel), thus you can run the training script only on a single GPU. Please see this appropriate section for more details. Other features such as gradient clipping, DeepSpeed, etc might not be supported out of the box. Please raise an issue on GitHub if you encounter such issue.

LOMO optimizer

The LOMO optimizers have been introduced in Full Parameter Fine-Tuning for Large Language Models with Limited Resources and AdaLomo: Low-memory Optimization with Adaptive Learning Rate. They both consist of an efficient full-parameter fine-tuning method. These optimizers fuse the gradient computation and the parameter update in one step to reduce memory usage. Supported optimizers for LOMO are "lomo" and "adalomo". First either install LOMO from pypi pip install lomo-optim or install it from source with pip install git+https://github.com/OpenLMLab/LOMO.git.

Below is a simple script to demonstrate how to fine-tune google/gemma-2b on IMDB dataset in full precision:

import torch
import datasets
from transformers import TrainingArguments, AutoTokenizer, AutoModelForCausalLM
import trl

train_dataset = datasets.load_dataset('imdb', split='train')

args = TrainingArguments(
    output_dir="./test-lomo",
    max_steps=1000,
    per_device_train_batch_size=4,
    optim="adalomo",
    gradient_checkpointing=True,
    logging_strategy="steps",
    logging_steps=1,
    learning_rate=2e-6,
    save_strategy="no",
    run_name="lomo-imdb",
)

model_id = "google/gemma-2b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True).to(0)

trainer = trl.SFTTrainer(
    model=model, 
    args=args,
    train_dataset=train_dataset,
    dataset_text_field='text',
    max_seq_length=1024,
)

trainer.train()

Accelerate and Trainer

The Trainer class is powered by Accelerate, a library for easily training PyTorch models in distributed environments with support for integrations such as FullyShardedDataParallel (FSDP) and DeepSpeed.

To use Accelerate with Trainer, run the accelerate.config command to set up training for your training environment. This command creates a config_file.yaml that’ll be used when you launch your training script. For example, some example configurations you can setup are:

compute_environment: LOCAL_MACHINE                                                                                             
distributed_type: MULTI_GPU                                                                                                    
downcast_bf16: 'no'
gpu_ids: all
machine_rank: 0 #change rank as per the node
main_process_ip: 192.168.20.1
main_process_port: 9898
main_training_function: main
mixed_precision: fp16
num_machines: 2
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

The accelerate_launch command is the recommended way to launch your training script on a distributed system with Accelerate and Trainer with the parameters specified in config_file.yaml. This file is saved to the Accelerate cache folder and automatically loaded when you run accelerate_launch.

For example, to run the run_glue.py training script with the FSDP configuration:

accelerate launch \
    ./examples/pytorch/text-classification/run_glue.py \
    --model_name_or_path google-bert/bert-base-cased \
    --task_name $TASK_NAME \
    --do_train \
    --do_eval \
    --max_seq_length 128 \
    --per_device_train_batch_size 16 \
    --learning_rate 5e-5 \
    --num_train_epochs 3 \
    --output_dir /tmp/$TASK_NAME/ \
    --overwrite_output_dir

You could also specify the parameters from the config_file.yaml file directly in the command line:

accelerate launch --num_processes=2 \
    --use_fsdp \
    --mixed_precision=bf16 \
    --fsdp_auto_wrap_policy=TRANSFORMER_BASED_WRAP  \
    --fsdp_transformer_layer_cls_to_wrap="BertLayer" \
    --fsdp_sharding_strategy=1 \
    --fsdp_state_dict_type=FULL_STATE_DICT \
    ./examples/pytorch/text-classification/run_glue.py
    --model_name_or_path google-bert/bert-base-cased \
    --task_name $TASK_NAME \
    --do_train \
    --do_eval \
    --max_seq_length 128 \
    --per_device_train_batch_size 16 \
    --learning_rate 5e-5 \
    --num_train_epochs 3 \
    --output_dir /tmp/$TASK_NAME/ \
    --overwrite_output_dir

Check out the Launching your Accelerate scripts tutorial to learn more about accelerate_launch and custom configurations.