👟 Trainer

An opinionated general purpose model trainer on PyTorch with a simple code base.

Installation

From Github:

git clone https://github.com/coqui-ai/Trainer
cd Trainer
make install

From PyPI:

pip install trainer

Prefer installing from Github as it is more stable.

Implementing a model

Subclass and overload the functions in the TrainerModel()

Training a model with auto-optimization

See the MNIST example.

Training a model with advanced optimization

With 👟 you can define the whole optimization cycle as you want as the in GAN example below. It enables more under-the-hood control and flexibility for more advanced training loops.

You just have to use the scaled_backward() function to handle mixed precision training.

...

def optimize(self, batch, trainer):
    imgs, _ = batch

    # sample noise
    z = torch.randn(imgs.shape[0], 100)
    z = z.type_as(imgs)

    # train discriminator
    imgs_gen = self.generator(z)
    logits = self.discriminator(imgs_gen.detach())
    fake = torch.zeros(imgs.size(0), 1)
    fake = fake.type_as(imgs)
    loss_fake = trainer.criterion(logits, fake)

    valid = torch.ones(imgs.size(0), 1)
    valid = valid.type_as(imgs)
    logits = self.discriminator(imgs)
    loss_real = trainer.criterion(logits, valid)
    loss_disc = (loss_real + loss_fake) / 2

    # step dicriminator
    _, _ = self.scaled_backward(loss_disc, None, trainer, trainer.optimizer[0])

    if trainer.total_steps_done % trainer.grad_accum_steps == 0:
        trainer.optimizer[0].step()
        trainer.optimizer[0].zero_grad()

    # train generator
    imgs_gen = self.generator(z)

    valid = torch.ones(imgs.size(0), 1)
    valid = valid.type_as(imgs)

    logits = self.discriminator(imgs_gen)
    loss_gen = trainer.criterion(logits, valid)

    # step generator
    _, _ = self.scaled_backward(loss_gen, None, trainer, trainer.optimizer[1])
    if trainer.total_steps_done % trainer.grad_accum_steps == 0:
        trainer.optimizer[1].step()
        trainer.optimizer[1].zero_grad()
    return {"model_outputs": logits}, {"loss_gen": loss_gen, "loss_disc": loss_disc}

...

See the GAN training example with Gradient Accumulation

Training with Batch Size Finder

see the test script here for training with batch size finder.

The batch size finder starts at a default BS(defaults to 2048 but can also be user defined) and searches for the largest batch size that can fit on your hardware. you should expect for it to run multiple trainings until it finds it. to use it instead of calling trainer.fit() youll call trainer.fit_with_largest_batch_size(starting_batch_size=2048) with starting_batch_size being the batch the size you want to start the search with. very useful if you are wanting to use as much gpu mem as possible.

Training with DDP

$ python -m trainer.distribute --script path/to/your/train.py --gpus "0,1"

We don't use .spawn() to initiate multi-gpu training since it causes certain limitations.

• Everything must the pickable.
• .spawn() trains the model in subprocesses and the model in the main process is not updated.
• DataLoader with N processes gets really slow when the N is large.

Training with Accelerate

Setting use_accelerate in TrainingArgs to True will enable training with Accelerate.

You can also use it for multi-gpu or distributed training.

CUDA_VISIBLE_DEVICES="0,1,2" accelerate launch --multi_gpu --num_processes 3 train_recipe_autoregressive_prompt.py

See the Accelerate docs.

Adding a callback

👟 Supports callbacks to customize your runs. You can either set callbacks in your model implementations or give them explicitly to the Trainer.

Please check trainer.utils.callbacks to see available callbacks.

Here is how you provide an explicit call back to a 👟Trainer object for weight reinitialization.

def my_callback(trainer):
    print(" > My callback was called.")

trainer = Trainer(..., callbacks={"on_init_end": my_callback})
trainer.fit()

Profiling example

• Create the torch profiler as you like and pass it to the trainer.

  import torch
  profiler = torch.profiler.profile(
      activities=[
          torch.profiler.ProfilerActivity.CPU,
          torch.profiler.ProfilerActivity.CUDA,
      ],
      schedule=torch.profiler.schedule(wait=1, warmup=1, active=3, repeat=2),
      on_trace_ready=torch.profiler.tensorboard_trace_handler("./profiler/"),
      record_shapes=True,
      profile_memory=True,
      with_stack=True,
  )
  prof = trainer.profile_fit(profiler, epochs=1, small_run=64)
  then run Tensorboard

• Run the tensorboard.

  tensorboard --logdir="./profiler/"

Supported Experiment Loggers

• Tensorboard - actively maintained
• ClearML - actively maintained
• MLFlow
• Aim
• WandDB

To add a new logger, you must subclass BaseDashboardLogger and overload its functions.

Anonymized Telemetry

We constantly seek to improve 🐸 for the community. To understand the community's needs better and address them accordingly, we collect stripped-down anonymized usage stats when you run the trainer.

Of course, if you don't want, you can opt out by setting the environment variable TRAINER_TELEMETRY=0.