PyTorch

PyTorch project is a Python package that provides GPU accelerated tensor computation and high level functionalities for building deep learning networks. For licensing details, see the PyTorch license doc on GitHub.

To monitor and debug your PyTorch models, consider using TensorBoard.

PyTorch is included in Databricks Runtime for Machine Learning. If you are using Databricks Runtime, see Install PyTorch for instructions on installing PyTorch.

Single node and distributed training

To test and migrate single-machine workflows, use a Single Node cluster.

For distributed training options for deep learning, see Distributed training.

Example notebook

PyTorch notebook

Get notebook

Install PyTorch

Databricks Runtime for ML

Databricks Runtime for Machine Learning includes PyTorch so you can create the cluster and start using PyTorch. For the version of PyTorch installed in the Databricks Runtime ML version you are using, see the release notes.

Databricks Runtime

Databricks recommends that you use the PyTorch included in Databricks Runtime for Machine Learning. However, if you must use the standard Databricks Runtime, PyTorch can be installed as a Databricks PyPI library. The following example shows how to install PyTorch 1.5.0:

• On GPU clusters, install pytorch and torchvision by specifying the following:

  • torch==1.5.0
  • torchvision==0.6.0

• On CPU clusters, install pytorch and torchvision by using the following Python wheel files:

  https://download.pytorch.org/whl/cpu/torch-1.5.0%2Bcpu-cp37-cp37m-linux_x86_64.whl
  
  https://download.pytorch.org/whl/cpu/torchvision-0.6.0%2Bcpu-cp37-cp37m-linux_x86_64.whl
  

Errors and troubleshooting for distributed PyTorch

The following sections describe common error messages and troubleshooting guidance for the classes: PyTorch DataParallel or PyTorch DistributedDataParallel. Most of these errors can likely be resolved with TorchDistributor, which is available on Databricks Runtime ML 13.0 and above. However, if TorchDistributor is not a viable solution, recommended solutions are also provided within each section.

The following is an example of how to use TorchDistributor:

from pyspark.ml.torch.distributor import TorchDistributor

def train_fn(learning_rate):
        # ...

num_processes=2
distributor = TorchDistributor(num_processes=num_processes, local_mode=True)

distributor.run(train_fn, 1e-3)

"process 0 terminated with exit code 1"

This error occurs when using notebooks, regardless of environment: Databricks, local machine, etc. To avoid this error, use torch.multiprocessing.start_processes with start_method=fork instead of torch.multiprocessing.spawn.

For example:

import torch

def train_fn(rank, learning_rate):
    # required setup, e.g. setup(rank)
        # ...

num_processes = 2
torch.multiprocessing.start_processes(train_fn, args=(1e-3,), nprocs=num_processes, start_method="fork")

"The server socket has failed to bind to [::]: {PORTNUMBER} (errno: 98 - Address already in use)."

This error appears when you restart the distributed training after interrupting the cell during training.

To fix the issue, restart the cluster. If restarting does not solve the problem, there might be an error in the training function code.

CUDA related errors

You can run into additional issues with CUDA since start_method="fork" is not CUDA-compatible. Using any .cuda commands in any cell might lead to failures. To avoid these errors, add the following check before you call torch.multiprocessing.start_method:

if torch.cuda.is_initialized():
    raise Exception("CUDA was initialized; distributed training will fail.") # or something similar

• Train a PyTorch model