ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Nano: update train pytorch key feature doc (#6906)

Browse source 
* update train doc

* update train page
This commit is contained in:
Junwei Deng 2022-12-19 15:51:19 +08:00 committed by GitHub
parent 217e8a1c2b
commit dea7b8f9f2
1 changed files with 180 additions and 59 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

239
docs/readthedocs/source/doc/Nano/Overview/pytorch_train.md
View file

@ -1,66 +1,47 @@
# PyTorch Training
BigDL-Nano can be used to accelerate PyTorch or PyTorch-Lightning applications on training workloads. The optimizations in BigDL-Nano are delivered through an extended version of PyTorch-Lightning `Trainer`. These optimizations are either enabled by default or can be easily turned on by setting a parameter or calling a method.
## Overview
BigDL-Nano can be used to accelerate **PyTorch** or **PyTorch-Lightning** applications on training workloads. These optimizations are either enabled by default or can be easily turned on by setting a parameter or calling a method.
We will briefly describe here the major features in BigDL-Nano for PyTorch training. You can find complete examples [here](https://github.com/intel-analytics/BigDL/tree/main/python/nano/notebooks/pytorch).
The optimizations in BigDL-Nano are delivered through
### Best Known Configurations
1) An extended version of PyTorch-Lightning `Trainer` for LightingModule and easy nn.Module.
When you run `source bigdl-nano-init`, BigDL-Nano will export a few environment variables, such as `OMP_NUM_THREADS` and `KMP_AFFINITY`, according to your current hardware. Empirically, these environment variables work best for most PyTorch applications. After setting these environment variables, you can just run your applications as usual (`python app.py`) and no additional changes are required.
2) An abstract `TorchNano` to accelerate raw or complex nn.Module.
### BigDL-Nano PyTorch Trainer
We will briefly describe here the major features in BigDL-Nano for PyTorch training. You can find complete how to guides for acceleration of [pytorch-lightning](https://bigdl.readthedocs.io/en/latest/doc/Nano/Howto/index.html#pytorch-lightning) and [pytorch]().
The PyTorch Trainer (`bigdl.nano.pytorch.Trainer`) is the place where we integrate most optimizations. It extends PyTorch Lightning's Trainer and has a few more parameters and methods specific to BigDL-Nano. The Trainer can be directly used to train a `LightningModule`.
For example,
```python
from pytorch_lightning import LightningModule
from bigdl.nano.pytorch import Trainer
class MyModule(LightningModule):
# LightningModule definition
from bigdl.nano.pytorch import Trainer
lightning_module = MyModule()
trainer = Trainer(max_epoch=10)
trainer.fit(lightning_module, train_loader)
## Best Known Environment Variables
When you successfully installed `bigdl-nano` (please refer to [installation guide](./install.html)) in a conda environment. You are **highly recommeneded** to run following command **once**.
```bash
source bigdl-nano-init
```
BigDL-Nano will export a few environment variables, such as `OMP_NUM_THREADS` and `KMP_AFFINITY`, according to your current hardware. Empirically, these environment variables work best for most PyTorch applications. After setting these environment variables, you can just run your applications as usual (e.g., `python app.py` or `jupyter notebook`).
For regular PyTorch modules, we also provide a "compile" method, that takes in a PyTorch module, an optimizer, and other PyTorch objects and "compiles" them into a `LightningModule`.
## Accelerate `nn.Module`'s training
`nn.Module` is the abstraction used in PyTorch for AI Model. It's common that users' model is easy enough to be handled by a regular training loop. In other cases, users may have highly customized training loop. Nano could support the acceleration for both cases.
### `nn.Module` with regular training loop
Most of the AI model defined in `nn.Module` could be trained in a similar regular training loop. Any `nn.Module` that
- Have only one output
- Need only 1 loss function and 1 optimizer (e.g., GAN might not applied)
- Have no special customized checkpoint/evaluation logic
could use `Trainer.compile` that takes in a PyTorch module, a loss, an optimizer, and other PyTorch objects and "compiles" them into a `LightningModule`. And then a `Trainer` instance could be used to train this compiled model.
For example,
```python
from bigdl.nano.pytorch import Trainer
lightning_module = Trainer.compile(pytorch_module, loss, optimizer, scheduler)
lightning_module = Trainer.compile(pytorch_module, loss, optimizer)
trainer = Trainer(max_epoch=10)
trainer.fit(lightning_module, train_loader)
```
#### Intel® Extension for PyTorch
`trainer.fit` will apply all the acceleration methods that could generally be applied to any models. While there are some optional acceleration method for which you could easily enable.
[Intel Extension for PyTorch](https://github.com/intel/intel-extension-for-pytorch) (a.k.a. IPEX) extends PyTorch with optimizations for an extra performance boost on Intel hardware. BigDL-Nano integrates IPEX through the `Trainer`. Users can turn on IPEX by setting `use_ipex=True`.
```python
from bigdl.nano.pytorch import Trainer
trainer = Trainer(max_epoch=10, use_ipex=True)
```
#### Multi-instance Training
When training on a server with dozens of CPU cores, it is often beneficial to use multiple training instances in a data-parallel fashion to make full use of the CPU cores. However, using PyTorch's DDP API is a little cumbersome and error-prone, and if not configured correctly, it will make the training even slow.
BigDL-Nano makes it very easy to conduct multi-instance training. You can just set the `num_processes` parameter in the `Trainer` constructor and BigDL-Nano will launch the specific number of processes to perform data-parallel training. Each process will be automatically pinned to a different subset of CPU cores to avoid conflict and maximize training throughput.
```python
from bigdl.nano.pytorch import Trainer
trainer = Trainer(max_epoch=10, num_processes=4)
```
Note that the effective batch size in multi-instance training is the `batch_size` in your `dataloader` times `num_processes` so the number of iterations of each epoch will be reduced `num_processes` fold. A common practice to compensate for that is to gradually increase the learning rate to `num_processes` times. You can find more details of this trick in this [paper](https://arxiv.org/abs/1706.02677) published by Facebook.
### BigDL-Nano PyTorch TorchNano
### `nn.Module` with customized training loop
The `TorchNano` (`bigdl.nano.pytorch.TorchNano`) class is what we use to accelerate raw pytorch code. By using it, we only need to make very few changes to accelerate custom training loop. For example,
@ -74,29 +55,170 @@ class MyNano(TorchNano) :
MyNano().train(...)
```
Our `TorchNano` also integrates IPEX and distributed training optimizations. For example,
## Accelerate `LightningModule`'s training
The PyTorch Trainer (`bigdl.nano.pytorch.Trainer`) extends PyTorch Lightning's Trainer and has a few more parameters and methods specific to BigDL-Nano. The Trainer can be directly used to train a `LightningModule`.
For example,
```python
from bigdl.nano.pytorch import TorchNano
from pytorch_lightning import LightningModule
from bigdl.nano.pytorch import Trainer
class MyNano(TorchNano):
def train(self, ...):
# define train loop
class MyModule(LightningModule):
# LightningModule definition
# enable IPEX optimizaiton
MyNano(use_ipex=True).train(...)
# enable IPEX and distributed training, using 'subprocess' backend
MyNano(use_ipex=True, num_processes=2, distributed_backend="subprocess").train(...)
lightning_module = MyModule()
trainer = Trainer(max_epoch=10)
trainer.fit(lightning_module, train_loader)
```
### Optimized Data Pipeline
## Optional Acceleration Methods
### Intel® Extension for PyTorch
[Intel Extension for PyTorch](https://github.com/intel/intel-extension-for-pytorch) (a.k.a. IPEX) extends PyTorch with optimizations for an extra performance boost on Intel hardware.
BigDL-Nano integrates IPEX in `Trainer` and `TorchNano`. Users can turn on IPEX by setting `use_ipex=True`.
```eval_rst
.. tabs::
.. tab:: Trainer
.. code-block:: python
from bigdl.nano.pytorch import Trainer
trainer = Trainer(max_epoch=10, use_ipex=True)
trainer.fit(...)
.. tab:: TorchNano
.. code-block:: python
from bigdl.nano.pytorch import TorchNano
class MyNano(TorchNano) :
def train(self, ...):
# copy your train loop here and make a few changes
MyNano(use_ipex=True).train(...)
```
### Multi-instance Training
When training on a server with dozens of CPU cores, it is often beneficial to use multiple training instances in a data-parallel fashion to make full use of the CPU cores. However, using PyTorch's DDP API is a little cumbersome and error-prone, and if not configured correctly, it will make the training even slow.
You can just set the `num_processes` parameter in the `Trainer` or `TorchNano` constructor and BigDL-Nano will launch the specific number of processes to perform data-parallel training. Each process will be automatically pinned to a different subset of CPU cores to avoid conflict and maximize training throughput.
```eval_rst
.. tabs::
.. tab:: Trainer
.. code-block:: python
from bigdl.nano.pytorch import Trainer
trainer = Trainer(max_epoch=10, num_processes=4)
trainer.fit(...)
.. tab:: TorchNano
.. code-block:: python
from bigdl.nano.pytorch import TorchNano
class MyNano(TorchNano) :
def train(self, ...):
# copy your train loop here and make a few changes
MyNano(num_processes=4).train(...)
```
Note that the effective batch size in multi-instance training is the `batch_size` in your `dataloader` times `num_processes` so the number of iterations of each epoch will be reduced `num_processes` fold. A common practice to compensate for that is to gradually increase the learning rate to `num_processes` times. You can find more details of this trick in this [paper](https://arxiv.org/abs/1706.02677) published by Facebook.
### BFloat16 Mixed Precision
BFloat16 Mixed Precison combines BFloat16 and FP32 during training, which could lead to increased performance and reduced memory usage. Compared to FP16 mixed precison, BFloat16 mixed precision has better numerical stability.
You could instantiate a BigDL-Nano `Trainer` or `TorchNano` with `precision='bf16'` to use BFloat16 mixed precision for training.
```eval_rst
.. tabs::
.. tab:: Trainer
.. code-block:: python
from bigdl.nano.pytorch import Trainer
trainer = Trainer(max_epochs=5, precision='bf16')
trainer.fit(...)
.. tab:: TorchNano
.. code-block:: python
from bigdl.nano.pytorch import TorchNano
class MyNano(TorchNano) :
def train(self, ...):
# copy your train loop here and make a few changes
MyNano(precision='bf16').train(...)
```
### Channels Last Memory Format
You could instantiate a BigDL-Nano `Trainer` or `TorchNano` with `channels_last=True` to use the channels last memory format, i.e. NHWC (batch size, height, width, channels), as an alternative way to store tensors in classic/contiguous NCHW order.
```eval_rst
.. tabs::
.. tab:: Trainer
.. code-block:: python
from bigdl.nano.pytorch import Trainer
trainer = Trainer(max_epochs=5, channels_last=True)
trainer.fit(...)
.. tab:: TorchNano
.. code-block:: python
from bigdl.nano.pytorch import TorchNano
class MyNano(TorchNano) :
def train(self, ...):
# copy your train loop here and make a few changes
MyNano(channels_last=True).train(...)
```
## Accelerate `torchvision` data processing
Computer Vision task often needs a data processing pipeline that sometimes constitutes a non-trivial part of the whole training pipeline.
Leveraging OpenCV and libjpeg-turbo, BigDL-Nano can accelerate computer vision data pipelines by providing a drop-in replacement of `torchvision`'s components such as `datasets` and `transforms`. Nano provides a patch API `patch_torch` to accelerate these functions.
Computer Vision task often needs a data processing pipeline that sometimes constitutes a non-trivial part of the whole training pipeline. Leveraging OpenCV and libjpeg-turbo, BigDL-Nano can accelerate computer vision data pipelines by providing a drop-in replacement of torch_vision's `datasets` and `transforms`.
```python
from bigdl.nano.pytorch.vision.datasets import ImageFolder
from bigdl.nano.pytorch.vision import transforms
from bigdl.nano.pytorch import patch_torch
patch_torch()
from torchvision.datasets import ImageFolder
from torchvision import transforms
data_transform = transforms.Compose([
transforms.Resize(256),
@ -109,5 +231,4 @@ data_transform = transforms.Compose([
train_set = ImageFolder(train_path, data_transform)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
trainer.fit(module, train_loader)
```