Run Effective Large Batch Contrastive Learning on Limited Memory GPU

Overview

Gradient Cache

Gradient Cache is a simple technique for unlimitedly scaling contrastive learning batch far beyond GPU memory constraint. This means training that used to take heavy hardware, e.g. 8 V100 GPU, can be done on a single GPU. In addition, Gradient Cache allow users to replace big RAM GPU with much more cost efficient high FLOP low RAM cards.

This repo holds a generic Pytorch implementation of Gradient Cache described in our paper Scaling Deep Contrastive Learning Batch Size under Memory Limited Setup .

@inproceedings{gao2021scaling,
     title={Scaling Deep Contrastive Learning Batch Size under Memory Limited Setup},
     author={Luyu Gao, Yunyi Zhang, Jiawei Han, Jamie Callan},
     booktitle ={Proceedings of the 6th Workshop on Representation Learning for NLP},
     year={2021},
}

Gradient Cache has also been integrated into dense passage retrieval (DPR). Checkout our GC-DPR toolkit.

Installation

The package depends only on pytorch>=1.6. To install, clone this repo and run pip.

git clone https://github.com/luyug/GradCache
cd GradCache
pip install .

For development,

pip install --editable .

Usage

Gradient caching functionalities are implemented in GradCache class. If you are developing a new project instead of patching an old one, also checkout our functional approach for a effort reduced approach.

Initialization

The class's __init__ method defines the cache and has several functional parameters *_fn for easy adjust of model behaviors. Alternatively you can also sub-class GradCache.

grad_cache.GradCache(  
  models: List[nn.Module],  
  chunk_sizes: Union[int, List[int]],  
  loss_fn: Callable[..., Tensor],  
  split_input_fn: Callable[[Any, int], Any] = None,  
  get_rep_fn: Callable[..., Tensor] = None,  
  fp16: bool = False,  
  scaler: GradScaler = None,  
)

models - A list of encoder models to be updated with with the Gradient Cache.

chunk_sizes - An integer indicating chunk size. Or a list of integers of chunk size for each model. This controls for each model the sub-batch size to run forward-backward pass and should be set based on available GPU memory. A value too small will leave the GPU under utilized.

loss_fn - A loss function that takes representation tensors of number equal to number of models in models and arbitrary numbers of keyword arguments. It should compute loss based on the input tensors, and in no case modify the input tensors' relations in the autograd graph, which are later relied upon to create the gradient cache.

split_input_fn - An optional function that split generic model input into chunks based on defined chunk_sizes. If not provided, this class will try its best to split the inputs of supported types. See split_inputs function.

get_rep_fn - An optional function that takes generic model output and return representation tensors. If not provided, the generic output is assumed to be the representation tensor.

fp16 - If True, run mixed precision training, which requires scaler to also be set.

scaler - A GradScaler object for automatic mixed precision training.

Cache Gradient Step

To run a cached gradient computatoin step, call cache_step function,

cache_step(  
  *model_inputs,  
  no_sync_except_last: bool = False,  
  **loss_kwargs  
)

Run a single gradient cache step. Upon function return, updates are computed for each model in self.models with gradient populated on the weights, as if the model_inputs are run as a huge single batch on sufficiently large hardware. Calling an GradCache object with __call__ will also invoke this function.

model_inputs - List of inputs to each encoder model. Should be in similar order as self.models.

no_sync_except_last - If True, under distributed setup, for each model, only trigger gradient reduction across processes for the last sub-batch's forward-backward pass. This could come in handy when dealing with a) large model, and/or b) non trivial number of sub-batches.

loss_kwargs - Additional keyword arguments to the loss function loss_fn. This is intended to enable flexible loss computation (thanks to dynamic graph in Pytorch) such as reduction, weighting, etc. Potentially, using loss_kwargs you can incorporate outputs from those encoder models not tracked by the cache.

Return - loss, the current steps loss scaler tensor (detached from the graph).

Natively Supported Input Types

  • x: Tensor - will be passed in as model(x)
  • x: List[Tensor] - will be passed in as model(*x)
  • x: Dict[str, Tensor] (or UserDict[str, Tensor]) - will be passed in as model(**x)
  • x: Tuple[List[Tensor], Dict[str, Tensor]] - will be passed in as model(*x[0], **x[1])

Other generic input are not fully supported, we perform model call using the following heuristics,

  • x: List[Any] - will be passed in as model(*x)
  • x: Dict[str, Any] - will be passed in as model(**x)
  • x: Tuple[List[Any], Dict[str, Any]] - will be passed in as model(*x[0], **x[1])

To run with them, split_input_fn should be specified during cache initialization to break these inputs into smaller batches. In some rare cases, you may also need to override get_input_tensors when its heuristic can not grab enough tensors that covers all cuda devices that hold some tensors in the input.

Example Usage with Huggingface Transformers

Learning a Bi-encoder

Say we want to learn a embedding space of labels and text. Consider the following four pairs. (In practice, you will have many more and much longer text entries.)

labels = ['fruit', 'meat', 'school', 'company']
texts = [
  'this is an apple', 
  'steak should be cooked medium rare', 
  'cmu is pittsburgh', 
  'apple sells laptop'
]

Initialize our encoder models,

from transformers import AutoTokenizer, TFAutoModel
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
encoder1 = AutoModel.from_pretrained("bert-base-uncased").cuda()
encoder2 = AutoModel.from_pretrained("bert-base-uncased").cuda()

Initialize the GradCache object,

from grad_cache import GradCache
from grad_cache.loss import SimpleContrastiveLoss

loss_fn = SimpleContrastiveLoss()
gc = GradCache(
  models=[encoder1, encoder2], 
  chunk_sizes=2, 
  loss_fn=loss_fn, 
  get_rep_fn=lambda v: v.pooler_output
)

Here we use the get_rep_fn argument to specify a function that takes generic Huggingface model output and return the actual representation tensor.

Create model input,

xx = tokenizer(tt, return_tensors='pt', padding=True)
yy = tokenizer(tt2, return_tensors='pt', padding=True)

Run a cache step,

gc(xx, yy, reduction='mean')

Here we use reduction='mean' as a loss_kwargs to control loss behavior. With a defined optimizer, the full gradient update can be done as,

optimizer.zero_grad()
gc(xx, yy, reduction='mean')
optimizer.step()

Use Tied Encoder?

This is naturally handled by the (magic of) dynamic graph. You pass shallow copies of the same encoder model to the GradCache init method.

tied_encoder = AutoModel.from_pretrained("bert-base-uncased").cuda()
gc = GradCache(
  models=[tied_encoder , tied_encoder], 
  chunk_sizes=2, 
  loss_fn=loss_fn, 
  get_rep_fn=lambda v: v.pooler_output
)

Under the hood, distinct hooks will be registered to make correct gradient computation.

Distributed Training with Multiple GPUs?

We expect cross process communication of representations to be handled by the loss_fn.

from grad_cache.loss import DistributedContrastiveLoss
loss_fn_dist = DistributedContrastiveLoss()

Properly wrap the the encoder models for gradient reduction,

encoder1_ddp = DistributedDataParallel(
	encoder1, device_ids=[local_rank], output_device=local_rank, find_unused_parameters=True)
encoder2_ddp = DistributedDataParallel(
	encoder2, device_ids=[local_rank], output_device=local_rank, find_unused_parameters=True)

You can initialize the cache use the distributed loss and the DDP models,

gc = GradCache(
  models=[encoder1_ddp, encoder2_ddp], 
  chunk_sizes=2, 
  loss_fn=loss_fn_dist, 
  get_rep_fn=lambda v: v.pooler_output
)

Run a cache step,

gc(xx, yy, no_sync_except_last=True, reduction='mean')

Set no_sync_except_last=True to avoid unnecessary gradient reduction.

Functional Approach

Decorators

If you are developing a new project, we recommend also checking out the decorators we have provided to create higher order functions for cache.

grad_cache.functional.cached(func: Callable[..., Tensor])

A decorator that takes a model call function into a cached compatible version.

func - A function that calls the model and return representation tensor.

Return - A function that returns 1) representation leaf tensors for cache construction, 2) a closure function for the 2nd forward and the cached backward. Call 2) with 1) as argument after calling backward on the loss Tensor.

grad_cache.functional.cat_input_tensor(func: Callable[..., Tensor])

A decorator that concatenates positional and keyword arguments of type List[Tensor] into a single Tensor on the 0th dimension. This can come in handy dealing with results of representation tensors from multiple cached forward.

func - A loss function

Return - Decorated loss function for cached results.

Usage

The functional decorators are particular useful if your data loader is emitting small batches, from which you can construct the big batch. Say you also want to do automatic mixed precision, we first define the model call function and loss function,

from grad_cache.functional import cached, cat_input_tensor

import torch
import torch.nn.functional as F
from torch.cuda.amp import autocast

@cached
@autocast()
def  call_model(model, input):
	return model(**input).pooler_output

@cat_input_tensor
@autocast()
def  contrastive_loss(x, y):
	target = torch.arange(0, y.size(0), int(y.size(0) / x.size(0)), device=x.device)
	scores = torch.matmul(x, y.transpose(0, 1))
	return F.cross_entropy(scores, target=target)

Say you have a DataLoader loader emitting small batches of tuple (xx, yy) of size (M * N) and that you want to train by aggregating 16 small batches to get a batch of (16M * 16N),

cache_x = []
cache_y = []
closures_x = []
closures_y = []

for step, sub_batch in enumerate(loader):  
    xx, yy = sub_batch
    rx, cx = call_model(bert, xx)
    ry, cy = call_model(bert, yy)
    
    cache_x.append(rx)
    cache_y.append(ry)
    closuresx.append(cx)
    closuresy.append(cy)
    
    if (step + 1) % 16 == 0:
        loss = contrastive_loss(cache_x, cache_y)
        scaler.scale(loss).backward()
        
	for f, r in zip(closuresx, cache_x):
            f(r)
        for f, r in zip(closuresy, cache_y):
            f(r)

        cache_x = []
        cache_y = []
        closures_x = []
        closures_y = []
	
        scaler.step(optimizer)
        scaler.update()
        optimizer.zero_grad()

Code Structure

grad_cache/grad_cache.py - Define the GradCache class. The code is under 300 lines including comments. For development, we encourage you to read through it.

grad_cache/functional.py - Define decorators to create higher order function for gradient caching from ordinary model call functions and loss functions.

Owner
Luyu Gao
NLP Research [email protected], CMU
Luyu Gao
Official implementation for (Show, Attend and Distill: Knowledge Distillation via Attention-based Feature Matching, AAAI-2021)

Show, Attend and Distill: Knowledge Distillation via Attention-based Feature Matching Official pytorch implementation of "Show, Attend and Distill: Kn

Clova AI Research 80 Dec 16, 2022
PyTorch implementation of the implicit Q-learning algorithm (IQL)

Implicit-Q-Learning (IQL) PyTorch implementation of the implicit Q-learning algorithm IQL (Paper) Currently only implemented for online learning. Offl

Sebastian Dittert 27 Dec 30, 2022
Examples of using f2py to get high-speed Fortran integrated with Python easily

f2py Examples Simple examples of using f2py to get high-speed Fortran integrated with Python easily. These examples are also useful to troubleshoot pr

Michael 35 Aug 21, 2022
InDuDoNet+: A Model-Driven Interpretable Dual Domain Network for Metal Artifact Reduction in CT Images

InDuDoNet+: A Model-Driven Interpretable Dual Domain Network for Metal Artifact Reduction in CT Images Hong Wang, Yuexiang Li, Haimiao Zhang, Deyu Men

Hong Wang 4 Dec 27, 2022
Companion code for the paper Theoretical characterization of uncertainty in high-dimensional linear classification

Companion code for the paper Theoretical characterization of uncertainty in high-dimensional linear classification Usage The required packages are lis

0 Feb 07, 2022
Synthesize photos from PhotoDNA using machine learning 🌱

Ribosome Synthesize photos from PhotoDNA. See the blog post for more information. Installation Dependencies You can install Python dependencies using

Anish Athalye 112 Nov 23, 2022
[CVPR'21] DeepSurfels: Learning Online Appearance Fusion

DeepSurfels: Learning Online Appearance Fusion Paper | Video | Project Page This is the official implementation of the CVPR 2021 submission DeepSurfel

Online Reconstruction 52 Nov 14, 2022
Attention-driven Robot Manipulation (ARM) which includes Q-attention

Attention-driven Robotic Manipulation (ARM) This codebase is home to: Q-attention: Enabling Efficient Learning for Vision-based Robotic Manipulation I

Stephen James 84 Dec 29, 2022
SemEval2022 Patronizing and Condescending Language (PCL) Detection

SemEval2022 Patronizing and Condescending Language (PCL) Detection This task is from SemEval 2022. What is Patronizing and Condescending Language (PCL

Daniel Saeedi 0 Aug 05, 2022
Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening

Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening Introduction This is an implementation of the model used for breast

757 Dec 30, 2022
How Do Adam and Training Strategies Help BNNs Optimization? In ICML 2021.

AdamBNN This is the pytorch implementation of our paper "How Do Adam and Training Strategies Help BNNs Optimization?", published in ICML 2021. In this

Zechun Liu 47 Sep 20, 2022
Security evaluation module with onnx, pytorch, and SecML.

🚀 🐼 🔥 PandaVision Integrate and automate security evaluations with onnx, pytorch, and SecML! Installation Starting the server without Docker If you

Maura Pintor 11 Apr 12, 2022
tf2onnx - Convert TensorFlow, Keras and Tflite models to ONNX.

tf2onnx converts TensorFlow (tf-1.x or tf-2.x), tf.keras and tflite models to ONNX via command line or python api.

Open Neural Network Exchange 1.8k Jan 08, 2023
ObjectDetNet is an easy, flexible, open-source object detection framework

Getting started with the ObjectDetNet ObjectDetNet is an easy, flexible, open-source object detection framework which allows you to easily train, resu

5 Aug 25, 2020
fklearn: Functional Machine Learning

fklearn: Functional Machine Learning fklearn uses functional programming principles to make it easier to solve real problems with Machine Learning. Th

nubank 1.4k Dec 07, 2022
SwinIR: Image Restoration Using Swin Transformer

SwinIR: Image Restoration Using Swin Transformer This repository is the official PyTorch implementation of SwinIR: Image Restoration Using Shifted Win

Jingyun Liang 2.4k Jan 08, 2023
Unofficial PyTorch reimplementation of the paper Swin Transformer V2: Scaling Up Capacity and Resolution

PyTorch reimplementation of the paper Swin Transformer V2: Scaling Up Capacity and Resolution [arXiv 2021].

Christoph Reich 122 Dec 12, 2022
Stacs-ci - A set of modules to enable integration of STACS with commonly used CI / CD systems

Static Token And Credential Scanner CI Integrations What is it? STACS is a YARA

STACS 18 Aug 04, 2022
一个免费开源一键搭建的通用验证码识别平台,大部分常见的中英数验证码识别都没啥问题。

captcha_server 一个免费开源一键搭建的通用验证码识别平台,大部分常见的中英数验证码识别都没啥问题。 使用方法 python = 3.8 以上环境 pip install -r requirements.txt -i https://pypi.douban.com/simple gun

Sml2h3 189 Dec 02, 2022
An implementation of quantum convolutional neural network with MindQuantum. Huawei, classifying MNIST dataset

关于实现的一点说明 山东大学 2020级 苏博南 www.subonan.com 文件说明 tools.py 这里面主要有两个函数: resize(a, lenb) 这其实是我找同学写的一个小算法hhh。给出一个$28\times 28$的方阵a,返回一个$lenb\times lenb$的方阵。因

ぼっけなす 2 Aug 29, 2022