This example activated DeepSpeed on the implementation of training a set of popular model architectures on ImageNet dataset. The models include ResNet, AlexNet, and VGG, and the

baseline implementation could be found at pytorch examples [Github repository](https://github.com/pytorch/examples/tree/main/imagenet). Parts of this README is also borrowed from the original repo. DeepSpeed activation allows for ease in

running the code in distributed manner, allowing for easily applying fp16 quantization benefitting Zero stage1 memory reduction.

Applying fp16 quantization and Zero stage 1 memory optimization we were able to reduce the required memory. The table bellow summarizes the results of running resnet 50 on a DGX-1

node (with 16 V100 GPUs):

| Optimization level | Allocated Memory (GB) | Mem. Consumption Improvement (%) |

|-------------------|-------------------|---------|

|Baseline | 1.66 | -|

|DS + fp16 | 1.04 | 37.3|

|Ds + fp16 + Zero 1 | 0.81 | 51.2|

Furthermore, the memory optimization had no adverse impact on accuracy, a point illustrated by the graph below.


- Install PyTorch ([pytorch.org](http://pytorch.org))

- `pip install -r requirements.txt`

- Download the ImageNet dataset from http://www.image-net.org/

- Then, move and extract the training and validation images to labeled subfolders, using [the following shell script](extract_ILSVRC.sh)

To train a model, run one of the bash files after setting the model name as well as the path to the ImageNet dataset:

run_ds.sh : baseline code with DeepSpeed activated

run_ds_fp16.sh : fp16 activated

run_ds_fp16_z1.sh: fp16 and Zero1 are activated

The default learning rate schedule starts at 0.1 and decays by a factor of 10 every 30 epochs. This is appropriate for ResNet and models with batch normalization, but too high for AlexNet and VGG. Use 0.01 as the initial learning rate for AlexNet or VGG:

ImageNet dataset is large and time-consuming to download. To get started quickly, run `main.py` using dummy data by "--dummy". It's also useful for training speed benchmark. Note that the loss or accuracy is useless in this case.

usage: deepspeed main.py [-h] [-a ARCH] [-j N] [--epochs N] [--start-epoch N] [-b N] [--lr LR] [--momentum M] [--wd W] [-p N] [--resume PATH] [-e] [--pretrained] [--world-size WORLD_SIZE]

[--num_gpus NUM GPU] [--num_nodes NUM NODES] [--seed SEED] [--gpu GPU] [--multiprocessing-distributed] [--deepspeed] [--deepspeed_config] [CONFIG FILE] [--dummy] [DIR]

positional arguments:

DIR path to dataset (default: imagenet)

CINFIG FILE path to deepspeed config file

optional arguments:

--deepspeed activated deepspeed library

--deepspeed_config path to the deepspeed config file

--help, -h show this help message and exit

--arch ARCH, -a ARCH model architecture: alexnet | convnext_base | convnext_large | convnext_small | convnext_tiny | densenet121 | densenet161 | densenet169 | densenet201 | efficientnet_b0 |

efficientnet_b1 | efficientnet_b2 | efficientnet_b3 | efficientnet_b4 | efficientnet_b5 | efficientnet_b6 | efficientnet_b7 | googlenet | inception_v3 | mnasnet0_5 | mnasnet0_75 |

mnasnet1_0 | mnasnet1_3 | mobilenet_v2 | mobilenet_v3_large | mobilenet_v3_small | regnet_x_16gf | regnet_x_1_6gf | regnet_x_32gf | regnet_x_3_2gf | regnet_x_400mf | regnet_x_800mf |

regnet_x_8gf | regnet_y_128gf | regnet_y_16gf | regnet_y_1_6gf | regnet_y_32gf | regnet_y_3_2gf | regnet_y_400mf | regnet_y_800mf | regnet_y_8gf | resnet101 | resnet152 | resnet18 |

resnet34 | resnet50 | resnext101_32x8d | resnext50_32x4d | shufflenet_v2_x0_5 | shufflenet_v2_x1_0 | shufflenet_v2_x1_5 | shufflenet_v2_x2_0 | squeezenet1_0 | squeezenet1_1 | vgg11 |

vgg11_bn | vgg13 | vgg13_bn | vgg16 | vgg16_bn | vgg19 | vgg19_bn | vit_b_16 | vit_b_32 | vit_l_16 | vit_l_32 | wide_resnet101_2 | wide_resnet50_2 (default: resnet18)

--workers N, -j N number of data loading workers (default: 4)

--epochs N number of total epochs to run

--start-epoch N manual epoch number (useful on restarts)

--batch-size N, -b N mini-batch size (default: 256), this is the total batch size of all GPUs on the current node when using Data Parallel or Distributed Data Parallel

--lr LR, --learning-rate LR

initial learning rate

--momentum M momentum

--wd W, --weight-decay W

weight decay (default: 1e-4)

--print-freq N, -p N print frequency (default: 10)

--resume PATH path to latest checkpoint (default: none)

--evaluate, -e evaluate model on validation set

--pretrained use pre-trained model

--world-size WORLD_SIZE

number of nodes for distributed training

--local_rank LOCAL_RANK

node rank for distributed training

--seed SEED seed for initializing training.

--gpu GPU GPU id to use.

--num_gpus NUM GPU Num GPUs to use.

--num_nodes NUM NODES Num nodes to use

--multiprocessing-distributed

use multi-processing distributed training to launch N processes per node, which has N GPUs. This is the fastest way to use PyTorch for either single node or multi node data parallel

training

--dummy use fake data to benchmark

Example of running resnet50 on single GPU:

deepspeed --num_nodes=1 --num_gpus=1 main.py -a resnet50 --deepspeed --deepspeed_config config/ds_config.json --dummy