tigerjeffx / transformers-in-pytorch

Transformers in Pytorch

what are provided in this repo?

• Model: Several transformer-based milestone models are reimplemented from scratch via pytorch
  • Vision Transformer (CV) : AN IMAGE IS WORTH 16X16 WORDS: TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE

  

  • Vanilla Transformer (NLP) : Attention Is All You Need

  

• Experiments: Conduct experiments on CV/NLP benchmark, respectively
  • Image Classification :
    • ImageNet1k

    

    • Cifar10

    

  • Neural Machine Translation :
    • Multi30k

    

• Pipeline: End-to-end pipeline
  • Conveniently Playing : integrate data processing and model training/validation into one-stop shop pipeline
  • Efficiently Training : accelerate training and evaluating via DistributeDataParallel(DDP) and Mixed Precision(fp16)
  • Neatly Reading : neat file structure, easy for reading but non-trivial
    • ./script → run train/eval
    • ./model → model implementation
    • ./data → data processing

Usage

1. Env Requirements

# Conda Env
python 3.6.10
torch 1.4.0+cu100
torchvision 0.5.0+cu100
torchtext 0.5.0
spacy 3.4.1
tqdm 4.63.0

# Apex (For mix precision training)
## run `gcc --version` 
gcc (GCC) 5.4.0
## apex installation
git clone https://github.com/NVIDIA/apex
cd apex
git checkout f3a960f80244cf9e80558ab30f7f7e8cbf03c0a0
rm -rf ./build
python setup.py install --cuda_ext --cpp_ext

# System Env 
## run `nvcc --version`
Cuda compilation tools, release 10.0, V10.0.130
# run `nvidia-smi`
Check your own gpu device status

2. Data Requirements

• multi30k, cifar10 could be automatically downloaded in pipeline
• imagenet1k(ILSVRC2012) need manual download (Guide for download imagenet1k)
  • Wait until three files download.
    • ILSVRC2012_devkit_t12.tar.gz (2.5M)
    • ILSVRC2012_img_train.tar (138G)
    • ILSVRC2012_img_val.tar (6.3G)
  • Run imagenet1k pipeline, ILSVRC2012_img_train.tar and ILSVRC2012_img_val.tar will be automatically unzipped and arranged in two directories 'data/ILSVRC2012/train' and 'data/ILSVRC2012/val'.
  • But the unzip process costs more than a few hours or you can do it faster by shell anyway.

# Guide for download imagenet1k
mkdir -p data/ILSVRC2012
cd data/ILSVRC2012
wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_devkit_t12.tar.gz --no-check-certificate
wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_train.tar --no-check-certificate
wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_val.tar --no-check-certificate

3. Run Experiments

3.1 Fine-tuning ViT on imagenet1k, cifar10

• Download pretrained ViT_B_16 model parameters from official storage.

cd data
curl -o ViT_B_16.npz https://storage.googleapis.com/vit_models/augreg/B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_224.npz
curl -o ViT_B_16_384.npz https://storage.googleapis.com/vit_models/augreg/B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_384.npz

• Before run experiments
  • Set CUDA env in script/run_img_cls_task.py/__main__ according to your GPU device
  • Adjust train/eval settings in script/run_img_cls_task.py/get_args() and launch the experiment

cd script

# run experiments on cifar10
# (4mins/epoch, 3.5hours totally | GPU device: P40×4)
python ./run_image_cls_task.py cifar10

# run experiments on imagenet1k
# (less than 5hours/epcoch, more than 10hours totally | GPU device: P40×4 )
python ./run_image_cls_task.py ILSVRC2012

# Tips:
# 1. Both DDP and FP16 Mixed Precision Training are adopted for accelerating
# 2. The ratio of acceleration depends on your specific GPU device

3.2 Train vanilla transformer from scratch on multi30k (en → de)

• Before run experiments
  • Set CUDA env in script/run_nmt_task.py/__main__ according to your GPU device
  • Adjust train/eval settings in script/run_nmt_task.py/get_args() and launch the experiment

# run experiments on multi30k (small dataset ,3mins total | GPU device : P40×4 | U can also fork and adjust the pipeline and run this experiments in a small capacity gpu device)
cd script

python ./run_nmt_task.py multi30k

# Tips:
# 1. DDP is adopted for accelerating
# 1. For inference, "greedy search" and "beam search" are also included in the nmt task pipeline.

4. Results

4.1 Fine-tuning ViT on imagenet1k, cifar10,

• This repo
  • Imagenet1k : ACC 84.9% (result on 50,000 val set for | resolution 384 | extra label smoothing confidence 0.9 | batch size 160, nearly 15,000 training steps)

  

  • Cifar10 : ACC 99.04% (resolution 224 | batch size 640, nearly 5500 training steps)

  

• Comparison to official result ViT Implementation by Google

  

4.2 Train vanilla transformer from scratch on multi30k (en → de)

• This repo
  • Multi30k : BLEU 38.6 (en→de | nearly 17M #Params | batch size 512, nearly 1200 training steps)

  

• Comparison to results in Dynamic Context-guided Capsule Network for Multimodal Machine Translation

  

Reference materials for further study

• Transformer Survey
  • Efficient Transformers: A Survey
  • A Survey of Transformers
• Vanilla Transformer Component Structures
  • self-attention
    • ON THE RELATIONSHIP BETWEEN SELF-ATTENTION AND CONVOLUTIONAL LAYERS
    • Online and Linear-Time Attention by Enforcing Monotonic Alignments
  • multi-head
    • Are Sixteen Heads Really Better than One?
    • Multi-head or Single-head? An Empirical Comparison for Transformer Training
  • feed forward network
    • Transformer Feed-Forward Layers Are Key-Value Memories
  • residual connection & layer norm
    • On Layer Normalization in the Transformer Architecture
    • PowerNorm: Rethinking Batch Normalization in Transformers
  • label smoothing related
    • When Does Label Smoothing Help?
    • On Calibration of Modern Neural Networks
    • Calibration of Pre-trained Transformers
    • Revisiting the Calibration of Modern Neural Networks
• Recent Transformer Milestone Work in CV
  • AN IMAGE IS WORTH 16X16 WORDS: TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE
  • Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
  • Masked Autoencoders Are Scalable Vision Learners
  • MetaFormer Is Actually What You Need for Vision