Graphcore / lxmert-gqa-uncased

Graphcore/lxmert-gqa-uncased

BERT (Bidirectional Encoder Representations from Transformers) is a transformers model which is designed to pretrain bidirectional representations from unlabeled texts. It enables easy and fast fine-tuning for different downstream task such as Sequence Classification, Named Entity Recognition, Question Answering, Multiple Choice and MaskedLM.

It was trained with two objectives in pretraining : Masked language modeling(MLM) and Next sentence prediction(NSP). First, MLM is different from traditional LM which sees the words one after another while BERT allows the model to learn a bidirectional representation. In addition to MLM, NSP is used for jointly pertaining text-pair representations.

It reduces the need of many engineering efforts for building task specific architectures through pre-trained representation. And achieves state-of-the-art performance on a large suite of sentence-level and token-level tasks.

Model description

LXMERT is a transformer model for learning vision-and-language cross-modality representations. It has a Transformer model that has three encoders: object relationship encoder, a language encoder, and a cross-modality encoder. It is pretrained via a combination of masked language modelling, visual-language text alignment, ROI-feature regression, masked visual-attribute modeling, masked visual-object modelling, and visual-question answering objectives. It achieves the state-of-the-art results on VQA anad GQA.

Paper link : LXMERT: Learning Cross-Modality Encoder Representations from Transformers

Intended uses & limitations

This model is a fine-tuned version of unc-nlp/lxmert-base-uncased on the Graphcore/gqa-lxmert dataset. It achieves the following results on the evaluation set:

• Loss: 1.9326
• Accuracy: 0.5934

Training and evaluation data

• Graphcore/gqa-lxmert dataset

Training procedure

Trained on 16 Graphcore Mk2 IPUs using optimum-graphcore .

Command line:

python examples/question-answering/run_vqa.py \
  --model_name_or_path unc-nlp/lxmert-base-uncased \
  --ipu_config_name Graphcore/lxmert-base-ipu \
  --dataset_name Graphcore/gqa-lxmert \
  --do_train \
  --do_eval \
  --max_seq_length 512 \
  --per_device_train_batch_size 1 \
  --num_train_epochs 4 \
  --dataloader_num_workers 64 \
  --logging_steps 5 \
  --learning_rate 1e-5 \
  --lr_scheduler_type linear \
  --loss_scaling 16384 \
  --weight_decay 0.01 \
  --warmup_ratio 0.1 \
  --output_dir /tmp/gqa/ \
  --dataloader_drop_last \
  --replace_qa_head \
  --pod_type pod16

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• distributed_type: IPU
• total_train_batch_size: 64
• total_eval_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 4.0
• training precision: Mixed Precision

Training results

***** train metrics *****
  "epoch": 4.0,
  "train_loss": 0.6123406731570221,
  "train_runtime": 29986.2288,
  "train_samples": 943000,
  "train_samples_per_second": 125.791,
  "train_steps_per_second": 1.965

***** eval metrics *****
  "eval_accuracy": 0.5933514030612245,
  "eval_loss": 1.9326171875,
  "eval_samples": 12576,

Framework versions

• Transformers 4.18.0.dev0
• Pytorch 1.10.0+cpu
• Datasets 2.0.0
• Tokenizers 0.11.6