README.md

eval_vl_glue/evaluation

Directory for our evalutaion experiments.

• glue_tasks fine-tunes the pretrained models on the GLUE tasks.
• analysis summarizes the fine-tuning results
• vl_tasks fine-tunes the pretrained models on the NLVR2, used for the sanity check.

glue_tasks

• run_glue.py is a script to fine-tune a model on the GLUE task. We modified run_glue.py in the Huggingface's transformers repository to support prediction with multiple validation sets (for MNLI matched and mismatched sets).

  The usage is basically the same as the original one.

  # Run from the repository root
  CUDA_VISIBLE_DEVICES=0 python -u evaluation/glue_tasks/run_glue.py \
      --model_name_or_path vl_models/pretrained/ctrl_vlibert \
      --task_name wnli \
      --do_train \
      --do_eval \
      --evaluation_strategy epoch \
      --save_strategy epoch \
      --max_seq_length 128 \
      --per_device_train_batch_size 64 \
      --per_device_eval_batch_size 128 \
      --learning_rate 2e-5 \
      --num_train_epochs 5 \
      --output_dir vl_models/finetuned/0/ctrl_vlibert/wnli \
      --fp16 \
      --disable_tqdm 1 \
      --seed 42
  

  This is an example of a run where ctrl_vlibert will be pre-trained on wnli and five checkpoints will be made in vl_models/finetuned/0/ctrl_vlibert/wnli.

  For pretrained models, we can get prediction files on the validation sets with the --do_dump_val option.

  # Run from the repository root and make sure that vl_models/prediction exists.
  python -u python -u evaluation/glue_tasks/run_glue.py \
      --model_name_or_path vl_models/finetuned/0/ctrl_vlibert/wnli \
      --task_name wnli \
      --do_dump_val 1 \
      --evaluation_strategy epoch \
      --save_strategy epoch \
      --max_seq_length 128 \
      --per_device_train_batch_size 1024 \
      --per_device_eval_batch_size 1024 \
      --dataloader_num_workers 1 \
      --learning_rate 2e-5 \
      --num_train_epochs 5 \
      --output_dir vl_models/prediction/0/ctrl_vlibert/wnli \
      --fp16 \
      --disable_tqdm 1 \
      --seed 42
  

• batch_run.sh is a batch version of run_glue.py to fine-tune with each model and task repeatedly. Run from the repository root directory.

  CUDA_VISIBLE_DEVICES=0 evaluation/glue_tasks/batch_run.sh
  

analysis

We used Jupyter Notebook. Please see each notebook for the detail.

• word_overlap.ipynb calculates word overlap between corpora (Figure 1)
• get_glue_score.ipynb calculates and summarizes models' GLUE score (Table 2 and 5)
• weight_similarity.ipynb calculates weight similarity between models (Table 3)
• error_analysis.ipynb deivides problems according to whether the models are successful in the problems or not (Table 4)

When you run those notebooks in your environment, install the Notebook packages:

pip install notebook ipywidgets
jupyter nbextension enable --py widgetsnbextension

vl_tasks

As a quick check for our conversion, we trained the models on NLVR2.
We used image features in the pre-processed data distributed in e-bug/volta/data.

Data creation

Before the nlvr2 training, the data creation is required:

1. Make a nlvr2 directory in /evaluation/vl_tasks.

2. Download annotation data (dev.json, test1.json, and train.json) from lil-lab/nlvr/nlvr2/data/ into nlvr2/annotations

3. Download the pre-processed data (image features) from e-bug/volta/data as nlvr2/nlvr2-feat.lmdb .

4. Convert the pre-processed data into pickle files.
   (We decided to convert the lmdb because using it was slow in our environment.)

   python convert_lmdb.py --src nlvr2/nlvr2-feat.lmdb --dest pickled
   

After that data creation, the directory structure will be like below:

vl_tasks
    + nlvr2
        + annotations
            - dev.json
            - test1.json
            - train.json
        + pickled
            + train
                (omitted)
            + test1
                (omitted)
            + dev
                (omitted)
        + nlvr2-feat.lmdb
            (omitted)
    - batch_run.sh
    - convert_lmdb.py
    - dataset_nlvr2.py
    - run_vl.py

The dataset_nlvr2.py script loads data assuming this structure.

Training

run_vl.py. We modified run_glue.py in the Huggingface's transformers repository to support image feature replacement when creating mini-batches (using set_format of the Dataset class).
Example of the usage for ctrl_vlibert:

python -u evaluation/vl_tasks/run_vl.py \
    --model_name_or_path vl_models/pretrained/ctrl_vlibert \
    --task_name nlvr2 \
    --task_dir evaluation/vl_tasks/nlvr2 \
    --do_train \
    --do_eval \
    --evaluation_strategy epoch \
    --save_strategy epoch \
    --max_seq_length 128 \
    --per_device_train_batch_size 64 \
    --per_device_eval_batch_size 128 \
    --learning_rate 2e-5 \
    --num_train_epochs 20 \
    --output_dir vl_models/vl_finetuned/0/ctrl_vlibert/nlvr2 \
    --fp16 \
    --disable_tqdm 1 \
    --seed 42

batch_run.sh is a batch version of run_vl.py.

Results

We fine-tuned the models by 20 epochs and obtained the simmilar results to those in the Table 2 in the volta paper.
We chose the best epoch according to eval_accuracy instead of eval_loss in this table, which is different from our GLUE evaluation.

model	task	epoch	eval_loss	*eval_accuracy
ctrl_lxmert	nlvr2	16.0	1.708	0.694
ctrl_uniter	nlvr2	10.0	0.937	0.722
ctrl_vilbert	nlvr2	18.0	2.042	0.724
ctrl_visual_bert	nlvr2	19.0	1.702	0.720
ctrl_vl_bert	nlvr2	10.0	0.899	0.724