Introduction

This is a repo for our paper "Self-training from Self-memory in Data-to-text Generation". We test our self-training models over two datasets, DART and E2E NLG.

If you want to check examples of other NLG datasets, see https://quest.ms.mff.cuni.cz/nlg/tabgenie.

Dependencies

• Python >= 3.8.x
• transformers >= 4.30.2
• torch >= 2.0.1+cu117 (not sure with previous versions)

Training information

We train our models on "t5-base" (without prefixes), "facebook/bart-base", and "google/flan-t5-base" on Windows, using Git Bash. All self-memory training methods only rely on "t5-base". We adapt the evaluation of DART and E2E NLG from two original packages (https://github.com/Yale-LILY/dart/tree/master/evaluation and https://github.com/tuetschek/e2e-metrics). Furthermore, we cut off the evaluation on several metrics due to errors. We also offer our evaluation version, which can be found in the file metrics.py.

Training parameters

Here is the list of training parameters:

• mode: train/test/generate
• epoch: the number of epochs
• batch_size: the size of batch
• train_file, var_file, test_file: URLs of training, validation, and test files
• decoding_type: greedy
• output_dir: the location of trained models
• dataset_name: the name of datasets (dart/e2e)
• source_prefix: helpful for T5 models. We DO NOT use prefixes such as "summarize: " in the experiments
• source_column, target_column: specify the fields of input and output in the training, validation, and test sets
• train_percent: the ratio of self-train data (30%).
• merge_new_data: merge self-memory with new data (1) or no merge self-memory with new data (0)
• self_train_t2d: self-train the T2D model
• same_data: train on the fixed data (1) or the random data (0)
• eval_metric: evaluation metric
• t2d_opt_metric: T2D optimization metric (osf = Overall Slot Filling, spm = simple phrase matching)
• no_self_mem: use self-memory (1) or not (0)
• same_data_type: use with no_self_mem = 1
• infer_multi_thread: use multi-threading in the inference time, (0 = No (default), 1 = Yes) # memory error may happen, not so much effective
• infer_max_workers: use with infer_multi_thread = 1 in the inference time, (default is 4) # memory error may happen, not so much effective
• same_data_size: same data size in each training epoch (0 = No, 1 = Yes (default))

DART

FULL TRAIN

python seq2seq.py --mode train --epoch 3 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --dataset_name "dart" --source_prefix "" --source_column "source" --target_column "target"

ACCELERATE FULL TRAIN

accelerate launch seq2seq.py --mode train --epoch 3 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --dataset_name "dart" --source_prefix "" --source_column "source" --target_column "target"

30% FIXED DATA + WITH NEW DATA + WITH SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 1 --self_train_t2d 1 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% FIXED DATA + WITH NEW DATA + WITHOUT SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 1 --self_train_t2d 0 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% FIXED DATA + WITHOUT NEW DATA + WITH SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 0 --self_train_t2d 1 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% FIXED DATA + WITHOUT NEW DATA + WITHOUT SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 0 --self_train_t2d 0 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% FIXED DATA + WITHOUT SELF MEMORY: TYPE 1

accelerate launch seq2seq.py --mode self_train --epoch 3 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --same_data 1 --eval_metric "eval_meteor" --no_self_mem 1 --same_data_type 1

30% FIXED DATA + WITHOUT SELF MEMORY: TYPE 2

accelerate launch seq2seq.py --mode self_train --epoch 3 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --same_data 1 --eval_metric "eval_meteor" --no_self_mem 1 --same_data_type 2

30% RANDOM DATA + WITH NEW DATA + WITH SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 1 --self_train_t2d 1 --same_data 0 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% RANDOM DATA + WITH NEW DATA + WITHOUT SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 1 --self_train_t2d 0 --same_data 0 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% RANDOM DATA + WITHOUT NEW DATA + WITH SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 0 --self_train_t2d 1 --same_data 0 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% RANDOM DATA + WITHOUT NEW DATA + WITHOUT SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 0 --self_train_t2d 0 --same_data 0 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% RANDOM DATA + WITHOUT SELF MEMORY: TYPE 3

python seq2seq.py --mode self_train --epoch 3 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --same_data 0 --eval_metric "eval_meteor" --no_self_mem 1

accelerate launch seq2seq.py --mode self_train --epoch 3 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --same_data 0 --eval_metric "eval_meteor" --no_self_mem 1

EVALUATION (use "dart_val")

python seq2seq.py --mode test --test_file "dataset/dart/test.json" --model_name "t5-base" --output_dir "output/xxx" --source_prefix "" --test_batch_size 16 --dataset_name "dart"

GENERATE

python seq2seq.py --mode generate --test_file "dataset/dart/test.json" --model_name "t5-base" --output_dir "output/xxx" --source_prefix "" --test_batch_size 16 --dataset_name "dart"

GENERATE with MULTITHREAD (speed up generation)

python seq2seq.py --mode generate --test_file "dataset/dart/test.json" --model_name "t5-base" --output_dir "output/xxx" --source_prefix "" --test_batch_size 16 --dataset_name "dart" --infer_max_workers 4 --infer_multi_thread 1

E2E NLG

FULLTRAIN

python seq2seq.py --mode train --epoch 3 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --dataset_name "e2e_nlg" --source_prefix "" --source_column "source" --target_column "target"

ACCELERATE FULL TRAIN

accelerate launch seq2seq.py --mode train --epoch 3 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --dataset_name "e2e_nlg" --source_prefix "" --source_column "source" --target_column "target"

30% FIXED DATA + WITH NEW DATA + WITH SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --merge_new_data 1 --self_train_t2d 1 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% FIXED DATA + WITH NEW DATA + WITHOUT SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --merge_new_data 1 --self_train_t2d 0 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% FIXED DATA + WITHOUT NEW DATA + WITH SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 0 --self_train_t2d 1 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% FIXED DATA + WITHOUT NEW DATA + WITHOUT SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --merge_new_data 0 --self_train_t2d 0 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% FIXED DATA + WITHOUT SELF MEMORY: TYPE 1

accelerate launch seq2seq.py --mode self_train --epoch 3 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --same_data 1 --eval_metric "eval_meteor" --no_self_mem 1 --same_data_type 1

30% FIXED DATA + WITHOUT SELF MEMORY: TYPE 2

accelerate launch seq2seq.py --mode self_train --epoch 3 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --same_data 1 --eval_metric "eval_meteor" --no_self_mem 1 --same_data_type 2

30% RANDOM DATA + WITH NEW DATA + WITH SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --merge_new_data 1 --self_train_t2d 1 --same_data 0 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% RANDOM DATA+ WITH NEW DATA + WITHOUT SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --merge_new_data 1 --self_train_t2d 0 --same_data 0 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% RANDOM DATA + WITHOUT NEW DATA + WITH SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 0 --self_train_t2d 1 --same_data 0 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% RANDOM DATA + WITHOUT NEW DATA + WITHOUT SELF-TRAIN T2D

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --merge_new_data 0 --self_train_t2d 0 --same_data 0 --eval_metric "eval_meteor" --t2d_opt_metric "osf"

30% RANDOM DATA + WITHOUT SELF MEMORY: TYPE 3

python seq2seq.py --mode self_train --epoch 3 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --same_data 0 --eval_metric "eval_meteor" --no_self_mem 1

accelerate launch seq2seq.py --mode self_train --epoch 3 --batch_size 4 --train_file "dataset/e2e_nlg/train.json" --var_file "dataset/e2e_nlg/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "e2e_nlg" --source_prefix "" --same_data 0 --eval_metric "eval_meteor" --no_self_mem 1

EVALUATION (use "e2e_val")

python seq2seq.py --mode test --test_file "dataset/e2e_nlg/test.json" --model_name "t5-base" --output_dir "output/xxx" --source_prefix "" --test_batch_size 16 --dataset_name "e2e_nlg"

GENERATE

python seq2seq.py --mode generate --test_file "dataset/e2e_nlg/test.json" --model_name "t5-base" --output_dir "output/xxx" --source_prefix "" --test_batch_size 16 --dataset_name "e2e_nlg"

GENERATE with MULTITHREAD (speed up generation)

python seq2seq.py --mode generate --test_file "dataset/e2e_nlg/test.json" --model_name "t5-base" --output_dir "output/xxx" --source_prefix "" --test_batch_size 16 --dataset_name "e2e_nlg" --infer_max_workers 4 --infer_multi_thread 1

OTHER SELF-TRAINING

If you already have pre-trained D2T and T2D models, you can continue to self-train them. For example, self-train on DART:

python seq2seq.py --mode self_train --epoch 1 --self_epoch 2 --batch_size 4 --train_file "dataset/dart/train.json" --var_file "dataset/dart/val.json" --model_name "t5-base" --output_dir "output/" --source_column "source" --target_column "target" --train_percent 30 --dataset_name "dart" --source_prefix "" --merge_new_data 0 --self_train_t2d 1 --same_data 1 --eval_metric "eval_meteor" --t2d_opt_metric "osf" --load_trained 1 --d2t_model_path "url_pretrained_model" --t2d_model_path "url_pretrained_model"

Minor errors

The inference time in self-training is sometimes very slow, and we are working to search for speed-up methods.

We do not apply "accelerate" in self-training.

Author Information

If you have any questions, please open issues or contact:

• [email protected]
• https://www.tahoangthang.com