ianyang66 / ADL2022-HW2

Applied Deep Learning (2022 Fall) HW2 at National Taiwan University (NTU) CSIE

Homework 2 - NTU ADL 2022 FALL

Reproduce testing process

Use run.sh to predict testing data

bash download.sh
bash run.sh /path/to/context.json /path/to/test.json /path/to/pred/prediction.csv

example:

bash download.sh
bash run.sh ./data/context.json ./data/test.json ./data/test_submission.csv

Reproduce training process

Install Spacy

You should install spacy and zh_core_web_md, or you can not use my function to compute exact match and f1 score in question answering.

bash spacy&dependency_install.sh

Context-Selection Data Preprocessing

We need to prepare proper data format for multiple-choice and question-answering (store preprocessed data to ./data/cs_train.json, ./data/cs_valid.json, ./data/qa_train.json, ./data/qa_valid.json)

bash preprocess_train.sh /path/to/train.json /path/to/valid.json /path/to/context.json

example:

bash preprocess_train.sh ./data/cs_train.json ./data/cs_valid.json ./data/context.json

Context-Selection train

bash train_cs.sh  /path/to/preprocessed_train.json /path/to/preprocessed_valid.json  /path/to/context.json

example:

bash train_cs.sh ./data/cs_train.json ./data/cs_valid.json ./data/context.json

Question-Answering train

bash train_qa.sh /path/to/preprocessed_train.json /path/to/preprocessed_valid.json  /path/to/context.json

example:

bash train_qa.sh ./data/qa_train.json ./data/qa_valid.json ./data/context.json

Experiment Result

Model Name (CS)	Model Name (QA)	Gradient Accumulation Steps (MC)	Gradient Accumulation Steps (QA)	Num Train Epochs (CS)	Num Train Epochs (QA)	lr scheduler type (CS)	lr scheduler type (QA)	Public Acc. on Kaggle
bert-base-chinese	bert-base-chinese	2	2	3	1	linear	linear	0.74231
hfl/chinese-roberta-wwm-ext	hfl/chinese-roberta-wwm-ext	2	2	3	3	linear	linear	0.76401
hfl/chinese-roberta-wwm-ext	hfl/chinese-roberta-wwm-ext	2	2	3	3	linear	cosine	0.783
hfl/chinese-roberta-wwm-ext	hfl/chinese-roberta-wwm-ext	2	2	3	20	linear	cosine	0.77215
hfl/chinese-roberta-wwm-ext	hfl/chinese-roberta-wwm-ext	2	2	3	20	linear	linear	0.7613
hfl/chinese-roberta-wwm-ext	hfl/chinese-pert-base	2	8	3	10	linear	linear	0.76672
hfl/chinese-roberta-wwm-ext	hfl/chinese-lert-base	2	8	3	10	linear	cosine	0.79023
hfl/chinese-roberta-wwm-ext	hfl/chinese-lert-base	2	8	3	10	linear	linear	0.79023
hfl/chinese-roberta-wwm-ext	hfl/chinese-lert-base	2	8	3	15	linear	linear	0.79023
hfl/chinese-roberta-wwm-ext	hfl/chinese-lert-base	2	8	3	20	linear	linear	0.78119
hfl/chinese-roberta-wwm-ext	hfl/chinese-lert-base	2	8	5	10	linear	linear	0.78661
hfl/chinese-roberta-wwm-ext	hfl/chinese-lert-large	2	8	3	6	linear	linear	0.79475
hfl/chinese-roberta-wwm-ext-large	hfl/chinese-lert-large	8	8	2	6	linear	linear	0.79385
hfl/chinese-macbert-base	hfl/chinese-lert-large	8	8	6	6	linear	linear	0.79927
hfl/chinese-macbert-large	hfl/chinese-lert-large	8	8	2	6	linear	linear	0.80831
hfl/chinese-macbert-large	hfl/chinese-lert-large	8	64	2	10	linear	linear	0.80289

Q5. bonus

Reproduce training&prediction process

Use train_intent.sh to train bert model, and it will also predict the testing data.

bash bonus/train_intent.sh /path/to/data_directory_contain_train.json&eval.json /path/to/test.json /path/to/prediction.csv

example:

bash bonus/train_intent.sh data_hw1/intent data_hw1/intent/test.json intent_pred.csv

Use train_slot.sh to train bert model, and it will also predict the testing data.

bash bonus/train_slot.sh /path/to/data_directory_contain_train.json&eval.json /path/to/test.json /path/to/prediction.csv

example:

bash bonus/train_slot.sh data_hw1/slot data_hw1/slot/test.json slot_pred.csv