X-MAML Reproduction

I found it is a little difficult for me to re-produce the results in the EMNLP2020 paper Zero-Shot Cross-Lingual Transfer with Meta Learning by the code in the original repository. So I re-write the code on my own.

Requirements

• Python version >= 3.6
• PyTorch version == 1.6.0
• transformers version == 3.0.2

Getting Started

1.Finetune Multilingual Pre-trained Model

At the very begining, it is necessary to finetune the multilingual pretrained models (mBERT, XLM-R, etc.) on the MNLI dataset, which is similar to XNLI but is in English. Otherwise, the pretrained model is unable to provide reasonable embeddings in the following procedure.

My script for the finetune procedure is:

cd finetune

data_path=path/to/dataset/
pretrain_model_dir=../cache_dir/bert-base-multilingual-cased/

python finetune.py \
    --data_path $data_path \
    --pretrain_model_dir $pretrain_model_dir \
    --per_gpu_train_batch_size 32 \
    --per_gpu_eval_batch_size 128 \
    --num_train_epochs 3 \
    --learning_rate 2e-5 \
    --warmup_ratio 0.1 \
    --seed 42

The finetuned model will be saved as ./finetune/finetune_saved_model/pytorch_model.bin.

If you want to load the finetuned model through model = BertForSequenceClassification.from_pretrained('./finetune/finetuned_saved_model'), you probably need to rename the parameters saved in ./finetune/finetune_saved_model/pytorch_model.bin:

from collections import OrderedDict

import torch

state_dict = torch.load('./finetune/finetuned_saved_model/pytorch_model.bin')
new_state_dict = OrderedDict()

for k, v in state_dict.items():
    new_state_dict[k.replace('model.', '')] = v

torch.save(new_state_dict, './finetune/finetuned_saved_model/pytorch_model.bin')

It is worth noting that you also need to copy the *.json and *.txt files in the ./cache_dir/bert-base-multilingual-cased/ to ./finetune/finetune_saved_model/.

You can evaluate the performance of finetuned model on the MNLI with following scripts:

cd finetune

data_path=path/to/dataset/
pretrain_model_dir=./finetune_saved_model

python eval_mnli.py \
    --data_path $data_path \
    --pretrain_model_dir $pretrain_model_dir \
    --per_gpu_train_batch_size 64 \
    --per_gpu_eval_batch_size 32 \
    --seed 42

If you wanna test it on the XNLI dataset, you can simply change the eval_mnli.py to eval_xnli.py.

2. Train the X-MAML Model

You can obtain the results under zero-shot learning with following script:

cd X-MAML

data_path=path/to/dataset/
pretrain_model_dir=../finetune/finetune_saved_model/

python xmaml.py \
    --data_path $data_path \
    --pretrain_model_dir $pretrain_model_dir \
    --num_train_iter 400 \
    --num_inner_iter 1 \
    --support_size 8 \
    --query_size 8 \
    --num_tasks_in_batch 2 \
    --num_accumulation_step 1 \
    --learning_rate 2e-5 \
    --aux_langs hi \
    --seed 42 \
    --lang ru

Adding --few_shot should give you the results under few-shot learning setting.

Results

Under the zero-shot learning setting, with hi as the xu language:

seed	ar	bg	de	el	en	es	fr	hi	ru	sw	th	tr	ur	vi	zh
42	65.77	71.35	73.08	68.36	82.18	75.41	75.63	-	71.51	48.78	55.75	63.03	61.53	72.19	73.26

There is merely some inconsistence between my results and those reported in the paper. Some further adjusting may be necessary.