hzlcodus / mixencoder

Prepare my model

move ./my_transformer/out_vector_modeling_bert.py to the dictionary models/bert under the path where lib transformers is installed, like

~/anaconda3/envs/your_envs_name/lib/python(3.8)/site-packages/transformers/models/bert`

Prepare data

All datasets are saved in ./dataset/, this dictionary is created by scripts automatically.

ubuntu

download data form here and put them into ./raw_data/ubuntu

then modify the root_path in ./raw_data/data_scripts/process_dstc.py before running the following command:

cd ./raw_data/data_scripts
python process_ubuntu.py

dstc7

download data (include augmented) and put them into ./raw_data/dstc， then

cd ./raw_data/data_scripts
python process_dstc.py

mnli

Data will be downloaded by scripts automatically.

hard negative msmarco (hard_msmarco)

data from here

Training

Commands vary with tasks and models. Some examples are provided in run_command. Readers can refer to nlp_main.py to make clear what arguments mean.

---

There are something should be paid attention to.

1. While training for classification tasks, arg: label_num should be set according to tasks.
2. While training PolyEncoder for matching tasks, arg: label_num should be set as 0.
3. While training for matching tasks, arg: step_max_query_num can be used to tackle cuda out of memory.

Test

Test will be automatically done after training.

If you want to do test but not train, appending following arguments to the command.

--no_train --do_test(do_val)

do_val means using dev data while do_test means using test data.

If you want to measure the running time of models' doing match in the way like real worlds, please replace do_test with do_real_test, like

--no_train --do_real_test --query_block_size 100

Arg: query_block_size controls the number of simultaneously processed queries. This should be set according to cuda memory. Our experiments show that this argument has little influence on testing time.