Neural architecture search (NAS) training script run_mlm.py
is based on run_mlm.py
of Model-Compression-Research-Package by IntelLabs.
Recommend python 3.9 or higher version.
pip install -r requirements.txt
pip install transformers==4.34.1
pip install datasets==2.14.7
Note: Please use transformers no higher than 4.34.1
Datasets are downloaded and processed using 🤗 Datasets
package.
The script run_mlm.py
can be used for NAS of 🤗 Transformers
models.
Search best model architectures based on google/mobilebert-uncased
on English Wikipedia and BookCorpus datasets using the following command:
python run_mlm.py \
--model_name_or_path google/mobilebert-uncased \
--datasets_name_config wikipedia:20200501.en bookcorpusopen \
--do_train --do_eval --nas \
--data_process_type segment_pair_nsp \
--max_seq_length 128 \
--dataset_cache_dir <DATA_CACHE_DIR> \
--output_dir <OUTPUT_DIR>
Search best model architectures based on prajjwal1/bert-tiny
on English Wikipedia and BookCorpus datasets using the following command:
python run_mlm.py \
--model_name_or_path prajjwal1/bert-tiny \
--datasets_name_config wikipedia:20200501.en bookcorpusopen \
--do_train --do_eval --nas \
--data_process_type segment_pair_nsp \
--max_seq_length 128 \
--dataset_cache_dir <DATA_CACHE_DIR> \
--output_dir <OUTPUT_DIR>