Train, dev and test sets are under data/controlled_text_reduction. If you want to use your own data, you first need to pre-process it. To do that, follow the steps described in preprocess_utils/README.md.

We suggest using conda to setup environment. You need to first replace prefix in environment.yml with your home path. With conda installed, create an environment called HighQu with:

conda env create -f environment.yml

For training HighQu with default hyperparameters, run:

conda activate HighQu
python main.py

To evaluate HighQu on the devset and test, run:

conda activate HighQu
python sample.py --dataset-test data/controlled_text_reduction/test_set.jsonl --dataset-dev  data/controlled_text_reduction/dev_set.jsonl --saved-model-path /path/to/ckpt

where /path/to/ckpt should point to the checkpoint that is evaluated. Additionally, you can evaluate only on the dev/test sets, by only passing the corresponding flag.

HighQu also supports multi-GPU training. To run a multi-GPU training, first run:

conda activate HighQu
pip install accelerate

Then, follow the instructions in accelerate and configure your default_config.yaml file to look as follows:

command_file: null
commands: null
compute_environment: LOCAL_MACHINE
deepspeed_config: {}
distributed_type: MULTI_GPU
downcast_bf16: 'no'
fsdp_config: {}
gpu_ids: <available_GPUs>
machine_rank: 0
main_process_ip: null
main_process_port: null
main_training_function: main
megatron_lm_config: {}
mixed_precision: fp16
num_machines: 1
num_processes: <number_of_available_GPUs>
rdzv_backend: static
same_network: true
tpu_name: null
tpu_zone: null
use_cpu: false

where <available_GPUs> should be replaced by a comma-delimited list of the avaliable GPUs (for example:gpu_ids: 0,1,2) and <number_of_available_GPUs> should be accordingly replaced, namely it should account for the total number of available GPUs (following the previous example: num_processes: 3).

Then, when the default_config.yaml is set, to run a multi GPU training of HighQu with default hyperparameters, run:

accelerate launch main.py