shism2 / zgen1 Goto Github PK

Hi, this describes our implementation for our paper: "Towards More Efficient Insertion Transformer with Fractional Positional Encoding".

Please refer to the paper for more details: [paper] [bib]

Clone this repo:

git clone https://github.com/zzsfornlp/zgen1/ src

Please refer to the scripts/setup_env.sh for details on setting up environments.

Before running anything, make sure to export the src directory to your $PYTHONPATH:

export PYTHONPATH=/your/path/to/src

Data and vocabs should be prepared before running. Here is an example for preparing MT data (Distilled WMT14-en-de), we simply use the prepared ones from fairseq.

Please refer to the script scripts/prepare-wmt14en2dev2.sh for more details.

Assuming the data is prepared at the directory of !!!!!MYDATA!!!!! (replace this in the command with the real data path!), use the following commands to train an FPE-based Insertion Model (here we are assuming using 2 GPUs, thus leading to a batch size of 12500*2=25K, please adjust the batch size according to the number of GPUs that you are using):

mkdir run
cd run
python -m zgen.main.run_ddp zgen.main.train log_file:_log log_last:1 conf_output:_confR  'train_end_test_conditions:[f"self.model.Milm.set_test_noi_penalty({-z/2.})" for z in range(10)]'  ilm.no_posi_embeds:1 ilm.incr_embed:yes ilm.incr_layers:6 ilm.incr_grow_hid:0 rollin_strategy:bt0 oracle_strategy:bt0 ilm.test_noi_penalty:-0.5 'conf_sbase:{"input_dir":"!!!!!MYDATA!!!!!","dec_mod":"ilm","dec_task":"wmt14en2de"}' max_uidx:300000 ilm.test_noi_penalty:-1. train0.batch_size:12500 train0.bitext_src_suffix:.enD train0.bitext_trg_suffix:.deD lrate_decrease_alpha:0.4

For testing with different batch sizes ($BSIZE), testing the speed with (here, test_certain_batches specifies how many batches to run for speed testing with repeated sampling, deleting this option will make it ordinary testing; using test_certain_batches will lead to a final eval error which does not matter, read the Test-Info: line for time information!):

python3 -m zgen.main.test _confR nn.dist_world_size:1 fp16:1 model_load_name:zmodel.bestn.m test0.input_dir:${MYDATA} test1.input_paths: test0.input_paths:wmt14en2de.test vocab_load_dir:${MYDATA}/voc_wmt14en2de/ test0.batch_size:$BSIZE test_certain_batches:100