-
Notifications
You must be signed in to change notification settings - Fork 7
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
4 changed files
with
489 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,266 @@ | ||
| # | ||
|
|
||
| # a single script handling the runnings of multiple modes | ||
|
|
||
| import argparse | ||
| import sys | ||
| import os | ||
|
|
||
| def printing(x, end='\n', file=sys.stderr, flush=True): | ||
| print(x, end=end, file=file, flush=flush) | ||
|
|
||
| def system(cmd, pp=True, ass=True, popen=False): | ||
| if pp: | ||
| printing("Executing cmd: %s" % cmd) | ||
| n = os.system(cmd) | ||
| if ass: | ||
| assert n==0, f"Executing previous cmd returns error {n}." | ||
| return n | ||
|
|
||
| def parse(): | ||
| parser = argparse.ArgumentParser(description='Process running confs.') | ||
| # basic settings | ||
| # todo(note): be sure to make the dirs relative to run_dir | ||
| parser.add_argument("--src_dir", type=str, default="../../src/") | ||
| parser.add_argument("--data_dir", type=str, default="../../data/UD_RUN/ud24/") | ||
| parser.add_argument("--run_dir", type=str, default=None) | ||
| parser.add_argument("--log_suffix", type=str, default=".log") | ||
| # env | ||
| parser.add_argument("--cur_run", type=int, default=1) | ||
| parser.add_argument("--rgpu", type=int, required=True) # -1 for CPU, >=0 for GPU | ||
| # languages | ||
| parser.add_argument("--cross", type=int, default=1) # whether cross-lingual (fix embeddings, etc) | ||
| parser.add_argument("--langs", type=str, nargs="+", required=True) | ||
| parser.add_argument("--train_cuts", type=str, nargs="+", required=True) | ||
| parser.add_argument("--dev_cuts", type=str, nargs="+", required=True) | ||
| # specific for cr-embeds and mbert features | ||
| parser.add_argument("--use_word", type=int, default=1) # whether use cr word embeds | ||
| parser.add_argument("--use_mb", type=int, default=0) | ||
| parser.add_argument("--mb_suffix", type=str, default="mb48.pkl") | ||
| parser.add_argument("--mb_fold", type=int, default=5) | ||
| # pos mode | ||
| parser.add_argument("--pos_mode", type=str, default="pred", choices=["pred", "nope", "gold"]) | ||
| # model | ||
| parser.add_argument("--model", type=str, choices=["g1", "ef"], required=True) | ||
| # enc type | ||
| parser.add_argument("--enc_type", type=str, default="att", choices=["att", "rnn"]) | ||
| # extras | ||
| parser.add_argument("--extras", type=str, default="") | ||
| # | ||
| args = parser.parse_args() | ||
| return args | ||
|
|
||
| def main(): | ||
| printing(f"Run with {sys.argv}") | ||
| args = parse() | ||
| for k in sorted(dir(args)): | ||
| if not k.startswith("_"): | ||
| printing(f"--{k} = {getattr(args, k)}") | ||
| # | ||
| RGPU = args.rgpu | ||
| CUR_RUN = args.cur_run | ||
| SRC_DIR = args.src_dir | ||
| DATA_DIR = args.data_dir | ||
| LOG_SUFFIX = args.log_suffix | ||
| # ===== step 1 | ||
| # mkdir and cd into it | ||
| run_dir = args.run_dir | ||
| if run_dir is None: | ||
| # todo(note): by default, simply changing the script name | ||
| fields = __file__.split(".") | ||
| assert fields[-1] == "py" | ||
| run_dir = ".".join(fields[:-1]) + "_run" + str(CUR_RUN) | ||
| system(f"mkdir -p {run_dir}") | ||
| os.chdir(run_dir) | ||
| # ===== step 2 | ||
| printing(f"Run in {run_dir} with RGPU={RGPU} CUR_LANG={args.langs} CUR_RUN={CUR_RUN}") | ||
| # | ||
| # ===== | ||
| # basis | ||
| base_opt = "conf_output:_conf init_from_pretrain:1 drop_embed:0.2 dropmd_embed:0.2 singleton_thr:2 fix_drop:1" | ||
| if args.cross: | ||
| # == for cross-lingual settings | ||
| # todo(note): cannot let 0. as freezed embeddings since this will make layer-norm NAN; assign a smaller range instead | ||
| base_opt += " emb_conf.dim_char:0 emb_conf.word_freeze:1 use_label0:1 pretrain_init_nohit:0.1" | ||
| # == | ||
| base_opt += " lrate.val:0.001 tconf.batch_size:32 lrate.m:0.75 min_save_epochs:100 max_epochs:300 patience:8 anneal_times:10" | ||
| base_opt += " loss_function:hinge margin.val:0.5 margin.max_val:0.5 margin.mode:linear margin.b:0. margin.k:0.01 reg_scores_lambda:0." | ||
| base_opt += " arc_space:512 lab_space:128 transform_act:elu biaffine_div:1. biaffine_init_ortho:1" | ||
| base_opt += " enc_lrf.val:1. enc_lrf.mode:none enc_lrf.start_bias:0 enc_lrf.b:0. enc_lrf.k:1. enc_optim.weight_decay:0." | ||
| # use the previous setting and fix all | ||
| base_opt += " dec_lrf.val:0. dec_lrf.mode:none dec_lrf.start_bias:0 dec_lrf.b:0. dec_lrf.k:1. dec_optim.weight_decay:0." | ||
| base_opt += " dec2_lrf.val:0. dec2_lrf.mode:none dec2_lrf.start_bias:0 dec2_lrf.b:0. dec2_lrf.k:1. dec2_optim.weight_decay:0." | ||
| # ===== | ||
| # which system | ||
| if args.model == "g1": | ||
| # todo(note): simply chaning this to "graph" is also ok | ||
| base_opt += " partype:g1" | ||
| elif args.model == "ef": | ||
| # ===== | ||
| # ef system | ||
| base_opt += " aabs.val:5. aabs.max_val:5. aabs.mode:linear aabs.start_bias:100 aabs.k:0.01" | ||
| arc_k = 1 | ||
| # system | ||
| base_opt += " partype:ef zero_extra_output_params:0 iconf.ef_mode:free tconf.ef_mode:free system_labeled:1" | ||
| base_opt += " chs_num:10 chs_att.out_act:linear chs_att.att_dropout:0.1 chs_att.head_count:2" | ||
| # loss | ||
| base_opt += " tconf.ending_mode:maxv cost0_weight.val:1. loss_div_weights:0 loss_div_fullbatch:1" | ||
| # features for ef | ||
| base_opt += " fdrop_chs:0. fdrop_par:0. use_label_feat:1 use_chs:1 use_par:1 ef_ignore_chs:func" | ||
| # search | ||
| # arc_k=1 # set previously | ||
| lab_k = 1 | ||
| base_opt += f" iconf.plain_k_arc:{arc_k} iconf.plain_k_label:{lab_k} iconf.plain_beam_size:{arc_k}" | ||
| base_opt += f" tconf.plain_k_arc:{arc_k} tconf.plain_k_label:{lab_k} tconf.plain_beam_size:{arc_k}" | ||
| else: | ||
| raise NotImplementedError(f"UNK model {args.model}") | ||
| # ===== | ||
| # which encoder | ||
| base_opt += " emb_proj_dim:512 enc_conf.enc_hidden:512" | ||
| if args.enc_type == "att": | ||
| # specific for self-att encoder | ||
| base_opt += " enc_conf.enc_rnn_layer:0 enc_conf.enc_att_layer:6 enc_conf.enc_att_add_wrapper:addnorm enc_conf.enc_att_conf.clip_dist:10 enc_conf.enc_att_conf.use_neg_dist:0 enc_conf.enc_att_conf.att_dropout:0.1 enc_conf.enc_att_conf.use_ranges:1 enc_conf.enc_att_fixed_ranges:5,10,20,30,50,100 enc_conf.enc_att_final_act:linear" | ||
| base_opt += " max_epochs:300 train_skip_length:80 tconf.batch_size:80 split_batch:2 biaffine_div:0. margin.b:0. reg_scores_lambda:0.01" | ||
| base_opt += " lrate.val:0.0001 lrate_warmup:-2 drop_embed:0.1 dropmd_embed:0. drop_hidden:0.1" | ||
| elif args.enc_type == "rnn": | ||
| pass | ||
| else: | ||
| raise NotImplementedError(f"UNK enc {args.enc_type}") | ||
| # ===== | ||
| # inputs | ||
| langs = args.langs | ||
| train_cuts = args.train_cuts | ||
| dev_cuts = args.dev_cuts | ||
| assert len(langs) == len(train_cuts) and len(langs) == len(dev_cuts) | ||
| printing(f"Langs: {langs}, train-cut: {train_cuts}, dev-cut: {dev_cuts}") | ||
| # pos mode | ||
| FILE_INFIX = "" | ||
| if args.pos_mode == "nope": | ||
| base_opt += " dim_extras:[] extra_names:[]" | ||
| elif args.pos_mode == "pred": | ||
| FILE_INFIX = ".ppos" | ||
| elif args.pos_mode == "gold": | ||
| pass | ||
| else: | ||
| raise NotImplementedError(f"UNK pos_mode {args.pos_mode}") | ||
| # inputs; todo(note): currently using gold for train; test take full lang0 test file | ||
| base_opt += f" train:" + ",".join([f"{DATA_DIR}/{cl}_train.conllu" for cl in langs]) | ||
| base_opt += f" dev:" + ",".join([f"{DATA_DIR}/{cl}_dev{FILE_INFIX}.conllu" for cl in langs]) | ||
| base_opt += f" test:{DATA_DIR}/{langs[0]}_test{FILE_INFIX}.conllu" | ||
| base_opt += f" pretrain_file:" + ",".join([f"{DATA_DIR}/wiki.multi.{cl}.filtered.vec" for cl in langs]) | ||
| base_opt += f" multi_source:1 cut_train:{','.join([str(z) for z in train_cuts])} cut_dev:{','.join([str(z) for z in dev_cuts])} word_rthres:200000" | ||
| # embeds and mbert features | ||
| if args.use_word: | ||
| pass | ||
| else: | ||
| base_opt += " dim_word:0" | ||
| if args.use_mb: | ||
| MB_SUFFIX = args.mb_suffix | ||
| MB_FOLD = args.mb_fold | ||
| base_opt += f" dim_auxes:[768] fold_auxes:[{MB_FOLD}]" | ||
| base_opt += f" aux_repr_train:" + ",".join([f"{DATA_DIR}/{cl}_train.{MB_SUFFIX}" for cl in langs]) | ||
| base_opt += f" aux_repr_dev:" + ",".join([f"{DATA_DIR}/{cl}_dev.{MB_SUFFIX}" for cl in langs]) | ||
| base_opt += f" aux_repr_test:{DATA_DIR}/{langs[0]}_test.{MB_SUFFIX}" | ||
| # todo(note): with bfeats, only smaller lrate works? | ||
| base_opt += " lrate.val:0.0001" | ||
| # different seed for different runnings | ||
| base_opt += f" niconf.random_seed:9347{CUR_RUN} niconf.random_cuda_seed:9349{CUR_RUN}" | ||
| # ===== | ||
| # get it run | ||
| # training | ||
| DEVICE = 0 if RGPU>=0 else -1 | ||
| system(f"CUDA_VISIBLE_DEVICES={RGPU} PYTHONPATH={SRC_DIR} python3 {SRC_DIR}/tasks/cmd.py zdpar.main.train device:{DEVICE} {base_opt} {args.extras} >_train{LOG_SUFFIX} 2>&1") | ||
| # testing | ||
| def _run_decode(decode_langs, which_set): | ||
| dec_opt = "" | ||
| dec_opt += f" test:{DATA_DIR}/${{cl}}_{which_set}{FILE_INFIX}.conllu" | ||
| dec_opt += f" test_extra_pretrain_files:{DATA_DIR}/wiki.multi.${{cl}}.filtered.vec" | ||
| dec_opt += f" output_file:crout_${{cl}}.{which_set}.out" | ||
| if args.model == "ef": | ||
| dec_opt += " aabs.mode:none" # ensure use max beam size | ||
| if args.use_mb: | ||
| dec_opt += f" aux_repr_test:{DATA_DIR}/${{cl}}_{which_set}.{MB_SUFFIX}" | ||
| system(f"for cl in {decode_langs}; do CUDA_VISIBLE_DEVICES={RGPU} PYTHONPATH={SRC_DIR} python3 {SRC_DIR}/tasks/cmd.py zdpar.main.test _conf conf_output: log_file: device:{DEVICE} {dec_opt} {args.extras}; done >_decode.{which_set}{LOG_SUFFIX} 2>&1") | ||
| # | ||
| DECODE_LANGS = "en ar eu zh 'fi' he hi it ja ko ru sv tr" | ||
| _run_decode(DECODE_LANGS, "dev") | ||
| _run_decode(DECODE_LANGS, "test") | ||
|
|
||
| if __name__ == '__main__': | ||
| main() | ||
|
|
||
| # example | ||
| # python3 runm.py --rgpu 2 --model g1 --langs ru --train_cuts 12000 --dev_cuts 1000 --run_dir ? | ||
|
|
||
| # running batch 1 (pos, word, mbert?) | ||
| """ | ||
| # general run of the basic ones | ||
| RGPU=2 | ||
| # ===== | ||
| # run single source | ||
| for cur_model in ef; do | ||
| for cur_run in 1 2 3; do | ||
| for use_mb in 0 1; do | ||
| DIR_NAME="z${cur_model}_ru_b${use_mb}_${cur_run}" | ||
| python3 runm.py --rgpu $RGPU --model ${cur_model} --langs ru --train_cuts 12000 --dev_cuts 1000 --use_word 1 --use_mb ${use_mb} --pos_mode pred --run_dir $DIR_NAME --cur_run ${cur_run} | ||
| done; done; done | ||
| # ===== | ||
| # run multi source | ||
| RGPU=2 | ||
| for cur_model in g1 ef; do | ||
| for cur_run in 1 2 3; do | ||
| for use_mb in 0 1; do | ||
| DIR_NAME="z${cur_model}_er_b${use_mb}_${cur_run}" | ||
| python3 runm.py --rgpu $RGPU --model ${cur_model} --langs en ru --train_cuts 6000 6000 --dev_cuts 500 500 --use_word 1 --use_mb ${use_mb} --pos_mode pred --run_dir $DIR_NAME --cur_run ${cur_run} | ||
| done; done; done | ||
| # ===== | ||
| # test other ef mode | ||
| python3 runm.py --rgpu 3 --model ef --langs en --train_cuts 12000 --dev_cuts 1000 --use_word 1 --use_mb 0 --pos_mode pred --run_dir test_t2d --cur_run 1 --extras "iconf.ef_mode:t2d tconf.ef_mode:t2d" | ||
| """ | ||
|
|
||
| # final running batch | ||
| """ | ||
| # ===== | ||
| # ----- | ||
| RGPU=2 | ||
| # run single source | ||
| for cur_model in g1 ef; do | ||
| for cur_run in 1 2 3 4 5; do | ||
| for use_mb in 1; do | ||
| for cl in en ru; do | ||
| DIR_NAME="z${cur_model}_${cl}_${cur_run}" | ||
| python3 runm.py --rgpu $RGPU --model ${cur_model} --langs ${cl} --train_cuts 12000 --dev_cuts 1000 --use_word 1 --use_mb ${use_mb} --pos_mode pred --run_dir $DIR_NAME --cur_run ${cur_run} | ||
| done; done; done; done | ||
| # ----- | ||
| RGPU=3 | ||
| # run multi source | ||
| for cur_model in g1 ef; do | ||
| for cur_run in 1 2 3 4 5; do | ||
| for use_mb in 1; do | ||
| DIR_NAME="z${cur_model}_er_${cur_run}" | ||
| python3 runm.py --rgpu $RGPU --model ${cur_model} --langs en ru --train_cuts 6000 6000 --dev_cuts 500 500 --use_word 1 --use_mb ${use_mb} --pos_mode pred --run_dir $DIR_NAME --cur_run ${cur_run} | ||
| done; done; done | ||
| # ----- | ||
| RGPU=3 | ||
| # run other ef mode | ||
| for ef_mode in t2d n2f l2r r2l; do | ||
| for cur_run in 1 2 3; do | ||
| for cur_arg in "--run_dir zana_en_${ef_mode}_${cur_run} --langs en --train_cuts 12000 --dev_cuts 1000" "--run_dir zana_ru_${ef_mode}_${cur_run} --langs ru --train_cuts 12000 --dev_cuts 1000" "--run_dir zana_er_${ef_mode}_${cur_run} --langs en ru --train_cuts 6000 6000 --dev_cuts 500 500"; do | ||
| python3 runm.py --rgpu $RGPU --model ef --use_word 1 --use_mb 1 --pos_mode pred --cur_run ${cur_run} --extras "iconf.ef_mode:${ef_mode} tconf.ef_mode:${ef_mode}" ${cur_arg} | ||
| done; done; done | ||
| # ----- | ||
| RGPU=2 | ||
| # run with various features for ef | ||
| for cur_feat in "00" "01" "10"; do | ||
| for cur_run in 1 2 3; do | ||
| for cur_arg in "--run_dir zfeat_en_${cur_feat}_${cur_run} --langs en --train_cuts 12000 --dev_cuts 1000" "--run_dir zfeat_ru_${cur_feat}_${cur_run} --langs ru --train_cuts 12000 --dev_cuts 1000" "--run_dir zfeat_er_${cur_feat}_${cur_run} --langs en ru --train_cuts 6000 6000 --dev_cuts 500 500"; do | ||
| if [[ ${cur_feat} == "00" ]]; then | ||
| cur_feat_arg="sl_conf.use_chs:0 sl_conf.use_par:0"; | ||
| elif [[ ${cur_feat} == "01" ]]; then | ||
| cur_feat_arg="sl_conf.use_chs:0 sl_conf.use_par:1"; | ||
| elif [[ ${cur_feat} == "10" ]]; then | ||
| cur_feat_arg="sl_conf.use_chs:1 sl_conf.use_par:0"; | ||
| fi | ||
| python3 runm.py --rgpu $RGPU --model ef --use_word 1 --use_mb 1 --pos_mode pred --cur_run ${cur_run} --extras "${cur_feat_arg}" ${cur_arg} | ||
| done; done; done | ||
| """ |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,77 @@ | ||
| #!/usr/bin/env bash | ||
|
|
||
| # train from g1 init | ||
|
|
||
| # ===== | ||
| function run_one () { | ||
| # check envs | ||
| echo "The environment variables: RGPU=${RGPU} CUR_LANG=${CUR_LANG} CUR_RUN=${CUR_RUN}"; | ||
|
|
||
| # ===== | ||
| # basis (same as basic g1 for architecture) | ||
| #base_opt="conf_output:_conf init_from_pretrain:1 drop_embed:0.33 dropmd_embed:0.33 singleton_thr:2 fix_drop:1" | ||
| base_opt="conf_output:_conf init_from_pretrain:1 drop_embed:0.2 dropmd_embed:0.2 singleton_thr:2 fix_drop:1" | ||
| base_opt+=" arc_space:512 lab_space:128 transform_act:elu biaffine_div:1. biaffine_init_ortho:1" | ||
|
|
||
| # ===== | ||
| # lrate and margin | ||
| if [[ -n "${G1MODEL}" ]]; then | ||
| # pre-load and pre-valid | ||
| base_opt+=" g1_pretrain_path:${G1MODEL} g1_pretrain_init:1 validate_first:1 lambda_g1_arc_testing:0. lambda_g1_lab_testing:0." | ||
| base_opt+=" lrate_warmup:0 lrate.val:0.0001 tconf.batch_size:32 split_batch:1 lrate.m:0.75 min_save_epochs:10 max_epochs:200 patience:5 anneal_times:8" | ||
| base_opt+=" margin.val:0.5 margin.max_val:0.5 margin.mode:linear margin.b:0.5 margin.k:0.01 reg_scores_lambda:0.01" | ||
| base_opt+=" enc_lrf.val:1. enc_lrf.mode:none enc_lrf.start_bias:0 enc_lrf.b:0. enc_lrf.k:1. enc_optim.weight_decay:0." | ||
| base_opt+=" mid_lrf.val:1. mid_lrf.mode:none mid_lrf.start_bias:0 mid_lrf.b:0. mid_lrf.k:1. mid_optim.weight_decay:0." | ||
| base_opt+=" dec_lrf.val:0. dec_lrf.mode:none dec_lrf.start_bias:0 dec_lrf.b:0. dec_lrf.k:1. dec_optim.weight_decay:0." | ||
| arc_k=5 | ||
| else | ||
| # train from start | ||
| base_opt+=" lrate_warmup:0 lrate.val:0.001 tconf.batch_size:32 split_batch:1 lrate.m:0.75 min_save_epochs:150 max_epochs:300 patience:8 anneal_times:10" | ||
| base_opt+=" margin.val:0.5 margin.max_val:0.5 margin.mode:linear margin.b:0. margin.k:0.01 reg_scores_lambda:0.01" | ||
| base_opt+=" enc_lrf.val:1. enc_lrf.mode:none enc_lrf.start_bias:0 enc_lrf.b:0. enc_lrf.k:1. enc_optim.weight_decay:0." | ||
| base_opt+=" mid_lrf.val:1. mid_lrf.mode:none mid_lrf.start_bias:0 mid_lrf.b:0. mid_lrf.k:1. mid_optim.weight_decay:0." | ||
| base_opt+=" dec_lrf.val:0. dec_lrf.mode:none dec_lrf.start_bias:0 dec_lrf.b:0. dec_lrf.k:1. dec_optim.weight_decay:0." | ||
| base_opt+=" aabs.val:5. aabs.max_val:5. aabs.mode:linear aabs.start_bias:100 aabs.k:0.01" | ||
| arc_k=1 | ||
| fi | ||
|
|
||
| # system | ||
| base_opt+=" partype:ef zero_extra_output_params:0 iconf.ef_mode:free tconf.ef_mode:free system_labeled:1" | ||
| base_opt+=" chs_num:10 chs_att.out_act:linear chs_att.att_dropout:0.1 chs_att.head_count:2" | ||
| # loss | ||
| base_opt+=" tconf.ending_mode:maxv cost0_weight.val:1. loss_div_weights:0 loss_div_fullbatch:1" | ||
| # features for ef | ||
| base_opt+=" fdrop_chs:0. fdrop_par:0. use_label_feat:1 use_chs:1 use_par:1 ef_ignore_chs:func" | ||
| # search | ||
| #arc_k=1 # set previously | ||
| lab_k=1 | ||
| base_opt+=" iconf.plain_k_arc:${arc_k} iconf.plain_k_label:${lab_k} iconf.plain_beam_size:${arc_k}" | ||
| base_opt+=" tconf.plain_k_arc:${arc_k} tconf.plain_k_label:${lab_k} tconf.plain_beam_size:${arc_k}" | ||
|
|
||
| # different POS modes | ||
| if [[ "${MPOS}" == "nope" ]]; then | ||
| # no pos input | ||
| base_opt+=" dim_extras:[] extra_names:[]" | ||
| elif [[ "${MPOS}" == "joint" ]]; then | ||
| # no pos input | ||
| base_opt+=" dim_extras:[] extra_names:[]" | ||
| # jpos (split one layer for jpos) | ||
| base_opt+=" enc_conf.enc_rnn_layer:2 jpos_multitask:1 jpos_enc.enc_rnn_layer:1" | ||
| # jpos lambda | ||
| base_opt+=" jpos_lambda:0.1" | ||
| fi | ||
|
|
||
| # run | ||
| SRC_DIR="../src/" | ||
| DATA_DIR="../data/UD_RUN/ud23/" | ||
| # | ||
| CUDA_VISIBLE_DEVICES=${RGPU} PYTHONPATH=${SRC_DIR} python3 ${SRC_DIR}/tasks/cmd.py zdpar.main.train train:${DATA_DIR}/${CUR_LANG}_train.conllu dev:${DATA_DIR}/${CUR_LANG}_dev.conllu test:${DATA_DIR}/${CUR_LANG}_test.conllu pretrain_file:${DATA_DIR}/wiki.${CUR_LANG}.filtered.vec device:0 ${base_opt} "niconf.random_seed:9347${CUR_RUN}" "niconf.random_cuda_seed:9349${CUR_RUN}" | ||
| } | ||
|
|
||
| # actual run | ||
| #G1MODEL=../zen_ef/zmodel.best CUR_LANG=en CUR_RUN=1 run_one # set init0=false | ||
| #G1MODEL=../zen_g1/zmodel.best CUR_LANG=en CUR_RUN=1 run_one | ||
| #G1MODEL='' CUR_LANG=en CUR_RUN=1 run_one | ||
| G1MODEL='' CUR_LANG=en CUR_RUN=1 MPOS=nope run_one | ||
|
|
||
| # RGPU=0 bash _go.sh |& tee log |
Oops, something went wrong.