model_conf

output_name="seq2seq"

init_model="./model_files/unimo_base_en"
data_path='./data/gigaword'

# hyper params
lr_scheduler="linear_warmup_decay"
use_fp16="False"
# Merge the ALLReduce times of a layer
use_fuse="True"
use_hierarchical_allreduce="True"
loss_scaling=12800

skip_steps=100
save_steps=20000
validation_steps=20000
label_smooth=0.1
weight_decay=0.01
max_seq_len=512

#decoding params
do_decode="true"
max_src_len=128
max_tgt_len=32
max_out_len=32
min_out_len=5
beam_size=5
length_penalty=0.6
block_trigram="False"
use_multi_gpu_test="True"

#adam optimizer
beta1=0.9
beta2=0.98
epsilon=1e-06

#data
tokenized_input="True"
continuous_position="False"

#dataset
train_set="train.tsv"
dev_set="dev.20k.tsv"
pred_set="test.tsv"
test_set="test.tsv"
do_train="true"
do_val="false"
do_test="true"
do_pred="false"

#evaluate
eval_script="bash ./src/eval/tasks/gigaword/eval.sh"
eval_mertrics="rouge-1,rouge-2,rouge-l"


## turning params
in_tokens="False"
pred_batch_size=32
epoch=10
BATCH_SIZE=("32")
LR_RATE=("3e-5")
DD_RAND_SEED=("1")
WARMUP_PROP=("0.06")

config_path="./model_files/config/unimo_base_en.json"
vocab_file="./model_files/dict/unimo_en.vocab.txt"
bpe_json="./model_files/dict/unimo_en.encoder.json"
bpe_file="./model_files/dict/unimo_en.vocab.bpe"