model_conf

output_name="seq2seq"

init_model="./model_files/unimo_large_en"
data_path='./data/coqa'

# hyper param
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=10000
validation_steps=10000
label_smooth=0.1
weight_decay=0.01
max_seq_len=512
random_seed=666

#for multi-turn dialog/qa
task_type="dialog"
role_type_size=3
turn_type_size=16

#decoding params
do_decode="true"
max_src_len=480
max_tgt_len=32
max_out_len=30
min_out_len=0
beam_size=3
length_penalty=0.0
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.tsv"
test_set="dev.tsv"
do_train="true"
do_val="true"
do_test="false"
do_pred="false"

#evaluate
eval_script="bash ./src/eval/tasks/coqa/eval.sh"
eval_mertrics="f1"


## turning params
in_tokens="False"
pred_batch_size=8
epoch=20
BATCH_SIZE=("8")
LR_RATE=("8e-6")
DD_RAND_SEED=("1")
WARMUP_PROP=("0.06")

config_path="./model_files/config/unimo_large_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"