-
Notifications
You must be signed in to change notification settings - Fork 0
/
config.yaml
53 lines (53 loc) · 1.35 KB
/
config.yaml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
# General (hyper)parameters
params:
seed: 777
do_train: True
do_eval: True
do_predict: True
src: en_XX # See https://huggingface.co/facebook/mbart-large-50
tgt: fa_IR # See https://huggingface.co/facebook/mbart-large-50
lower_case: True # Tokenizer
normalization: True # Tokenizer
checkpoint: facebook/mbart-large-50
max_len: 100
# dataloader (bs = batch size)
train_bs: 8
dev_bs: 8
test_bs: 8
# bs per device
device_train_bs: 8
device_eval_bs: 8
epoch: 2
# log, save, and evaluate every n steps
logging_step: 1000
save_steps: 1000
eval_steps: 1000
save: 30
early_stop: 5
weight_decay: 0
warmup_steps: 0
fp16: False # Only can be used on CUDA devices
lr: 5e-06
evaluation: steps
best_model: True
optim: adafactor # adafactor, adamw_torch default:adamw_hf
hidden_dropout_prob: 0.1
# eval metric
metric_best_model: eval_loss
predict_with_generate: True # Otherwise you get: out of range integral type conversion attempted
use_mps_device: True # Using Apple Silicon chip
# Dataset setup
dataset:
train_path: datasets/EN-FA/train.csv
dev_path: datasets/EN-FA/dev.csv
test_path: datasets/EN-FA/test.csv
predict_path: datasets/EN-FA/en-fa.txt # Inference
# Splitter setup
splitter:
active: True
path: datasets/EN-FA/en-fa.txt
split: 0.003
# MLFlow config
mlflow:
exp_name: EN-FA-MT
params: "_1"