File size: 1,006 Bytes
fdc4101
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d33873d
446c112
fdc4101
565e5fc
 
d33873d
55bd44c
565e5fc
55bd44c
fdc4101
 
d33873d
 
fdc4101
 
 
e867aeb
446c112
e867aeb
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
MODEL="t5-base-dutch"

MODEL_DIR="${HOME}/${MODEL}"

mkdir -p "${MODEL_DIR}/runs"

# T5 paper lr 0.01 with batch size 128
# We have a batch size of 8 devices * 32 = 256, so lr = 0.01/2

./run_t5_mlm_flax_custom_dataset.py \
    --output_dir="${MODEL_DIR}" \
    --model_type="t5" \
    --config_name="flax-community/${MODEL}" \
    --tokenizer_name="${MODEL_DIR}" \
    --preprocessing_num_workers="96" \
    --do_train --do_eval \
    --adafactor \
    --max_seq_length="512" \
    --per_device_train_batch_size="32" \
    --per_device_eval_batch_size="32" \
    --learning_rate="5e-3" \
    --dtype="bfloat16" \
    --overwrite_output_dir \
    --num_train_epochs="3" \
    --logging_steps="50" \
    --save_steps="2000" \
    --eval_steps="10000000" \
    --resume_from_checkpoint="${MODEL_DIR}/ckpt-16000" \
    --warmup_steps="3413" \
    --push_to_hub



#git add pytorch_model.bin
#git commit -m "Update pytorch model after training"
#git push origin main

#    --gradient_accumulation_steps="2" \