|
#!/bin/bash |
|
|
|
SCRIPT_DIR=. |
|
MODEL_DIR=/mnt/disks/data-1/models/training_v4_unfreeze |
|
|
|
IMAGE_ENCODER="openai/clip-vit-base-patch32" |
|
TEXT_ENCODER="flax-community/indonesian-roberta-base" |
|
|
|
python ${SCRIPT_DIR}/run_hybrid_clip.py \ |
|
--output_dir ${MODEL_DIR} \ |
|
--overwrite_output_dir \ |
|
--tokenizer_name=${TEXT_ENCODER} \ |
|
--train_file="../data/train_dataset_v6.json" \ |
|
--validation_file="../data/val_dataset_v6.json" \ |
|
--do_train --do_eval \ |
|
--num_train_epochs="20" --max_seq_length 96 \ |
|
--per_device_train_batch_size="64" \ |
|
--per_device_eval_batch_size="64" \ |
|
--learning_rate="0.00001" --warmup_ratio 0.1 --weight_decay 0.0 \ |
|
--preprocessing_num_workers 16 \ |
|
--exp_name training_v4_unfreeze \ |
|
--text_model_name_or_path=${TEXT_ENCODER} \ |
|
--vision_model_name_or_path=${IMAGE_ENCODER} \ |
|
--eval_steps 500 \ |
|
--logging_steps 100 \ |
|
--save_steps 500 \ |
|
--save_total_limit 5 \ |
|
--log_wandb \ |
|
--run_from_checkpoint="/mnt/disks/data-1/models/training_v4/ckpt-70999" |
|
|
|
|
|
|
