MODEL_DIR=./output
BASE_L_MODEL=cardiffnlp/twitter-roberta-base
BASE_LL_MODEL=cardiffnlp/twitter-roberta-base-emoji
# BASE_LL_MODEL=cardiffnlp/twitter-roberta-base-emotion
# BASE_LL_MODEL=cardiffnlp/twitter-roberta-base-sentiment
# BASE_L_MODEL=roberta-base
# BASE_L_MODEL=bert-large-cased



BASE_V_MODEL=openai/clip-vit-base-patch32

# BASE_V_MODEL=google/vit-base-patch16-384
# BASE_V_MODEL=google/vit-base-patch32-384
# BASE_V_MODEL=timm/vit_huge_patch14_224_in21k

#USE_TORCH=false
python run_hybrid_clip.py \
    --output_dir ${MODEL_DIR} \
    --text_model_name_or_path=${BASE_LL_MODEL}  \
    --vision_model_name_or_path=${BASE_V_MODEL} \
    --tokenizer_name=${BASE_L_MODEL}  \
    --train_file="/home/ceyda/data/train.json" \
    --validation_file="/home/ceyda/data/val.json"  \
    --do_train --do_eval \
    --num_train_epochs="40" --max_seq_length 128 \
    --per_device_train_batch_size="32" \
    --per_device_eval_batch_size="8" \
    --learning_rate="1e-5" --warmup_steps="150" --weight_decay 0.1 \
    --overwrite_output_dir \
    --preprocessing_num_workers 32 
    # --push_to_hub