You can retrain the base decoder of ChartVLM and reproduce our results via the following steps:

cd base_decoder/train

The following datasets are used in our paper:

• ChartQA [Dataset Page]
• PlotQA [Dataset Page]
• Chart2Text [Dataset Page]
• SimChart9K [Download]

• In order to speed up the data i/o during the training process for base_decoder, we choose to preprocess the downloaded chart data, saving as the .npy format.

• You have to first preprocess the data before starting the training process (This should be the absolute path of the downloaded datasets)

cd tools/data_preprocess/
# Change the root path for the downloaded ChartQA or Chart2Text dataset
python data_preprocess_ChartQA_Chart2Text.py

cd tools/data_preprocess/
# Change the root path for the downloaded PlotQA dataset
python data_preprocess_PlotQA.py

cd tools/data_preprocess/
# Change the root path for the downloaded SimChart9K dataset
python data_preprocess_SimChart9K.py

# return to the 'tools' directory
cd ..

• Train the Base Model using multi-GPU

sh scripts/dist_train.sh 8 \
--config ./cfgs/image_to_csv_base_merge_all_trained.yaml \
--VAL_PER_EPOCH 0

• Train the Base Model using multi-machines

sh scripts/slurm_train.sh ${PARTITION} ${JOB_NAME} ${NUM_NODES} \
--cfg_file ./cfgs/image_to_csv_base_merge_all_trained.yaml \
--VAL_PER_EPOCH 0

• Train the Large Model using multi-GPU

sh scripts/dist_train.sh 8 \
--config ./cfgs/image_to_csv_large_merge_all_trained.yaml \
--VAL_PER_EPOCH 0

• Train the Large Model using multi-machines

sh scripts/slurm_train.sh ${PARTITION} ${JOB_NAME} ${NUM_NODES} \
--cfg_file ./cfgs/image_to_csv_large_merge_all_trained.yaml \
--VAL_PER_EPOCH 0

• Evaluate the Model using multi-GPU, SCRM metric, and 1280 output tokens

sh scripts/dist_test.sh 4 \
--config ./cfgs/image_to_csv_large_merge_all_trained.yaml \
--criterion csv_metric \
--num_token 1280