In order to perform model inference clone this repository and please download the finetuned model and config from here. Create a directory called 'checkpoints' inside the 'document_classification' directory and place the 'pytorch_model.bin' and 'config.json' files inside the 'checkpoints' directory. The directory structure should look like as follows:
Document-Classification-LayoutLMv3
|
|--document_classifiation
| |-checkpoints
| |-Dataset
| |-all the .py files
|
|
|--test_imgs
|--Dockerfile
Once the checkpoints have been placed in the correct directories, the docker image must be built using the following command.
docker build -t doc_classification:latest .
After the docker image has been built, it can be used for inference by first running the docker using the following command:
docker run --name document_classification --gpus all -itd -v d:/zain_dev/python_dev/Document-Classification-LayoutLMv3/document_classification:/document_classification -v d:/zain_dev/python_dev/Document-Classification-LayoutLMv3/test_imgs:/images_dir doc_classification:latest
Note that two volumes are mounted here. The first volume contains all the code required to run the mode, the second volume is a directory where all the test images must need to be placed. The docker is made detachable and interactable.
Once the docker has been been run, the docker exec command is used to execute commands. In this case the training script is executed using the python interpretter using the following command.
docker exec document_classification python train.py
The model checkpoints are stored inside the checkpoints directory.
When running for the first time the base model will be downloaded, which will be fine tuned on custom datasetl.
The docker exec command is used to make inference on files that have been placed inside the test_imgs directory. The name of the image is specified using an environment variable called TEST_IMG_NAME. The model inference is given below:
docker exec -e "TEST_IMG_NAME=doc_000051.png" document_classification python inference.py
The model prediction is logged to the screen.