Sample config.yaml and json file for inference only

[SinaHonari]

I want to run your model on some images I have (not your datasets) and the yaml files of the dataset do not work and I cannot find the right config for them. Also the jason sample you provide, does not clearly show how to set the values. I would appreciate if you can provide a simple examples (e.g. taking two images and comparing them) for both config.yaml and json files.

[zhiyuanyou]

Hello,

For single image assessment,

1. Construct meta file (let us name it dir_path/single1.json) like:

[
    {
        "id": set a unique id of each sample, required,
        "image_ref": reference image path [if full-reference] / null [if non-reference],
        "image_A": image path,
        "image_B": null,
        "query": input question like "Could you assess the overall quality of the image and elaborate on your evaluation?", 
    }, 
    ...
]

2. Construct inference shell like:

#!/bin/bash
src_dir=dir_path/DepictQA/src
export PYTHONPATH=$src_dir:$PYTHONPATH
export CUDA_VISIBLE_DEVICES=$1

python $src_dir/infer.py \
    --meta_path dir_path/single1.json \
                dir_path/single2.json \
    --dataset_name name1 \
                   name2 \
    --task_name quality_single_A_noref [if non-reference]  / quality_single_A [if full-reference] \
    --batch_size 16 \

For pair-wise image comparison:

1. Construct meta file (let us name it dir_path/pair1.json) like:

[
    {
        "id": set a unique id of each sample, required,
        "image_ref": reference image path [if full-reference] / null [if non-reference],
        "image_A": image A path,
        "image_B": image B path,
        "query": input question like "Could you assess the overall quality of the image and elaborate on your evaluation?", 
    }, 
    ...
]

Construct the inference shell like:

#!/bin/bash
src_dir=dir_path/DepictQA/src
export PYTHONPATH=$src_dir:$PYTHONPATH
export CUDA_VISIBLE_DEVICES=$1

python $src_dir/infer.py \
    --meta_path dir_path/pair1.json \
                dir_path/pair2.json \
    --dataset_name name1 \
                   name2 \
    --task_name quality_compare_noref [if non-reference]  / quality_compare [if full-reference] \
    --batch_size 16 \

You can reuse any existing config files, and the command line parameters (--meta_path, --dataset_name, --task_name, --batch_size) will overwrite the config file.

[SinaHonari]

I'm using it as above to compare a pair of images in terms of noise, and asking the model to output which image is cleaner?, but it yields a null output.

I have my json file as below named data_comp1.json

[
    {
        "id": "000, 111",
        "image_ref": null,
        "image_A": "DepictQA/test_data/test_1/noisy_image/img_A.png",
        "image_B": "DepictQA/test_data/test_1/noisy_image/img_B.png",
        "query": "look at image_A and image_B and explain which image is cleaner or has hess noise and explain why?"
    }
]

The path to both files exist and there is no reference image.

Here is the bash file

#!/bin/bash
src_dir=code_base/DepictQA/src
export PYTHONPATH=$src_dir:$PYTHONPATH
export CUDA_VISIBLE_DEVICES=$1

python -m pdb $src_dir/infer.py \
    --meta_path code_base/DepictQA/test_data/test_1/data_comp1.json \
    --dataset_name comp1 \
    --task_name quality_compare_noref \
    --batch_size 2 \

Do you see an error in the format of the messages?