TianyuZhang commited on
Commit
eac4450
1 Parent(s): f19ec6c

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +55 -15
README.md CHANGED
@@ -56,24 +56,64 @@ We found that OCR and text-based processing become ineffective in VCR as accurat
56
  - **Paper:** [VCR: Visual Caption Restoration](https://arxiv.org/abs/2406.06462)
57
  - **Point of Contact:** [Tianyu Zhang](mailto:tianyu.zhang@mila.quebec)
58
 
59
- ## Evaluation
60
 
61
- We recommend you to evaluate your model with [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval). Before evaluating, please refer to the doc of `lmms-eval`.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
62
 
63
- ```console
64
- pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git
 
 
 
 
 
 
 
 
 
 
 
 
 
65
 
66
- # We use MiniCPM-Llama3-V-2_5 and vcr_wiki_en_easy as an example
67
- python3 -m accelerate.commands.launch \
68
- --num_processes=8 \
69
- -m lmms_eval \
70
- --model minicpm_v \
71
- --model_args pretrained="openbmb/MiniCPM-Llama3-V-2_5" \
72
- --tasks vcr_wiki_en_easy \
73
- --batch_size 1 \
74
- --log_samples \
75
- --log_samples_suffix MiniCPM-Llama3-V-2_5_vcr_wiki_en_easy \
76
- --output_path ./logs/
 
 
 
 
 
 
 
 
 
 
 
77
  ```
78
 
79
  `lmms-eval` supports the following VCR `--tasks` settings:
 
56
  - **Paper:** [VCR: Visual Caption Restoration](https://arxiv.org/abs/2406.06462)
57
  - **Point of Contact:** [Tianyu Zhang](mailto:tianyu.zhang@mila.quebec)
58
 
59
+ # Model Evaluation
60
 
61
+ ## Method 1 (recommended): use the evaluation script
62
+ ```bash
63
+ git clone https://github.com/tianyu-z/VCR.git
64
+ ```
65
+ ### Open-source evaluation
66
+ We support open-source model_id:
67
+ ```python
68
+ ["openbmb/MiniCPM-Llama3-V-2_5",
69
+ "OpenGVLab/InternVL-Chat-V1-5",
70
+ "internlm/internlm-xcomposer2-vl-7b",
71
+ "HuggingFaceM4/idefics2-8b",
72
+ "Qwen/Qwen-VL-Chat",
73
+ "THUDM/cogvlm2-llama3-chinese-chat-19B",
74
+ "THUDM/cogvlm2-llama3-chat-19B",
75
+ "echo840/Monkey-Chat",]
76
+ ```
77
+ For the models not on list, they are not intergated with huggingface, please refer to their github repo to create the evaluation pipeline.
78
 
79
+ ```bash
80
+ # We use HuggingFaceM4/idefics2-8b and vcr_wiki_en_easy as an example
81
+ # Inference from the VLMs and save the results to {model_id}_{difficulty}_{language}.json
82
+ cd src/evaluation
83
+ python3 inference.py --dataset_handler "vcr-org/VCR-wiki-en-easy-test" --model_id "HuggingFaceM4/idefics2-8b" --device "cuda" --dtype "bf16" --save_interval 50 --resume True
84
+
85
+ # Evaluate the results and save the evaluation metrics to {model_id}_{difficulty}_{language}_evaluation_result.json
86
+ python3 evaluation_metrics.py --model_id HuggingFaceM4/idefics2-8b --output_path . --json_filename "HuggingFaceM4_idefics2-8b_en_easy.json" --dataset_handler "vcr-org/VCR-wiki-en-easy-test"
87
+
88
+ # To get the mean score of all the `{model_id}_{difficulty}_{language}_evaluation_result.json` in `jsons_path` (and the std, confidence interval if `--bootstrap`) of the evaluation metrics
89
+ python3 gather_results.py --jsons_path .
90
+ ```
91
+
92
+ ### Close-source evaluation
93
+ We provide the evaluation script for the close-source model: `GPT-4o`, `GPT-4-Turbo`, `Claude-3-Opus` in the `evaluation` folder.
94
 
95
+ You need an API Key, a pre-saved testing dataset and specify the path of the data saving the paper
96
+ ```bash
97
+ cd src/evaluation
98
+ # save the testing dataset to the path
99
+ python3 save_image_from_dataset.py --output_path .
100
+
101
+ # Inference Put your API key and Image Path in the evaluation script (e.g. gpt-4o.py)
102
+ python3 gpt-4o.py
103
+
104
+ # Evaluate the results and save the evaluation metrics to {model_id}_{difficulty}_{language}_evaluation_result.json
105
+ python3 evaluation_metrics.py --model_id gpt4o --output_path . --json_filename "gpt4o_en_easy.json" --dataset_handler "vcr-org/VCR-wiki-en-easy-test"
106
+
107
+ # To get the mean score of all the `{model_id}_{difficulty}_{language}_evaluation_result.json` in `jsons_path` (and the std, confidence interval if `--bootstrap`) of the evaluation metrics
108
+ python3 gather_results.py --jsons_path .
109
+ ```
110
+
111
+ ## Method 2: use lmms-eval framework
112
+ You may need to incorporate the inference method of your model if the lmms-eval framework does not support it. For details, please refer to [here](https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/docs/model_guide.md)
113
+ ```bash
114
+ pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git
115
+ # We use HuggingFaceM4/idefics2-8b and vcr_wiki_en_easy as an example
116
+ python3 -m accelerate.commands.launch --num_processes=8 -m lmms_eval --model idefics2 --model_args pretrained="HuggingFaceM4/idefics2-8b" --tasks vcr_wiki_en_easy --batch_size 1 --log_samples --log_samples_suffix HuggingFaceM4_idefics2-8b_vcr_wiki_en_easy --output_path ./logs/
117
  ```
118
 
119
  `lmms-eval` supports the following VCR `--tasks` settings: