Update instructions for ScienceQA

Author: haotian-liu

.gitignore

@@ -6,6 +6,7 @@ dist
 
 # Log
 *.log
+*.log.*
 *.json
 
 # Data

README.md

@@ -174,7 +174,57 @@ python summarize_gpt_review.py
 
 ### ScienceQA
 
-Please see ScienceQA [repo](https://github.com/lupantech/ScienceQA) for setting up the dataset.  You may either use the official ScienceQA evaluation script or [our script](eval_science_qa.py) to evaluate the model.
+#### Prepare Data
+1. Please see ScienceQA [repo](https://github.com/lupantech/ScienceQA) for setting up the dataset.
+2. Generate ScienceQA dataset for LLaVA conversation-style format.
+
+```Shell
+python scripts/convert_sqa_to_llava \
+    convert_to_llava \
+    --base-dir /path/to/ScienceQA/data/scienceqa \
+    --split {train,val,minival,test,minitest}
+```
+
+#### Evaluation
+
+1. Download our pretrained LLaVA-13B (delta) weights for ScienceQA dataset [here](https://huggingface.co/liuhaotian/LLaVA-13b-delta-v0-science_qa).  Convert the delta weights to actual weights following instructions [here](https://github.com/haotian-liu/LLaVA#llava-13b), and make sure to modify the command accordingly for ScienceQA.
+
+2. Generate LLaVA responses on ScienceQA dataset
+
+```Shell
+python -m llava.eval.model_vqa_science \
+    --model-name /path/to/LLaVA-13b-v0-science_qa \
+    --question-file /path/to/ScienceQA/data/scienceqa/llava_test.json \
+    --image-folder /path/to/ScienceQA/data/scienceqa/images/test \
+    --answers-file vqa/results/ScienceQA/test_llava-13b.jsonl
+```
+
+Alternatively, you may evaluate this with multiple GPUs, and concatenate the generated jsonl files.
+
+```Shell
+CHUNKS=8
+CHUNK_IDX=0
+CUDA_VISIBLE_DEVICES=CHUNK_IDX python model_vqa_science.py \
+    --model-name /path/to/LLaVA-13b-v0-science_qa \
+    --question-file /path/to/ScienceQA/data/scienceqa/llava_test.json \
+    --image-folder /path/to/ScienceQA/data/scienceqa/images/test \
+    --answers-file vqa/results/ScienceQA/test_llava-13b-chunk${CHUNKS}_${CHUNK_IDX}.jsonl \
+    --num-chunks $CHUNKS \
+    --chunk-idx $CHUNK_IDX
+
+# after running this for all chunks, concatenate the results
+cat {...} > vqa/results/ScienceQA/test_llava-13b.jsonl
+```
+
+3. Evaluate the generated responses
+
+```Shell
+python eval_science_qa.py \
+    --base-dir /path/to/ScienceQA/data/scienceqa \
+    --result-file vqa/results/ScienceQA/test_llava-13b.jsonl \
+    --output-file vqa/results/ScienceQA/test_llava-13b_output.json \
+    --result-file vqa/results/ScienceQA/test_llava-13b_result.json \
+```
 
 ## Fine-tuning
 ### Data

llava/eval/model_vqa_science.py

@@ -292,11 +292,11 @@ def eval_model(args):
     parser = argparse.ArgumentParser()
     parser.add_argument("--model-name", type=str, default="facebook/opt-350m")
     parser.add_argument("--image-folder", type=str, default="")
-    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--question-file", type=str, default="tables/question.json")
     parser.add_argument("--answers-file", type=str, default="answer.jsonl")
     parser.add_argument("--mm-projector", type=str, default=None)
     parser.add_argument("--vision-tower", type=str, default=None)
-    parser.add_argument("--conv-mode", type=str, default="default")
+    parser.add_argument("--conv-mode", type=str, default="simple")
     parser.add_argument("--num-chunks", type=int, default=1)
     parser.add_argument("--chunk-idx", type=int, default=0)
     parser.add_argument("--answer-prompter", action="store_true")

scripts/convert_sqa_to_llava.py

@@ -0,0 +1,88 @@
+import json
+import os
+import fire
+import re
+from convert_sqa_to_llava_base_prompt import build_prompt_chatbot
+
+
+def convert_to_llava(base_dir, split, prompt_format="QCM-LEPA"):
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+
+    split_problems = build_prompt_chatbot(
+        problems, split_indices, prompt_format,
+        use_caption=False, is_test=False)
+
+    target_format = []
+    for prob_id, (input, output) in split_problems.items():
+        if input.startswith('Question: '):
+            input = input.replace('Question: ', '')
+        if output.startswith('Answer: '):
+            output = output.replace('Answer: ', '')
+
+        raw_prob_data = problems[prob_id]
+        if raw_prob_data['image'] is None:
+            target_format.append({
+                "id": prob_id,
+                "conversations": [
+                    {'from': 'human', 'value': f"{input}"},
+                    {'from': 'gpt', 'value': f"{output}"},
+                ],
+            })
+
+        else:
+            target_format.append({
+                "id": prob_id,
+                "image": os.path.join(prob_id, raw_prob_data['image']),
+                "conversations": [
+                    {'from': 'human', 'value': f"{input}\n<image>"},
+                    {'from': 'gpt', 'value': f"{output}"},
+                ],
+            })
+
+    print(f'Number of samples: {len(target_format)}')
+
+    with open(os.path.join(base_dir, f"llava_{split}_{prompt_format}.json"), "w") as f:
+        json.dump(target_format, f, indent=2)
+
+
+def convert_to_jsonl(base_dir, split, prompt_format="QCM-LEPA"):
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+
+    split_problems = build_prompt_chatbot(
+        problems, split_indices, prompt_format,
+        use_caption=False, is_test=False)
+
+    writer = open(os.path.join(base_dir, f"scienceqa_{split}_{prompt_format}.jsonl"), "w")
+    for prob_id, (input, output) in split_problems.items():
+        if input.startswith('Question: '):
+            input = input.replace('Question: ', '')
+        if output.startswith('Answer: '):
+            output = output.replace('Answer: ', '')
+
+        raw_prob_data = problems[prob_id]
+        if raw_prob_data['image'] is None:
+            data = {
+                "id": prob_id,
+                "instruction": f"{input}",
+                "output": f"{output}",
+            }
+
+        else:
+            data = {
+                "id": prob_id,
+                "image": os.path.join(prob_id, raw_prob_data['image']),
+                "instruction": f"{input}\n<image>",
+                "output": f"{output}",
+            }
+        writer.write(json.dumps(data) + '\n')
+    writer.close()
+
+
+def main(task, **kwargs):
+    globals()[task](**kwargs)
+
+
+if __name__ == "__main__":
+    fire.Fire(main)