Upload LLaVA-Next-3D/data_precessing/llm_analyze.py with huggingface_hub

LLaVA-Next-3D/data_precessing/llm_analyze.py

@@ -0,0 +1,55 @@
+import pdb
+import json
+from tqdm import tqdm
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import argparse
+import os
+import re
+
+def process_data(split):
+    model = AutoModelForCausalLM.from_pretrained("/mnt/petrelfs/wangzehan/cache/DeepSeek-R1-Distill-Qwen-14B", device_map="auto", torch_dtype="auto")
+    tokenizer = AutoTokenizer.from_pretrained("/mnt/petrelfs/wangzehan/cache/DeepSeek-R1-Distill-Qwen-14B")
+
+    with open('extra_data/annotation/scanrefer_vg_train_llava_style.json', 'r') as f:
+        all_data = json.load(f)
+
+    data = all_data[split*3667:(split+1)*3667]
+
+    if os.path.exists(f'extra_data/annotation/scanrefer_extra_info_{split}.jsonl'):
+        with open(f'extra_data/annotation/scanrefer_extra_info_{split}.jsonl', 'r') as f_extra_info:
+            count = sum(1 for _ in f_extra_info)
+        data = data[count:]
+
+    prompt = '''Suppose you are locating one target object described by a sentence from a video. Please analyze what the target object is, and simplify the core information that help locating the object.
+    Sentence: <SENTENCE>.
+    Only output json format as: {target_obj: "", core_info: "1.It is XXX, 2. It is XXX"}'''
+
+    for item in tqdm(data):
+        sentence = item['conversations'][0]['value'].split('\n')[-1]
+
+        messages = [
+            {"role": "user", "content": prompt.replace("<SENTENCE>", sentence)},
+        ]
+
+        input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
+        output_ids = model.generate(input_ids,
+                                    max_new_tokens=4096,
+                                    temperature=0.5)
+
+        content = tokenizer.decode(output_ids[0])
+        match = re.search(r'```json\n(.*?)\n```', content, re.DOTALL)
+        if match:
+            match_results = match.group(1)
+            match_results = json.loads(match_results)
+            item['extra_info'] = match_results
+        else:
+            item['extra_info'] = None
+
+        with open(f'extra_data/annotation/scanrefer_extra_info_{split}.jsonl', 'a') as f_extra_info:
+            f_extra_info.write(json.dumps(item) + '\n')
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--split', type=int, default=0)
+    args = parser.parse_args()
+    process_data(args.split)