script/Janus-Pro-7B-1-shot.py

# -*- coding: utf-8 -*-


import os
import gc
import random
import torch
from tqdm import tqdm
from transformers import AutoModelForCausalLM
from janus.models import MultiModalityCausalLM, VLChatProcessor
from janus.utils.io import load_pil_images  


chinese_to_english = {
    'strawberry', 'tomato', 'guava', 'dragon fruit',
    'orange',   'pear',     'lychee',   'mango',
    'kiwi',    'papaya', 'apple',    'grape',
    'pomegranate','peach', 'banana',   'pomelo'
}

stage_to_english = {
    'unripe', 'mature', 'pest-damaged', 'rotten'
}

recommendation_map = {
    'unripe': 'keep for further growth',
    'mature': 'picking it',
    'pest-damaged': 'try to recover it',
    'rotten': 'discard it'
}

score_map = {
    'unripe': 30, 'mature': 85, 'pest-damaged': 20, 'rotten': 5
}


model_path  = "deepseek-ai/Janus-Pro-7B"
root_folder = "../data"            
output_root = "result_1shot"
os.makedirs(output_root, exist_ok=True)

print("🚀 Loading Janus‑Pro‑7B ...")
vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(model_path, cache_dir=".")
tokenizer = vl_chat_processor.tokenizer
vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(
    model_path, cache_dir=".", trust_remote_code=True
)
vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
print("✅ Model ready.\n")


question = (
    "1. Identify the type of fruit or crop shown in the image.  \n"
    "2. Determine its current growth stage. (Options: unripe, mature, pest-damaged, rotten)  \n"
    "3. Recommend the farmer’s next action. (Options: keep for further growth, try to recover it, discard it)  \n"
    "4. Evaluate the consumer’s willingness to consume this fruit, from 1 (very unlikely) to 100 (very likely).\n"
    "Please respond in the following format:\n"
    "Type: [Fruit/Crop Name]  Growth Stage: [unripe / mature / pest-damaged / rotten]  "
    "Recommendation: [keep for further growth / pick it /  try to recover it / discard it]  Consumer Score: [1-100]"
)



def build_stage_example(fruit_dir: str, fruit_cn: str, stage_cn: str):
   
    stage_path = os.path.join(fruit_dir, stage_cn)
    imgs = [f for f in os.listdir(stage_path)
            if f.lower().endswith((".png", ".jpg", ".jpeg", ".bmp"))]
    if not imgs:
        return []

    img_path = os.path.join(stage_path, random.choice(imgs))

    fruit_en = chinese_to_english.get(fruit_cn, fruit_cn)
    stage_en  = stage_to_english[stage_cn]

    assistant_reply = (
        f"Type: {fruit_en}  Growth Stage: {stage_en}  "
        f"Recommendation: {recommendation_map[stage_en]}  "
        f"Consumer Score: {score_map[stage_en]}"
    )

  
    return [
        {
            "role": "<|User|>",
            "content": f"<image_placeholder>\n{question}",
            "images": [img_path],
        },
        {
            "role": "<|Assistant|>",
            "content": assistant_reply,
        },
    ]

for fruit_cn in os.listdir(root_folder):
    fruit_path = os.path.join(root_folder, fruit_cn)
    if not os.path.isdir(fruit_path):
        continue

    for stage_cn in os.listdir(fruit_path):
        stage_path = os.path.join(fruit_path, stage_cn)
        if not os.path.isdir(stage_path):
            continue

        img_files = [f for f in os.listdir(stage_path)
                     if f.lower().endswith((".png", ".jpg", ".jpeg", ".bmp"))]


       
        example_1shot = build_stage_example(fruit_path, fruit_cn, stage_cn)
        if not example_1shot:
           
            continue

        output_file = os.path.join(output_root, f"{fruit_cn}_{stage_cn}.txt")
        with open(output_file, "w", encoding="utf-8") as fout:
            for img_name in tqdm(img_files, desc=f"{fruit_cn}/{stage_cn}"):
                img_path = os.path.join(stage_path, img_name)

               
                if any(img_path in m.get("images", []) for m in example_1shot):
                    continue

                conversation = example_1shot + [
                    {
                        "role": "<|User|>",
                        "content": f"<image_placeholder>\n{question}",
                        "images": [img_path],
                    },
                    {"role": "<|Assistant|>", "content": ""},
                ]

                try:
                    with torch.no_grad():
                        pil_imgs = load_pil_images(conversation)
                        prep_in  = vl_chat_processor(
                            conversations=conversation,
                            images=pil_imgs,
                            force_batchify=True
                        ).to(vl_gpt.device)
                        
                        seq_len = len(prep_in.input_ids[0])
                        

                        tail_text = tokenizer.decode(prep_in.input_ids[0][-120:])
                       
                        del pil_imgs

                        embeds = vl_gpt.prepare_inputs_embeds(**prep_in)
                        outputs = vl_gpt.language_model.generate(
                            inputs_embeds=embeds,
                            attention_mask=prep_in.attention_mask,
                            pad_token_id=tokenizer.eos_token_id,
                            bos_token_id=tokenizer.bos_token_id,
                            eos_token_id=tokenizer.eos_token_id,
                            max_new_tokens=256,
                            do_sample=False,
                            use_cache=True,
                        )
                        answer = tokenizer.decode(outputs[0], skip_special_tokens=True)

               
                    print(f"\n🖼️ {img_name}\n{answer.strip()}")
                    fout.write("="*25 + " IMAGE START " + "="*25 + "\n")
                    fout.write(f"🖼️ Image Name: {img_name}\n")
                    fout.write(f"📝 Answer:\n{answer.strip()}\n")
                    fout.write("="*25 + " IMAGE END " + "="*25 + "\n\n")

                except Exception as e:
                    print(f"[ERROR] {fruit_cn}/{stage_cn}/{img_name}: {e}")
                    fout.write("="*25 + " IMAGE START " + "="*25 + "\n")
                    fout.write(f"🖼️ Image Name: {img_name}\n")
                    fout.write(f"❌ ERROR: {e}\n")
                    fout.write("="*25 + " IMAGE END " + "="*25 + "\n\n")

                finally:
                    for var in ("prep_in", "embeds", "outputs"):
                        if var in locals():
                            del locals()[var]
                    torch.cuda.empty_cache()