inference.py

# This Python 3 environment comes with many helpful analytics libraries installed
# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python
# For example, here's several helpful packages to load

import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)

# Input data files are available in the read-only "../input/" directory
# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory

import os
for dirname, _, filenames in os.walk('/kaggle/input'):
    for filename in filenames:
        print(os.path.join(dirname, filename))

# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using "Save & Run All" 
# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session


get_ipython().getoutput("pip install nltk")



import torch
from transformers import BlipProcessor, BlipForConditionalGeneration
from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
from tqdm import tqdm


device = "cuda" if torch.cuda.is_available() else "cpu"

model_name = "utkarshpise/blip-vrs"

processor = BlipProcessor.from_pretrained(model_name)
model = BlipForConditionalGeneration.from_pretrained(model_name)

model.to(device)
model.eval()

print(" Model loaded")


def evaluate_model(model, loader, processor, device):
    model.eval()

    total_loss = 0
    preds = []
    refs = []

    with torch.no_grad():
        for batch in tqdm(loader, desc="Evaluating"):
            batch = {k: v.to(device) for k, v in batch.items()}

            # 🔥 LM LOSS
            outputs = model(**batch)
            loss = outputs.loss
            total_loss += loss.item()

            # 🔥 Generate captions
            generated_ids = model.generate(
                pixel_values=batch["pixel_values"],
                max_length=50,
                num_beams=5
            )

            pred = processor.batch_decode(generated_ids, skip_special_tokens=True)
            ref = processor.batch_decode(batch["labels"], skip_special_tokens=True)

            preds.extend(pred)
            refs.extend(refs if False else ref)

    avg_loss = total_loss / len(loader)

    bleu_scores = []
    for p, r in zip(preds, refs):
        score = sentence_bleu([r.split()], p.split(), smoothing_function=smooth)
        bleu_scores.append(score)

    bleu = sum(bleu_scores) / len(bleu_scores)

    return avg_loss, bleu



import kagglehub
import os

path = kagglehub.dataset_download("ayaanmustafa/vrs-bench")

print("Dataset path:", path)
print("Files:", os.listdir(path))


import json

DATA_JSON = os.path.join(path, "/kaggle/input/datasets/ayaanmustafa/vrs-bench/vrsbench_data/VRSBench_train.json")  # adjust if needed

with open(DATA_JSON) as f:
    data = json.load(f)


print(type(data))
print(data[0])


import json
import os

# Paths
ann_dir = "/kaggle/input/datasets/ayaanmustafa/vrs-bench/vrsbench_data/Annotations_train/Annotations_train"
img_dir = "/kaggle/input/datasets/ayaanmustafa/vrs-bench/vrsbench_data/Images_train/Images_train"

output_path = "/kaggle/working/blip_train_vrs.json"
converted = []

for file in os.listdir(ann_dir):
    if file.endswith(".json"):
        
        # LIMIT TO 1000
        if len(converted) >= 1000:
            break

        with open(os.path.join(ann_dir, file)) as f:
            item = json.load(f)

        # Skip if no caption
        if "caption" not in item:
            continue

        image_name = item["image"]
        image_path = os.path.join(img_dir, image_name)

        converted.append({
            "image": image_path,
            "caption": item["caption"]
        })

# Save
with open(output_path, "w") as f:
    json.dump(converted, f, indent=2)

print("Total samples:", len(converted))


import json

with open("/kaggle/working/blip_train_vrs.json") as f:
    data = json.load(f)

print(data[0])


from torch.utils.data import Dataset
from PIL import Image

class VRSDataset(Dataset):
    def __init__(self, samples, processor):
        self.samples = samples
        self.processor = processor

    def __len__(self):
        return len(self.samples)

    def __getitem__(self, idx):
        item = self.samples[idx]

        image = Image.open(item["image"]).convert("RGB")
        caption = item["caption"]

        encoding = self.processor(
            images=image,
            text=caption,
            padding="max_length",
            truncation=True,
            return_tensors="pt"
        )

        encoding = {k: v.squeeze(0) for k, v in encoding.items()}
        encoding["labels"] = encoding["input_ids"]

        return encoding


import json

with open("/kaggle/working/blip_train_vrs.json") as f:
    samples = json.load(f)

print("Samples:", len(samples))


from torch.utils.data import DataLoader

dataset = VRSDataset(samples, processor)

test_loader = DataLoader(
    dataset,
    batch_size=8,
    shuffle=False,
    num_workers=0   
)

print("Batches:", len(test_loader))


from nltk.translate.bleu_score import SmoothingFunction

smooth = SmoothingFunction().method1


loss, bleu = evaluate_model(model, test_loader, processor, device)

print("LM Loss:", loss)
print("BLEU:", bleu)



from huggingface_hub import login

login()



Evaluating: 100%|██████████| 125/125 [06:48<00:00,  3.27s/it]
LM Loss: 0.1737535742521286
BLEU: 0.10278809043174526