scripts/to_parquet.py

# import os
# import pandas as pd
# from datasets import Dataset, Image, Features, Value


# df = pd.read_json("data/test.jsonl", lines=True)


# def get_full_path(file_name):
#     return os.path.join(os.getcwd(), "data", file_name)

# df['image_path'] = df['file_name'].apply(get_full_path)

# # 3. binary
# def __force_embed_image(example):
    
#     with open(example["image_path"], "rb") as f:
#         image_bytes = f.read()
#     return {"image": {"bytes": image_bytes}}


# ds = Dataset.from_pandas(df)


# ds = ds.map(__force_embed_image, remove_columns=["image_path"], num_proc=1) 


# ds = ds.cast_column("image", Image())


# ds.to_parquet("violin-test.parquet")




import os
import pandas as pd
from datasets import Dataset, Image, Features, Value

# 1. Load Data
df = pd.read_json("data/test.jsonl", lines=True)

# 2. Embedding Function
def __force_embed_images(example):
    image_cols = ["ground_truth", "image1_path", "image2_path"]
    data_root = os.path.abspath("data")
    
    for col in image_cols:
        val = example.get(col)
        
        if val and isinstance(val, str):
            full_path = os.path.join(data_root, val.replace('\\', '/'))
            if os.path.exists(full_path):
                with open(full_path, "rb") as f:
                    img_bytes = f.read()
                example[col] = {"bytes": img_bytes}
            else:
                example[col] = None
        else:
            example[col] = None
    return example

# 3. Define the TARGET Features

target_features = Features({
    "id": Value("string"),
    "prompt": Value("string"),
    "task": Value("int64"),
    "ground_truth": Image(),
    "color_1": Value("string"),
    "color_2": Value("string"),
    "hex_val_1": Value("string"),
    "hex_val_2": Value("string"),
    "direction": Value("string"),
    "shape": Value("string"),
    "position": Value("string"),
    "size_ratio": Value("string"),
    "center_x": Value("string"),
    "center_y": Value("string"),
    "mask_type": Value("string"),
    "image_id": Value("string"),
    "image1_path": Image(),
    "image2_path": Image(),
})

# 4. Create Dataset without initial features

ds = Dataset.from_pandas(df)

# 5. EXECUTE MAP WITH FEATURES

ds = ds.map(
    __force_embed_images, 
    features=target_features, 
    num_proc=1
)

# 6. Save to Parquet
ds.to_parquet("violin-test.parquet")
print("Success! All images manually embedded into parquet.")