code/fine_tuning/use_vlm_ft_gemma3.py

import httpx
import re
import requests
import time

from enum import Enum
from pydantic import BaseModel, Field
from typing_extensions import List
from typing import Literal, Optional

from requests.exceptions import ConnectionError

from PIL import Image

VLM_TEMPERATURE         = 0
######################################################################
######################################################################

class WeightUnit(Enum):
    GRAMM           = "Gramm"
    KILOGRAM        = "Kilogramm"
    MILLILITER      = "Milliliter"
    LITER           = "Liter"
    WASCHLADUNGEN   = "Waschladungen"
    BLATT           = "Blatt"
    STUECK          = "Stück"

class YesNo(Enum):
    YES = "yes"
    NO  = "no"

class product_promotion_data(BaseModel):
    """Collection of product and promotion data of an product advertisement."""
    brand: str = Field(description="The brand associated with the product")
    product_category: List[str] = Field(description="List of categories associated with the product.")
    price: float = Field(description="The promotional price.")
    regular_price: Optional[float] = Field(default=None, description="The regular price of the promotion.")
    relative_discount: Optional[int] = Field(default=None, description="The relative discount of the promotion.")
    absolute_discount: Optional[float] = Field(default=None, description="The absolute discount of the promotion.")
    GTINs: List[str] = Field(description="List of the GTINs for the products.")
    weight_number: float = Field(description="Only the numerical weight specication.")
    # weight_unit: WeightUnit = Field(description="Only the weight unit.")
    weight_unit: Literal["Gramm", "Kilogramm", "Milliliter", "Liter", "Waschladungen", "Blatt", "Stück"] = Field(description="Only the weight unit.")
    # different_types: YesNo = Field(description="If promotion offer different sorts.")
    different_types: Literal["yes", "no"] = Field(description="If promotion offer different sorts.")

######################################################################
######################################################################

def convert_items_to_strings(prediction):
    if isinstance(prediction, str):
        return prediction
    elif isinstance(prediction, list):
        return ', '.join(prediction)
    else:
        return str(prediction)

def get_output_results(dict_output, dict_result):
    for key, value in dict_output.items():
        if key == 'brand':
            dict_result['brand'] = convert_items_to_strings(dict_output['brand'])
        elif key == 'product_category':
            dict_result['product_category'] = convert_items_to_strings(dict_output['product_category'])
        elif key == 'price':
            dict_result['price'] = convert_items_to_strings(dict_output['price'])
        elif key == 'regular_price':
            dict_result['regular_price'] = convert_items_to_strings(dict_output['regular_price'])
        elif key == 'relative_discount':
            dict_result['relative_discount'] = convert_items_to_strings(dict_output['relative_discount'])
        elif key == 'absolute_discount':
            dict_result['absolute_discount'] = convert_items_to_strings(dict_output['absolute_discount'])
        elif key == 'GTINs':
            dict_result['GTINs'] = convert_items_to_strings(dict_output['GTINs'])
        elif key == 'weight_number':
            dict_result['weight_number'] = convert_items_to_strings(dict_output['weight_number'])
        elif key == 'weight_unit':
            dict_result['weight_unit'] = convert_items_to_strings(dict_output['weight_unit'])
        elif key == 'different_types':
            dict_result['different_types'] = convert_items_to_strings(dict_output['different_types'])
    return dict_result


def prompt(query_image, task, dict_log):
    system_message = "You are an assistant for question-answering tasks."
    dict_log['system_message'] = system_message

    human_message_text = "Do the user-provided task on the input image. \
    The answer must be provided in JSON format. \
    The task is: " + task + ".\
    If there is no information of a target, return NaN." 
    dict_log['human_message_text'] = human_message_text

    input_messages = [
            {
                "role": "system",
                "content": [{"type": "text", "text": system_message}],
            },
            {
                "role": "user",
                "content": [
                    {
                        "type": "image",
                        "image": query_image,
                    },
                    {
                        "type": "text",
                        "text": human_message_text,
                    },
                ],
            },
        ]

    return dict_log, input_messages

def process_vision_info(messages: list[dict]) -> list[Image.Image]:
    image_inputs = []
    # Iterate through each conversation
    for msg in messages:
        # Get content (ensure it's a list)
        content = msg.get("content", [])
        if not isinstance(content, list):
            content = [content]

        # Check each content element for images
        for element in content:
            if isinstance(element, dict) and (
                "image" in element or element.get("type") == "image"
            ):
                # Get the image and convert to RGB
                if "image" in element:
                    image = element["image"]
                else:
                    image = element
                image_inputs.append(image.convert("RGB"))
    return image_inputs


def get_dict_from_output_text(output_text):
    # Remove the surrounding braces:
    trimmed = output_text[0].strip('{}').strip()

    # Find all keys with their start positions:
    # Key pattern: word characters followed by colon
    matches = list(re.finditer(r'(\b\w+\b)\s*:', trimmed))

    data = {}
    for i, match in enumerate(matches):
        key = match.group(1)
        start = match.end()  # position after colon
        
        # end is start of next key or end of string
        if i+1 < len(matches):
            end = matches[i+1].start()
        else:
            end = len(trimmed)
        
        # The value is substring from start:end
        value = trimmed[start:end].strip().rstrip(',')
        
        # Clean value - strip whitespace and trailing commas
        value = value.strip()

        data[key] = value
    
    return data


def do_request(ft_model, processor, pil_image, task, dict_log, dict_result, dict_result_cost):
    dict_log, messages = prompt(pil_image, task, dict_log)

    text = processor.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )

    # Process the image and text
    image_inputs = process_vision_info(messages)

    # Tokenize the text and process the images
    inputs = processor(
        text=[text],
        images=image_inputs,
        padding=True,
        return_tensors="pt",
    )

    # Move the inputs to the device
    inputs = inputs.to(ft_model.device)

    stop_token_ids = [processor.tokenizer.eos_token_id, processor.tokenizer.convert_tokens_to_ids("<end_of_turn>")]

    try:
        start_time = time.time()
        try:
            # Generate the output
            generated_ids = ft_model.generate(**inputs, max_new_tokens=256, top_p=1.0, do_sample=True, temperature=0.8, eos_token_id=stop_token_ids, disable_compile=True)
        except:
            print("FAILED")
        elapsed_time = time.time() - start_time
        dict_result_cost['elapsed_time_[s]'] = float("{:.2f}".format(elapsed_time))

        # Trim the generation and decode the output to text
        generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
        output_text = processor.batch_decode(
            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
        )

        if len(output_text) == 1:
            dict_output = get_dict_from_output_text(output_text)
            dict_result = get_output_results(dict_output, dict_result)
            print('dict_result')
            print(dict_result)
    except ConnectionError as e:
        print(f"Connection error occurred: {e}")
        return dict_log, dict_result, dict_result_cost
    except requests.exceptions.RequestException as e:
        print(f"An error occurred: {e}")
        return dict_log, dict_result, dict_result_cost
    except ValueError as ve:
        print(f"Validation error: {ve}")
        return dict_log, dict_result, dict_result_cost
    except httpx.HTTPStatusError as e:
        print(f"HTTPStatusError: {e}")
        time.sleep(60)
        return dict_log, dict_result, dict_result_cost

    return dict_log, dict_result, dict_result_cost