app.py

import os 
import torch
import openai
import requests
import gradio as gr
import transformers
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor
from peft import PeftModel

## CoT prompts

def _add_markup(table):
    try:
        parts = [p.strip() for p in table.splitlines(keepends=False)]
        if parts[0].startswith('TITLE'):
            result = f"Title: {parts[0].split(' | ')[1].strip()}\n"
            rows = parts[1:]
        else:
            result = ''
            rows = parts
        prefixes = ['Header: '] + [f'Row {i+1}: ' for i in range(len(rows) - 1)]
        return result + '\n'.join(prefix + row for prefix, row in zip(prefixes, rows))
    except:
        # just use the raw table if parsing fails
        return table


_TABLE = """Year | Democrats | Republicans | Independents
2004 | 68.1% | 45.0% | 53.0%
2006 | 58.0% | 42.0% | 53.0%
2007 | 59.0% | 38.0% | 45.0%
2009 | 72.0% | 49.0% | 60.0%
2011 | 71.0% | 51.2% | 58.0%
2012 | 70.0% | 48.0% | 53.0%
2013 | 72.0% | 41.0% | 60.0%"""

_INSTRUCTION = 'Read the table below to answer the following questions.'


_TEMPLATE = f"""First read an example then the complete question for the second table.
------------
{_INSTRUCTION}
{_add_markup(_TABLE)}
Q: In which year republicans have the lowest favor rate?
A: Let's find the column of republicans. Then let's extract the favor rates, they [45.0, 42.0, 38.0, 49.0, 51.2, 48.0, 41.0]. The smallest number is 38.0, that's Row 3.  Row 3 is year 2007. The answer is 2007.
Q: What is the sum of Democrats' favor rates of 2004, 2012, and 2013?
A: Let's find the rows of years 2004, 2012, and 2013. We find Row 1, 6, 7. The favor dates of Demoncrats on that 3 rows are 68.1, 70.0, and 72.0. 68.1+70.0+72=210.1. The answer is 210.1.
Q: By how many points do Independents surpass Republicans in the year of 2011?
A: Let's find the row with year = 2011. We find Row 5. We extract Independents and Republicans' numbers. They are 58.0 and 51.2. 58.0-51.2=6.8. The answer is 6.8.
Q: Which group has the overall worst performance?
A: Let's sample a couple of years. In Row 1, year 2004, we find Republicans having the lowest favor rate 45.0 (since 45.0<68.1, 45.0<53.0). In year 2006, Row 2, we find Republicans having the lowest favor rate 42.0 (42.0<58.0, 42.0<53.0). The trend continues to other years. The answer is Republicans.
Q: Which party has the second highest favor rates in 2007?
A: Let's find the row of year 2007, that's Row 3. Let's extract the numbers on Row 3: [59.0, 38.0, 45.0]. 45.0 is the second highest. 45.0 is the number of Independents. The answer is Independents.
{_INSTRUCTION}"""


## alpaca-lora

# debugging...
assert (
    "LlamaTokenizer" in transformers._import_structure["models.llama"]
), "LLaMA is now in HuggingFace's main branch.\nPlease reinstall it: pip uninstall transformers && pip install git+https://github.com/huggingface/transformers.git"
from transformers import LlamaTokenizer, LlamaForCausalLM, GenerationConfig

tokenizer = LlamaTokenizer.from_pretrained("decapoda-research/llama-7b-hf")

BASE_MODEL = "decapoda-research/llama-7b-hf"
LORA_WEIGHTS = "tloen/alpaca-lora-7b"

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

try:
    if torch.backends.mps.is_available():
        device = "mps"
except:
    pass

if device == "cuda":
    model = LlamaForCausalLM.from_pretrained(
        BASE_MODEL,
        load_in_8bit=False,
        torch_dtype=torch.float16,
        device_map="auto",
    )
    model = PeftModel.from_pretrained(
        model, LORA_WEIGHTS, torch_dtype=torch.float16, force_download=True
    )
elif device == "mps":
    model = LlamaForCausalLM.from_pretrained(
        BASE_MODEL,
        device_map={"": device},
        torch_dtype=torch.float16,
    )
    model = PeftModel.from_pretrained(
        model,
        LORA_WEIGHTS,
        device_map={"": device},
        torch_dtype=torch.float16,
    )
else:
    model = LlamaForCausalLM.from_pretrained(
        BASE_MODEL, device_map={"": device}, low_cpu_mem_usage=True
    )
    model = PeftModel.from_pretrained(
        model,
        LORA_WEIGHTS,
        device_map={"": device},
    )


if device != "cpu":
    model.half()
model.eval()
if torch.__version__ >= "2":
    model = torch.compile(model)


## FLAN-UL2
HF_TOKEN = os.environ.get("API_TOKEN", None)
API_URL = "https://api-inference.huggingface.co/models/google/flan-ul2"
headers = {"Authorization": f"Bearer {HF_TOKEN}"}
def query(payload):
	response = requests.post(API_URL, headers=headers, json=payload)
	return response.json()

## OpenAI models
openai.api_key = os.environ.get("OPENAI_TOKEN", None) 
def set_openai_api_key(api_key):
    if api_key and api_key.startswith("sk-") and len(api_key) > 50:
        openai.api_key = api_key

def get_response_from_openai(prompt, model="gpt-3.5-turbo", max_output_tokens=256):
  messages = [{"role": "assistant", "content": prompt}]
  response = openai.ChatCompletion.create(
      model=model,
      messages=messages,
      temperature=0.7,
      max_tokens=max_output_tokens,
      top_p=1,
      frequency_penalty=0,
      presence_penalty=0,
  )
  ret = response.choices[0].message['content']
  return ret

## deplot models
model_deplot = Pix2StructForConditionalGeneration.from_pretrained("google/deplot", torch_dtype=torch.bfloat16).to(0)
processor_deplot = Pix2StructProcessor.from_pretrained("google/deplot")

def evaluate(
    table,
    question,
    llm="alpaca-lora",
    input=None,
    temperature=0.1,
    top_p=0.75,
    top_k=40,
    num_beams=4,
    max_new_tokens=128,
    **kwargs,
):
    prompt_0shot = _INSTRUCTION + "\n" + _add_markup(table) + "\n" + "Q: " + question + "\n" + "A:"
    prompt = _TEMPLATE + "\n" + _add_markup(table) + "\n" + "Q: " + question + "\n" + "A:"
    if llm == "alpaca-lora":
        inputs = tokenizer(prompt, return_tensors="pt")
        input_ids = inputs["input_ids"].to(device)
        generation_config = GenerationConfig(
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            num_beams=num_beams,
            **kwargs,
        )
        with torch.no_grad():
            generation_output = model.generate(
                input_ids=input_ids,
                generation_config=generation_config,
                return_dict_in_generate=True,
                output_scores=True,
                max_new_tokens=max_new_tokens,
            )
        s = generation_output.sequences[0]
        output = tokenizer.decode(s)
    elif llm == "flan-ul2":
        output = query({"inputs": prompt_0shot})[0]["generated_text"]
    elif llm == "gpt-3.5-turbo":
        try:
            output = get_response_from_openai(prompt_0shot)
        except:
            output = "<Remember to input your OpenAI API key ☺>"
    else:
        RuntimeError(f"No such LLM: {llm}")
        
    return output


def process_document(image, question, llm):
    # image = Image.open(image)
    inputs = processor_deplot(images=image, text="Generate the underlying data table for the figure below:", return_tensors="pt").to(0, torch.bfloat16)
    predictions = model_deplot.generate(**inputs, max_new_tokens=512)
    table = processor_deplot.decode(predictions[0], skip_special_tokens=True).replace("<0x0A>", "\n")

    # send prompt+table to LLM
    res = evaluate(table, question, llm=llm)
    if llm == "alpaca-lora":
        return [table, res.split("A:")[-1]]
    else:
        return [table, res]

theme = gr.themes.Monochrome(
    primary_hue="indigo",
    secondary_hue="blue",
    neutral_hue="slate",
    radius_size=gr.themes.sizes.radius_sm,
    font=[gr.themes.GoogleFont("Open Sans"), "ui-sans-serif", "system-ui", "sans-serif"],
)

with gr.Blocks(theme=theme) as demo:
    with gr.Column():
      gr.Markdown(
            """<h1><center>DePlot+LLM: Multimodal chain-of-thought reasoning on plots</center></h1>
            <p>
            This is a demo of DePlot+LLM for QA and summarisation. <a href='https://arxiv.org/abs/2212.10505' target='_blank'>DePlot</a> is an image-to-text model that converts plots and charts into a textual sequence. The sequence then is used to prompt LLM for chain-of-thought reasoning. The current underlying LLMs are <a href='https://huggingface.co/spaces/tloen/alpaca-lora' target='_blank'>alpaca-lora</a>, <a href='https://huggingface.co/google/flan-ul2' target='_blank'>flan-ul2</a>, and <a href='https://openai.com/blog/chatgpt' target='_blank'>gpt-3.5-turbo</a>. To use it, simply upload your image and type a question or instruction and click 'submit', or click one of the examples to load them. Read more at the links below.
            </p>
            """
            )

    with gr.Row():
      with gr.Column(scale=2):
        input_image = gr.Image(label="Input Image", type="pil", interactive=True)
        #input_image.style(height=512, width=512)
        instruction = gr.Textbox(placeholder="Enter your instruction/question...", label="Question/Instruction")
        llm = gr.Dropdown(["alpaca-lora", "flan-ul2", "gpt-3.5-turbo"], label="LLM")
        openai_api_key_textbox = gr.Textbox(value='', 
                                            placeholder="Paste your OpenAI API key (sk-...) and hit Enter (if using OpenAI models, otherwise leave empty)",
                                            show_label=False, lines=1, type='password')          
        submit = gr.Button("Submit", variant="primary")
  
      with gr.Column(scale=2):  
        with gr.Accordion("Show intermediate table", open=False):
          output_table = gr.Textbox(lines=8, label="Intermediate Table")
        output_text = gr.Textbox(lines=8, label="Output")

    gr.Examples(
        examples=[
            ["deplot_case_study_6.png", "Rank the four methods according to model performances. By how much does deplot outperform the second strongest approach on average across the two sets?  Show the computation.", "gpt-3.5-turbo"],
            ["deplot_case_study_4.png", "What are the acceptance rates? And how does the acceptance change over the years?", "gpt-3.5-turbo"],
            ["deplot_case_study_m1.png", "Summarise the chart for me please.", "gpt-3.5-turbo"],
            ["deplot_case_study_m1.png", "What is the sum of numbers of Indonesia and Ireland? Remember to think step by step.", "alpaca-lora"],
            ["deplot_case_study_3.png", "By how much did China's growth rate drop? Think step by step.", "alpaca-lora"],
            ["deplot_case_study_4.png", "How many papers are submitted in 2020?", "flan-ul2"],
            ["deplot_case_study_5.png", "Which sales channel has the second highest portion?", "flan-ul2"],
            #["deplot_case_study_x2.png", "Summarise the chart for me please.", "alpaca-lora"],
            #["deplot_case_study_4.png", "How many papers are submitted in 2020?", "alpaca-lora"],
            #["deplot_case_study_m1.png", "Summarise the chart for me please.", "alpaca-lora"],
            #["deplot_case_study_4.png", "acceptance rate = # accepted / #submitted . What is the acceptance rate of 2010?", "flan-ul2"],
            #["deplot_case_study_m1.png", "Summarise the chart for me please.", "flan-ul2"],
        ],
        cache_examples=True,
        inputs=[input_image, instruction, llm],
        outputs=[output_table, output_text],
        fn=process_document
    )

    gr.Markdown(
            """<p style='text-align: center'><a href='https://arxiv.org/abs/2212.10505' target='_blank'>DePlot: One-shot visual language reasoning by plot-to-table translation</a></p>"""
    )
    openai.api_key = ""
    openai_api_key_textbox.change(set_openai_api_key,
                                      inputs=[openai_api_key_textbox],
                                      outputs=[])
    openai_api_key_textbox.submit(set_openai_api_key,
                                      inputs=[openai_api_key_textbox],
                                      outputs=[])
    submit.click(process_document, inputs=[input_image, instruction, llm], outputs=[output_table, output_text])
    instruction.submit(
        process_document, inputs=[input_image, instruction, llm], outputs=[output_table, output_text]
    )

demo.queue(concurrency_count=1).launch()