app.py

import spaces
import gradio as gr
from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast, LlavaForConditionalGeneration, TextIteratorStreamer
import torch
import torch.amp.autocast_mode
from PIL import Image
import torchvision.transforms.functional as TVF
from threading import Thread
from typing import Generator


MODEL_PATH = "fancyfeast/llama-joycaption-alpha-two-vqa-test-1"
TITLE = "<h1><center>JoyCaption Alpha Two - VQA Test - (2024-11-25a)</center></h1>"
DESCRIPTION = """
<div>
<p>🚨🚨🚨 BY USING THIS SPACE YOU AGREE THAT YOUR QUERIES (but not images) <i>MAY</i> BE LOGGED AND COLLECTED ANONYMOUSLY 🚨🚨🚨</p>
<p>🧪🧪🧪 This an experiment to see how well JoyCaption Alpha Two can learn to answer questions about images and follow instructions.
I've only finetuned it on 600 examples, so it is highly experimental, very weak, broken, and volatile.  But for only training 600 examples,
I thought it was performing surprisingly well and wanted to share. 🧪🧪🧪</p>
<p>Unlike JoyCaption Alpha Two, you can ask this finetune questions about the image, like "What is he holding in his hand?", "Where might this be?",
and "What are they doing?".  It can also follow instructions, like "Write me a poem about this image",
"Write a caption but don't use any ambigious language, and make sure you mention that the image is from Instagram.", and
"Output JSON with the following properties: 'skin_tone', 'hair_style', 'hair_length', 'clothing', 'background'." Remember that this was only finetuned on
600 VQA/instruction examples, so it is _very_ limited right now.  Expect it to frequently fallback to its base behavior of just writing image descriptions.
Expect accuracy to be lower.  Expect glitches.  Despite that, I've found that it will follow most queries I've tested it with, even outside its training,
with enough coaxing and re-rolling.</p>
<p>About the 🚨🚨🚨 above: this space will log all prompts sent to it.  The only thing this space logs is the text query; no images, no user data, etc.
I cannot see what images you send, and frankly, I don't want to.  But knowing what kinds of instructions and queries users want JoyCaption to handle will
help guide me in building JoyCaption's VQA dataset.  I've found out the hard way that almost all public VQA datasets are garbage and don't do a good job of
training and exercising visual understanding.  Certainly not good enough to handle the complicated instructions that will allow JoyCaption users to guide and
direct how JoyCaption writes descriptions and captions.  So I'm building my own dataset, that will be made public.  So, with peace and love, this space logs the text
queries.  As always, the model itself is completely public and free to use outside of this space.  And, of course, I have no control nor access to what HuggingFace,
which are graciously hosting this space, log.</p>
</div>
"""

PLACEHOLDER = """
"""



# Load model
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, use_fast=True)
assert isinstance(tokenizer, PreTrainedTokenizer) or isinstance(tokenizer, PreTrainedTokenizerFast), f"Expected PreTrainedTokenizer, got {type(tokenizer)}"

model = LlavaForConditionalGeneration.from_pretrained(MODEL_PATH, torch_dtype="bfloat16", device_map=0)
assert isinstance(model, LlavaForConditionalGeneration), f"Expected LlavaForConditionalGeneration, got {type(model)}"


def trim_off_prompt(input_ids: list[int], eoh_id: int, eot_id: int) -> list[int]:
	# Trim off the prompt
	while True:
		try:
			i = input_ids.index(eoh_id)
		except ValueError:
			break
		
		input_ids = input_ids[i + 1:]
	
	# Trim off the end
	try:
		i = input_ids.index(eot_id)
	except ValueError:
		return input_ids
	
	return input_ids[:i]

end_of_header_id = tokenizer.convert_tokens_to_ids("<|end_header_id|>")
end_of_turn_id = tokenizer.convert_tokens_to_ids("<|eot_id|>")
assert isinstance(end_of_header_id, int) and isinstance(end_of_turn_id, int)


@spaces.GPU()
@torch.no_grad()
def chat_joycaption(message: dict, history, temperature: float, max_new_tokens: int) -> Generator[str, None, None]:
	torch.cuda.empty_cache()

	print(message)

	# Prompts are always stripped in training for now
	prompt = message['text'].strip()

	# Load image
	if "files" not in message or len(message["files"]) != 1:
		yield "ERROR: This model requires exactly one image as input."
		return
	
	image = Image.open(message["files"][0])
	
	# Log the prompt
	print(f"Prompt: {prompt}")

	# Preprocess image
	# NOTE: I found the default processor for so400M to have worse results than just using PIL directly
	if image.size != (384, 384):
		image = image.resize((384, 384), Image.LANCZOS)
	image = image.convert("RGB")
	pixel_values = TVF.pil_to_tensor(image)

	convo = [
		{
			"role": "system",
			"content": "You are a helpful image captioner.",
		},
		{
			"role": "user",
			"content": prompt,
		},
	]

	# Format the conversation
	convo_string = tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = True)
	assert isinstance(convo_string, str)

	# Tokenize the conversation
	convo_tokens = tokenizer.encode(convo_string, add_special_tokens=False, truncation=False)

	# Repeat the image tokens
	input_tokens = []
	for token in convo_tokens:
		if token == model.config.image_token_index:
			input_tokens.extend([model.config.image_token_index] * model.config.image_seq_length)
		else:
			input_tokens.append(token)
	
	input_ids = torch.tensor(input_tokens, dtype=torch.long)
	attention_mask = torch.ones_like(input_ids)

	# Move to GPU
	input_ids = input_ids.unsqueeze(0).to("cuda")
	attention_mask = attention_mask.unsqueeze(0).to("cuda")
	pixel_values = pixel_values.unsqueeze(0).to("cuda")

	# Normalize the image
	pixel_values = pixel_values / 255.0
	pixel_values = TVF.normalize(pixel_values, [0.5], [0.5])
	pixel_values = pixel_values.to(torch.bfloat16)

	streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)

	generate_kwargs = dict(
		input_ids=input_ids,
		pixel_values=pixel_values,
		attention_mask=attention_mask,
		max_new_tokens=max_new_tokens,
		do_sample=True,
		suppress_tokens=None,
		use_cache=True,
		temperature=temperature,
		top_k=None,
		top_p=0.9,
		streamer=streamer,
	)

	if temperature == 0:
		generate_kwargs["do_sample"] = False
	
	t = Thread(target=model.generate, kwargs=generate_kwargs)
	t.start()

	outputs = []
	for text in streamer:
		outputs.append(text)
		yield "".join(outputs)


chatbot=gr.Chatbot(height=450, placeholder=PLACEHOLDER, label='Gradio ChatInterface', type="messages")
textbox = gr.MultimodalTextbox(file_types=["image"], file_count="single")

with gr.Blocks() as demo:
	gr.HTML(TITLE)
	gr.Markdown(DESCRIPTION)
	gr.ChatInterface(
		fn=chat_joycaption,
		chatbot=chatbot,
		fill_height=True,
		multimodal=True,
		textbox=textbox,
		additional_inputs_accordion=None,#gr.Accordion(label="⚙️ Parameters", open=False, render=False),
		additional_inputs=[
			gr.Slider(minimum=0,
						maximum=1, 
						step=0.1,
						value=0.6, 
						label="Temperature", 
						render=False),
			gr.Slider(minimum=128, 
						maximum=4096,
						step=1,
						value=1024, 
						label="Max new tokens", 
						render=False ),
			],
		examples=[
			['How to setup a human base on Mars? Give short answer.'],
			['Explain theory of relativity to me like I’m 8 years old.'],
			['What is 9,000 * 9,000?'],
			['Write a pun-filled happy birthday message to my friend Alex.'],
			['Justify why a penguin might make a good king of the jungle.']
			],
		cache_examples=False,
    )


if __name__ == "__main__":
    demo.launch()