Spaces:
Running
on
Zero
Running
on
Zero
fancyfeast
commited on
Commit
β’
2fb728d
1
Parent(s):
0ba5137
Initial commit
Browse files- app.py +199 -0
- requirements.txt +6 -0
app.py
ADDED
@@ -0,0 +1,199 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import spaces
|
2 |
+
import gradio as gr
|
3 |
+
from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast, LlavaForConditionalGeneration, TextIteratorStreamer
|
4 |
+
import torch
|
5 |
+
import torch.amp.autocast_mode
|
6 |
+
from PIL import Image
|
7 |
+
import torchvision.transforms.functional as TVF
|
8 |
+
from threading import Thread
|
9 |
+
from typing import Generator
|
10 |
+
|
11 |
+
|
12 |
+
MODEL_PATH = "fancyfeast/llama-joycaption-alpha-two-vqa-test-1"
|
13 |
+
TITLE = "<h1><center>JoyCaption Alpha Two - VQA Test - (2024-11-25a)</center></h1>"
|
14 |
+
DESCRIPTION = """
|
15 |
+
<div>
|
16 |
+
<p>π¨π¨π¨ BY USING THIS SPACE YOU AGREE THAT YOUR QUERIES (but not images) <i>MAY</i> BE LOGGED AND COLLECTED ANONYMOUSLY π¨π¨π¨</p>
|
17 |
+
<p>π§ͺπ§ͺπ§ͺ This an experiment to see how well JoyCaption Alpha Two can learn to answer questions about images and follow instructions.
|
18 |
+
I've only finetuned it on 600 examples, so it is highly experimental, very weak, broken, and volatile. But for only training 600 examples,
|
19 |
+
I thought it was performing surprisingly well and wanted to share. π§ͺπ§ͺπ§ͺ</p>
|
20 |
+
<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?",
|
21 |
+
and "What are they doing?". It can also follow instructions, like "Write me a poem about this image",
|
22 |
+
"Write a caption but don't use any ambigious language, and make sure you mention that the image is from Instagram.", and
|
23 |
+
"Output JSON with the following properties: 'skin_tone', 'hair_style', 'hair_length', 'clothing', 'background'." Remember that this was only finetuned on
|
24 |
+
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.
|
25 |
+
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,
|
26 |
+
with enough coaxing and re-rolling.</p>
|
27 |
+
<p>About the π¨π¨π¨ 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.
|
28 |
+
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
|
29 |
+
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
|
30 |
+
training and exercising visual understanding. Certainly not good enough to handle the complicated instructions that will allow JoyCaption users to guide and
|
31 |
+
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
|
32 |
+
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,
|
33 |
+
which are graciously hosting this space, log.</p>
|
34 |
+
</div>
|
35 |
+
"""
|
36 |
+
|
37 |
+
PLACEHOLDER = """
|
38 |
+
"""
|
39 |
+
|
40 |
+
|
41 |
+
|
42 |
+
# Load model
|
43 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, use_fast=True)
|
44 |
+
assert isinstance(tokenizer, PreTrainedTokenizer) or isinstance(tokenizer, PreTrainedTokenizerFast), f"Expected PreTrainedTokenizer, got {type(tokenizer)}"
|
45 |
+
|
46 |
+
model = LlavaForConditionalGeneration.from_pretrained(MODEL_PATH, torch_dtype="bfloat16", device_map=0)
|
47 |
+
assert isinstance(model, LlavaForConditionalGeneration), f"Expected LlavaForConditionalGeneration, got {type(model)}"
|
48 |
+
|
49 |
+
|
50 |
+
def trim_off_prompt(input_ids: list[int], eoh_id: int, eot_id: int) -> list[int]:
|
51 |
+
# Trim off the prompt
|
52 |
+
while True:
|
53 |
+
try:
|
54 |
+
i = input_ids.index(eoh_id)
|
55 |
+
except ValueError:
|
56 |
+
break
|
57 |
+
|
58 |
+
input_ids = input_ids[i + 1:]
|
59 |
+
|
60 |
+
# Trim off the end
|
61 |
+
try:
|
62 |
+
i = input_ids.index(eot_id)
|
63 |
+
except ValueError:
|
64 |
+
return input_ids
|
65 |
+
|
66 |
+
return input_ids[:i]
|
67 |
+
|
68 |
+
end_of_header_id = tokenizer.convert_tokens_to_ids("<|end_header_id|>")
|
69 |
+
end_of_turn_id = tokenizer.convert_tokens_to_ids("<|eot_id|>")
|
70 |
+
assert isinstance(end_of_header_id, int) and isinstance(end_of_turn_id, int)
|
71 |
+
|
72 |
+
|
73 |
+
@spaces.GPU()
|
74 |
+
@torch.no_grad()
|
75 |
+
def chat_joycaption(message: dict, history, temperature: float, max_new_tokens: int) -> Generator[str, None, None]:
|
76 |
+
torch.cuda.empty_cache()
|
77 |
+
|
78 |
+
# Prompts are always stripped in training for now
|
79 |
+
prompt = message['text'].strip()
|
80 |
+
|
81 |
+
# Load image
|
82 |
+
if "files" not in message or len(message["files"]) != 1:
|
83 |
+
raise ValueError("This model requires exactly one image as input.")
|
84 |
+
|
85 |
+
image = Image.open(message["files"][0])
|
86 |
+
|
87 |
+
# Log the prompt
|
88 |
+
print(f"Prompt: {prompt}")
|
89 |
+
|
90 |
+
# Preprocess image
|
91 |
+
# NOTE: I found the default processor for so400M to have worse results than just using PIL directly
|
92 |
+
if image.size != (384, 384):
|
93 |
+
image = image.resize((384, 384), Image.LANCZOS)
|
94 |
+
image = image.convert("RGB")
|
95 |
+
pixel_values = TVF.pil_to_tensor(image)
|
96 |
+
|
97 |
+
convo = [
|
98 |
+
{
|
99 |
+
"role": "system",
|
100 |
+
"content": "You are a helpful image captioner.",
|
101 |
+
},
|
102 |
+
{
|
103 |
+
"role": "user",
|
104 |
+
"content": prompt,
|
105 |
+
},
|
106 |
+
]
|
107 |
+
|
108 |
+
# Format the conversation
|
109 |
+
convo_string = tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = True)
|
110 |
+
assert isinstance(convo_string, str)
|
111 |
+
|
112 |
+
# Tokenize the conversation
|
113 |
+
convo_tokens = tokenizer.encode(convo_string, add_special_tokens=False, truncation=False)
|
114 |
+
|
115 |
+
# Repeat the image tokens
|
116 |
+
input_tokens = []
|
117 |
+
for token in convo_tokens:
|
118 |
+
if token == model.config.image_token_index:
|
119 |
+
input_tokens.extend([model.config.image_token_index] * model.config.image_seq_length)
|
120 |
+
else:
|
121 |
+
input_tokens.append(token)
|
122 |
+
|
123 |
+
input_ids = torch.tensor(input_tokens, dtype=torch.long)
|
124 |
+
attention_mask = torch.ones_like(input_ids)
|
125 |
+
|
126 |
+
# Move to GPU
|
127 |
+
input_ids = input_ids.unsqueeze(0).to("cuda")
|
128 |
+
attention_mask = attention_mask.unsqueeze(0).to("cuda")
|
129 |
+
pixel_values = pixel_values.unsqueeze(0).to("cuda")
|
130 |
+
|
131 |
+
# Normalize the image
|
132 |
+
pixel_values = pixel_values / 255.0
|
133 |
+
pixel_values = TVF.normalize(pixel_values, [0.5], [0.5])
|
134 |
+
pixel_values = pixel_values.to(torch.bfloat16)
|
135 |
+
|
136 |
+
generate_kwargs = dict(
|
137 |
+
input_ids=input_ids,
|
138 |
+
pixel_values=pixel_values,
|
139 |
+
attention_mask=attention_mask,
|
140 |
+
max_new_tokens=max_new_tokens,
|
141 |
+
do_sample=True,
|
142 |
+
suppress_tokens=None,
|
143 |
+
use_cache=True,
|
144 |
+
temperature=temperature,
|
145 |
+
top_k=None,
|
146 |
+
top_p=0.9,
|
147 |
+
streamer=streamer,
|
148 |
+
)
|
149 |
+
|
150 |
+
if temperature == 0:
|
151 |
+
generate_kwargs["do_sample"] = False
|
152 |
+
|
153 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
|
154 |
+
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
155 |
+
t.start()
|
156 |
+
|
157 |
+
outputs = []
|
158 |
+
for text in streamer:
|
159 |
+
outputs.append(text)
|
160 |
+
yield "".join(outputs)
|
161 |
+
|
162 |
+
|
163 |
+
chatbot=gr.Chatbot(height=450, placeholder=PLACEHOLDER, label='Gradio ChatInterface')
|
164 |
+
|
165 |
+
with gr.Blocks() as demo:
|
166 |
+
gr.HTML(TITLE)
|
167 |
+
gr.Markdown(DESCRIPTION)
|
168 |
+
gr.ChatInterface(
|
169 |
+
fn=chat_joycaption,
|
170 |
+
chatbot=chatbot,
|
171 |
+
fill_height=True,
|
172 |
+
additional_inputs_accordion=gr.Accordion(label="βοΈ Parameters", open=False, render=False),
|
173 |
+
additional_inputs=[
|
174 |
+
gr.Slider(minimum=0,
|
175 |
+
maximum=1,
|
176 |
+
step=0.1,
|
177 |
+
value=0.6,
|
178 |
+
label="Temperature",
|
179 |
+
render=False),
|
180 |
+
gr.Slider(minimum=128,
|
181 |
+
maximum=4096,
|
182 |
+
step=1,
|
183 |
+
value=1024,
|
184 |
+
label="Max new tokens",
|
185 |
+
render=False ),
|
186 |
+
],
|
187 |
+
examples=[
|
188 |
+
['How to setup a human base on Mars? Give short answer.'],
|
189 |
+
['Explain theory of relativity to me like Iβm 8 years old.'],
|
190 |
+
['What is 9,000 * 9,000?'],
|
191 |
+
['Write a pun-filled happy birthday message to my friend Alex.'],
|
192 |
+
['Justify why a penguin might make a good king of the jungle.']
|
193 |
+
],
|
194 |
+
cache_examples=False,
|
195 |
+
)
|
196 |
+
|
197 |
+
|
198 |
+
if __name__ == "__main__":
|
199 |
+
demo.launch()
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
huggingface_hub==0.23.4
|
2 |
+
accelerate
|
3 |
+
torch
|
4 |
+
transformers==4.45.2
|
5 |
+
sentencepiece
|
6 |
+
torchvision
|